From 62af464227dafa1c55e0535bcb24346326748f46 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 24 Jun 2025 18:26:30 +0300 Subject: [PATCH 01/12] batch : fix check for empty sequences in memory (#14364) * batch : fix check for empty sequences in memory ggml-ci * cont : reuse the var ggml-ci --- src/llama-batch.cpp | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 401e11364..91b1d6078 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -244,11 +244,13 @@ bool llama_batch_allocr::init( continue; } - if (memory) { + const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1; + + if (p0 >= 0) { bool ok = true; if (batch.token) { - if (seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + if (seq_pos_min(s) != p0 + 1) { ok = false; } } else { @@ -256,7 +258,7 @@ bool llama_batch_allocr::init( // for embeddings (typically used as vision input), we allow them to have repeating positions // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 - if (seq_pos_min(s) != memory->seq_pos_max(s) && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) { ok = false; } } @@ -267,7 +269,7 @@ bool llama_batch_allocr::init( " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" " it is required that the sequence positions remain consecutive: Y = X + 1\n", - __func__, s, s, memory->seq_pos_max(s), s, seq_pos_min(s)); + __func__, s, s, p0, s, seq_pos_min(s)); return false; } From 73e53dc834c0a2336cd104473af6897197b96277 Mon Sep 17 00:00:00 2001 From: lhez Date: Tue, 24 Jun 2025 11:46:25 -0700 Subject: [PATCH 02/12] opencl: ref count `ggml_backend_opencl_context` and refactor profiling (#14254) * Move profiling info into `ggml_backend_opencl_context` * Add `enqueue_ndrange_kernel` to launch kernel --- ggml/src/ggml-opencl/ggml-opencl.cpp | 777 +++++++++------------------ 1 file changed, 241 insertions(+), 536 deletions(-) diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index 628e574f0..96e8a8588 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -231,6 +231,71 @@ static ggml_cl_compiler_version get_adreno_cl_compiler_version(const char *drive return { type, major, minor, patch }; } +// Profiling +struct ProfilingInfo { + std::string op_name; + std::string kernel_name; + + cl_kernel kernel; + cl_event evt; + + cl_ulong cmd_queued; + cl_ulong cmd_submit; + cl_ulong cmd_start; + cl_ulong cmd_end; + cl_ulong overhead_start; + cl_ulong overhead_end; + // For the times below, see spec for clGetEventProfilingInfo + // The time kernel spent in cmd queue - SUBMIT - QUEUED + cl_ulong cmd_queued_duration_ns; + // The time kernel spent for submission - START - SUBMIT + cl_ulong cmd_submit_duration_ns; + // Kernel execution time in nanoseconds - END - START + cl_ulong cmd_duration_ns; + // The time for the kernel to complete - COMPLETE - END + cl_ulong cmd_complete_duration_ns; + // Total time to finish the kernel - COMPELTE - QUEUED + cl_ulong cmd_total_duration_ns; + // Global and local work sizes. + size_t global_size[3]; + size_t local_size[3]; + // Op output size. + size_t output_size[4]; +}; + +static void populateProfilingInfo( + ProfilingInfo& info, cl_event evt, cl_kernel kernel, cl_uint work_dim, + size_t global_size[3], size_t local_size[3], + const ggml_tensor * tensor) { + info.op_name = tensor->name; + info.kernel = kernel; + info.evt = evt; + + // 0 means not specified, e.g., 2D workgroup, or NULL for driver to choose + info.local_size[0] = 0; + info.local_size[1] = 0; + info.local_size[2] = 0; + + info.global_size[0] = 0; + info.global_size[1] = 0; + info.global_size[2] = 0; + + if (local_size) { + for (cl_uint i = 0; i < work_dim; ++i) { + info.local_size[i] = local_size[i]; + } + } + + for (cl_uint i = 0; i < work_dim; ++i) { + info.global_size[i] = global_size[i]; + } + + info.output_size[0] = tensor->ne[0]; + info.output_size[1] = tensor->ne[1]; + info.output_size[2] = tensor->ne[2]; + info.output_size[3] = tensor->ne[3]; +} + struct ggml_backend_opencl_context; // backend device context @@ -254,6 +319,8 @@ struct ggml_backend_opencl_device_context { // backend context struct ggml_backend_opencl_context { + int ref_count; + cl_device_id device; std::string device_name; @@ -369,6 +436,108 @@ struct ggml_backend_opencl_context { cl_kernel kernel_timestep_embedding; cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat; + std::vector profiling_info; + + void write_profiling_info() { + FILE * fperf = fopen("cl_profiling.csv", "w"); + if (!fperf) { + GGML_LOG_ERROR("Failed to open cl_profiling.csv\n"); + return; + } + + // Populate profiling info + for (ProfilingInfo & info : profiling_info) { + cl_ulong cmd_queued; + cl_ulong cmd_submit; + cl_ulong cmd_start; + cl_ulong cmd_end; + cl_ulong cmd_complete; + + CL_CHECK(clWaitForEvents(1, &info.evt)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL)); + CL_CHECK(clReleaseEvent(info.evt)); + + char kernel_name[512]; + CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME, + sizeof(kernel_name), kernel_name, NULL)); + info.kernel_name = kernel_name; + + info.cmd_queued = cmd_queued; + info.cmd_submit = cmd_submit; + info.cmd_start = cmd_start; + info.cmd_end = cmd_end; + + info.cmd_queued_duration_ns = cmd_submit - cmd_queued; + info.cmd_submit_duration_ns = cmd_start - cmd_submit; + info.cmd_duration_ns = cmd_end - cmd_start; + info.cmd_complete_duration_ns = cmd_complete - cmd_end; + info.cmd_total_duration_ns = cmd_complete - cmd_queued; + } + + // Dump a csv + float total_kernel_time = 0; + fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n"); + for (const ProfilingInfo & info : profiling_info) { + total_kernel_time += info.cmd_duration_ns/1.e6f; + fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n", + info.op_name.c_str(), info.kernel_name.c_str(), + info.cmd_queued_duration_ns/1.e6f, + info.cmd_submit_duration_ns/1.e6f, + info.cmd_duration_ns/1.e6f, + info.cmd_complete_duration_ns/1.e6f, + info.cmd_total_duration_ns/1.e6f, + info.global_size[0], info.global_size[1], info.global_size[2], + info.local_size[0], info.local_size[1], info.local_size[2], + info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]); + } + fclose(fperf); + + GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time); + + // Dump a simple chrome trace + FILE* ftrace = fopen("cl_trace.json", "w"); + if (!ftrace) { + GGML_LOG_ERROR("Failed to open cl_trace.json\n"); + return; + } + + fprintf(ftrace, "[\n"); + for (const ProfilingInfo & info : profiling_info) { + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", + info.kernel_name.c_str(), info.cmd_queued/1000); + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", + info.kernel_name.c_str(), info.cmd_submit/1000); + + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", + info.kernel_name.c_str(), info.cmd_start/1000); + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", + info.kernel_name.c_str(), info.cmd_end/1000); + } + fclose(ftrace); + } + + void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) { +#ifdef GGML_OPENCL_PROFILING + cl_event evt; + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, &evt)); + + profiling_info.emplace_back(); + populateProfilingInfo(profiling_info.back(), evt, kernel, work_dim, global_work_size, local_work_size, tensor); +#else + GGML_UNUSED(tensor); + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, NULL)); +#endif + } + #ifdef GGML_OPENCL_USE_ADRENO_KERNELS // Transpose kernels cl_program program_transpose; @@ -395,47 +564,20 @@ struct ggml_backend_opencl_context { cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096; cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096; #endif // GGML_OPENCL_USE_ADRENO_KERNELS + + void free() { + ref_count--; + if (ref_count == 0) { +#ifdef GGML_OPENCL_PROFILING + write_profiling_info(); +#endif + } + } }; // All registered devices with a default device in the front. static std::vector g_ggml_backend_opencl_devices; -// Profiling -#ifdef GGML_OPENCL_PROFILING -struct ProfilingInfo { - std::string op_name; - std::string kernel_name; - - cl_kernel kernel; - cl_event evt; - - cl_ulong cmd_queued; - cl_ulong cmd_submit; - cl_ulong cmd_start; - cl_ulong cmd_end; - cl_ulong overhead_start; - cl_ulong overhead_end; - // For the times below, see spec for clGetEventProfilingInfo - // The time kernel spent in cmd queue - SUBMIT - QUEUED - cl_ulong cmd_queued_duration_ns; - // The time kernel spent for submission - START - SUBMIT - cl_ulong cmd_submit_duration_ns; - // Kernel execution time in nanoseconds - END - START - cl_ulong cmd_duration_ns; - // The time for the kernel to complete - COMPLETE - END - cl_ulong cmd_complete_duration_ns; - // Total time to finish the kernel - COMPELTE - QUEUED - cl_ulong cmd_total_duration_ns; - // Global and local work sizes. - size_t global_size[3]; - size_t local_size[3]; - // Op output size. - size_t output_size[4]; -}; - -std::vector g_profiling_info; -#endif - inline std::string read_file(const std::string &path) { std::ifstream ifs(path); if (!ifs) { @@ -1669,6 +1811,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { backend_ctx->device = dev_ctx->device; backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN; + // ref_count get increased in ggml_backend_opencl_device_init + // This function is also used to retrieve backend context, so we don't want + // to increase ref_count for each call. We only want to increase ref_count + // when the associated device is initialized + backend_ctx->ref_count = 0; + if (strstr(dev_ctx->device_name.c_str(), "Adreno") || strstr(dev_ctx->device_name.c_str(), "Qualcomm") || strstr(dev_ctx->device_version.c_str(), "Adreno")) { @@ -1841,93 +1989,22 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { return dev_ctx->backend_ctx; } -static void ggml_cl2_free(void) { -#ifdef GGML_OPENCL_PROFILING - FILE * fperf = fopen("cl_profiling.csv", "w"); - if (!fperf) { - GGML_LOG_ERROR("Failed to open cl_profiling.csv\n"); - return; +static void ggml_cl2_free(ggml_backend_t backend) { + ggml_backend_opencl_context * ctx = (ggml_backend_opencl_context *) backend->context; + ctx->free(); + + // The CL context is shared by all backends, release it if all backends have been released + bool should_release_opencl = true; + for (auto device : g_ggml_backend_opencl_devices) { + ggml_backend_opencl_device_context * ctx_dev = (ggml_backend_opencl_device_context *) device.context; + if (ctx_dev->backend_ctx->ref_count > 0) { + should_release_opencl = false; + } } - // Populate profiling info - for (ProfilingInfo & info : g_profiling_info) { - cl_ulong cmd_queued; - cl_ulong cmd_submit; - cl_ulong cmd_start; - cl_ulong cmd_end; - cl_ulong cmd_complete; - - CL_CHECK(clWaitForEvents(1, &info.evt)); - CL_CHECK(clGetEventProfilingInfo( - info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL)); - CL_CHECK(clGetEventProfilingInfo( - info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL)); - CL_CHECK(clGetEventProfilingInfo( - info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL)); - CL_CHECK(clGetEventProfilingInfo( - info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL)); - CL_CHECK(clGetEventProfilingInfo( - info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL)); - CL_CHECK(clReleaseEvent(info.evt)); - - char kernel_name[512]; - CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME, - sizeof(kernel_name), kernel_name, NULL)); - info.kernel_name = kernel_name; - - info.cmd_queued = cmd_queued; - info.cmd_submit = cmd_submit; - info.cmd_start = cmd_start; - info.cmd_end = cmd_end; - - info.cmd_queued_duration_ns = cmd_submit - cmd_queued; - info.cmd_submit_duration_ns = cmd_start - cmd_submit; - info.cmd_duration_ns = cmd_end - cmd_start; - info.cmd_complete_duration_ns = cmd_complete - cmd_end; - info.cmd_total_duration_ns = cmd_complete - cmd_queued; + if (should_release_opencl) { + CL_CHECK(clReleaseContext(ctx->context)); } - - // Dump a csv - float total_kernel_time = 0; - fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n"); - for (const ProfilingInfo & info : g_profiling_info) { - total_kernel_time += info.cmd_duration_ns/1.e6f; - fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n", - info.op_name.c_str(), info.kernel_name.c_str(), - info.cmd_queued_duration_ns/1.e6f, - info.cmd_submit_duration_ns/1.e6f, - info.cmd_duration_ns/1.e6f, - info.cmd_complete_duration_ns/1.e6f, - info.cmd_total_duration_ns/1.e6f, - info.global_size[0], info.global_size[1], info.global_size[2], - info.local_size[0], info.local_size[1], info.local_size[2], - info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]); - } - fclose(fperf); - - GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time); - - // Dump a simple chrome trace - FILE* ftrace = fopen("cl_trace.json", "w"); - if (!ftrace) { - GGML_LOG_ERROR("Failed to open cl_trace.json\n"); - return; - } - - fprintf(ftrace, "[\n"); - for (const ProfilingInfo & info : g_profiling_info) { - fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", - info.kernel_name.c_str(), info.cmd_queued/1000); - fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", - info.kernel_name.c_str(), info.cmd_submit/1000); - - fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", - info.kernel_name.c_str(), info.cmd_start/1000); - fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", - info.kernel_name.c_str(), info.cmd_end/1000); - } - fclose(ftrace); -#endif } //------------------------------------------------------------------------------ @@ -2011,9 +2088,7 @@ static const char * ggml_backend_opencl_name(ggml_backend_t backend) { } static void ggml_backend_opencl_free(ggml_backend_t backend) { - ggml_cl2_free(); - - GGML_UNUSED(backend); + ggml_cl2_free(backend); } static void ggml_backend_opencl_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { @@ -2899,6 +2974,8 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct static ggml_backend_t ggml_backend_opencl_device_init(ggml_backend_dev_t dev, const char * params) { ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(dev); + // Getting a new reference to the backend, increase ref_count + backend_ctx->ref_count++; ggml_backend_t backend = new ggml_backend { /* .guid = */ ggml_backend_opencl_guid(), @@ -3159,31 +3236,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso #define dump_tensor(tensor) #endif -//------------------------------------------------------------------------------ -// Profiling utility -//------------------------------------------------------------------------------ -#ifdef GGML_OPENCL_PROFILING -static void populateProfilingInfo( - ProfilingInfo& info, cl_event evt, cl_kernel kernel, - size_t global_size[3], size_t local_size[3], - const ggml_tensor * tensor) { - info.op_name = tensor->name; - info.kernel = kernel; - info.evt = evt; - - info.local_size[0] = local_size[0]; - info.local_size[1] = local_size[1]; - info.local_size[2] = local_size[2]; - info.global_size[0] = global_size[0]; - info.global_size[1] = global_size[1]; - info.global_size[2] = global_size[2]; - info.output_size[0] = tensor->ne[0]; - info.output_size[1] = tensor->ne[1]; - info.output_size[2] = tensor->ne[2]; - info.output_size[3] = tensor->ne[3]; -} -#endif - //------------------------------------------------------------------------------ // Ops //------------------------------------------------------------------------------ @@ -3227,7 +3279,6 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c const cl_ulong nb2 = dst ? dst->nb[2] : 0; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -3271,15 +3322,7 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c size_t global_work_size[] = {(size_t)ne10, (size_t)ne11, 1}; size_t local_work_size[] = {1, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -3321,7 +3364,6 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const const cl_ulong nb3 = dst ? dst->nb[3] : 0; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -3396,29 +3438,13 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } else { unsigned int nth = MIN(64, ne0); size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } } @@ -3461,7 +3487,6 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const const cl_ulong nb3 = dst ? dst->nb[3] : 0; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -3536,29 +3561,13 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } else { unsigned int nth = MIN(64, ne0); size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } } @@ -3598,7 +3607,6 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const const cl_ulong nb3 = dst->nb[3]; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -3661,29 +3669,13 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)n, 1, 1}; size_t local_work_size[] = {64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } else { unsigned int nth = MIN(64, ne0); size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } } @@ -3723,7 +3715,6 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const const cl_ulong nb3 = dst->nb[3]; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -3786,29 +3777,13 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)n, 1, 1}; size_t local_work_size[] = {64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } else { unsigned int nth = MIN(64, ne0); size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } } @@ -3821,7 +3796,6 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -3848,15 +3822,7 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)n, 1, 1}; size_t local_work_size[] = {64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -3868,7 +3834,6 @@ static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -3895,15 +3860,7 @@ static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, size_t global_work_size[] = {(size_t)n, 1, 1}; size_t local_work_size[] = {64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -3915,7 +3872,6 @@ static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -3947,15 +3903,7 @@ static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -3967,7 +3915,6 @@ static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -3992,15 +3939,7 @@ static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4012,7 +3951,6 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -4044,15 +3982,7 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4064,7 +3994,6 @@ static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, cons UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -4096,15 +4025,7 @@ static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, cons local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4116,7 +4037,6 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -4157,15 +4077,7 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4177,7 +4089,6 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; //ggml_backend_opencl_device_context * dev_ctx = // (ggml_backend_opencl_device_context *)backend->device->context; @@ -4241,15 +4152,7 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c // This is local memory - the size depends on subgroup size. CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth/sgs, NULL)); -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4261,7 +4164,6 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -4300,15 +4202,7 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, size_t global_work_size[] = {(size_t)n_groups*sgs, 1, 1}; size_t local_work_size[] = {(size_t)sgs, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4320,7 +4214,6 @@ static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const UNUSED(src1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -4397,16 +4290,7 @@ static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const } if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return; - -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) { @@ -4419,7 +4303,6 @@ static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, con UNUSED(src1_shape_def); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; if (backend_ctx->kernel_repeat == nullptr) { GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__); @@ -4467,15 +4350,7 @@ static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, con size_t global_work_size[] = { gws0, gws1, gws2 }; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, (size_t[3]){0,0,0}, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst); } static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) { @@ -4488,7 +4363,6 @@ static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_t GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; if (backend_ctx->kernel_pad == nullptr) { GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__); @@ -4533,15 +4407,7 @@ static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_t local_work_size_ptr = nullptr; } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) { @@ -4553,7 +4419,6 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg GGML_ASSERT(dst->type == GGML_TYPE_F32); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0); cl_kernel kernel = nullptr; @@ -4644,17 +4509,7 @@ static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, gg local_work_size_ptr = nullptr; } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - size_t profiling_gws[3] = {global_work_size[0], 1, 1}; - size_t profiling_lws[3] = {local_work_size_ptr ? local_work_size[0] : 0, 1, 1}; - populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4732,7 +4587,7 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con global_work_size[1] = d_ne1; global_work_size[2] = d_ne2; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL)); + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst); } } } else { @@ -4782,7 +4637,7 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con d_ne2 > 0 ? (size_t)d_ne2 : 1, d_ne3 > 0 ? (size_t)d_ne3 : 1 }; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size_nc, NULL, 0, NULL, NULL)); + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size_nc, NULL, dst); } } @@ -4795,7 +4650,6 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor GGML_ASSERT(dst->type == GGML_TYPE_F32); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; if (backend_ctx->kernel_timestep_embedding == nullptr) { GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__); @@ -4828,17 +4682,7 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor size_t global_work_size[] = {gws0, gws1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, &evt)); // Pass 2 for 2D problem - - g_profiling_info.emplace_back(); - size_t profiling_gws[3] = {global_work_size[0], global_work_size[1], 1}; - size_t profiling_lws[3] = {0,0,0}; // Reflects NULL LWS - populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, NULL)); // Pass 2 for 2D problem -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst); } static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -4853,7 +4697,6 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -5058,15 +4901,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co static_cast(padded_height_B) }; - #ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_size_t, local_size_t, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_size_t, local_size_t, dst); - #else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_size_t, local_size_t, 0, NULL, NULL)); - #endif + backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst); } else { // no need to transpose B in other cases // create an image for B from sub_buffer @@ -5188,16 +5023,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co // enqueue kernel with profiling // <--------------------------------------------> // - #ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); - // enqueue kernel without profiling - #else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); - #endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); // <--------------------------------------------> // // deallocate sub buffers and images @@ -5277,15 +5103,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co global_work_size[2] = (size_t)ne12*ne13; } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); return; } #else // GGML_OPENCL_SOA_Q @@ -5515,15 +5333,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co size_t global_work_size[] = {(size_t)(ne01 + ndst-1)/ndst*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13}; size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } else if (src0t == GGML_TYPE_Q4_K) { GGML_ASSERT(false && "not implemented"); } else if (src0t == GGML_TYPE_Q3_K) { @@ -5534,30 +5344,14 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co size_t global_work_size[] = {(size_t)(ne01+1)/2*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13}; size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } else { int64_t ny = (ne11 + nrows - 1)/nrows; size_t global_work_size[] = {(size_t)ne01*nth0, (size_t)ny*nth1, (size_t)ne12*ne13}; size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } } @@ -5574,7 +5368,6 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, GGML_ASSERT(src2->extra); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra; @@ -5680,15 +5473,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123}; size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -5701,7 +5486,6 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons GGML_ASSERT(ggml_is_contiguous(src0)); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; float scale; memcpy(&scale, dst->op_params, sizeof(scale)); @@ -5730,15 +5514,7 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -5775,7 +5551,6 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -5840,15 +5615,7 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, src1); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src1); } static void ggml_cl_dup(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -5871,7 +5638,6 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr const int ne02 = src0 ? src0->ne[2] : 0; ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -5895,15 +5661,7 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr size_t global_work_size[] = {(size_t)ne00*ne01*ne02/8, 1, 1}; size_t local_work_size[] = {64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } else { kernel = backend_ctx->kernel_diag_mask_inf; @@ -5923,15 +5681,7 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr local_work_size_ptr = nullptr; // Let driver choose the work-group sizes. } -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst); } } @@ -5951,7 +5701,6 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c } ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -6031,15 +5780,7 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6051,7 +5792,6 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const GGML_ASSERT(dst->extra); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; @@ -6217,15 +5957,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6240,7 +5972,6 @@ static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, con GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -6309,15 +6040,7 @@ static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, con size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC}; size_t local_work_size[] = {256, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6332,7 +6055,6 @@ static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, co GGML_ASSERT(ggml_is_contiguous(src0)); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -6364,15 +6086,7 @@ static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, co size_t global_work_size[] = {(size_t)ne00_padded, (size_t)nrows, (size_t)1}; size_t local_work_size[] = {(size_t)ne00_padded, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -6386,7 +6100,6 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c GGML_ASSERT(ggml_is_contiguous(src0)); ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; - cl_command_queue queue = backend_ctx->queue; ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; @@ -6427,15 +6140,7 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)64, 1, 1}; -#ifdef GGML_OPENCL_PROFILING - cl_event evt; - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); - - g_profiling_info.emplace_back(); - populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); -#else - CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); -#endif + backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst); } //------------------------------------------------------------------------------ From 2bf9d539dd158345e3a3b096e16474af535265b4 Mon Sep 17 00:00:00 2001 From: Anton Mitkov Date: Wed, 25 Jun 2025 17:09:55 +0100 Subject: [PATCH 03/12] sycl: GGML_SYCL_DISABLE_OPT on by default for all Intel Devices (#13973) --- docs/backend/SYCL.md | 2 +- ggml/src/ggml-sycl/common.hpp | 25 +------------------------ ggml/src/ggml-sycl/ggml-sycl.cpp | 6 ++---- ggml/src/ggml-sycl/sycl_hw.cpp | 4 +++- ggml/src/ggml-sycl/sycl_hw.hpp | 3 +++ 5 files changed, 10 insertions(+), 30 deletions(-) diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index 249e73451..6e9b88935 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -757,7 +757,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512 | Name | Value | Function | |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------| | GGML_SYCL_DEBUG | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG | -| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase | +| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) | | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. | | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. | | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.
Recommended to use when --split-mode = layer | diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index 753b4af14..4e7449d06 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -199,7 +199,7 @@ struct sycl_device_info { // size_t smpb; // max. shared memory per block bool vmm; // virtual memory support size_t total_vram; - sycl_hw_info hw_info; + //sycl_hw_info hw_info; \\ device id and aarch, currently not used optimize_feature opt_feature; }; @@ -286,29 +286,6 @@ struct ggml_tensor_extra_gpu { void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector streams={}); -inline optimize_feature check_gpu_optimize_feature(syclex::architecture &arch) { - optimize_feature opt; - - opt.reorder = - (arch == syclex::architecture::intel_gpu_dg1 || - arch == syclex::architecture::intel_gpu_acm_g10 || - arch == syclex::architecture::intel_gpu_acm_g11 || - arch == syclex::architecture::intel_gpu_acm_g12 || - arch == syclex::architecture::intel_gpu_pvc || - arch == syclex::architecture::intel_gpu_pvc_vg || - arch == syclex::architecture::intel_gpu_mtl_u || - arch == syclex::architecture::intel_gpu_mtl_s || - arch == syclex::architecture::intel_gpu_mtl_h || - arch == syclex::architecture::intel_gpu_arl_u || - arch == syclex::architecture::intel_gpu_arl_s || - arch == syclex::architecture::intel_gpu_arl_h || - arch == syclex::architecture::intel_gpu_bmg_g21 || - arch == syclex::architecture::intel_gpu_lnl_m - ); - - return opt; -} - namespace sycl_ex = sycl::ext::oneapi::experimental; struct ggml_backend_sycl_context { int device; diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index f25a96a62..9cb36ae99 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -83,9 +83,7 @@ static ggml_sycl_device_info ggml_sycl_init() { info.devices[i].cc = 100 * prop.get_major_version() + 10 * prop.get_minor_version(); - info.devices[i].hw_info = get_device_hw_info(&device); - info.devices[i].opt_feature = check_gpu_optimize_feature(info.devices[i].hw_info.arch); - + info.devices[i].opt_feature.reorder = !device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu); info.max_work_group_sizes[i] = prop.get_max_work_group_size(); } @@ -195,7 +193,7 @@ static void ggml_check_sycl() try { if (!initialized) { g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); - g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1); + g_ggml_sycl_disable_optimize = get_sycl_env("GGML_SYCL_DISABLE_OPT", 0); g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1); g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0); g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0); diff --git a/ggml/src/ggml-sycl/sycl_hw.cpp b/ggml/src/ggml-sycl/sycl_hw.cpp index da121ffc2..704114003 100644 --- a/ggml/src/ggml-sycl/sycl_hw.cpp +++ b/ggml/src/ggml-sycl/sycl_hw.cpp @@ -1,6 +1,7 @@ #include "sycl_hw.hpp" - +// TODO: currently not used +/* sycl_hw_info get_device_hw_info(sycl::device *device_ptr) { sycl_hw_info res; int32_t id = device_ptr->get_info(); @@ -11,3 +12,4 @@ sycl_hw_info get_device_hw_info(sycl::device *device_ptr) { return res; } +*/ diff --git a/ggml/src/ggml-sycl/sycl_hw.hpp b/ggml/src/ggml-sycl/sycl_hw.hpp index bf689450c..36b140bf0 100644 --- a/ggml/src/ggml-sycl/sycl_hw.hpp +++ b/ggml/src/ggml-sycl/sycl_hw.hpp @@ -10,6 +10,8 @@ namespace syclex = sycl::ext::oneapi::experimental; +// TODO: currently not used +/* struct sycl_hw_info { syclex::architecture arch; int32_t device_id; @@ -18,6 +20,7 @@ struct sycl_hw_info { bool is_in_vector(std::vector &vec, int item); sycl_hw_info get_device_hw_info(sycl::device *device_ptr); +*/ #endif // SYCL_HW_HPP From b193d5306912a2adae0fde7481819f6ee0941bc6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Wed, 25 Jun 2025 23:26:51 +0200 Subject: [PATCH 04/12] ggml : do not output unprintable characters on GGUF load failure (#14381) --- ggml/src/gguf.cpp | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp index a0a318a29..5ffd12b8b 100644 --- a/ggml/src/gguf.cpp +++ b/ggml/src/gguf.cpp @@ -335,7 +335,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par for (uint32_t i = 0; i < magic.size(); i++) { if (magic[i] != GGUF_MAGIC[i]) { - GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]); + char c0 = isprint(magic[0]) ? magic[0] : '?'; + char c1 = isprint(magic[1]) ? magic[1] : '?'; + char c2 = isprint(magic[2]) ? magic[2] : '?'; + char c3 = isprint(magic[3]) ? magic[3] : '?'; + GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, c0, c1, c2, c3); gguf_free(ctx); return nullptr; } From 60ef23d6c14d325d83eae5752e5de39ad268e9b0 Mon Sep 17 00:00:00 2001 From: Aaron Teo Date: Thu, 26 Jun 2025 05:49:04 +0800 Subject: [PATCH 05/12] ggml-cpu: enable IBM NNPA Vector Intrinsics (#14317) * ggml-cpu: add nnpa compile flag Signed-off-by: Aaron Teo (cherry picked from commit 4a9f60c201573128f73a65999b3e5cc497fae5c1) * ggml-cpu: add fp16->fp32 nnpa first Signed-off-by: Aaron Teo (cherry picked from commit 8d4a7987f9c1887f716be96250f2caeee0253929) * ggml-cpu: add fp32->fp16 Signed-off-by: Aaron Teo (cherry picked from commit 0ff0d6516247a41d2ade42b42cf0d676a4dd1627) * ggml-cpu: better variable names Signed-off-by: Aaron Teo (cherry picked from commit 2f58bbcbb89c183340e252362b2a40651f573f1f) * docs: update s390x docs Signed-off-by: Aaron Teo (cherry picked from commit 01b929491b50071a5d0572235dcf5a449da70aa7) * ggml-cpu: add debugging prints to see if dlf16 is correct Signed-off-by: Aaron Teo * ggml-cpu: fix print vs printf Signed-off-by: Aaron Teo * ggml-cpu: fix float placeholder Signed-off-by: Aaron Teo * ggml-cpu: ensure fp16 and fp32 load and stores are called Signed-off-by: Aaron Teo * ggml-cpu: fp16 load ensured to hit Signed-off-by: Aaron Teo * ggml-cpu: remove sigint from fp16 store for some reason, the function is not getting a hit when debugged with gdb. we will need to investigate further Signed-off-by: Aaron Teo * ggml-cpu: activate nnpa for ggml_cpu_fp16_to_fp32 Signed-off-by: Aaron Teo * ggml-cpu: nnpa activate ggml_cpu_fp16_to_fp32 for 8 elements Signed-off-by: Aaron Teo * ggml-cpu: nnpa switch to vec_xst test Signed-off-by: Aaron Teo * ggml-cpu: switch to vec_xst for 4 element loops also Signed-off-by: Aaron Teo * ggml-cpu: rework noop Signed-off-by: Aaron Teo * ggml-cpu: remove noop, general code cleanup Signed-off-by: Aaron Teo * ggml-cpu: clarify variable naming Signed-off-by: Aaron Teo * ggml-cpu: activate nnpa for ggml_cpu_fp32_to_fp16 Signed-off-by: Aaron Teo * ggml-cpu: add breakpoint for debugging Signed-off-by: Aaron Teo * ggml-cpu: test fix for conversion failure Signed-off-by: Aaron Teo * ggml-cpu: disable fp32->fp16 nnpa conversions for now there are some conversion failures in nnpa that requires the eyes of an ibm stsm. will create a separate pr to introduce the fp32->fp16 change. Signed-off-by: Aaron Teo * ggml-cpu: switch to elif macro Signed-off-by: Aaron Teo * ggml-cpu: reattempt fp32->fp16 Signed-off-by: Aaron Teo * ggml-cpu: fix typo Signed-off-by: Aaron Teo * ggml-cpu: reattempt fp32->fp16 Signed-off-by: Aaron Teo * ggml-cpu: fix compiler types Signed-off-by: Aaron Teo * ggml-cpu: change to typedef vector types Signed-off-by: Aaron Teo * ggml-cpu: add 4 element loops for fp32->fp16 Signed-off-by: Aaron Teo * ggml-cpu: clarified vector naming Signed-off-by: Aaron Teo * ggml-cpu: bring back fp32->fp16 store nnpa Signed-off-by: Aaron Teo * ggml-cpu: activate nnpa fp32->fp16 or fp16->fp32 compute Signed-off-by: Aaron Teo * ggml-cpu: add nnpa macro check in ggml-impl Signed-off-by: Aaron Teo * ggml-cpu: add missing __func__ Signed-off-by: Aaron Teo * ggml-cpu: diagnose why __NNPA__ macro is not being defined Signed-off-by: Aaron Teo * ggml-cpu: import vecintrin.h to fix compiler errors Signed-off-by: Aaron Teo * ggml-cpu: update macro tests Signed-off-by: Aaron Teo * ggml-cpu: move s390x typedef to own header file Signed-off-by: Aaron Teo * Revert "ggml-cpu: move s390x typedef to own header file" This reverts commit 157f856c34589566151630e294563a420702db39. Signed-off-by: Aaron Teo * ggml-cpu: switch to importing ggml-cpu-impl instead Signed-off-by: Aaron Teo * ggml-cpu: fix macro declaration Signed-off-by: Aaron Teo * ggml-cpu: test more macros Signed-off-by: Aaron Teo * ggml-cpu: add debug prints Signed-off-by: Aaron Teo * ggml-cpu: bruteforce macro definitions Signed-off-by: Aaron Teo * ggml-cpu: move macro definitions Signed-off-by: Aaron Teo * ggml-cpu: add ggml-impl.h to cmakelists Signed-off-by: Aaron Teo * ggml-cpu: switch to private macros Signed-off-by: Aaron Teo * ggml-cpu: move s390x typedef to own header file Signed-off-by: Aaron Teo (cherry picked from commit 157f856c34589566151630e294563a420702db39) * ggml-cpu: move things around Signed-off-by: Aaron Teo * ggml-cpu: bring back compile macros Signed-off-by: Aaron Teo * ggml-cpu: switch to quotes for import Signed-off-by: Aaron Teo * ggml-cpu: add compiler error macro Signed-off-by: Aaron Teo * ggml-cpu: add s390x detection in ggml-src Signed-off-by: Aaron Teo * ggml-cpu: bring back compile definitions Signed-off-by: Aaron Teo * ggml-cpu: undo cmakelists work Signed-off-by: Aaron Teo * Revert "ggml-cpu: move s390x typedef to own header file" This reverts commit 18d79e1a30b39d9aaa0bd58400c5cf2c32135c9a. Signed-off-by: Aaron Teo * ggml-cpu: remove typedefs.h Signed-off-by: Aaron Teo * ggml-cpu: remove typedef from cmakelists Signed-off-by: Aaron Teo * ggml-cpu: add ggml-impl.h future notes Signed-off-by: Aaron Teo * ggml-cpu: add todo comment for future reference Signed-off-by: Aaron Teo * ggml-cpu: clarify naming of dlf16 Signed-off-by: Aaron Teo * ggml-cpu: remove unnecessary target compile definitions Signed-off-by: Aaron Teo * ggml-cpu: move nnpa fp16->fp32 and fp32->fp16 to simd-mappings Signed-off-by: Aaron Teo * ggml: refactor fp32->fp16 and fp16->fp32 simd to ggml-cpu Signed-off-by: Aaron Teo * docs: update broken huggingface link for s390x Signed-off-by: Aaron Teo * ggml-cpu: fix duplicate func names during compile Signed-off-by: Aaron Teo * Revert "ggml-cpu: fix duplicate func names during compile" This reverts commit fbb733451f27677063b914d4f6c9a9841d45b38d. Signed-off-by: Aaron Teo * Revert "ggml: refactor fp32->fp16 and fp16->fp32 simd to ggml-cpu" This reverts commit bd288e8fa52b5244f65cee21cb61062f1a9e0ca5. Signed-off-by: Aaron Teo * ggml: refactor fp16<->fp32 simd to ggml-cpu Signed-off-by: Aaron Teo * ggml-cpu: fix missing simd-mappings.h import in quants.c Signed-off-by: Aaron Teo * ggml-cpu: fix missing simd-mappings.h within repack Signed-off-by: Aaron Teo * ggml-cpu: fix amx mmq missing simd-mappings.h Signed-off-by: Aaron Teo * ggml-cpu: attempt at fixing loongarch failing build Signed-off-by: Aaron Teo * ggml-cpu: move nnpa together with other fp16<->fp32 simd Signed-off-by: Aaron Teo * ggml-cpu: fix wrong refactor of ggml-base ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164176555 Signed-off-by: Aaron Teo * ggml: remove dependency on ggml-cpu from ggml-base Signed-off-by: Aaron Teo * ggml-cpu: rename all fp16<->fp32 macros to prefix with ggml_cpu ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164449406 Signed-off-by: Aaron Teo * ggml-cpu: remove mistaken fallback macro fallback logic was already implemented but i was too sleepy to realise Signed-off-by: Aaron Teo * ggml: move ggml_table_f32_f16 to ggml-cpu ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164775006 Signed-off-by: Aaron Teo * ggml-cpu: move ggml_table_f32_f16 back to ggml-base due to ci failures Signed-off-by: Aaron Teo * Revert "ggml-cpu: move ggml_table_f32_f16 back to ggml-base due to ci failures" This reverts commit 32a3533564bdb7902cefb9c89b1c9e956a81ce29. Signed-off-by: Aaron Teo * Revert "ggml: move ggml_table_f32_f16 to ggml-cpu" This reverts commit 9e40d984ad27d7b60392fb2b7548885201864fe4. Signed-off-by: Aaron Teo * ggml: move ggml_table_f32_f16 to ggml-cpu ref: https://github.com/ggml-org/llama.cpp/pull/14317#discussion_r2164775006 Signed-off-by: Aaron Teo (cherry picked from commit 9e40d984ad27d7b60392fb2b7548885201864fe4) * ggml: move ggml_table_f32_f16 to ggml-cpu.c Signed-off-by: Aaron Teo * ggml-cpu: extern c ggml_table_f32_f16 + chore docs Signed-off-by: Aaron Teo * ggml-cpu: dedup ggml_table_f32_f16 from simd-mappings.h we rely on the variable declaration in ggml-cpu.c instead Signed-off-by: Aaron Teo * Revert "ggml-cpu: dedup ggml_table_f32_f16 from simd-mappings.h" This reverts commit f71b21d2f74f5e03ec0c2b4fefd3cbf395aecf16. Signed-off-by: Aaron Teo * ggml-cpu: bring back ggml_table_f32_f16 Signed-off-by: Aaron Teo * Revert "ggml-cpu: bring back ggml_table_f32_f16" This reverts commit 2dce119178bed5ef5c8398c4230ddd14fef80e49. Signed-off-by: Aaron Teo * fix ggml time initialization * fix f32_f16 table init * remove extra line --------- Signed-off-by: Aaron Teo Co-authored-by: slaren --- docs/build-s390x.md | 41 +++- docs/build.md | 4 + ggml/CMakeLists.txt | 1 + ggml/include/ggml-cpu.h | 1 + ggml/src/ggml-cpu/CMakeLists.txt | 8 + ggml/src/ggml-cpu/amx/mmq.cpp | 19 +- ggml/src/ggml-cpu/arch/arm/quants.c | 217 +++++++++--------- ggml/src/ggml-cpu/arch/arm/repack.cpp | 25 ++- ggml/src/ggml-cpu/arch/loongarch/quants.c | 105 ++++----- ggml/src/ggml-cpu/arch/powerpc/quants.c | 111 ++++----- ggml/src/ggml-cpu/arch/riscv/quants.c | 83 +++---- ggml/src/ggml-cpu/arch/riscv/repack.cpp | 47 ++-- ggml/src/ggml-cpu/arch/s390/quants.c | 57 ++--- ggml/src/ggml-cpu/arch/wasm/quants.c | 59 ++--- ggml/src/ggml-cpu/arch/x86/quants.c | 165 +++++++------- ggml/src/ggml-cpu/arch/x86/repack.cpp | 39 ++-- ggml/src/ggml-cpu/common.h | 5 +- ggml/src/ggml-cpu/ggml-cpu-impl.h | 12 +- ggml/src/ggml-cpu/ggml-cpu.c | 75 +++++-- ggml/src/ggml-cpu/ggml-cpu.cpp | 3 + ggml/src/ggml-cpu/llamafile/sgemm.cpp | 5 +- ggml/src/ggml-cpu/ops.cpp | 96 ++++---- ggml/src/ggml-cpu/quants.c | 49 ++-- ggml/src/ggml-cpu/repack.cpp | 29 +-- ggml/src/ggml-cpu/simd-mappings.h | 244 +++++++++++++++++--- ggml/src/ggml-cpu/vec.cpp | 4 +- ggml/src/ggml-cpu/vec.h | 90 ++++---- ggml/src/ggml-impl.h | 262 ++++++---------------- ggml/src/ggml.c | 11 - 29 files changed, 1005 insertions(+), 862 deletions(-) diff --git a/docs/build-s390x.md b/docs/build-s390x.md index f44038c58..bb6eae784 100644 --- a/docs/build-s390x.md +++ b/docs/build-s390x.md @@ -28,8 +28,9 @@ cmake --build build --config Release -j $(nproc) ``` **Notes**: -- For faster repeated compilation, install [ccache](https://ccache.dev/) -- By default, VXE/VXE2 is enabled. To disable it (not recommended): + +- For faster repeated compilation, install [ccache](https://ccache.dev/) +- By default, VXE/VXE2 is enabled. To disable it (not recommended): ```bash cmake -S . -B build \ @@ -41,18 +42,29 @@ cmake --build build --config Release -j $(nproc) cmake --build build --config Release -j $(nproc) ``` -- For debug builds: +- By default, NNPA is enabled when available. To disable it (not recommended): + + ```bash + cmake -S . -B build \ + -DCMAKE_BUILD_TYPE=Release \ + -DGGML_BLAS=ON \ + -DGGML_BLAS_VENDOR=OpenBLAS \ + -DGGML_NNPA=OFF + + cmake --build build --config Release -j $(nproc) + ``` + +- For debug builds: ```bash cmake -S . -B build \ -DCMAKE_BUILD_TYPE=Debug \ -DGGML_BLAS=ON \ -DGGML_BLAS_VENDOR=OpenBLAS - cmake --build build --config Debug -j $(nproc) ``` -- For static builds, add `-DBUILD_SHARED_LIBS=OFF`: +- For static builds, add `-DBUILD_SHARED_LIBS=OFF`: ```bash cmake -S . -B build \ @@ -70,7 +82,7 @@ All models need to be converted to Big-Endian. You can achieve this in three cas 1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)** - You can find popular models pre-converted and verified at [s390x Ready Models](hf.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08). + You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08). These models and their respective tokenizers are verified to run correctly on IBM Z & LinuxONE. @@ -101,27 +113,33 @@ All models need to be converted to Big-Endian. You can achieve this in three cas ``` For example, + ```bash python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf ``` **Notes:** + - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2. ## IBM Accelerators ### 1. SIMD Acceleration -Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14 or EC13. In such systems, the APIs can still run but will use a scalar implementation. +Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation. -### 2. zDNN Accelerator +### 2. NNPA Vector Intrinsics Acceleration -*Only available in IBM z16 or later system. No direction at the moment.* +Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation. -### 3. Spyre Accelerator +### 3. zDNN Accelerator -*No direction at the moment.* +_Only available in IBM z16 or later system. No direction at the moment._ + +### 4. Spyre Accelerator + +_No direction at the moment._ ## Performance Tuning @@ -154,4 +172,3 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl 2. **Other Questions** Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com). - diff --git a/docs/build.md b/docs/build.md index 20a6f606e..2e0b5d970 100644 --- a/docs/build.md +++ b/docs/build.md @@ -557,6 +557,10 @@ ninja To read documentation for how to build on Android, [click here](./android.md) +## IBM Z & LinuxONE + +To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md) + ## Notes about GPU-accelerated backends The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`. diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 4e7399f9e..215eb2348 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -131,6 +131,7 @@ option(GGML_RVV "ggml: enable rvv" ON) option(GGML_RV_ZFH "ggml: enable riscv zfh" OFF) option(GGML_XTHEADVECTOR "ggml: enable xtheadvector" OFF) option(GGML_VXE "ggml: enable vxe" ON) +option(GGML_NNPA "ggml: enable nnpa" ON) option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF) set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM") diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index de77a875e..e3b79d09b 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -101,6 +101,7 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_riscv_v (void); GGML_BACKEND_API int ggml_cpu_has_vsx (void); GGML_BACKEND_API int ggml_cpu_has_vxe (void); + GGML_BACKEND_API int ggml_cpu_has_nnpa (void); GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void); GGML_BACKEND_API int ggml_cpu_has_llamafile (void); diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 71b1d67b8..671fad4d2 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -448,6 +448,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) # TODO: Separation to determine activation of VX/VXE/VXE2 if (${S390X_M} MATCHES "8561|8562") + set(GGML_NNPA OFF) message(STATUS "z15 target") list(APPEND ARCH_FLAGS -march=z15) elseif (${S390X_M} MATCHES "3931") @@ -464,7 +465,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name) endif() if (GGML_VXE) + message(STATUS "VX/VXE/VXE2 enabled") list(APPEND ARCH_FLAGS -mvx -mzvector) + list(APPEND ARCH_DEFINITIONS GGML_VXE) + endif() + + if (GGML_NNPA) + message(STATUS "NNPA enabled") + list(APPEND ARCH_DEFINITIONS GGML_NNPA) endif() elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm") message(STATUS "Wasm detected") diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp index cec34eb64..47c61b881 100644 --- a/ggml/src/ggml-cpu/amx/mmq.cpp +++ b/ggml/src/ggml-cpu/amx/mmq.cpp @@ -8,6 +8,7 @@ #include "mmq.h" #include "ggml-impl.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "quants.h" #include "ggml-quants.h" #include @@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_ // Quantize these floats const float iscale = 127.f / amax; - y[i].d = GGML_FP32_TO_FP16(1 / iscale); + y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale); const float id = ( amax != 0.0f ) ? iscale : 0.f; const __m512 vscale = _mm512_set1_ps(id); @@ -1090,7 +1091,7 @@ struct acc_C { const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); for (int m = 0; m < nr; ++m) { - const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); __m512 vsum; @@ -1113,8 +1114,8 @@ struct acc_C { const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half)))); for (int m = 0; m < nr; ++m) { - const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); - const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s)); + const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d)); + const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); __m512 vsum; @@ -1137,7 +1138,7 @@ struct acc_C { const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset))); for (int m = 0; m < nr; ++m) { - const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d)); + const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d)); const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N)); __m512 vsum; @@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni for (int k = 0; k < 8; ++k) { va[k] = _mm512_set1_epi32(a_ptr[k]); } - vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d)); - vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s)); + vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d)); + vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s)); } // load b @@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnniqs + 16); float32_t _scale[4] = { - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d) }; float32x4_t scale = vld1q_f32(_scale); @@ -274,10 +275,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // dot product sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4, svdot_s32(svdup_n_s32(0), qx0ls, qy0l), - svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4, svdot_s32(svdup_n_s32(0), qx1ls, qy1l), - svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); @@ -313,9 +314,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // dot product sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), - svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), - svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); @@ -354,9 +355,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // dot product sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32, - svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32, - svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1)); @@ -404,8 +405,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -423,7 +424,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -464,10 +465,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i]; float32_t summs_t[4] = { - GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), - GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s), - GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s), - GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s) + GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s), + GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s), + GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s), + GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s) }; summs0 = vaddq_f32(summs0, vld1q_f32(summs_t)); @@ -490,10 +491,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // mmla into int32x4_t float32_t _scale[4] = { - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d) }; float32x4_t scale = vld1q_f32(_scale); @@ -539,7 +540,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0]; const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1]; - summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); + summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s); const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -562,8 +563,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); - sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); - sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; @@ -582,7 +583,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -666,10 +667,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -694,7 +695,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -739,8 +740,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8x16_t m4b = vdupq_n_u8(0x0F); - summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); - summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); + summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s); + summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s); // extract the 5th bit via lookup table ((b) << 4) memcpy(&qh0, x0->qh, sizeof(qh0)); @@ -784,10 +785,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), - ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), - ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; @@ -812,7 +813,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -864,10 +865,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); float32_t _scale[4] = { - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d) + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d), + GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d) }; float32x4_t scale = vld1q_f32(_scale); @@ -934,10 +935,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16, svdot_s32(svdup_n_s32(0), qx0_0, qy0_0), - svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16, svdot_s32(svdup_n_s32(0), qx1_0, qy1_0), - svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1)); @@ -960,9 +961,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs); sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), - svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), - svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1)); @@ -1002,8 +1003,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64); // scale creation - const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d); - const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d); + const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d); + const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d); // duplicate deq1 in first half of vector and deq2 in second half of vector const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2); @@ -1043,11 +1044,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), - ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d)); + ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d)); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), - ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d)); + ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d)); } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); @@ -1059,7 +1060,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -1217,7 +1218,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo const int16x8_t ysum0 = vld1q_s16(y[i].bsums); const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8); - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; #if defined(__ARM_FEATURE_DOTPROD) sumi0 = vaddq_s32(sumi0, sumi1); @@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } } - sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); + sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); } *s = sumf; @@ -1362,7 +1363,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo const int16x8_t ysum0 = vld1q_s16(y[i].bsums); const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8); - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; #if defined(__ARM_FEATURE_DOTPROD) sumi0 = vaddq_s32(sumi0, sumi1); @@ -1393,7 +1394,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); sumf += (float) sumi * d; } @@ -1425,9 +1426,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi switch (vector_length) { case 128: for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); svfloat32_t d_broad = svdup_n_f32((float32_t)d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; @@ -1570,9 +1571,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi case 256: case 512: for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); svfloat32_t d_broad = svdup_n_f32((float32_t)d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; @@ -1671,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sum = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1742,8 +1743,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -1805,7 +1806,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q3_sv = x[i].qs; const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask; @@ -1981,7 +1982,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].hmask; @@ -2112,7 +2113,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2258,18 +2259,18 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)), vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins)))); const float32x4_t dmins = { - GGML_FP16_TO_FP32(x0->dmin) * y0->d, - GGML_FP16_TO_FP32(x0->dmin) * y1->d, - GGML_FP16_TO_FP32(x1->dmin) * y0->d, - GGML_FP16_TO_FP32(x1->dmin) * y1->d, + GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d, + GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d, + GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d, + GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d, }; vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins); const float32x4_t superblock_scale = { - GGML_FP16_TO_FP32(x0->d) * y0->d, - GGML_FP16_TO_FP32(x0->d) * y1->d, - GGML_FP16_TO_FP32(x1->d) * y0->d, - GGML_FP16_TO_FP32(x1->d) * y1->d, + GGML_CPU_FP16_TO_FP32(x0->d) * y0->d, + GGML_CPU_FP16_TO_FP32(x0->d) * y1->d, + GGML_CPU_FP16_TO_FP32(x1->d) * y0->d, + GGML_CPU_FP16_TO_FP32(x1->d) * y1->d, }; vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale); } @@ -2289,8 +2290,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); @@ -2377,8 +2378,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); @@ -2478,9 +2479,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2520,8 +2521,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8)); @@ -2630,9 +2631,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2827,10 +2828,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32); const float32x4_t superblock_scale = { - GGML_FP16_TO_FP32(x0->d) * y0->d, - GGML_FP16_TO_FP32(x0->d) * y1->d, - GGML_FP16_TO_FP32(x1->d) * y0->d, - GGML_FP16_TO_FP32(x1->d) * y1->d, + GGML_CPU_FP16_TO_FP32(x0->d) * y0->d, + GGML_CPU_FP16_TO_FP32(x0->d) * y1->d, + GGML_CPU_FP16_TO_FP32(x1->d) * y0->d, + GGML_CPU_FP16_TO_FP32(x1->d) * y1->d, }; visum = vsubq_s32(visum, vibias); @@ -2858,7 +2859,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi svuint8_t q6h_1, q6h_2, q6h_3, q6h_4; for (int i = 0; i < nb; ++i) { - const float d_all = GGML_FP16_TO_FP32(x[i].d); + const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -3011,7 +3012,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d_all = GGML_FP16_TO_FP32(x[i].d); + const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -3128,7 +3129,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -3199,7 +3200,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; float sumf1 = 0, sumf2 = 0; @@ -3234,7 +3235,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; @@ -3284,7 +3285,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; const uint8x8_t scales8 = vld1_u8(x[i].scales); @@ -3329,7 +3330,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT sc = x[i].scales; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3398,7 +3399,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -3458,7 +3459,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const int8_t * q8 = y[i].qs; const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; @@ -3521,7 +3522,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3557,7 +3558,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3630,7 +3631,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; @@ -3691,7 +3692,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT signs = x[i].signs; @@ -3786,7 +3787,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } - sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3); + sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3); } *s = sumf; @@ -3817,7 +3818,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 4; } - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); } *s = sumf; @@ -3905,7 +3906,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } - sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); + sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2)); } *s = sumf; @@ -3952,7 +3953,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qh += 2; } - sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); + sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); } *s = sumf; @@ -4003,13 +4004,13 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]); sumf += - GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) + - GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2); + GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) + + GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2); } #endif for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -4071,7 +4072,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v } - sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); + sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); } *s = sumf; @@ -4079,7 +4080,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 39a0dd301..2f8bc9e25 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -6,6 +6,7 @@ #include "ggml-impl.h" #include "ggml-cpu.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "traits.h" #include @@ -51,7 +52,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < 8; j++) { @@ -102,7 +103,7 @@ void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTR const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < QK8_0 * 4; j++) { @@ -145,7 +146,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < 4; j++) { @@ -221,7 +222,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < QK8_0 * 4; j++) { @@ -311,7 +312,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -399,7 +400,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -514,7 +515,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -608,7 +609,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -1117,7 +1118,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -1570,7 +1571,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -2039,7 +2040,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -2147,7 +2148,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } diff --git a/ggml/src/ggml-cpu/arch/loongarch/quants.c b/ggml/src/ggml-cpu/arch/loongarch/quants.c index f2ea96572..9e33fb322 100644 --- a/ggml/src/ggml-cpu/arch/loongarch/quants.c +++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -474,7 +475,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Quantize these floats const float d = max_scalar / 127.f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id ); @@ -548,7 +549,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Quantize these floats const float d = max_scalar / 127.f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const __m256 mul = __lasx_xvreplfr2vr_s( id ); @@ -576,7 +577,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Compute the sum of the quants and set y[i].s const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3)); const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7)); - y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1))); + y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1))); // Convert int32 to int16 ni0 = lsx_packs_w( ni0, ni1 ); @@ -667,7 +668,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { /* Compute combined scale for the block */ - const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); + const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) ); __m256i qx = bytes_from_nibbles_32(x[ib].qs); @@ -699,7 +700,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi for (; ib + 1 < nb; ib += 2) { // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); + const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) ); const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0); @@ -717,7 +718,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) ); + const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) ); const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0); @@ -766,7 +767,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -797,10 +798,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { - const float d0 = GGML_FP16_TO_FP32(x[ib].d); - const float d1 = GGML_FP16_TO_FP32(y[ib].d); + const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d); + const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); const __m256 d0v = __lasx_xvreplfr2vr_s( d0 ); const __m256 d1v = __lasx_xvreplfr2vr_s( d1 ); @@ -834,7 +835,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -865,7 +866,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { /* Compute combined scale for the block */ - const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME + const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME __m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i bxhi = bytes_from_bits_32(x[ib].qh); @@ -902,7 +903,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -934,16 +935,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { - const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d)); + const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d)); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); __m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i bxhi = bytes_from_bits_32(x[ib].qh); bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10)); qx = __lasx_xvor_v(qx, bxhi); - const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d)); + const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d)); const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0); const __m256 q = mul_sum_us8_pairs_float(qx, qy); @@ -973,7 +974,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -1003,7 +1004,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { // Compute combined scale for the block - const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0); __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0); @@ -1023,7 +1024,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -1047,8 +1048,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1116,8 +1117,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -1170,7 +1171,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q3 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; // Set up scales @@ -1294,7 +1295,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1330,8 +1331,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -1438,9 +1439,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1477,8 +1478,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8_t * GGML_RESTRICT q5 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -1593,9 +1594,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1624,7 +1625,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -1713,7 +1714,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1780,7 +1781,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = __lasx_xvldi(0); @@ -1820,7 +1821,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; @@ -1895,7 +1896,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1980,7 +1981,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT sc = x[i].scales; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2049,7 +2050,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); @@ -2108,7 +2109,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const int8_t * q8 = y[i].qs; const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; @@ -2168,7 +2169,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2213,7 +2214,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2279,7 +2280,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; @@ -2340,7 +2341,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT signs = x[i].signs; @@ -2451,7 +2452,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum); accum1 += d * sumi1; } @@ -2484,7 +2485,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 4; } - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); } *s = sumf; @@ -2530,9 +2531,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const __m256i p_1 = lasx_madd_h(p16_1, mone); const __m256i p_2 = lasx_madd_h(p16_2, mone); - accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)), + accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)), __lasx_xvffint_s_w(p_1), accum1); - accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)), + accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)), __lasx_xvffint_s_w(p_2), accum2); } @@ -2540,7 +2541,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v #endif for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -2595,7 +2596,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v sumi1 = __lasx_xvadd_w(p_1, sumi1); sumi2 = __lasx_xvadd_w(p_2, sumi2); } - accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), + accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum); } @@ -2604,7 +2605,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/arch/powerpc/quants.c b/ggml/src/ggml-cpu/arch/powerpc/quants.c index ce4e47a86..053d5cbdc 100644 --- a/ggml/src/ggml-cpu/arch/powerpc/quants.c +++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -67,7 +68,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float id = d ? 1.0f/d : 0.0f; const vector float vid = vec_splats(id); - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); for (int j = 0; j < 8; j++) { const vector float v = vec_round(vec_mul(srcv[j], vid)); @@ -112,7 +113,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float id = d ? 1.0f/d : 0.0f; const vector float vid = vec_splats(id); - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); vector int accv = vec_splats(0); @@ -127,7 +128,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i accv = vec_add(accv, vec_sld(accv, accv, 4)); accv = vec_add(accv, vec_sld(accv, accv, 8)); - y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0)); + y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0)); } #else @@ -170,8 +171,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); @@ -214,7 +215,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -249,12 +250,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m)); - vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f}; + vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m)); + vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f}; vsumf0 = vec_madd(vxmin, vys, vsumf0); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); @@ -291,7 +292,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -326,8 +327,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])}; @@ -379,7 +380,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -415,12 +416,12 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m)); - vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f}; + vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m)); + vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f}; vsumf0 = vec_madd(vxmin, vys, vsumf0); vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])}; @@ -470,7 +471,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -502,8 +503,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); vector signed char q8x0 = vec_xl( 0, x[ib].qs); @@ -542,7 +543,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -574,11 +575,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin)); vector float vdmin = vec_mul(vxmin, vyd); vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); @@ -708,8 +709,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -770,7 +771,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -962,7 +963,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1005,11 +1006,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin)); vector float vdmin = vec_mul(vxmin, vyd); vector signed short q8ysums0 = vec_xl( 0, y[i].bsums); @@ -1177,9 +1178,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1222,11 +1223,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); - vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin)); + vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin)); vector float vdmin = vec_mul(vxmin, vyd); UNUSED(kmask1); @@ -1394,9 +1395,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1432,7 +1433,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -1591,7 +1592,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1659,7 +1660,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -1742,7 +1743,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; @@ -1790,7 +1791,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -1871,7 +1872,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT sc = x[i].scales; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1939,7 +1940,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -2033,7 +2034,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const int8_t * q8 = y[i].qs; const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; @@ -2096,7 +2097,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -2176,7 +2177,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2236,7 +2237,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -2329,7 +2330,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT signs = x[i].signs; @@ -2394,7 +2395,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo vector float vsumf3 = vec_splats(0.0f); for (int i = 0; i < nb; ++i) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d)); vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); @@ -2505,7 +2506,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 4; } - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); } *s = sumf; @@ -2546,8 +2547,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v __builtin_prefetch(y[ib].qs, 0, 1); - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d)); - vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d)); + vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d)); vector float vd = vec_mul(vxd, vyd); vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs); @@ -2582,7 +2583,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v #endif for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -2620,7 +2621,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v for (int ibl = 0; ibl < nb; ++ibl) { - vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d)); + vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d)); vector float vyd = vec_splats(y[ibl].d); vector float vd = vec_mul(vxd, vyd); @@ -2697,7 +2698,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/arch/riscv/quants.c b/ggml/src/ggml-cpu/arch/riscv/quants.c index 6f3aa94fb..8b64d8adc 100644 --- a/ggml/src/ggml-cpu/arch/riscv/quants.c +++ b/ggml/src/ggml-cpu/arch/riscv/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -45,7 +46,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); @@ -85,7 +86,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); @@ -102,7 +103,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // set y[i].s int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); - y[i].s = GGML_FP32_TO_FP16(sum*d); + y[i].s = GGML_CPU_FP32_TO_FP16(sum*d); } #else @@ -160,7 +161,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } #endif @@ -177,7 +178,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -225,7 +226,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } #endif @@ -242,7 +243,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -293,7 +294,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); - sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } #endif @@ -316,7 +317,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -366,7 +367,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } #endif @@ -389,7 +390,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -427,7 +428,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } #endif @@ -438,7 +439,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -465,8 +466,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8_t * q2 = x[i].qs; const int8_t * q8 = y[i].qs; const uint8_t * sc = x[i].scales; - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); uint8_t *patmp = atmp; int vsums; int tmp; @@ -569,8 +570,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const int8_t * q8 = y[i].qs; const uint8_t * sc = x[i].scales; - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); size_t vl = 16; @@ -644,8 +645,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8_t * q2 = x[i].qs; const int8_t * q8 = y[i].qs; const uint8_t * sc = x[i].scales; - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); uint8_t *patmp = atmp; int vsums; int tmp; @@ -750,8 +751,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -916,7 +917,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi q3 += 32; q8 += 128; scale += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; sumf += d * isum; } @@ -1017,7 +1018,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; sumf += d*sum_t; @@ -1134,7 +1135,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi q3 += 32; q8 += 128; scale += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; sumf += d * isum; } break; @@ -1202,7 +1203,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1239,8 +1240,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int tmp, tmp2, sumi; __asm__ __volatile__( @@ -1361,8 +1362,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi size_t vl = 8; - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); @@ -1422,8 +1423,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi break; case 128: for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int tmp, tmp2, sumi; __asm__ __volatile__( @@ -1580,9 +1581,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1627,8 +1628,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8_t * GGML_RESTRICT hm = x[i].qh; const int8_t * GGML_RESTRICT q8 = y[i].qs; - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl); vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl); @@ -1749,9 +1750,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1778,7 +1779,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -1862,7 +1863,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi case 256: for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q6 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -1943,7 +1944,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi case 128: for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * restrict q6 = x[i].ql; const uint8_t * restrict qh = x[i].qh; @@ -2058,7 +2059,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp index 0882b4102..45c91a694 100644 --- a/ggml/src/ggml-cpu/arch/riscv/repack.cpp +++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp @@ -6,6 +6,7 @@ #include "ggml-impl.h" #include "ggml-cpu.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "traits.h" #include @@ -90,16 +91,16 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4); // vector version needs Zvfhmin extension - const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d); + const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d); const float b_scales[8] = { - GGML_FP16_TO_FP32(b_ptr[l].d[0]), - GGML_FP16_TO_FP32(b_ptr[l].d[1]), - GGML_FP16_TO_FP32(b_ptr[l].d[2]), - GGML_FP16_TO_FP32(b_ptr[l].d[3]), - GGML_FP16_TO_FP32(b_ptr[l].d[4]), - GGML_FP16_TO_FP32(b_ptr[l].d[5]), - GGML_FP16_TO_FP32(b_ptr[l].d[6]), - GGML_FP16_TO_FP32(b_ptr[l].d[7]) + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7]) }; const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4); const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4); @@ -129,7 +130,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -181,20 +182,20 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo // vector version needs Zvfhmin extension const float a_scales[4] = { - GGML_FP16_TO_FP32(a_ptr[l].d[0]), - GGML_FP16_TO_FP32(a_ptr[l].d[1]), - GGML_FP16_TO_FP32(a_ptr[l].d[2]), - GGML_FP16_TO_FP32(a_ptr[l].d[3]) + GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]), + GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]), + GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]), + GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3]) }; const float b_scales[8] = { - GGML_FP16_TO_FP32(b_ptr[l].d[0]), - GGML_FP16_TO_FP32(b_ptr[l].d[1]), - GGML_FP16_TO_FP32(b_ptr[l].d[2]), - GGML_FP16_TO_FP32(b_ptr[l].d[3]), - GGML_FP16_TO_FP32(b_ptr[l].d[4]), - GGML_FP16_TO_FP32(b_ptr[l].d[5]), - GGML_FP16_TO_FP32(b_ptr[l].d[6]), - GGML_FP16_TO_FP32(b_ptr[l].d[7]) + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]), + GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7]) }; const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4); @@ -382,7 +383,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c index 26bd90875..a840219a4 100644 --- a/ggml/src/ggml-cpu/arch/s390/quants.c +++ b/ggml/src/ggml-cpu/arch/s390/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -49,7 +50,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f / d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); for (int j = 0; j < 8; j++) { const __vector float v = vec_mul(srcv[j], vec_splats(id)); @@ -94,7 +95,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f / d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); __vector int32_t acc = vec_splats(0); @@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i acc = vec_add(acc, vi); } - y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3])); + y[i].s = GGML_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3])); } #else GGML_UNUSED(nb); @@ -164,7 +165,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_); const __vector float v_xy = vec_float(vec_unpackh(v_xy_)); - const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); acc = vec_madd(v_xy, v_d, acc); } @@ -185,7 +186,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -219,7 +220,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi __builtin_prefetch(x[ib].qs, 0, 1); __builtin_prefetch(y[ib].qs, 0, 1); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); const uint8x16_t v_x = vec_xl(0, x[ib].qs); const int8x16_t v_xl = (const int8x16_t)(v_x & v_m); @@ -231,7 +232,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); const float32x4_t v_xy = vec_float(v_xy_); - const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); acc = vec_madd(v_xy, v_d, acc); } @@ -252,7 +253,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); const float32x4_t v_xy = vec_float(v_xy_); - const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); acc = vec_madd(v_xy, v_d, acc); } @@ -305,7 +306,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -348,7 +349,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sum = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * restrict x0l = x[i].qs; const uint8_t * restrict x0h = x[i].hmask; @@ -497,7 +498,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -537,8 +538,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); @@ -647,9 +648,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -698,8 +699,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); @@ -819,9 +820,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -859,7 +860,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi int8x16_t v_y[4]; for (int i = 0; i < nb; ++i) { - const float d_all = GGML_FP16_TO_FP32(x[i].d); + const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT x0l = x[i].ql; const uint8_t * GGML_RESTRICT x0h = x[i].qh; @@ -1004,7 +1005,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1071,7 +1072,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi // float sumf = 0; // for (int i = 0; i < nb; ++i) { -// const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; +// const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; // const uint16_t * GGML_RESTRICT q2 = x[i].qs; // const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1121,7 +1122,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi // float sumf = 0.f; // for (int i = 0; i < nb; ++i) { -// const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; +// const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; // const uint16_t * GGML_RESTRICT q2 = x[i].qs; // const int8_t * GGML_RESTRICT q8 = y[i].qs; // int32_t bsum = 0; @@ -1182,12 +1183,12 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs); const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); - sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]); + sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]); } #endif for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -1257,7 +1258,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2; } - sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); + sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); } *s = sumf; @@ -1265,7 +1266,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/arch/wasm/quants.c b/ggml/src/ggml-cpu/arch/wasm/quants.c index 4ec97f533..b0904d8a3 100644 --- a/ggml/src/ggml-cpu/arch/wasm/quants.c +++ b/ggml/src/ggml-cpu/arch/wasm/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -65,7 +66,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); for (int j = 0; j < 8; j++) { const v128_t v = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id)); @@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i const float d = amax / ((1 << 7) - 1); const float id = d ? 1.0f/d : 0.0f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); v128_t accv = wasm_i32x4_splat(0); @@ -126,7 +127,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i accv = wasm_i32x4_add(accv, vi); } - y[i].s = GGML_FP32_TO_FP16( + y[i].s = GGML_CPU_FP32_TO_FP16( d * (wasm_i32x4_extract_lane(accv, 0) + wasm_i32x4_extract_lane(accv, 1) + wasm_i32x4_extract_lane(accv, 2) + @@ -324,8 +325,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi ); // Accumulate results with scaling - float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); - float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d); + float scale0 = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d); + float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0))); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1))); @@ -348,7 +349,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -428,7 +429,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi wasm_i32x4_dot_i16x8(v0lfh, v1lh)), wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), - wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); + wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)))); } sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + @@ -454,7 +455,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -491,7 +492,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi const block_q5_1 * GGML_RESTRICT x0 = &x[ib]; const block_q8_1 * GGML_RESTRICT y0 = &y[ib]; - summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); + summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s); const v128_t m4b = wasm_i8x16_splat(0x0F); @@ -538,7 +539,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi wasm_i32x4_dot_i16x8(v0lfh, v1lh)), wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), - wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d)))); + wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d)))); } sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + @@ -564,7 +565,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -620,7 +621,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1)); // Convert to float and accumulate - const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); + const float scale = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d); sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale))); } @@ -635,7 +636,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -746,8 +747,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi isum += wasm_i32x4_extract_lane(isum_vec, 0); } - const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf += dall * isum - dmin * summs; } @@ -768,8 +769,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -880,7 +881,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi } // Accumulate results - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const v128_t v_d = wasm_f32x4_splat(d); v128_t v_sum = wasm_f32x4_add( wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d), @@ -957,7 +958,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -991,8 +992,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Corrected sign const uint8_t * GGML_RESTRICT q4 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1136,9 +1137,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1170,8 +1171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi float sumf = 0; for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Fixed sign const uint8_t * GGML_RESTRICT q5 = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -1331,9 +1332,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1420,7 +1421,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi wasm_v128_store(&aux32[0], acc0); wasm_v128_store(&aux32[4], acc1); - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) { sums[l] += d * aux32[l]; } @@ -1470,7 +1471,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c index e3f722b52..e7527c00a 100644 --- a/ggml/src/ggml-cpu/arch/x86/quants.c +++ b/ggml/src/ggml-cpu/arch/x86/quants.c @@ -3,6 +3,7 @@ #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "simd-mappings.h" #include "../../quants.h" #include "../../ggml-cpu-impl.h" @@ -256,9 +257,9 @@ static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_ // quad fp16 delta calculation static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) { - // GGML_FP16_TO_FP32 is faster than Intel F16C - return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)), - _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0))); + // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C + return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)), + _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0))); } #endif #elif defined(__SSSE3__) @@ -305,7 +306,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Quantize these floats const float d = maxScalar / 127.f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; const __m256 mul = _mm256_set1_ps( id ); @@ -401,7 +402,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Quantize these floats const float d = max_scalar / 127.f; - y[i].d = GGML_FP32_TO_FP16(d); + y[i].d = GGML_CPU_FP32_TO_FP16(d); const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f; const __m256 mul = _mm256_set1_ps( id ); @@ -425,7 +426,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i #if defined(__AVX2__) // Compute the sum of the quants and set y[i].s - y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)))); + y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)))); // Convert int32 to int16 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 @@ -455,7 +456,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i // Compute the sum of the quants and set y[i].s const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); - y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1))); + y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1))); // Convert int32 to int16 ni0 = _mm_packs_epi32( ni0, ni1 ); @@ -552,7 +553,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); + const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) ); __m256i qx = bytes_from_nibbles_32(x[ib].qs); @@ -613,7 +614,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 0 and 1 - const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) ); + const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) ); const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs); @@ -631,7 +632,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0); // Compute combined scale for the block 2 and 3 - const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) ); + const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) ); const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs); @@ -680,7 +681,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -711,10 +712,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { - const float d0 = GGML_FP16_TO_FP32(x[ib].d); - const float d1 = GGML_FP16_TO_FP32(y[ib].d); + const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d); + const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); const __m256 d0v = _mm256_set1_ps( d0 ); const __m256 d1v = _mm256_set1_ps( d1 ); @@ -752,7 +753,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -783,7 +784,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); __m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i bxhi = bytes_from_bits_32(x[ib].qh); @@ -807,7 +808,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { /* Compute combined scale for the block */ - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs); const __m256i bxhi = bytes_from_bits_32(x[ib].qh); @@ -851,7 +852,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -883,16 +884,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { - const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d)); + const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d)); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); __m256i qx = bytes_from_nibbles_32(x[ib].qs); __m256i bxhi = bytes_from_bits_32(x[ib].qh); bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); qx = _mm256_or_si256(qx, bxhi); - const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d)); + const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d)); const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs); const __m256 q = mul_sum_us8_pairs_float(qx, qy); @@ -910,9 +911,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { - const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d)); + const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d)); - summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s); __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs); const __m256i bxhi = bytes_from_bits_32(x[ib].qh); @@ -926,7 +927,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi bxh = _mm_or_si128(bxh, bxhih); bx_0 = MM256_SET_M128I(bxh, bxl); - const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d)); + const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d)); const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs); const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0); @@ -956,7 +957,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -986,7 +987,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi // Main loop for (; ib < nb; ++ib) { // Compute combined scale for the block - const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs); __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs); @@ -1025,7 +1026,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -1144,7 +1145,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums); - const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); + const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d)); sumi0 = _mm256_sub_epi16(sumi0, ysum); sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2)); @@ -1190,7 +1191,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } } - sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); + sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); } *s = sumf; @@ -1244,7 +1245,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums); - const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)); + const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d)); sumi0 = _mm256_add_epi16(sumi0, sumi1); sumi0 = _mm256_sub_epi16(sumi0, ysum); @@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo } } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); sumf += (float) sumi * d; } @@ -1299,8 +1300,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1366,8 +1367,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1477,8 +1478,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -1533,7 +1534,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q3 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1638,7 +1639,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q3 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -1824,7 +1825,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -1862,8 +1863,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -1928,8 +1929,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q4 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2049,9 +2050,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2092,8 +2093,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi const uint8_t * GGML_RESTRICT q5 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -2170,8 +2171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); const uint8_t * GGML_RESTRICT q5 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2311,9 +2312,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2344,7 +2345,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -2422,7 +2423,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int i = 0; i < nb; ++i) { - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); const uint8_t * GGML_RESTRICT q4 = x[i].ql; const uint8_t * GGML_RESTRICT qh = x[i].qh; @@ -2555,7 +2556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -2622,7 +2623,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = _mm256_setzero_si256(); @@ -2663,7 +2664,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; __m128i sumi1_0 = _mm_setzero_si128(); @@ -2717,7 +2718,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; @@ -2792,7 +2793,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -2913,7 +2914,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3035,7 +3036,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT sc = x[i].scales; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3104,7 +3105,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); @@ -3177,7 +3178,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); @@ -3253,7 +3254,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const int8_t * q8 = y[i].qs; const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; @@ -3313,7 +3314,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3358,7 +3359,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3414,7 +3415,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -3480,7 +3481,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; @@ -3565,7 +3566,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; @@ -3648,7 +3649,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT signs = x[i].signs; @@ -3753,7 +3754,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum); accum1 += d * sumi1; @@ -3801,7 +3802,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2; } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum); accum1 += d * sumi1; @@ -3835,7 +3836,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 4; } - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); } *s = sumf; @@ -3947,7 +3948,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 8; qh += 4; } - const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); + const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16)); accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1); accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2); @@ -4033,7 +4034,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qs += 8; qh += 4; } - const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16)); + const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16)); accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1); accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2); @@ -4083,7 +4084,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo qh += 2; } - sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); + sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); } *s = sumf; @@ -4129,9 +4130,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const __m256i p_1 = _mm256_madd_epi16(p16_1, mone); const __m256i p_2 = _mm256_madd_epi16(p16_2, mone); - accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)), + accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)), _mm256_cvtepi32_ps(p_1), accum1); - accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)), + accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)), _mm256_cvtepi32_ps(p_2), accum2); } @@ -4164,7 +4165,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v #endif for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -4219,7 +4220,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v sumi1 = _mm256_add_epi32(p_1, sumi1); sumi2 = _mm256_add_epi32(p_2, sumi2); } - accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), + accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum); } @@ -4267,7 +4268,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v } __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0); __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1); - accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d), + accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum); } @@ -4276,7 +4277,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v #else float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp index e7635a294..c00c1e541 100644 --- a/ggml/src/ggml-cpu/arch/x86/repack.cpp +++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp @@ -6,6 +6,7 @@ #include "ggml-impl.h" #include "ggml-cpu.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "traits.h" #include @@ -39,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) { float tmp[16]; for (int i = 0; i < 8; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]); } for (int i = 0; i < 8; i++) { - tmp[i + 8] = GGML_FP16_TO_FP32(y[i]); + tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]); } return _mm512_loadu_ps(tmp); @@ -54,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) { _mm_storeu_si128((__m128i*)tmphalf, x); for (int i = 0; i < 4; i++) { - tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]); - tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]); - tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]); - tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]); + tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]); + tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]); + tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]); } return _mm512_loadu_ps(tmp); @@ -67,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) { float tmp[8]; for (int i = 0; i < 8; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]); } return _mm256_loadu_ps(tmp); @@ -76,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) { float tmp[8]; for (int i = 0; i < 4; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); - tmp[i + 4] = GGML_FP16_TO_FP32(x[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]); + tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]); } return _mm256_loadu_ps(tmp); @@ -88,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask)); for (int i = 0; i < 8; i++) { - tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]); } return _mm256_loadu_ps(tmp); @@ -211,7 +212,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f; // Store the scale for the individual block - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); // Store the values in blocks of eight values - Aim is to use these later for block interleaving srcv[row_iter][0] = v0; @@ -297,7 +298,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < QK8_0 * 4; j++) { @@ -647,7 +648,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask); // Load and convert to FP32 scale from block_q8_0 - const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d)); + const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d)); // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs)); @@ -706,7 +707,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -972,13 +973,13 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi2 = sumi2 * scales_1[j]; sumi += sumi1 + sumi2; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; } } for (int sb = 0; sb < 8; sb++) { uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; for (int j = 0; j < ncols_interleaved; j++) { - sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; + sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; } } } @@ -1755,7 +1756,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -3259,7 +3260,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo sumi2 = sumi2 * scales_1[j]; sumi += sumi1 + sumi2; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; } } } @@ -3268,7 +3269,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo for(int m = 0; m < 4; m++) { const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); for(int j = 0; j < ncols_interleaved; j++) { - sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; + sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; } } } diff --git a/ggml/src/ggml-cpu/common.h b/ggml/src/ggml-cpu/common.h index 5624176cc..353563dc3 100644 --- a/ggml/src/ggml-cpu/common.h +++ b/ggml/src/ggml-cpu/common.h @@ -4,6 +4,7 @@ #include "traits.h" #include "ggml-cpu-impl.h" #include "ggml-impl.h" +#include "simd-mappings.h" #ifdef __cplusplus @@ -12,11 +13,11 @@ // convenience functions/macros for use in template calls // note: these won't be required after the 'traits' lookup table is used. static inline ggml_fp16_t f32_to_f16(float x) { - return GGML_FP32_TO_FP16(x); + return GGML_CPU_FP32_TO_FP16(x); } static inline float f16_to_f32(ggml_fp16_t x) { - return GGML_FP16_TO_FP32(x); + return GGML_CPU_FP16_TO_FP32(x); } static inline ggml_bf16_t f32_to_bf16(float x) { diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h index 73a8f9398..d839cf5c5 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-impl.h +++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h @@ -62,11 +62,17 @@ struct ggml_compute_params { #if defined(__s390x__) && defined(__VEC__) #ifndef __VXE__ #define __VXE__ -#endif +#endif // __VXE__ #ifndef __VXE2__ #define __VXE2__ -#endif -#endif +#endif // __VXE2__ +#endif // __s390x__ && __VEC__ + +#if defined(__s390x__) && defined(GGML_NNPA) +#ifndef __NNPA__ +#define __NNPA__ +#endif // __NNPA__ +#endif // __s390x__ && GGML_NNPA #if defined(__ARM_FEATURE_SVE) #include diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 1d3cd009a..7cae96f4b 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -72,6 +72,9 @@ #define UNUSED GGML_UNUSED #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0) +// precomputed f32 table for f16 (256 KB) (simd-mappings.h) +float ggml_table_f32_f16[1 << 16]; + #if defined(__ARM_ARCH) struct ggml_arm_arch_features_type { int sve_cnt; @@ -736,7 +739,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) { { assert(tensor->nb[0] == sizeof(ggml_fp16_t)); for (int i = 0; i < n; i++) { - ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value)); + ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value)); } } break; case GGML_TYPE_BF16: @@ -795,7 +798,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) { { assert(tensor->nb[0] == sizeof(ggml_fp16_t)); for (int i = 0; i < n; i++) { - ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value)); + ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value)); } } break; case GGML_TYPE_BF16: @@ -846,7 +849,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) { case GGML_TYPE_F16: { GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t)); - return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); + return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); } case GGML_TYPE_BF16: { @@ -891,7 +894,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) { case GGML_TYPE_F16: { GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t)); - ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value); + ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value); } break; case GGML_TYPE_BF16: { @@ -920,7 +923,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i case GGML_TYPE_I32: return ((int32_t *) data)[0]; case GGML_TYPE_F16: - return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); + return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); case GGML_TYPE_BF16: return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]); case GGML_TYPE_F32: @@ -947,7 +950,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, } break; case GGML_TYPE_F16: { - ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value); + ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value); } break; case GGML_TYPE_BF16: { @@ -985,7 +988,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) { } case GGML_TYPE_F16: { - return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); + return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]); } case GGML_TYPE_BF16: { @@ -1024,7 +1027,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) { } break; case GGML_TYPE_F16: { - ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value); + ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value); } break; case GGML_TYPE_BF16: { @@ -1051,7 +1054,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, case GGML_TYPE_I32: return ((int32_t *) data)[0]; case GGML_TYPE_F16: - return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); + return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]); case GGML_TYPE_BF16: return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]); case GGML_TYPE_F32: @@ -1078,7 +1081,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, } break; case GGML_TYPE_F16: { - ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value); + ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value); } break; case GGML_TYPE_BF16: { @@ -3141,9 +3144,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) { __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); _mm_storel_epi64((__m128i *)(y + i), y_vec); } +#elif defined(__NNPA__) + for (; i + 7 < n; i += 8) { + float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0)); + float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4)); + uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0); + uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0); + vec_xst(v_y, 0, (ggml_fp16_t *)(y + i)); + } + for (; i + 3 < n; i += 4) { + float32x4_t v_x = vec_xl(0, (const float *)(x + i)); + float32x4_t v_zero = vec_splats(0.0f); + uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0); + uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0); + vec_xst(v_y, 0, (ggml_fp16_t *)(y + i)); + } #endif for (; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16(x[i]); } } @@ -3167,9 +3185,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) { __m128 y_vec = _mm_cvtph_ps(x_vec); _mm_storeu_ps(y + i, y_vec); } +#elif defined(__NNPA__) + for (; i + 7 < n; i += 8) { + uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i)); + uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0); + float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0); + float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0); + vec_xst(v_yh, 0, (float *)(y + i + 0)); + vec_xst(v_yl, 0, (float *)(y + i + 4)); + } + for (; i + 3 < n; i += 4) { + uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i)); + uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0); + float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0); + vec_xst(v_yh, 0, (float *)(y + i)); + } #endif + for (; i < n; ++i) { - y[i] = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP16_TO_FP32(x[i]); } } @@ -3369,6 +3403,14 @@ int ggml_cpu_has_vxe(void) { #endif } +int ggml_cpu_has_nnpa(void) { +#if defined(GGML_NNPA) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_neon(void) { #if defined(__ARM_ARCH) && defined(__ARM_NEON) return 1; @@ -3418,7 +3460,7 @@ int ggml_cpu_has_sme(void) { } void ggml_cpu_init(void) { - // needed to initialize f16 tables + // needed to initialize ggml_time { struct ggml_init_params params = { 0, NULL, false }; struct ggml_context * ctx = ggml_init(params); @@ -3439,9 +3481,10 @@ void ggml_cpu_init(void) { uint16_t u16; ggml_fp16_t fp16; } u = {i}; - float f = GGML_FP16_TO_FP32(u.fp16); - ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); - ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f)); + float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16); + ggml_table_f32_f16[i] = f; + ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f)); + ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f)); } const uint64_t t_end = ggml_time_us(); UNUSED(t_end); diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index 735ef3f01..a98866a2d 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -578,6 +578,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r if (ggml_cpu_has_vxe()) { features.push_back({ "VXE", "1" }); } + if (ggml_cpu_has_nnpa()) { + features.push_back({ "NNPA", "1" }); + } if (ggml_cpu_has_wasm_simd()) { features.push_back({ "WASM_SIMD", "1" }); } diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index 7ed3874af..ed61869a5 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -52,6 +52,7 @@ #include "ggml-impl.h" #include "ggml-cpu-impl.h" #include "ggml-quants.h" +#include "simd-mappings.h" #include #include @@ -73,7 +74,7 @@ namespace { inline float unhalf(ggml_fp16_t d) { - return GGML_FP16_TO_FP32(d); + return GGML_CPU_FP16_TO_FP32(d); } //////////////////////////////////////////////////////////////////////////////////////////////////// @@ -252,7 +253,7 @@ template <> inline float32x4_t load(const ggml_fp16_t * p) { float tmp[4]; for (int i = 0; i < 4; i++) { - tmp[i] = GGML_FP16_TO_FP32(p[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(p[i]); } return vec_xl(0, (const float *)(tmp)); diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index eff4a53e3..8531baf6c 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -108,7 +108,7 @@ static void ggml_compute_forward_dup_f16( for (int i01 = ir0; i01 < ir1; i01++) { const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); for (int i00 = 0; i00 < ne00; i00++) { - dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); + dst_ptr[id] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]); id++; } } @@ -130,7 +130,7 @@ static void ggml_compute_forward_dup_f16( const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); for (int i00 = 0; i00 < ne00; i00++) { - src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); + src0_f32[i00] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]); } quantize_row_q(src0_f32, dst_ptr + id, ne00); @@ -156,7 +156,7 @@ static void ggml_compute_forward_dup_f16( for (int i00 = 0; i00 < ne00; i00++) { const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); + dst_ptr[id] = GGML_CPU_FP16_TO_FP32(*src0_ptr); id++; } } @@ -267,7 +267,7 @@ static void ggml_compute_forward_dup_f16( const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); + *(float *) dst_ptr = GGML_CPU_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); if (++i10 == ne0) { i10 = 0; @@ -372,7 +372,7 @@ static void ggml_compute_forward_dup_bf16( for (int i01 = ir0; i01 < ir1; i01++) { const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); for (int i00 = 0; i00 < ne00; i00++) { - dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); + dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); id++; } } @@ -473,7 +473,7 @@ static void ggml_compute_forward_dup_bf16( for (int i00 = 0; i00 < ne00; i00++) { const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); + dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); id++; } } @@ -566,7 +566,7 @@ static void ggml_compute_forward_dup_bf16( const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); + *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); if (++i10 == ne0) { i10 = 0; @@ -765,7 +765,7 @@ static void ggml_compute_forward_dup_f32( for (int i00 = 0; i00 < ne00; i00++) { const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); + dst_ptr[id] = GGML_CPU_FP32_TO_FP16(*src0_ptr); id++; } } @@ -878,7 +878,7 @@ static void ggml_compute_forward_dup_f32( const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); + *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(*(const float *) src0_ptr); if (++i10 == ne0) { i10 = 0; @@ -1419,7 +1419,7 @@ static void ggml_compute_forward_add1_f16_f32( ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v); } } } @@ -1435,7 +1435,7 @@ static void ggml_compute_forward_add1_f16_f16( GGML_ASSERT(ggml_is_scalar(src1)); // scalar to add - const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); + const float v = GGML_CPU_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); const int ith = params->ith; const int nth = params->nth; @@ -1467,7 +1467,7 @@ static void ggml_compute_forward_add1_f16_f16( ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v); } } } @@ -1889,7 +1889,7 @@ static void ggml_compute_forward_sum_f16( } } } - ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); + ((ggml_fp16_t *) dst->data)[0] = GGML_CPU_FP32_TO_FP16(sum); } static void ggml_compute_forward_sum_bf16( @@ -2660,7 +2660,7 @@ static void ggml_compute_forward_gelu_f16( #ifndef NDEBUG for (int k = 0; k < nc; k++) { const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); GGML_UNUSED(v); assert(!isnan(v)); assert(!isinf(v)); @@ -2763,7 +2763,7 @@ static void ggml_compute_forward_gelu_erf_f16( #ifndef NDEBUG for (int k = 0; k < nc; k++) { const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); GGML_UNUSED(v); assert(!isnan(v)); assert(!isinf(v)); @@ -2866,7 +2866,7 @@ static void ggml_compute_forward_gelu_quick_f16( #ifndef NDEBUG for (int k = 0; k < nc; k++) { const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); GGML_UNUSED(v); assert(!isnan(v)); assert(!isinf(v)); @@ -2969,7 +2969,7 @@ static void ggml_compute_forward_silu_f16( #ifndef NDEBUG for (int k = 0; k < nc; k++) { const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k]; - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); GGML_UNUSED(v); assert(!isnan(v)); assert(!isinf(v)); @@ -3163,7 +3163,7 @@ static void ggml_compute_forward_silu_back_f16( #ifndef NDEBUG for (int k = 0; k < nc; k++) { const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); GGML_UNUSED(v); assert(!isnan(v)); assert(!isinf(v)); @@ -4500,7 +4500,7 @@ static void ggml_compute_forward_get_rows_back_f32_f16( for (int j = 0; j < nc; ++j) { ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; - ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); + ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_CPU_FP16_TO_FP32(v); } } } @@ -4792,7 +4792,7 @@ static void ggml_compute_forward_soft_max_f32( if (mp_f32) { if (use_f16) { for (int i = 0; i < nc; ++i) { - wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]); + wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]); } } else { for (int i = 0; i < nc; ++i) { @@ -5018,8 +5018,8 @@ static void ggml_compute_forward_clamp_f16( ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); for (int i = 0; i < nc; i++) { - float v = GGML_FP16_TO_FP32(src0_ptr[i]); - dst_ptr[i] = GGML_FP32_TO_FP16(MAX(MIN(v, max), min)); + float v = GGML_CPU_FP16_TO_FP32(src0_ptr[i]); + dst_ptr[i] = GGML_CPU_FP32_TO_FP16(MAX(MIN(v, max), min)); } } } @@ -5476,11 +5476,11 @@ static void ggml_compute_forward_rope_f16( const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims]); + const float x0 = GGML_CPU_FP16_TO_FP32(src[0]); + const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } else { for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { @@ -5492,11 +5492,11 @@ static void ggml_compute_forward_rope_f16( const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + const float x0 = GGML_CPU_FP16_TO_FP32(src[0]); + const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims/2]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } } else { @@ -5507,11 +5507,11 @@ static void ggml_compute_forward_rope_f16( const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[1]); + const float x0 = GGML_CPU_FP16_TO_FP32(src[0]); + const float x1 = GGML_CPU_FP16_TO_FP32(src[1]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } @@ -5525,11 +5525,11 @@ static void ggml_compute_forward_rope_f16( const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims]); + const float x0 = GGML_CPU_FP16_TO_FP32(src[0]); + const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]); - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); } } else { for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { @@ -5640,7 +5640,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( for (int64_t i11 = 0; i11 < ne11; i11++) { const float * const src = (float *)((char *) src1->data + i11*nb11); for (int64_t i10 = 0; i10 < ne10; i10++) { - dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]); + dst_data[i10*ne11 + i11] = GGML_CPU_FP32_TO_FP16(src[i10]); } } } @@ -5933,7 +5933,7 @@ static void ggml_compute_forward_im2col_f16( if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; } else { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); + dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]); } } } @@ -6109,7 +6109,7 @@ void ggml_compute_forward_conv_transpose_2d( const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11); ggml_fp16_t * dst_data = wdata + i11*ne10*ne12; for (int i10 = 0; i10 < ne10; i10++) { - dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]); + dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]); } } } @@ -6358,7 +6358,7 @@ static void ggml_compute_forward_pool_1d_sk_p0( case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); } for (int ki = 0; ki < k; ++ki) { - const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); switch (op) { case GGML_OP_POOL_AVG: drow[i] += srow_j; break; case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break; @@ -6450,7 +6450,7 @@ void ggml_compute_forward_pool_2d( for (int kx = 0; kx < k0; ++kx) { int j = ix + kx; if (j < 0 || j >= src->ne[0]) continue; - const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); switch (op) { case GGML_OP_POOL_AVG: *out += srow_j; break; case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break; @@ -6538,7 +6538,7 @@ void ggml_compute_forward_pool_2d_back( } const float val = dst->type == GGML_TYPE_F32 ? - ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); + ((const float *) drowf)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); if (val <= maxval) { continue; } @@ -6558,7 +6558,7 @@ void ggml_compute_forward_pool_2d_back( if (dst->type == GGML_TYPE_F32) { ((float *) drow)[j] += grad0; } else { - ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); + ((ggml_fp16_t *) drow)[j] = GGML_CPU_FP32_TO_FP16(grad0 + GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); } } else if (op == GGML_OP_POOL_AVG) { const float grad = grad0 / ka; @@ -6577,7 +6577,7 @@ void ggml_compute_forward_pool_2d_back( if (dst->type == GGML_TYPE_F32) { ((float *) drow)[j] += grad; } else { - ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); + ((ggml_fp16_t *) drow)[j] += GGML_CPU_FP32_TO_FP16(grad); } } } @@ -7142,7 +7142,7 @@ static void ggml_compute_forward_flash_attn_ext_f16( // loop over n_kv and n_head_kv // ref: https://arxiv.org/pdf/2112.05682.pdf for (int64_t ic = 0; ic < nek1; ++ic) { - const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f; + const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f; if (mv == -INFINITY) { continue; } @@ -7210,7 +7210,7 @@ static void ggml_compute_forward_flash_attn_ext_f16( if (v->type == GGML_TYPE_F16) { for (int64_t d = 0; d < DV; ++d) { - VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]); + VKQ32[d] = GGML_CPU_FP16_TO_FP32(VKQ16[d]); } } diff --git a/ggml/src/ggml-cpu/quants.c b/ggml/src/ggml-cpu/quants.c index d2e705f28..ee35ab42f 100644 --- a/ggml/src/ggml-cpu/quants.c +++ b/ggml/src/ggml-cpu/quants.c @@ -2,6 +2,7 @@ #include "ggml-common.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "ggml-quants.h" #include "quants.h" @@ -137,7 +138,7 @@ void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c } int sumi = sumi0 + sumi1; - sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d); + sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); } *s = sumf; @@ -174,7 +175,7 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -217,7 +218,7 @@ void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; } *s = sumf; @@ -260,7 +261,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c } int sumi = sumi0 + sumi1; - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); + sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); } *s = sumf; @@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c sumi += x[ib].qs[j]*y[ib].qs[j]; } - sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)); + sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); } *s = sumf; @@ -342,7 +343,7 @@ void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, } } - sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d); + sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); } *s = sumf; @@ -372,7 +373,7 @@ void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, } } - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); sumf += (float) sumi * d; } @@ -405,8 +406,8 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c summs += y[i].bsums[j] * (sc[j] >> 4); } - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); int isum = 0; int is = 0; @@ -504,7 +505,7 @@ void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -577,9 +578,9 @@ void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -657,9 +658,9 @@ void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; - const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; sumf -= dmin * sumi; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -714,7 +715,7 @@ void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; q8 += 8; a += 8; } - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; } for (int l = 0; l < 8; ++l) sumf += sums[l]; @@ -739,7 +740,7 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; @@ -778,7 +779,7 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint16_t * GGML_RESTRICT q2 = x[i].qs; const uint8_t * GGML_RESTRICT sc = x[i].scales; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -829,7 +830,7 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, float sumf = 0; for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const int8_t * q8 = y[i].qs; const uint8_t * qs = x[i].qs; const uint8_t * qh = x[i].qh; @@ -882,7 +883,7 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT q3 = x[i].qs; const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; const int8_t * GGML_RESTRICT q8 = y[i].qs; @@ -924,7 +925,7 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, float sumf = 0.f; for (int i = 0; i < nb; ++i) { - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; const uint8_t * GGML_RESTRICT qs = x[i].qs; const uint8_t * GGML_RESTRICT qh = x[i].qh; const uint8_t * GGML_RESTRICT signs = x[i].signs; @@ -1002,7 +1003,7 @@ void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, qs += 4; } - sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); + sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); } *s = sumf; @@ -1063,7 +1064,7 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, qh += 2; } - sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); + sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); } *s = sumf; @@ -1087,7 +1088,7 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, float sumf = 0; for (; ib < nb; ++ib) { - const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); + const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); int sumi1 = 0, sumi2 = 0; for (int j = 0; j < QK4_NL/2; ++j) { sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; @@ -1113,7 +1114,7 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, float sumf = 0; for (int ibl = 0; ibl < nb; ++ibl) { - const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d; + const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; uint16_t h = x[ibl].scales_h; const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 692c53e01..72ee93a5a 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -6,6 +6,7 @@ #include "ggml-impl.h" #include "ggml-cpu.h" #include "ggml-cpu-impl.h" +#include "simd-mappings.h" #include "traits.h" #include "arch-fallback.h" @@ -72,7 +73,7 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < QK8_0 * 4; j++) { @@ -110,7 +111,7 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG const float d = amax / ((1 << 7) - 1); id[row_iter] = d ? 1.0f / d : 0.0f; - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); + y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d); } for (int j = 0; j < QK8_0 * 4; j++) { @@ -236,7 +237,7 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -280,7 +281,7 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -325,7 +326,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -396,13 +397,13 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, sumi2 = sumi2 * scales_1[j]; sumi += sumi1 + sumi2; } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; } } for (int sb = 0; sb < 8; sb++) { uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; for (int j = 0; j < ncols_interleaved; j++) { - sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; + sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; } } } @@ -449,7 +450,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } } @@ -500,7 +501,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -555,7 +556,7 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -609,7 +610,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } @@ -688,7 +689,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, sumi2 = sumi2 * scales_1[j]; sumi += sumi1 + sumi2; } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; } } } @@ -697,7 +698,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, for(int m = 0; m < 4; m++) { const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); for(int j = 0; j < ncols_interleaved; j++) { - sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; + sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; } } } @@ -753,7 +754,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); } } } diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h index e42364c59..b68ac0dd6 100644 --- a/ggml/src/ggml-cpu/simd-mappings.h +++ b/ggml/src/ggml-cpu/simd-mappings.h @@ -2,10 +2,167 @@ #include "ggml-cpu-impl.h" +#ifdef __ARM_FEATURE_SVE +#include +#endif // __ARM_FEATURE_SVE + +#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__) +// if YCM cannot find , make a symbolic link to it, for example: +// +// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ +// +#include +#endif + +#if defined(__F16C__) +#include +#endif + +#ifdef __cplusplus +extern "C" { +#endif + // // simd mappings // +// FP16 to FP32 conversion + +// 16-bit float +// on Arm, we use __fp16 +// on x86, we use uint16_t +// +// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616 +// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843 +// +#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x) + + #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x) + + static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) { + __fp16 tmp; + memcpy(&tmp, &h, sizeof(ggml_fp16_t)); + return (float)tmp; + } + + static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + __fp16 tmp = f; + memcpy(&res, &tmp, sizeof(ggml_fp16_t)); + return res; + } +#elif defined(__F16C__) + #ifdef _MSC_VER + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) + #else + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) + #endif +#elif defined(__POWER9_VECTOR__) + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x) + /* the inline asm below is about 12% faster than the lookup method */ + #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x) + #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x) + + static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) { + float f; + double d; + __asm__( + "mtfprd %0,%2\n" + "xscvhpdp %0,%0\n" + "frsp %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=f"(f): + /* in */ "r"(h)); + return f; + } + + static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) { + double d; + ggml_fp16_t r; + __asm__( /* xscvdphp can work on double or single precision */ + "xscvdphp %0,%2\n" + "mffprd %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=r"(r): + /* in */ "f"(f)); + return r; + } +#elif defined(__riscv) && defined(__riscv_zfhmin) + static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) { + float f; + __asm__( + "fmv.h.x %[f], %[h]\n\t" + "fcvt.s.h %[f], %[f]" + : [f] "=&f" (f) + : [h] "r" (h) + ); + return f; + } + + static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + __asm__( + "fcvt.h.s %[f], %[f]\n\t" + "fmv.x.h %[h], %[f]" + : [h] "=&r" (res) + : [f] "f" (f) + ); + return res; + } + + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x) + #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x) + #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x) +#elif defined(__NNPA__) + #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x) + #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x) + + #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x) + #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x) + + static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) { + uint16x8_t v_h = vec_splats(h); + uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0); + return vec_extend_to_fp32_hi(v_hd, 0)[0]; + } + + static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) { + float32x4_t v_f = vec_splats(f); + float32x4_t v_zero = vec_splats(0.0f); + uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0); + uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0); + return vec_extract(v_h, 0); + } +#endif + +// precomputed f32 table for f16 (256 KB) +// defined in ggml-cpu.c, initialized in ggml_cpu_init() +extern float ggml_table_f32_f16[1 << 16]; + +// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, +// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON. +// This is also true for POWER9. +#if !defined(GGML_CPU_FP16_TO_FP32) +inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { + uint16_t s; + memcpy(&s, &f, sizeof(uint16_t)); + return ggml_table_f32_f16[s]; +} + +#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) +#endif + +#if !defined(GGML_CPU_FP32_TO_FP16) +#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) +#endif + + // we define a common set of C macros which map to specific intrinsics based on the current architecture // we then implement the fundamental computation operations below using only these macros // adding support for new architectures requires to define the corresponding SIMD macros @@ -415,7 +572,7 @@ static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) { float tmp[8]; for (int i = 0; i < 8; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]); } return _mm256_loadu_ps(tmp); @@ -426,7 +583,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) { _mm256_storeu_ps(arr, y); for (int i = 0; i < 8; i++) - x[i] = GGML_FP32_TO_FP16(arr[i]); + x[i] = GGML_CPU_FP32_TO_FP16(arr[i]); } #define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x) #define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y) @@ -574,10 +731,10 @@ static inline unsigned char ggml_endian_byte(int i) { inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) { float tmp[4]; - tmp[0] = GGML_FP16_TO_FP32(p[0]); - tmp[1] = GGML_FP16_TO_FP32(p[1]); - tmp[2] = GGML_FP16_TO_FP32(p[2]); - tmp[3] = GGML_FP16_TO_FP32(p[3]); + tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]); + tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]); + tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]); + tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]); return wasm_v128_load(tmp); } @@ -587,10 +744,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) { wasm_v128_store(tmp, x); - p[0] = GGML_FP32_TO_FP16(tmp[0]); - p[1] = GGML_FP32_TO_FP16(tmp[1]); - p[2] = GGML_FP32_TO_FP16(tmp[2]); - p[3] = GGML_FP32_TO_FP16(tmp[3]); + p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]); + p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]); + p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]); + p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]); } #define GGML_F16x4 v128_t @@ -690,10 +847,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) { static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) { float tmp[4]; - tmp[0] = GGML_FP16_TO_FP32(x[0]); - tmp[1] = GGML_FP16_TO_FP32(x[1]); - tmp[2] = GGML_FP16_TO_FP32(x[2]); - tmp[3] = GGML_FP16_TO_FP32(x[3]); + tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]); + tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]); + tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]); + tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]); return _mm_loadu_ps(tmp); } @@ -703,10 +860,10 @@ static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) { _mm_storeu_ps(arr, y); - x[0] = GGML_FP32_TO_FP16(arr[0]); - x[1] = GGML_FP32_TO_FP16(arr[1]); - x[2] = GGML_FP32_TO_FP16(arr[2]); - x[3] = GGML_FP32_TO_FP16(arr[3]); + x[0] = GGML_CPU_FP32_TO_FP16(arr[0]); + x[1] = GGML_CPU_FP32_TO_FP16(arr[1]); + x[2] = GGML_CPU_FP32_TO_FP16(arr[2]); + x[3] = GGML_CPU_FP32_TO_FP16(arr[3]); } #define GGML_F32Cx4 __m128 @@ -828,7 +985,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { #define GGML_F32x4_ZERO __lsx_vldi(0) #define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) #define GGML_F32x4_LOAD(x) __lsx_vld((x), 0) -#define GGML_F32x4_STORE((x),(y)) __lsx_vst((y), (x), 0) +#define GGML_F32x4_STORE(x, y) __lsx_vst(y, x, 0) #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a) #define GGML_F32x4_ADD __lsx_vfadd_s #define GGML_F32x4_MUL __lsx_vfmul_s @@ -874,10 +1031,10 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { float tmp[4]; - tmp[0] = GGML_FP16_TO_FP32(x[0]); - tmp[1] = GGML_FP16_TO_FP32(x[1]); - tmp[2] = GGML_FP16_TO_FP32(x[2]); - tmp[3] = GGML_FP16_TO_FP32(x[3]); + tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]); + tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]); + tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]); + tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]); return __lsx_vld(tmp, 0); } @@ -887,10 +1044,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { __lsx_vst(y, arr, 0); - x[0] = GGML_FP32_TO_FP16(arr[0]); - x[1] = GGML_FP32_TO_FP16(arr[1]); - x[2] = GGML_FP32_TO_FP16(arr[2]); - x[3] = GGML_FP32_TO_FP16(arr[3]); + x[0] = GGML_CPU_FP32_TO_FP16(arr[0]); + x[1] = GGML_CPU_FP32_TO_FP16(arr[1]); + x[2] = GGML_CPU_FP32_TO_FP16(arr[2]); + x[3] = GGML_CPU_FP32_TO_FP16(arr[3]); } #define GGML_F32Cx4 __m128 @@ -922,7 +1079,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { #define GGML_F32_STEP 32 #define GGML_F32_EPR 4 -#define GGML_F32x4 __vector float +#define GGML_F32x4 float32x4_t #define GGML_F32x4_ZERO vec_splats(0.0f) #define GGML_F32x4_SET1 vec_splats #define GGML_F32x4_LOAD(p) vec_xl(0, p) @@ -962,28 +1119,45 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { #define GGML_F16_STEP GGML_F32_STEP #define GGML_F16_EPR GGML_F32_EPR -static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) { +static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) { +#if defined(__NNPA__) + uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x); + uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0); + return vec_extend_to_fp32_hi(v_xd, 0); +#else float tmp[4]; for (int i = 0; i < 4; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); + tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]); } // note: keep type-cast here to prevent compiler bugs // see: https://github.com/ggml-org/llama.cpp/issues/12846 return vec_xl(0, (const float *)(tmp)); +#endif } -static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) { +static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) { +#if defined(__NNPA__) + float32x4_t v_zero = vec_splats(0.0f); + uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0); + uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0); + + x[0] = vec_extract(v_x, 0); + x[1] = vec_extract(v_x, 1); + x[2] = vec_extract(v_x, 2); + x[3] = vec_extract(v_x, 3); +#else float arr[4]; // note: keep type-cast here to prevent compiler bugs // see: https://github.com/ggml-org/llama.cpp/issues/12846 - vec_xst(y, 0, (float *)(arr)); + vec_xst(v_y, 0, (float *)(arr)); for (int i = 0; i < 4; i++) { - x[i] = GGML_FP32_TO_FP16(arr[i]); + x[i] = GGML_CPU_FP32_TO_FP16(arr[i]); } +#endif } #define GGML_F16_VEC GGML_F32x4 @@ -1004,3 +1178,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) { #define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR) #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR) #endif + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp index f7614568e..5e34d79a1 100644 --- a/ggml/src/ggml-cpu/vec.cpp +++ b/ggml/src/ggml-cpu/vec.cpp @@ -219,11 +219,11 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G // leftovers for (int i = np; i < n; ++i) { - sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); + sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i])); } #else for (int i = 0; i < n; ++i) { - sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); + sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i])); } #endif diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h index 09dbade21..84f6c0e6d 100644 --- a/ggml/src/ggml-cpu/vec.h +++ b/ggml/src/ggml-cpu/vec.h @@ -58,7 +58,7 @@ inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { for (int i = 0; i < n; ++i) { - z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i])); + z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i])); } } inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } @@ -67,7 +67,7 @@ inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { for (int i = 0; i < n; ++i) { - z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i])); + z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i])); } } inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; } @@ -75,20 +75,20 @@ inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; } inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i])); + y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_FP16_TO_FP32(x[i])); } } inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; } inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { for (int i = 0; i < n; ++i) { - z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i])); + z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i])); } } inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; } inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { for (int i = 0; i < n; ++i) { - z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i])); + z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i])); } } @@ -131,13 +131,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG // leftovers for (int i = np; i < n; ++i) { for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { - sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); + sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i])); } } #else for (int i = 0; i < n; ++i) { for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { - sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); + sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i])); } } #endif @@ -280,12 +280,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, // leftovers for (int i = np; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v); } #else // scalar for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v); } #endif } @@ -430,12 +430,12 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float // leftovers for (int i = np; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v); } #else // scalar for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v); } #endif } @@ -444,103 +444,103 @@ inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16(v*v); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16(v*v); } } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); } inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); } inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f)); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f)); } } inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f); + y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f); } } inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); } inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i]))); } } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : 0.f); } } inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f)); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f)); } } inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); } inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i])))); + y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_FP16_TO_FP32(x[i])))); } } // TODO: optimize performance inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f))); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f))); } } inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f))); + y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f))); } } inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); } inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i]))); + y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i]))); } } @@ -562,9 +562,9 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float xi = GGML_FP16_TO_FP32(x[i]); + float xi = GGML_CPU_FP16_TO_FP32(x[i]); float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV)); - y[i] = GGML_FP32_TO_FP16(res); + y[i] = GGML_CPU_FP32_TO_FP16(res); } } @@ -577,9 +577,9 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { } else if (x[i] >= 10.0f) { y[i] = x[i]; } else { - ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]); memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]); + y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]); } } } @@ -613,9 +613,9 @@ inline static float ggml_gelu_quick_f32(float x) { inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { uint16_t t; for (int i = 0; i < n; ++i) { - ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]); memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); + y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); } } #else @@ -628,8 +628,8 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { for (int i = 0; i < n; ++i) { - float v = GGML_FP16_TO_FP32(x[i]); - y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v)))); + float v = GGML_CPU_FP16_TO_FP32(x[i]); + y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v)))); } } @@ -638,8 +638,8 @@ inline static float ggml_silu_f32(float x) { return x/(1.0f + expf(-x)); } inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) { - float v = GGML_FP16_TO_FP32(x); - return GGML_FP32_TO_FP16(v/(1.0f + expf(-v))); + float v = GGML_CPU_FP16_TO_FP32(x); + return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v))); } #if __FINITE_MATH_ONLY__ @@ -888,9 +888,9 @@ inline static float ggml_silu_backward_f32(float x, float dy) { } inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) { - const float v = GGML_FP16_TO_FP32(x); + const float v = GGML_CPU_FP16_TO_FP32(x); const float s = 1.0f/(1.0f + expf(-v)); - return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s))); + return GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s))); } inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { @@ -928,7 +928,7 @@ inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) { float sum = 0.0f; for (int i = 0; i < n; ++i) { - sum += GGML_FP16_TO_FP32(x[i]); + sum += GGML_CPU_FP16_TO_FP32(x[i]); } *s = sum; } diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index 6dc5ce0d9..57761644f 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -317,203 +317,81 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1); GGML_API void * ggml_aligned_malloc(size_t size); GGML_API void ggml_aligned_free(void * ptr, size_t size); -// FP16 to FP32 conversion +// FP16 <-> FP32 +// ref: https://github.com/Maratyszcza/FP16 -// 16-bit float -// on Arm, we use __fp16 -// on x86, we use uint16_t -// -// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616 -// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843 -// -#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) - #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - - #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - - static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - __fp16 tmp; - memcpy(&tmp, &h, sizeof(ggml_fp16_t)); - return (float)tmp; - } - - static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - ggml_fp16_t res; - __fp16 tmp = f; - memcpy(&res, &tmp, sizeof(ggml_fp16_t)); - return res; - } - -#elif defined(__F16C__) - - #ifdef _MSC_VER - #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) - #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) - #else - #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) - #endif - -#elif defined(__POWER9_VECTOR__) - - #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - /* the inline asm below is about 12% faster than the lookup method */ - #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) - #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - - static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - float f; - double d; - __asm__( - "mtfprd %0,%2\n" - "xscvhpdp %0,%0\n" - "frsp %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=f"(f): - /* in */ "r"(h)); - return f; - } - - static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - double d; - ggml_fp16_t r; - __asm__( /* xscvdphp can work on double or single precision */ - "xscvdphp %0,%2\n" - "mffprd %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=r"(r): - /* in */ "f"(f)); - return r; - } - -#elif defined(__riscv) && defined(__riscv_zfhmin) - - static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - float f; - __asm__( - "fmv.h.x %[f], %[h]\n\t" - "fcvt.s.h %[f], %[f]" - : [f] "=&f" (f) - : [h] "r" (h) - ); - return f; - } - - static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - ggml_fp16_t res; - __asm__( - "fcvt.h.s %[f], %[f]\n\t" - "fmv.x.h %[h], %[f]" - : [h] "=&r" (res) - : [f] "f" (f) - ); - return res; - } - - #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) - #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -#else - - // FP16 <-> FP32 - // ref: https://github.com/Maratyszcza/FP16 - - static inline float fp32_from_bits(uint32_t w) { - union { - uint32_t as_bits; - float as_value; - } fp32; - fp32.as_bits = w; - return fp32.as_value; - } - - static inline uint32_t fp32_to_bits(float f) { - union { - float as_value; - uint32_t as_bits; - } fp32; - fp32.as_value = f; - return fp32.as_bits; - } - - static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - const uint32_t w = (uint32_t) h << 16; - const uint32_t sign = w & UINT32_C(0x80000000); - const uint32_t two_w = w + w; - - const uint32_t exp_offset = UINT32_C(0xE0) << 23; - #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) - const float exp_scale = 0x1.0p-112f; - #else - const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); - #endif - const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; - - const uint32_t magic_mask = UINT32_C(126) << 23; - const float magic_bias = 0.5f; - const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; - - const uint32_t denormalized_cutoff = UINT32_C(1) << 27; - const uint32_t result = sign | - (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); - return fp32_from_bits(result); - } - - static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) - const float scale_to_inf = 0x1.0p+112f; - const float scale_to_zero = 0x1.0p-110f; - #else - const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); - const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); - #endif - float base = (fabsf(f) * scale_to_inf) * scale_to_zero; - - const uint32_t w = fp32_to_bits(f); - const uint32_t shl1_w = w + w; - const uint32_t sign = w & UINT32_C(0x80000000); - uint32_t bias = shl1_w & UINT32_C(0xFF000000); - if (bias < UINT32_C(0x71000000)) { - bias = UINT32_C(0x71000000); - } - - base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; - const uint32_t bits = fp32_to_bits(base); - const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); - const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); - const uint32_t nonsign = exp_bits + mantissa_bits; - return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); - } - - #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) - #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - -#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) - -// precomputed f32 table for f16 (256 KB) -// defined in ggml.c, initialized in ggml_init() -GGML_API float ggml_table_f32_f16[1 << 16]; - -// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, -// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. -// This is also true for POWER9. -#if !defined(GGML_FP16_TO_FP32) -inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { - uint16_t s; - memcpy(&s, &f, sizeof(uint16_t)); - return ggml_table_f32_f16[s]; +static inline float fp32_from_bits(uint32_t w) { + union { + uint32_t as_bits; + float as_value; + } fp32; + fp32.as_bits = w; + return fp32.as_value; } -#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) -#endif +static inline uint32_t fp32_to_bits(float f) { + union { + float as_value; + uint32_t as_bits; + } fp32; + fp32.as_value = f; + return fp32.as_bits; +} -#if !defined(GGML_FP32_TO_FP16) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + const uint32_t w = (uint32_t) h << 16; + const uint32_t sign = w & UINT32_C(0x80000000); + const uint32_t two_w = w + w; + + const uint32_t exp_offset = UINT32_C(0xE0) << 23; +#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) + const float exp_scale = 0x1.0p-112f; +#else + const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); #endif + const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; + + const uint32_t magic_mask = UINT32_C(126) << 23; + const float magic_bias = 0.5f; + const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; + + const uint32_t denormalized_cutoff = UINT32_C(1) << 27; + const uint32_t result = sign | + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); + return fp32_from_bits(result); +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { +#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L) + const float scale_to_inf = 0x1.0p+112f; + const float scale_to_zero = 0x1.0p-110f; +#else + const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); + const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); +#endif + float base = (fabsf(f) * scale_to_inf) * scale_to_zero; + + const uint32_t w = fp32_to_bits(f); + const uint32_t shl1_w = w + w; + const uint32_t sign = w & UINT32_C(0x80000000); + uint32_t bias = shl1_w & UINT32_C(0xFF000000); + if (bias < UINT32_C(0x71000000)) { + bias = UINT32_C(0x71000000); + } + + base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; + const uint32_t bits = fp32_to_bits(base); + const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); + const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); + const uint32_t nonsign = exp_bits + mantissa_bits; + return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); +} + +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + +#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) /** * Converts brain16 to float32. diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index f8e7c595b..ee605977f 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -61,9 +61,6 @@ #define m512i(p) (__m512i)(p) #endif -// precomputed f32 table for f16 (256 KB) (ggml-impl.h) -float ggml_table_f32_f16[1 << 16]; - #if defined(__linux__) || \ defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \ (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH) @@ -1422,14 +1419,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { // initialize time system (required on Windows) ggml_time_init(); - for (int i = 0; i < (1 << 16); ++i) { - union { - uint16_t u16; - ggml_fp16_t fp16; - } u = {i}; - ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16); - } - is_first_call = false; } From 716301d1b03c31875ec3b24526c48c8b1bd0fd8c Mon Sep 17 00:00:00 2001 From: R0CKSTAR Date: Thu, 26 Jun 2025 12:11:59 +0800 Subject: [PATCH 06/12] musa: enable fp16 mma (all) and cublas on qy2 (#13842) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * musa: enable fp16 mma (all) and cublas on qy2 Signed-off-by: Xiaodong Ye * Update ggml/src/ggml-cuda/ggml-cuda.cu Co-authored-by: Johannes Gäßler * Address review comments Signed-off-by: Xiaodong Ye * Address review comments Signed-off-by: Xiaodong Ye * musa: disable MUL_MAT_ID (q2_k × f32) due to precision issues Signed-off-by: Xiaodong Ye --------- Signed-off-by: Xiaodong Ye Co-authored-by: Johannes Gäßler --- ggml/src/ggml-cuda/common.cuh | 25 +++++++++++++------------ ggml/src/ggml-cuda/fattn-wmma-f16.cu | 4 ++++ ggml/src/ggml-cuda/ggml-cuda.cu | 25 +++++++++++++++---------- ggml/src/ggml-musa/mudnn.cuh | 4 ++-- 4 files changed, 34 insertions(+), 24 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index f6127aeee..ea2035502 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -76,11 +76,9 @@ #define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1) // Moore Threads -#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210) - -#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 -#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 -#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD +#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 +#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 +#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD) #define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2) @@ -203,9 +201,9 @@ typedef float2 dfloat2; #define FAST_FP16_AVAILABLE #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610 -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA +#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA) #define FP16_MMA_AVAILABLE -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA +#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA) #if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4))) #define FP16_MMA_AVAILABLE @@ -219,9 +217,9 @@ typedef float2 dfloat2; #define CP_ASYNC_AVAILABLE #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE -#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1) +#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220) #define FLASH_ATTN_AVAILABLE -#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1) +#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220) static bool fp16_available(const int cc) { return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL; @@ -233,7 +231,8 @@ static bool fast_fp16_available(const int cc) { // To be used for feature selection of external libraries, e.g. cuBLAS. static bool fast_fp16_hardware_available(const int cc) { - return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc); + return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc) || + (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2); } // Any FP16 tensor core instructions are available for ggml code. @@ -242,7 +241,8 @@ static bool fp16_mma_available(const int cc) { return false; #else if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) || - GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) { + GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || + GGML_CUDA_CC_IS_MTHREADS(cc)) { return true; } else if (GGML_CUDA_CC_IS_RDNA4(cc)) { #if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12) @@ -259,7 +259,8 @@ static bool fp16_mma_available(const int cc) { // To be used for feature selection of external libraries, e.g. cuBLAS. static bool fp16_mma_hardware_available(const int cc) { return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || - GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc); + GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) || + (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2); } static bool bf16_mma_hardware_available(const int cc) { diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu index c5668adb1..f3b794c36 100644 --- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu +++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu @@ -9,7 +9,11 @@ #ifdef FP16_MMA_AVAILABLE #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) #include +#ifdef GGML_USE_MUSA +namespace wmma = mtmusa::wmma; +#else // GGML_USE_MUSA namespace wmma = nvcuda::wmma; +#endif // GGML_USE_MUSA #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE) #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers #include diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index b3e6833c3..b30c13c62 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -1227,9 +1227,12 @@ static void ggml_cuda_op_mul_mat_cublas( const int cc = ggml_cuda_info().devices[id].cc; + const bool supports_bf16 = GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) || + (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2); + const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT; - if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) { + if (supports_bf16 && src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) { ggml_cuda_pool_alloc src1_as_bf16(ctx.pool(id)); if (src1->type != GGML_TYPE_BF16) { const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type); @@ -1257,7 +1260,7 @@ static void ggml_cuda_op_mul_mat_cublas( const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16); to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream); - } else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) { + } else if (fast_fp16_hardware_available(cc) && use_fp16) { // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 ggml_cuda_pool_alloc src0_as_f16(ctx.pool(id)); if (src0->type != GGML_TYPE_F16) { @@ -3061,9 +3064,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g return false; } #ifdef GGML_USE_MUSA - if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 && - !ggml_is_transposed(a) && !ggml_is_transposed(b)) { - return false; + const int cc = ggml_cuda_info().devices[dev_ctx->device].cc; + if (b->ne[2]*b->ne[3] > 1 && !ggml_is_transposed(a) && !ggml_is_transposed(b)) { + if (GGML_CUDA_CC_IS_QY1(cc) && op->op == GGML_OP_MUL_MAT && + a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) { + return false; + } + if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID && + a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) { + return false; + } } #endif // GGML_USE_MUSA switch (a->type) { @@ -3090,11 +3100,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g case GGML_TYPE_IQ4_NL: case GGML_TYPE_IQ4_XS: case GGML_TYPE_BF16: -#ifdef GGML_USE_MUSA - if (a->type == GGML_TYPE_Q3_K) { - return false; - } -#endif // GGML_USE_MUSA return true; default: return false; diff --git a/ggml/src/ggml-musa/mudnn.cuh b/ggml/src/ggml-musa/mudnn.cuh index a63be5755..c30128561 100644 --- a/ggml/src/ggml-musa/mudnn.cuh +++ b/ggml/src/ggml-musa/mudnn.cuh @@ -1,7 +1,7 @@ #pragma once -#include "../include/ggml.h" -#include "../ggml-cuda/common.cuh" +#include "ggml-cuda/common.cuh" +#include "ggml.h" // Asynchronously copies data from src tensor to dst tensor using the provided context. // Returns a musaError_t indicating success or failure. From bf5bcd0b857db420235e03639f0a5f218a7f8cf8 Mon Sep 17 00:00:00 2001 From: Aaron Teo Date: Thu, 26 Jun 2025 18:41:41 +0800 Subject: [PATCH 07/12] docs: update s390x documentation + add faq (#14389) * docs: update s390x documentation + add faq Signed-off-by: Aaron Teo * docs: add s390x z17 build q&a Signed-off-by: Aaron Teo --------- Signed-off-by: Aaron Teo --- docs/build-s390x.md | 76 +++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 74 insertions(+), 2 deletions(-) diff --git a/docs/build-s390x.md b/docs/build-s390x.md index bb6eae784..4c9ebb271 100644 --- a/docs/build-s390x.md +++ b/docs/build-s390x.md @@ -16,7 +16,7 @@ cd llama.cpp ## CPU Build with BLAS -Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. +Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment. ```bash cmake -S . -B build \ @@ -82,12 +82,18 @@ All models need to be converted to Big-Endian. You can achieve this in three cas 1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)** + ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff) + You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08). - These models and their respective tokenizers are verified to run correctly on IBM Z & LinuxONE. + These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system. 2. **Convert safetensors model to GGUF Big-Endian directly (recommended)** + ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28) + + The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case. + ```bash python3 convert_hf_to_gguf.py \ --outfile model-name-be.f16.gguf \ @@ -108,6 +114,10 @@ All models need to be converted to Big-Endian. You can achieve this in three cas 3. **Convert existing GGUF Little-Endian model to Big-Endian** + ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff) + + The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case. + ```bash python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG ``` @@ -163,6 +173,22 @@ It is strongly recommended to disable SMT via the kernel boot parameters as it n IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS. +## Frequently Asked Questions (FAQ) + +1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?` + + Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`. + + You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian. + +2. I'm getting extremely poor performance when running inference on a model + + Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration. + +3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17` + + Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue. + ## Getting Help on IBM Z & LinuxONE 1. **Bugs, Feature Requests** @@ -172,3 +198,49 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl 2. **Other Questions** Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com). + +## Appendix A: Hardware Support Matrix + +| | Support | Minimum Compiler Version | +| ------- | ------- | ------------------------ | +| IBM z15 | ✅ | | +| IBM z16 | ✅ | | +| IBM z17 | ✅ | GCC 15.1.0 | + +- ✅ - supported and verified to run as intended +- 🚫 - unsupported, we are unlikely able to provide support + +## Appendix B: SIMD Support Matrix + +| | VX/VXE/VXE2 | NNPA | zDNN | Spyre | +| ---------- | ----------- | ---- | ---- | ----- | +| FP32 | ✅ | ✅ | ❓ | ❓ | +| FP16 | ✅ | ✅ | ❓ | ❓ | +| BF16 | 🚫 | 🚫 | ❓ | ❓ | +| Q4_0 | ✅ | ✅ | ❓ | ❓ | +| Q4_1 | ✅ | ✅ | ❓ | ❓ | +| Q5_0 | 🚫 | 🚫 | ❓ | ❓ | +| Q5_1 | 🚫 | 🚫 | ❓ | ❓ | +| Q8_0 | ✅ | ✅ | ❓ | ❓ | +| Q2_K | 🚫 | 🚫 | ❓ | ❓ | +| Q3_K | ✅ | ✅ | ❓ | ❓ | +| Q4_K | ✅ | ✅ | ❓ | ❓ | +| Q5_K | ✅ | ✅ | ❓ | ❓ | +| Q6_K | ✅ | ✅ | ❓ | ❓ | +| TQ1_0 | 🚫 | 🚫 | ❓ | ❓ | +| TQ2_0 | 🚫 | 🚫 | ❓ | ❓ | +| IQ2_XXS | 🚫 | 🚫 | ❓ | ❓ | +| IQ2_XS | 🚫 | 🚫 | ❓ | ❓ | +| IQ2_S | 🚫 | 🚫 | ❓ | ❓ | +| IQ3_XXS | 🚫 | 🚫 | ❓ | ❓ | +| IQ3_S | 🚫 | 🚫 | ❓ | ❓ | +| IQ1_S | 🚫 | 🚫 | ❓ | ❓ | +| IQ1_M | 🚫 | 🚫 | ❓ | ❓ | +| IQ4_NL | ✅ | ✅ | ❓ | ❓ | +| IQ4_XS | ✅ | ✅ | ❓ | ❓ | +| FP32->FP16 | 🚫 | ✅ | ❓ | ❓ | +| FP16->FP32 | 🚫 | ✅ | ❓ | ❓ | + +- ✅ - acceleration available +- 🚫 - acceleration unavailable, will still run using scalar implementation +- ❓ - acceleration unknown, please contribute if you can test it yourself From 5783ae43599400b723b5da0569c1f848419ff3c7 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 26 Jun 2025 15:50:15 +0300 Subject: [PATCH 08/12] metal : batch rows copy in a single threadgroup (#14384) * metal : batch rows copy in a single threadgroup ggml-ci * metal : handle some edge cases when threadgroup size is not a power of 2 ggml-ci --- ggml/src/ggml-metal/ggml-metal.m | 43 ++++++++++++++++++++++++---- ggml/src/ggml-metal/ggml-metal.metal | 11 +++++-- 2 files changed, 45 insertions(+), 9 deletions(-) diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index 19f4d59e5..248fa378e 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -2450,6 +2450,7 @@ static bool ggml_metal_encode_node( nth *= 2; } + nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup); nth = MIN(nth, ne00); ggml_metal_kargs_sum_rows args = { @@ -3780,6 +3781,7 @@ static bool ggml_metal_encode_node( nth *= 2; } + nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup); nth = MIN(nth, ne00/4); ggml_metal_kargs_rms_norm args = { @@ -3816,6 +3818,7 @@ static bool ggml_metal_encode_node( nth *= 2; } + nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup); nth = MIN(nth, ne00/4); ggml_metal_kargs_l2_norm args = { @@ -3888,6 +3891,7 @@ static bool ggml_metal_encode_node( nth *= 2; } + nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup); nth = MIN(nth, ne00/4); ggml_metal_kargs_norm args = { @@ -4974,8 +4978,39 @@ static bool ggml_metal_encode_node( default: GGML_ABORT("not implemented"); } + GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); + + // TODO: support + //const int32_t nk00 = ne00/ggml_blck_size(dst->type); + const int32_t nk00 = ne00; + + int nth = 32; // SIMD width + + while (nth < nk00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) { + nth *= 2; + } + + nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup); + + // when rows are small, we can batch them together in a single threadgroup + int nrptg = 1; + + // TODO: relax this constraint in the future + if (ggml_blck_size(src0->type) == 1 && ggml_blck_size(dst->type) == 1) { + if (nth > nk00) { + nrptg = (nth + nk00 - 1)/nk00; + nth = nk00; + + if (nrptg*nth > (int) pipeline.maxTotalThreadsPerThreadgroup) { + nrptg--; + } + } + } + + nth = MIN(nth, nk00); + ggml_metal_kargs_cpy args = { - /*.ne00 =*/ ne00, + /*.ne00 =*/ nk00, /*.ne01 =*/ ne01, /*.ne02 =*/ ne02, /*.ne03 =*/ ne03, @@ -4998,11 +5033,7 @@ static bool ggml_metal_encode_node( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); - int nth = MIN(1024, ne00/ggml_blck_size(src0->type)); - - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; - + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nrptg - 1)/nrptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, nrptg, 1)]; } break; case GGML_OP_SET: { diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index 3da19879b..f02827606 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -4306,11 +4306,16 @@ kernel void kernel_cpy( device const char * src0, device char * dst, uint3 tgpig[[threadgroup_position_in_grid]], + uint tiitg[[thread_index_in_threadgroup]], ushort3 tpitg[[thread_position_in_threadgroup]], - ushort3 ntg[[threads_per_threadgroup]]) { + ushort3 tptg[[threads_per_threadgroup]]) { const int i03 = tgpig[2]; const int i02 = tgpig[1]; - const int i01 = tgpig[0]; + const int i01 = tgpig[0]*tptg.y + tiitg/tptg.x; + + if (i01 >= args.ne01) { + return; + } const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00; @@ -4321,7 +4326,7 @@ kernel void kernel_cpy( device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); - for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) { + for (int64_t i00 = tiitg%tptg.x; i00 < args.ne00; i00 += tptg.x) { device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00); dst_data[i00] = (T1) src[0]; } From e8215dbb96b8fb94a24c29cdd228166fb972dbfc Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 26 Jun 2025 15:51:19 +0300 Subject: [PATCH 09/12] metal : add special-case mat-vec mul for ne00 == 4 (#14385) ggml-ci --- ggml/src/ggml-metal/ggml-metal.m | 25 +++++++++-- ggml/src/ggml-metal/ggml-metal.metal | 64 ++++++++++++++++++++++++++++ tests/test-backend-ops.cpp | 64 +++++++++++++++------------- 3 files changed, 121 insertions(+), 32 deletions(-) diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index 248fa378e..d8d30cc0b 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -211,11 +211,14 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32, GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, + GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, + GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, + GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4, GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, @@ -1175,11 +1178,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32, rwkv_wkv6_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32, rwkv_wkv7_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4, mul_mv_f32_f32_c4, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4, mul_mv_bf16_f32_c4, use_bfloat); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && use_bfloat); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && use_bfloat); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && use_bfloat); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4, mul_mv_f16_f32_c4, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, has_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, has_simdgroup_reduction); @@ -3111,14 +3117,23 @@ static bool ggml_metal_encode_node( nsg = 1; nr0 = 1; nr1 = 4; - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline; + if (ne00 == 4) { + nr0 = 32; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline; + } } break; case GGML_TYPE_F16: { nsg = 1; nr0 = 1; if (src1t == GGML_TYPE_F32) { - if (ne11 * ne12 < 4) { + if (ne00 == 4) { + nr0 = 32; + nr1 = 4; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4].pipeline; + } else if (ne11 * ne12 < 4) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline; } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline; @@ -3137,7 +3152,11 @@ static bool ggml_metal_encode_node( nsg = 1; nr0 = 1; if (src1t == GGML_TYPE_F32) { - if (ne11 * ne12 < 4) { + if (ne00 == 4) { + nr0 = 32; + nr1 = 4; + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4].pipeline; + } else if (ne11 * ne12 < 4) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline; } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline; diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index f02827606..5f004a856 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -2532,6 +2532,70 @@ template [[host_name("kernel_mul_mv_bf16_f32")]] kernel mul_mv_t kernel_mul_mv< template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv; #endif +template +void kernel_mul_mv_c4_impl( + args_t args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig, + ushort tiisg) { + const int r0 = tgpig.x*32 + tiisg; + const int rb = tgpig.y*N_MV_T_T; + const int im = tgpig.z; + + if (r0 >= args.ne01) { + return; + } + + const uint i12 = im%args.ne12; + const uint i13 = im/args.ne12; + + const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + + device const T04 * x = (device const T04 *) (src0 + offset0); + + device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1; + + for (int row = 0; row < N_MV_T_T; ++row) { + int r1 = rb + row; + if (r1 >= args.ne11) { + break; + } + + const uint64_t offset1 = r1*args.nb11 + (i12 )*args.nb12 + (i13 )*args.nb13; + + device const T14 * y = (device const T14 *) (src1 + offset1); + + dst_f32[(uint64_t)r1*args.ne0 + r0] = dot((float4) x[0], (float4) y[0]); + } +} + +template +kernel void kernel_mul_mv_c4( + constant ggml_metal_kargs_mul_mv & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]]) { + kernel_mul_mv_c4_impl( + args, + src0, + src1, + dst, + tgpig, + tiisg); +} + +typedef decltype(kernel_mul_mv_c4) mul_mv_c4_t; + +template [[host_name("kernel_mul_mv_f32_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4; +template [[host_name("kernel_mul_mv_f16_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4; +#if defined(GGML_METAL_USE_BF16) +template [[host_name("kernel_mul_mv_bf16_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4; +#endif + template kernel void kernel_mul_mv_1row( constant ggml_metal_kargs_mul_mv & args, diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 7be7f2205..615c2dc00 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -4252,39 +4252,45 @@ static std::vector> make_test_cases_eval() { #if 1 for (ggml_type type_a : base_types) { for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) { - // test cases without permutation - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1, 1}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {1, 1}, {1, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 1}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 2}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 2}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 2}, {1, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {3, 2}, {2, 2})); + std::vector ks = { 256 }; + if (ggml_blck_size(type_a) == 1) { + ks.push_back(4); + } + for (auto k : ks) { + // test cases without permutation + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {1, 1}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {1, 1}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {1, 1}, {1, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 1}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 1}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {1, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {3, 2}, {2, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {1, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 1}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {1, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {2, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 1}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 1}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {1, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {2, 2})); - // test cases with permutation - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 1, 3, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 256, {2, 3}, {1, 1}, {0, 3, 2, 1})); + // test cases with permutation + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {2, 3}, {1, 1}, {0, 2, 1, 3})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {2, 3}, {1, 1}, {0, 1, 3, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, k, {2, 3}, {1, 1}, {0, 3, 2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 2, 1, 3})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 1, 3, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 256, {2, 3}, {1, 1}, {0, 3, 2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 2, 1, 3})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 1, 3, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, k, {2, 3}, {1, 1}, {0, 3, 2, 1})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2})); - test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 2, 1, 3})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 1, 3, 2})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 3, 2, 1})); + } // test cases with large ne00/ne10 to cover stream-k fixup test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 1024, {3, 2}, {1, 1})); From b25346221dadb9101aa9dda55431dde4d3596943 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Thu, 26 Jun 2025 15:01:14 +0200 Subject: [PATCH 10/12] llama : return mistral-v7-tekken as default template only (#14390) --- src/llama-model.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 9b19da984..c2835ce67 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -14377,7 +14377,7 @@ const char * llama_model_chat_template(const llama_model * model, const char * n // do not extend this list unless absolutely necessary // Mistral-Small-2503 does not have built-in chat template llama_vocab_pre_type pre_type = model->vocab.get_pre_type(); - if (pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) { + if (!name && pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) { return "mistral-v7-tekken"; } From a01047b041aa04aeea351933658433ed004516ab Mon Sep 17 00:00:00 2001 From: bandoti <141645996+bandoti@users.noreply.github.com> Date: Thu, 26 Jun 2025 13:46:53 -0300 Subject: [PATCH 11/12] cmake: regen vulkan shaders when shaders-gen sources change (#14398) * Add shaders-gen sources as target deps --- ggml/src/ggml-vulkan/CMakeLists.txt | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 39f022f33..0bf4cb14f 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -143,7 +143,8 @@ if (Vulkan_FOUND) -DCMAKE_BUILD_TYPE=$ ${VULKAN_SHADER_GEN_CMAKE_ARGS} - BUILD_COMMAND ${CMAKE_COMMAND} --build . --config $ + BUILD_COMMAND ${CMAKE_COMMAND} --build . --config $ + BUILD_ALWAYS TRUE # NOTE: When DESTDIR is set using Makefile generators and # "make install" triggers the build step, vulkan-shaders-gen @@ -164,6 +165,14 @@ if (Vulkan_FOUND) file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp") + # Because external projects do not provide source-level tracking, + # the vulkan-shaders-gen sources need to be explicitly added to + # ensure that changes will cascade into shader re-generation. + + file(GLOB _ggml_vk_shaders_gen_sources + CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.cpp" + "${_ggml_vk_input_dir}/*.h") + add_custom_command( OUTPUT ${_ggml_vk_header} ${_ggml_vk_source} @@ -177,6 +186,7 @@ if (Vulkan_FOUND) --no-clean DEPENDS ${_ggml_vk_shader_files} + ${_ggml_vk_shaders_gen_sources} vulkan-shaders-gen COMMENT "Generate vulkan shaders" From 8846aace4934ad29651ea61b8c7e3f6b0556e3d2 Mon Sep 17 00:00:00 2001 From: Xuan-Son Nguyen Date: Thu, 26 Jun 2025 19:34:02 +0200 Subject: [PATCH 12/12] model : gemma3n text-only (#14400) * gemma3n * add llm_graph_input_one --- convert_hf_to_gguf.py | 124 +++++++- gguf-py/gguf/constants.py | 75 +++++ gguf-py/gguf/gguf_writer.py | 18 ++ gguf-py/gguf/tensor_mapping.py | 64 ++++ src/llama-arch.cpp | 54 ++++ src/llama-arch.h | 17 ++ src/llama-graph.cpp | 23 +- src/llama-graph.h | 16 +- src/llama-hparams.h | 6 + src/llama-kv-cache-unified.cpp | 30 +- src/llama-model.cpp | 517 +++++++++++++++++++++++++++++++++ src/llama-model.h | 22 ++ src/llama-quant.cpp | 9 +- 13 files changed, 960 insertions(+), 15 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index bbf8b30ff..4f2339a02 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -310,6 +310,8 @@ class ModelBase: gguf.MODEL_TENSOR.POSNET_NORM2, gguf.MODEL_TENSOR.V_ENC_EMBD_POS, gguf.MODEL_TENSOR.A_ENC_EMBD_POS, + gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF, + gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF, ) ) or not new_name.endswith(".weight") @@ -320,7 +322,11 @@ class ModelBase: self.match_model_tensor_name(new_name, key, bid) for key in ( gguf.MODEL_TENSOR.TOKEN_EMBD, + gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD, gguf.MODEL_TENSOR.OUTPUT, + gguf.MODEL_TENSOR.ALTUP_ROUTER, + gguf.MODEL_TENSOR.LAUREL_L, + gguf.MODEL_TENSOR.LAUREL_R, ) ): if self.ftype in ( @@ -921,13 +927,16 @@ class TextModel(ModelBase): tokenizer = SentencePieceProcessor() tokenizer.LoadFromFile(str(tokenizer_path)) - vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + vocab_size = self.find_hparam([ + "vocab_size_per_layer_input", # gemma3n + "vocab_size", + ], optional=True) or tokenizer.vocab_size() tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size - for token_id in range(tokenizer.vocab_size()): + for token_id in range(vocab_size): piece = tokenizer.IdToPiece(token_id) text = piece.encode("utf-8") score = tokenizer.GetScore(token_id) @@ -942,6 +951,10 @@ class TextModel(ModelBase): elif tokenizer.IsByte(token_id): toktype = SentencePieceTokenTypes.BYTE + if token_id >= vocab_size: + logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}') + break + tokens[token_id] = text scores[token_id] = score toktypes[token_id] = toktype @@ -4217,6 +4230,7 @@ class Gemma2Model(TextModel): @ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration") class Gemma3Model(TextModel): model_arch = gguf.MODEL_ARCH.GEMMA3 + norm_shift = 1.0 # Gemma3RMSNorm adds 1.0 to the norm value def set_vocab(self): self._set_vocab_sentencepiece() @@ -4238,9 +4252,8 @@ class Gemma3Model(TextModel): self.gguf_writer.add_value_length(hparams.get("head_dim", 256)) self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers - # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3 + # attn_logit_softcapping is removed in Gemma3 assert hparams.get("attn_logit_softcapping") is None - assert hparams.get("final_logit_softcapping") is None self.gguf_writer.add_sliding_window(hparams["sliding_window"]) self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4)) if hparams.get("rope_scaling") is not None: @@ -4252,7 +4265,7 @@ class Gemma3Model(TextModel): def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused - if name.startswith("language_model."): + if "language_model." in name: name = name.replace("language_model.", "") elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \ @@ -4267,8 +4280,9 @@ class Gemma3Model(TextModel): # ref code in Gemma3RMSNorm # output = output * (1.0 + self.weight.float()) + # note: this is not the case on gemma3n if name.endswith("norm.weight"): - data_torch = data_torch + 1 + data_torch = data_torch + self.norm_shift return [(self.map_tensor_name(name), data_torch)] @@ -4325,6 +4339,104 @@ class Gemma3VisionModel(MmprojModel): return [] # skip other tensors +@ModelBase.register("Gemma3nForConditionalGeneration") +class Gemma3NModel(Gemma3Model): + model_arch = gguf.MODEL_ARCH.GEMMA3N + norm_shift = 0.0 # same value with Gemma3p5RMSNorm scale_shift on python code + + _altup_proj: list[Tensor] = [] + _altup_unembd: list[Tensor] = [] + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + assert self.hparams["altup_num_inputs"] == 4, "Current conversion only supports 4 altup inputs" + self._altup_proj = [ + torch.Tensor(), # to be replaced + torch.Tensor(), # to be replaced + torch.Tensor(), # to be replaced + ] + self._altup_unembd = [ + torch.Tensor(), # to be replaced + torch.Tensor(), # to be replaced + torch.Tensor(), # to be replaced + ] + + def set_vocab(self): + with open(self.dir_model / "chat_template.jinja") as f: + # quick hack to make sure chat template is added + self.gguf_writer.add_chat_template(f.read()) + super().set_vocab() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_altup_active_idx(self.hparams["altup_active_idx"]) + self.gguf_writer.add_altup_num_inputs(self.hparams["altup_num_inputs"]) + self.gguf_writer.add_embedding_length_per_layer_input(self.hparams["hidden_size_per_layer_input"]) + self.gguf_writer.add_shared_kv_layers(self.hparams["num_kv_shared_layers"]) + + activation_sparsity_scale = [] + for s in self.hparams["activation_sparsity_pattern"]: + normal_dist = torch.distributions.normal.Normal(0, 1) + std_multiplier = normal_dist.icdf(torch.tensor(s, dtype=torch.float32)) + activation_sparsity_scale.append(std_multiplier.item()) + self.gguf_writer.add_activation_sparsity_scale(activation_sparsity_scale) + + sliding_window_pattern = [] + for t in self.hparams["layer_types"]: + sliding_window_pattern.append(t == "sliding_attention") + self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern) + + def _stack_matrices(self, matrices: list[Tensor]) -> Tensor | None: + has_all = all(m.numel() > 0 for m in matrices) + if not has_all: + return None + else: + return torch.stack(matrices, dim=0) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if name.endswith("_scale"): + name = name + ".weight" + + # TODO: implement self.prediction_coefs.weight.clamp_(...) + + if "language_model." not in name: + return [] # skip non-language model tensors + + if "altup_unembed_projections" in name: + data_torch = data_torch.to(device="cpu") + if ".0." in name: + self._altup_unembd[0] = data_torch + elif ".1." in name: + self._altup_unembd[1] = data_torch + elif ".2." in name: + self._altup_unembd[2] = data_torch + else: + raise ValueError(f"Unknown name: {name}") + out = self._stack_matrices(self._altup_unembd) + if out is not None: + return [(self.map_tensor_name("model.altup_unembed_projections.weight"), out)] + else: + return [] + + if "altup_projections" in name: + data_torch = data_torch.to(device="cpu") + if ".0." in name: + self._altup_proj[0] = data_torch + elif ".1." in name: + self._altup_proj[1] = data_torch + elif ".2." in name: + self._altup_proj[2] = data_torch + else: + raise ValueError(f"Unknown name: {name}") + out = self._stack_matrices(self._altup_proj) + if out is not None: + return [(self.map_tensor_name("model.altup_projections.weight"), out)] + else: + return [] + + return super().modify_tensors(data_torch, name, bid) + + @ModelBase.register("Starcoder2ForCausalLM") class StarCoder2Model(TextModel): model_arch = gguf.MODEL_ARCH.STARCODER2 diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 0429b0aaf..fb75143b0 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -118,6 +118,10 @@ class Keys: EMBEDDING_SCALE = "{arch}.embedding_scale" TOKEN_SHIFT_COUNT = "{arch}.token_shift_count" INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step" + ACTIVATION_SPARSITY_SCALE = "{arch}.activation_sparsity_scale" + ALTUP_ACTIVE_IDX = "{arch}.altup.active_idx" + ALTUP_NUM_INPUTS = "{arch}.altup.num_inputs" + EMBD_LENGTH_PER_LAYER_INP = "{arch}.embedding_length_per_layer_input" class Attention: HEAD_COUNT = "{arch}.attention.head_count" @@ -142,6 +146,8 @@ class Keys: SCALE = "{arch}.attention.scale" KEY_LENGTH_MLA = "{arch}.attention.key_length_mla" VALUE_LENGTH_MLA = "{arch}.attention.value_length_mla" + SHARED_KV_LAYERS = "{arch}.attention.shared_kv_layers" + SLIDING_WINDOW_PATTERN = "{arch}.attention.sliding_window_pattern" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" @@ -314,6 +320,7 @@ class MODEL_ARCH(IntEnum): GEMMA = auto() GEMMA2 = auto() GEMMA3 = auto() + GEMMA3N = auto() STARCODER2 = auto() RWKV6 = auto() RWKV6QWEN2 = auto() @@ -399,6 +406,22 @@ class MODEL_TENSOR(IntEnum): ATTN_Q_NORM = auto() ATTN_K_NORM = auto() LAYER_OUT_NORM = auto() + PER_LAYER_TOKEN_EMBD = auto() # gemma3n + PER_LAYER_MODEL_PROJ = auto() # gemma3n + PER_LAYER_INP_GATE = auto() # gemma3n + PER_LAYER_PROJ = auto() # gemma3n + PER_LAYER_PROJ_NORM = auto() # gemma3n + PER_LAYER_POST_NORM = auto() # gemma3n + ALTUP_PROJ = auto() # gemma3n + ALTUP_UNEMBD_PROJ = auto() # gemma3n + ALTUP_CORRECT_COEF = auto() # gemma3n + ALTUP_CORRECT_SCALE = auto() # gemma3n + ALTUP_PREDICT_COEF = auto() # gemma3n + ALTUP_ROUTER = auto() # gemma3n + ALTUP_ROUTER_NORM = auto() # gemma3n + LAUREL_L = auto() # gemma3n + LAUREL_R = auto() # gemma3n + LAUREL_POST_NORM = auto() # gemma3n SSM_IN = auto() SSM_CONV1D = auto() SSM_X = auto() @@ -597,6 +620,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.GEMMA2: "gemma2", MODEL_ARCH.GEMMA3: "gemma3", + MODEL_ARCH.GEMMA3N: "gemma3n", MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.RWKV6: "rwkv6", MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2", @@ -682,6 +706,22 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps", MODEL_TENSOR.FFN_EXP_PROBS_B: "blk.{bid}.exp_probs_b", MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm", + MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: "per_layer_token_embd", # gemma3n + MODEL_TENSOR.PER_LAYER_MODEL_PROJ: "per_layer_model_proj", # gemma3n + MODEL_TENSOR.PER_LAYER_PROJ_NORM: "per_layer_proj_norm", # gemma3n + MODEL_TENSOR.ALTUP_UNEMBD_PROJ: "altup_unembd_proj", # gemma3n + MODEL_TENSOR.ALTUP_PROJ: "altup_proj", # gemma3n + MODEL_TENSOR.PER_LAYER_INP_GATE: "blk.{bid}.inp_gate", # gemma3n + MODEL_TENSOR.PER_LAYER_PROJ: "blk.{bid}.proj", # gemma3n + MODEL_TENSOR.PER_LAYER_POST_NORM: "blk.{bid}.post_norm", # gemma3n + MODEL_TENSOR.ALTUP_CORRECT_COEF: "blk.{bid}.altup_correct_coef", # gemma3n + MODEL_TENSOR.ALTUP_CORRECT_SCALE: "blk.{bid}.altup_correct_scale", # gemma3n + MODEL_TENSOR.ALTUP_PREDICT_COEF: "blk.{bid}.altup_predict_coef", # gemma3n + MODEL_TENSOR.ALTUP_ROUTER: "blk.{bid}.altup_router", # gemma3n + MODEL_TENSOR.ALTUP_ROUTER_NORM: "blk.{bid}.altup_router_norm", # gemma3n + MODEL_TENSOR.LAUREL_L: "blk.{bid}.laurel_l", # gemma3n + MODEL_TENSOR.LAUREL_R: "blk.{bid}.laurel_r", # gemma3n + MODEL_TENSOR.LAUREL_POST_NORM: "blk.{bid}.laurel_post_norm", # gemma3n MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in", MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d", MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x", @@ -1486,6 +1526,41 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_PRE_NORM, MODEL_TENSOR.FFN_POST_NORM, ], + MODEL_ARCH.GEMMA3N: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + # altup / laurel + MODEL_TENSOR.PER_LAYER_TOKEN_EMBD, + MODEL_TENSOR.PER_LAYER_MODEL_PROJ, + MODEL_TENSOR.PER_LAYER_INP_GATE, + MODEL_TENSOR.PER_LAYER_PROJ, + MODEL_TENSOR.PER_LAYER_PROJ_NORM, + MODEL_TENSOR.PER_LAYER_POST_NORM, + MODEL_TENSOR.ALTUP_PROJ, + MODEL_TENSOR.ALTUP_UNEMBD_PROJ, + MODEL_TENSOR.ALTUP_CORRECT_COEF, + MODEL_TENSOR.ALTUP_CORRECT_SCALE, + MODEL_TENSOR.ALTUP_PREDICT_COEF, + MODEL_TENSOR.ALTUP_ROUTER, + MODEL_TENSOR.ALTUP_ROUTER_NORM, + MODEL_TENSOR.LAUREL_L, + MODEL_TENSOR.LAUREL_R, + MODEL_TENSOR.LAUREL_POST_NORM, + ], MODEL_ARCH.STARCODER2: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index b9b63d052..d32cd479a 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -672,6 +672,18 @@ class GGUFWriter: def add_decoder_start_token_id(self, id: int) -> None: self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id) + def add_embedding_length_per_layer_input(self, value: int) -> None: + self.add_uint32(Keys.LLM.EMBD_LENGTH_PER_LAYER_INP.format(arch=self.arch), value) + + def add_altup_active_idx(self, val: int) -> None: + self.add_uint32(Keys.LLM.ALTUP_ACTIVE_IDX.format(arch=self.arch), val) + + def add_altup_num_inputs(self, val: int) -> None: + self.add_uint32(Keys.LLM.ALTUP_NUM_INPUTS.format(arch=self.arch), val) + + def add_activation_sparsity_scale(self, values: Sequence[float]) -> None: + self.add_array(Keys.LLM.ACTIVATION_SPARSITY_SCALE.format(arch=self.arch), values) + def add_head_count(self, count: int | Sequence[int]) -> None: if isinstance(count, int): self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) @@ -702,6 +714,12 @@ class GGUFWriter: def add_clamp_kqv(self, value: float) -> None: self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value) + def add_shared_kv_layers(self, value: float) -> None: + self.add_float32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value) + + def add_sliding_window_pattern(self, value: Sequence[bool]) -> None: + self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value) + def add_logit_scale(self, value: float) -> None: self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 79f044d2a..b30f77dbe 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -480,6 +480,70 @@ class TensorNameMap: "encoder.layer.{bid}.layer_norm_2" # jina-v2-code ), + MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: ( + "model.embed_tokens_per_layer", # gemma3n + ), + + MODEL_TENSOR.PER_LAYER_MODEL_PROJ: ( + "model.per_layer_model_projection", # gemma3n + ), + + MODEL_TENSOR.PER_LAYER_PROJ_NORM: ( + "model.per_layer_projection_norm", # gemma3n + ), + + MODEL_TENSOR.ALTUP_PROJ: ( + "model.altup_projections", # gemma3n + ), + + MODEL_TENSOR.ALTUP_UNEMBD_PROJ: ( + "model.altup_unembed_projections", # gemma3n + ), + + MODEL_TENSOR.PER_LAYER_INP_GATE: ( + "model.layers.{bid}.per_layer_input_gate", # gemma3n + ), + + MODEL_TENSOR.PER_LAYER_PROJ: ( + "model.layers.{bid}.per_layer_projection", # gemma3n + ), + + MODEL_TENSOR.PER_LAYER_POST_NORM: ( + "model.layers.{bid}.post_per_layer_input_norm", # gemma3n + ), + + MODEL_TENSOR.ALTUP_CORRECT_COEF: ( + "model.layers.{bid}.altup.correction_coefs", # gemma3n + ), + + MODEL_TENSOR.ALTUP_CORRECT_SCALE: ( + "model.layers.{bid}.altup.correct_output_scale", # gemma3n + ), + + MODEL_TENSOR.ALTUP_PREDICT_COEF: ( + "model.layers.{bid}.altup.prediction_coefs", # gemma3n + ), + + MODEL_TENSOR.ALTUP_ROUTER: ( + "model.layers.{bid}.altup.modality_router", # gemma3n + ), + + MODEL_TENSOR.ALTUP_ROUTER_NORM: ( + "model.layers.{bid}.altup.router_norm", # gemma3n + ), + + MODEL_TENSOR.LAUREL_L: ( + "model.layers.{bid}.laurel.linear_left", # gemma3n + ), + + MODEL_TENSOR.LAUREL_R: ( + "model.layers.{bid}.laurel.linear_right", # gemma3n + ), + + MODEL_TENSOR.LAUREL_POST_NORM: ( + "model.layers.{bid}.laurel.post_laurel_norm", # gemma3n + ), + MODEL_TENSOR.SSM_IN: ( "model.layers.{bid}.in_proj", "backbone.layers.{bid}.mixer.in_proj", diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 8dadef204..435e3b9ba 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -42,6 +42,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_GEMMA2, "gemma2" }, { LLM_ARCH_GEMMA3, "gemma3" }, + { LLM_ARCH_GEMMA3N, "gemma3n" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_XVERSE, "xverse" }, @@ -932,6 +933,42 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, }, }, + { + LLM_ARCH_GEMMA3N, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + { LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" }, + { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" }, + { LLM_TENSOR_PER_LAYER_PROJ_NORM, "per_layer_proj_norm" }, + { LLM_TENSOR_ALTUP_UNEMBD_PROJ, "altup_unembd_proj" }, + { LLM_TENSOR_ALTUP_PROJ, "altup_proj" }, + { LLM_TENSOR_PER_LAYER_INP_GATE, "blk.%d.inp_gate" }, + { LLM_TENSOR_PER_LAYER_PROJ, "blk.%d.proj" }, + { LLM_TENSOR_PER_LAYER_POST_NORM, "blk.%d.post_norm" }, + { LLM_TENSOR_ALTUP_CORRECT_COEF, "blk.%d.altup_correct_coef" }, + { LLM_TENSOR_ALTUP_CORRECT_SCALE, "blk.%d.altup_correct_scale" }, + { LLM_TENSOR_ALTUP_PREDICT_COEF, "blk.%d.altup_predict_coef" }, + { LLM_TENSOR_ALTUP_ROUTER, "blk.%d.altup_router" }, + { LLM_TENSOR_ALTUP_ROUTER_NORM, "blk.%d.altup_router_norm" }, + { LLM_TENSOR_LAUREL_L, "blk.%d.laurel_l" }, + { LLM_TENSOR_LAUREL_R, "blk.%d.laurel_r" }, + { LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" }, + }, + }, { LLM_ARCH_STARCODER2, { @@ -1749,6 +1786,23 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, {LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}}, {LLM_TENSOR_FFN_EXP_PROBS_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + // altup / laurel (gemma 3n) + {LLM_TENSOR_PER_LAYER_TOKEN_EMBD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}}, + {LLM_TENSOR_PER_LAYER_MODEL_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_PER_LAYER_PROJ_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}}, + {LLM_TENSOR_ALTUP_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ALTUP_UNEMBD_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_PER_LAYER_INP_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_PER_LAYER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_PER_LAYER_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_ALTUP_CORRECT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ALTUP_CORRECT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_ALTUP_PREDICT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ALTUP_ROUTER, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ALTUP_ROUTER_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_LAUREL_L, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_LAUREL_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_LAUREL_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // this tensor is loaded for T5, but never used {LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}}, {LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_INPUT, GGML_OP_IM2COL}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 5b0230c15..9181ad053 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -46,6 +46,7 @@ enum llm_arch { LLM_ARCH_GEMMA, LLM_ARCH_GEMMA2, LLM_ARCH_GEMMA3, + LLM_ARCH_GEMMA3N, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, @@ -269,6 +270,22 @@ enum llm_tensor { LLM_TENSOR_LAYER_OUT_NORM, LLM_TENSOR_POST_ATTN_NORM, LLM_TENSOR_POST_MLP_NORM, + LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n + LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n + LLM_TENSOR_PER_LAYER_INP_GATE, // gemma3n + LLM_TENSOR_PER_LAYER_PROJ, // gemma3n + LLM_TENSOR_PER_LAYER_PROJ_NORM, // gemma3n + LLM_TENSOR_PER_LAYER_POST_NORM, // gemma3n + LLM_TENSOR_ALTUP_PROJ, // gemma3n + LLM_TENSOR_ALTUP_UNEMBD_PROJ, // gemma3n + LLM_TENSOR_ALTUP_CORRECT_COEF, // gemma3n + LLM_TENSOR_ALTUP_CORRECT_SCALE, // gemma3n + LLM_TENSOR_ALTUP_PREDICT_COEF, // gemma3n + LLM_TENSOR_ALTUP_ROUTER, // gemma3n + LLM_TENSOR_ALTUP_ROUTER_NORM, // gemma3n + LLM_TENSOR_LAUREL_L, // gemma3n + LLM_TENSOR_LAUREL_R, // gemma3n + LLM_TENSOR_LAUREL_POST_NORM, // gemma3n LLM_TENSOR_SSM_IN, LLM_TENSOR_SSM_CONV1D, LLM_TENSOR_SSM_X, diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 48589a50a..71ee431a9 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -350,6 +350,12 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) { } } +void llm_graph_input_one::set_input(const llama_ubatch *) { + GGML_ASSERT(one && ggml_nelements(one) == 1); + float f_one = 1.0f; + ggml_backend_tensor_set(one, &f_one, 0, sizeof(float)); +} + // // llm_graph_context // @@ -1267,8 +1273,14 @@ ggml_tensor * llm_graph_context::build_attn( // these nodes are added to the graph together so that they are not reordered // by doing so, the number of splits in the graph is reduced ggml_build_forward_expand(gf, q_cur); - ggml_build_forward_expand(gf, k_cur); - ggml_build_forward_expand(gf, v_cur); + + if (k_cur) { + ggml_build_forward_expand(gf, k_cur); + } + + if (v_cur) { + ggml_build_forward_expand(gf, v_cur); + } const auto * mctx_iswa = static_cast(mctx); @@ -1276,9 +1288,12 @@ ggml_tensor * llm_graph_context::build_attn( const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base(); - // store to KV cache - { + // optionally store to KV cache + if (k_cur) { ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il)); + } + + if (v_cur) { ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il)); } diff --git a/src/llama-graph.h b/src/llama-graph.h index b433f266d..4b1ec354d 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -329,6 +329,17 @@ public: const llama_memory_hybrid_context * mctx; }; +// TODO: remove this when ggml_scale_add is implemented +class llm_graph_input_one : public llm_graph_input_i { +public: + llm_graph_input_one() {} + virtual ~llm_graph_input_one() = default; + + void set_input(const llama_ubatch *) override; + + ggml_tensor * one = nullptr; // F32 +}; + // // llm_graph_result // @@ -589,14 +600,15 @@ struct llm_graph_context { llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const; + // note: if k_cur or v_cur are not provided, they will not be stored in the memory ggml_tensor * build_attn( llm_graph_input_attn_kv_unified_iswa * inp, ggml_cgraph * gf, ggml_tensor * wo, ggml_tensor * wo_b, ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] - ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] - ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] + ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional + ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional ggml_tensor * kq_b, ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] float kq_scale, diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 7b315a9a7..e85afe145 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -143,6 +143,12 @@ struct llama_hparams { uint32_t n_attn_temp_floor_scale = 8192; float f_attn_temp_scale = 0.1; + // gemma3n altup + uint32_t n_altup = 4; // altup_num_inputs + uint32_t i_altup_act = 0; // altup_active_idx + uint32_t laurel_rank = 64; + uint32_t n_embd_altup = 256; + // needed by encoder-decoder models (e.g. T5, FLAN-T5) // ref: https://github.com/ggerganov/llama.cpp/pull/8141 llama_token dec_start_token_id = LLAMA_TOKEN_NULL; diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp index b506d32ed..8517b722a 100644 --- a/src/llama-kv-cache-unified.cpp +++ b/src/llama-kv-cache-unified.cpp @@ -33,13 +33,19 @@ llama_kv_cache_unified::llama_kv_cache_unified( GGML_ASSERT(kv_size % n_pad == 0); + // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE] + auto n_layer_cache = hparams.n_layer; + if (model.arch == LLM_ARCH_GEMMA3N) { + n_layer_cache = 20; + } + // create a context for each buffer type std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { ggml_init_params params = { - /*.mem_size =*/ size_t(2u*hparams.n_layer*ggml_tensor_overhead()), + /*.mem_size =*/ size_t(2u*n_layer_cache*ggml_tensor_overhead()), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; @@ -62,7 +68,7 @@ llama_kv_cache_unified::llama_kv_cache_unified( cells.resize(kv_size); - for (uint32_t il = 0; il < hparams.n_layer; il++) { + for (uint32_t il = 0; il < n_layer_cache; il++) { if (filter && !filter(il)) { LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il); continue; @@ -102,6 +108,26 @@ llama_kv_cache_unified::llama_kv_cache_unified( layers.push_back({ il, k, v }); } + // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE] + if (model.arch == LLM_ARCH_GEMMA3N) { + LLAMA_LOG_DEBUG("%s: GEMMA3N: reuse layers [%d, %d]\n", __func__, n_layer_cache, hparams.n_layer - 1); + + for (uint32_t il = n_layer_cache; il < hparams.n_layer; il++) { + if (filter && !filter(il)) { + LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il); + continue; + } + + const bool is_swa = hparams.is_swa(il); + const uint32_t il_reuse = n_layer_cache - (is_swa ? 2 : 1); + + GGML_ASSERT(map_layer_ids.find(il_reuse) != map_layer_ids.end()); + map_layer_ids[il] = map_layer_ids[il_reuse]; + + LLAMA_LOG_DEBUG("%s: layer %3d: reuse layer %d, isw = %d\n", __func__, il, il_reuse, is_swa); + } + } + // allocate tensors and initialize the buffers to avoid NaNs in the padding for (auto it : ctx_map) { auto * buft = it.first; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index c2835ce67..fc39195ed 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -103,6 +103,8 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)"; case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; + case LLM_TYPE_E2B: return "E2B"; + case LLM_TYPE_E4B: return "E4B"; default: return "?B"; } } @@ -1017,6 +1019,24 @@ void llama_model::load_hparams(llama_model_loader & ml) { ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0))) : 1.0f / std::sqrt(float(hparams.n_embd_head_k)); } break; + case LLM_ARCH_GEMMA3N: + { + hparams.swa_type = LLAMA_SWA_TYPE_STANDARD; + hparams.set_swa_pattern(5); + + hparams.rope_freq_base_train_swa = 10000.0f; + hparams.rope_freq_scale_train_swa = 1.0f; + hparams.f_attention_scale = 1.0f; + + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 30: type = LLM_TYPE_E2B; break; + case 35: type = LLM_TYPE_E4B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_STARCODER2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -2950,6 +2970,62 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); } } break; + case LLM_ARCH_GEMMA3N: + { + const int64_t n_altup = hparams.n_altup; + const int64_t laurel_rank = hparams.laurel_rank; + const int64_t n_embd_altup = hparams.n_embd_altup; + + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + tok_embd_per_layer = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_altup * n_layer, n_vocab}, 0); + + altup_proj = create_tensor(tn(LLM_TENSOR_ALTUP_PROJ, "weight"), {n_embd, n_embd, n_altup - 1}, 0); + altup_unembd_proj = create_tensor(tn(LLM_TENSOR_ALTUP_UNEMBD_PROJ, "weight"), {n_embd, n_embd, n_altup - 1}, 0); + per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight"), {n_embd, n_embd_altup * n_layer}, 0); + per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight"), {n_embd_altup}, 0); + + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + + // altup & laurel + layer.per_layer_inp_gate = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE, "weight", i), {n_embd, n_embd_altup}, 0); + layer.per_layer_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ, "weight", i), {n_embd_altup, n_embd}, 0); + layer.per_layer_post_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0); + layer.altup_correct_coef = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_COEF, "weight", i), {n_altup, n_altup}, 0); + layer.altup_correct_scale = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_SCALE, "weight", i), {n_embd}, 0); + layer.altup_predict_coef = create_tensor(tn(LLM_TENSOR_ALTUP_PREDICT_COEF, "weight", i), {n_altup, n_altup * n_altup}, 0); + layer.altup_router = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER, "weight", i), {n_embd, n_altup}, 0); + layer.altup_router_norm = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER_NORM, "weight", i), {n_embd}, 0); + layer.laurel_l = create_tensor(tn(LLM_TENSOR_LAUREL_L, "weight", i), {n_embd, laurel_rank}, 0); + layer.laurel_r = create_tensor(tn(LLM_TENSOR_LAUREL_R, "weight", i), {laurel_rank, n_embd}, 0); + layer.laurel_post_norm = create_tensor(tn(LLM_TENSOR_LAUREL_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; case LLM_ARCH_STARCODER2: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -8980,6 +9056,442 @@ struct llm_build_gemma3_iswa : public llm_graph_context { } }; +struct llm_build_gemma3n_iswa : public llm_graph_context { + const llama_model & model; + ggml_cgraph * gf; + + const int64_t n_embd_head; + const int64_t n_embd_altup; + const int64_t n_altup; + const int i_altup_act; + const int n_layer_kv = 20; // number of layers having KV [KV_REUSE] + const int n_layer_sparsity = 10; // number of layers using activation sparsity + const float f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95) + + ggml_tensor * one; // containing single element 1.0f + + llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) + : llm_graph_context(params), + model(model), + gf(gf), + n_embd_head(model.hparams.n_embd_head_k), + n_embd_altup(model.hparams.n_embd_altup), + n_altup(model.hparams.n_altup), + i_altup_act(model.hparams.i_altup_act) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // TODO: remove this when ggml_scale_add is implemented + one = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + { + auto inp = std::make_unique(); + inp->one = one; + res->add_input(std::move(inp)); + } + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // TODO: is causal == true correct? might need some changes + auto * inp_attn = build_attn_inp_kv_unified_iswa(); + + // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer] + ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs()); + + // inpL now has only 1 altup, project it to the rest of the altups + // these "added" altups will be concat to the last dim of inpL + { + ggml_tensor * target_magnitude = calc_magnitude(inpL); + ggml_tensor * inp_repeated = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1); + ggml_tensor * altup_added = ggml_mul_mat(ctx0, model.altup_proj, inp_repeated); // shape: [n_embd, n_tokens, n_altup - 1] + ggml_tensor * new_magnitude = calc_magnitude(altup_added); + altup_added = ggml_div(ctx0, + ggml_mul(ctx0, altup_added, target_magnitude), + new_magnitude); + inpL = ggml_concat(ctx0, inpL, altup_added, 2); // shape: [n_embd, n_tokens, n_altup] + cb(inpL, "inp_stacked", -1); + } + + // inpL now has shape: [n_embd, n_tokens, n_altup] + // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer] + + for (int il = 0; il < n_layer; ++il) { + // this block is made to be closely resemble Gemma3p5DecoderLayer on python code + const bool has_kv = (il < n_layer_kv); + + const float freq_base_l = model.get_rope_freq_base (cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + ggml_tensor * cur = inpL; // [n_embd, n_tokens, n_altup] + ggml_tensor * predictions = altup_predict(cur, il); // [n_embd, n_tokens, n_altup] + + // predicted value will go through self-attention and laurel + ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); // [n_embd, n_tokens] + cur = active_prediction; + cb(cur, "active_prediction", il); + + // norm + cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // laurel + ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens] + + // self-attention + if (has_kv) { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps); + + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + cb(Vcur, "Vcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur_pos", il); + cb(Kcur, "Kcur_pos", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il); + } else { + // no KV layers + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur_pos", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + cur = ggml_add(ctx0, cur, active_prediction); // [n_embd, n_tokens] + cb(cur, "attn_gated", il); + + ggml_tensor * attn_laurel = ggml_scale(ctx0, + ggml_add(ctx0, cur, laurel_out), + 1.0f / sqrtf(2.0f)); // [n_embd, n_tokens] + cb(attn_laurel, "attn_laurel", il); + + cur = build_norm(attn_laurel, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + ggml_tensor * up_proj = build_lora_mm(model.layers[il].ffn_up, cur); + ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur); + + if (il < n_layer_sparsity) { + // apply activation sparsity + gate_proj = gaussian_topk(gate_proj); + } + gate_proj = ggml_gelu(ctx0, gate_proj); + + cur = ggml_mul(ctx0, up_proj, gate_proj); + cur = build_lora_mm(model.layers[il].ffn_down, cur); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", il); + + ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel); // [n_embd, n_tokens] + cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il); + + ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il); // [n_embd, n_tokens, n_altup] + + ggml_tensor * first_prediction; // [n_embd, n_tokens] + { + first_prediction = view_2d_slice(corrected, i_altup_act); // [n_embd, n_tokens] + first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale); + first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction); + first_prediction = ggml_gelu(ctx0, first_prediction); // [n_embd_altup, n_tokens] + cb(first_prediction, "first_prediction_gated", il); + ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_altup, n_tokens] + first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer); // [n_embd_altup, n_tokens] + cb(first_prediction, "first_prediction_scaled", il); + + first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction); // [n_embd, n_tokens] + first_prediction = build_norm(first_prediction, + model.layers[il].per_layer_post_norm, NULL, + LLM_NORM_RMS, il); + cb(first_prediction, "first_prediction_out", il); + } + + // equivalent to python code: corrected_predictions[1:] += first_prediction + { + ggml_tensor * slice_first = view_2d_slice(corrected, 0); + ggml_tensor * slice_rest = ggml_view_3d(ctx0, corrected, n_embd, n_tokens, n_altup - 1, + ggml_row_size(corrected->type, n_embd), + ggml_row_size(corrected->type, n_embd*n_tokens), + n_embd*n_tokens*ggml_element_size(corrected)); + ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction); // [n_embd, n_tokens, n_altup - 1] + corrected = ggml_concat(ctx0, slice_first, tmp, 2); // [n_embd, n_tokens, n_altup] + } + + cur = corrected; // [n_embd, n_tokens, n_altup] + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; // [n_embd, n_tokens, n_altup] + + // cur now has multiple altup(s), we want to merge them back to 1 altup + { + ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act)); // [n_embd, n_tokens] + // do a view to skip the first slice (active altup) + ggml_tensor * alt_slice = ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, + ggml_row_size(cur->type, n_embd), + ggml_row_size(cur->type, n_embd*n_tokens), + n_embd*n_tokens*ggml_element_size(cur)); + ggml_tensor * altup_unembd = ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice); // shape: [n_embd, n_tokens, n_altup - 1] + ggml_tensor * new_magnitude = calc_magnitude(altup_unembd); + altup_unembd = ggml_div(ctx0, + ggml_mul(ctx0, altup_unembd, target_magnitude), + new_magnitude); + cb(altup_unembd, "altup_unembd", -1); + + // equivalent to torch.mean(hidden_states, dim=0) + cur = view_2d_slice(cur, 0); // [n_embd, n_tokens] + for (int i = 0; i < n_altup - 1; ++i) { + cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i)); + } + cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup)); // [n_embd, n_tokens] + cb(cur, "unembd_merged", -1); + } + + // cur now has shape: [n_embd, n_tokens] + + // TODO: move this to right after the last KV layer + { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + { + // final logit soft-capping + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + ggml_tensor * calc_magnitude(ggml_tensor * x) { + return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x))); + } + + // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim + ggml_tensor * view_2d_slice(ggml_tensor * x, int idx) { + GGML_ASSERT(idx < (int)x->ne[2]); + return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], + ggml_row_size(x->type, x->ne[0]), + idx * x->ne[0] * x->ne[1] * ggml_element_size(x)); + } + + // equivalent to get_per_layer_inputs() in python code + // output shape: [n_embd_altup, n_layer, n_tokens] + ggml_tensor * get_per_layer_inputs() { + auto inp = std::make_unique(); + ggml_tensor * inp_per_layer; + if (ubatch.token) { + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens); + ggml_set_input(inp->tokens); + res->t_tokens = inp->tokens; + inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens); + inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens); + inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float)n_embd_altup)); + cb(inp_per_layer, "inp_per_layer_selected", -1); + } else { + GGML_ABORT("TODO: support embd input"); + } + res->add_input(std::move(inp)); + return inp_per_layer; + } + + // equivalent to project_per_layer_inputs() in python code + // this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim + // output shape: [n_embd_altup, n_tokens, n_layer] + ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) { + const float per_layer_projection_scale = 1.0f / sqrtf((float)n_embd); + const float per_layer_input_scale = 1.0f / sqrtf(2.0f); + + ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds); + per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale); + per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens); + per_layer_proj = build_norm(per_layer_proj, + model.per_layer_proj_norm, NULL, + LLM_NORM_RMS, -1); // [n_embd_altup, n_layer, n_tokens] + cb(per_layer_proj, "per_layer_proj", -1); + + inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj); + inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale); + cb(inp_per_layer, "inp_per_layer", -1); + + // permute to shape: [n_embd_altup, n_tokens, n_layer] + inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3)); + return inp_per_layer; + } + + // input cur shape: [n_altup, n_tokens] + // output shape: [n_altup, n_tokens] + ggml_tensor * laurel(ggml_tensor * cur, int il) { + ggml_tensor * tmp = cur; + tmp = build_lora_mm(model.layers[il].laurel_l, tmp); + tmp = build_lora_mm(model.layers[il].laurel_r, tmp); + tmp = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il); + tmp = ggml_add(ctx0, tmp, cur); + cb(tmp, "laurel_out", il); + return tmp; + } + + // input x shape: [n_embd, n_tokens] + // output shape: [n_embd, n_tokens] + ggml_tensor * gaussian_topk(ggml_tensor * x) { + ggml_tensor * mean = ggml_mean(ctx0, x); + ggml_tensor * std = ggml_sqrt(ctx0, ggml_scale(ctx0, + ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))), + 1.0f / (float)(x->ne[0] - 1) + )); + ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul)); + return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x)); + } + + // + // altup functions + // + + // equivalent to compute_router_modalities() in python code + // input x shape: [n_embd, n_tokens] + // output shape: [n_altup, n_tokens] + ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il) { + ggml_tensor * router_inputs = build_norm(x, + model.layers[il].altup_router_norm, NULL, + LLM_NORM_RMS, il); + + // router_input_scale + router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float)n_embd); + + ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs); + return ggml_tanh(ctx0, output); // [n_altup, n_tokens] + } + + // input cur shape: [n_embd, n_tokens, n_altup] + // output shape: [n_embd, n_tokens, n_altup] + ggml_tensor * altup_predict(ggml_tensor * cur, int il) { + ggml_tensor * activated = view_2d_slice(cur, i_altup_act); // [n_embd, n_tokens] + ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] + cb(modalities, "modalities", il); + + ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities); + cb(all_coefs, "all_coefs", il); + // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor) + all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens); + + // permute to [n_altup, n_embd, n_tokens] + ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3)); + ggml_tensor * predictions = ggml_mul_mat(ctx0, cur_permuted, all_coefs); // [n_altup, n_embd, n_tokens] + + // final shape must be the same as cur: [n_embd, n_tokens, n_altup] + predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3)); + predictions = ggml_add(ctx0, predictions, cur); + cb(predictions, "predictions", il); + + return predictions; + } + + // input predictions shape: [n_embd, n_tokens, n_altup] + // input activated shape: [n_embd, n_tokens] + // output shape: [n_embd, n_tokens, n_altup] + ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) { + ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] + cb(modalities, "modalities", il); + + ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); + ggml_tensor * innovation = ggml_sub(ctx0, activated, active_prediction); // [n_embd, n_tokens] + cb(innovation, "innovation", il); + + ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens] + all_coefs = ggml_add(ctx0, all_coefs, one); + cb(all_coefs, "all_coefs", il); + all_coefs = ggml_cont(ctx0, ggml_transpose(ctx0, all_coefs)); // [n_tokens, n_altup] + all_coefs = ggml_reshape_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup] + + innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1); + ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup] + corrected = ggml_add(ctx0, corrected, predictions); // [n_embd, n_tokens, n_altup] + cb(corrected, "corrected", il); + + return corrected; + } +}; + // TODO: move up next to build_starcoder struct llm_build_starcoder2 : public llm_graph_context { llm_build_starcoder2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { @@ -13974,6 +14486,10 @@ llm_graph_result_ptr llama_model::build_graph( { llm = std::make_unique(*this, params, gf); } break; + case LLM_ARCH_GEMMA3N: + { + llm = std::make_unique(*this, params, gf); + } break; case LLM_ARCH_STARCODER2: { llm = std::make_unique(*this, params, gf); @@ -14295,6 +14811,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: case LLM_ARCH_GEMMA3: + case LLM_ARCH_GEMMA3N: case LLM_ARCH_STARCODER2: case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: diff --git a/src/llama-model.h b/src/llama-model.h index 06e6c6879..40063b790 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -95,6 +95,8 @@ enum llm_type { LLM_TYPE_17B_128E, // llama4 Maverick LLM_TYPE_30B_A3B, LLM_TYPE_235B_A22B, + LLM_TYPE_E2B, + LLM_TYPE_E4B, }; std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type); @@ -316,6 +318,19 @@ struct llama_layer { struct ggml_tensor * ffn_up_scale = nullptr; struct ggml_tensor * ffn_down_scale = nullptr; + // altup & laurel + struct ggml_tensor * per_layer_inp_gate = nullptr; + struct ggml_tensor * per_layer_proj = nullptr; + struct ggml_tensor * per_layer_post_norm = nullptr; + struct ggml_tensor * altup_correct_coef = nullptr; + struct ggml_tensor * altup_correct_scale = nullptr; + struct ggml_tensor * altup_predict_coef = nullptr; + struct ggml_tensor * altup_router = nullptr; + struct ggml_tensor * altup_router_norm = nullptr; + struct ggml_tensor * laurel_l = nullptr; + struct ggml_tensor * laurel_r = nullptr; + struct ggml_tensor * laurel_post_norm = nullptr; + struct llama_layer_posnet posnet; struct llama_layer_convnext convnext; @@ -354,6 +369,13 @@ struct llama_model { struct ggml_tensor * conv1d = nullptr; struct ggml_tensor * conv1d_b = nullptr; + // gemma3n altup + struct ggml_tensor * tok_embd_per_layer = nullptr; + struct ggml_tensor * altup_proj = nullptr; + struct ggml_tensor * altup_unembd_proj = nullptr; + struct ggml_tensor * per_layer_model_proj = nullptr; + struct ggml_tensor * per_layer_proj_norm = nullptr; + std::vector layers; llama_model_params params; diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index 43229e193..f4b5713d7 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -223,7 +223,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t new_type = GGML_TYPE_Q6_K; } } - } else if (name == "token_embd.weight") { + } else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") { if (qs.params->token_embedding_type < GGML_TYPE_COUNT) { new_type = qs.params->token_embedding_type; } else { @@ -830,6 +830,13 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: // NOTE: can't use LLM_TN here because the layer number is not known quantize &= name.find("ffn_gate_inp.weight") == std::string::npos; + // these are very small (e.g. 4x4) + quantize &= name.find("altup") == std::string::npos; + quantize &= name.find("laurel") == std::string::npos; + + // these are not too big so keep them as it is + quantize &= name.find("per_layer_model_proj") == std::string::npos; + // do not quantize positional embeddings and token types (BERT) quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD, "weight"); quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");