mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-11 09:34:37 +00:00
vulkan: Add bfloat16 support (#12554)
* vulkan: Add bfloat16 support This adds bfloat16 matrix multiply support based on VK_KHR_shader_bfloat16. The extension is required for coopmat multiply support, but matrix-vector multiply trivially promotes bf16 to fp32 and doesn't require the extension. The copy/get_rows shaders also don't require the extension. It's probably possible to fall back to non-coopmat and promote to fp32 when the extension isn't supported, but this change doesn't do that. The coopmat support also requires a glslc that supports the extension, which currently requires a custom build. * vulkan: Support bf16 tensors without the bf16 extension or coopmat support Compile a variant of the scalar mul_mm shader that will promote the bf16 values to float, and use that when either the bf16 extension or the coopmat extensions aren't available. * vulkan: bfloat16 fixes (really works without bfloat16 support now) * vulkan: fix spirv-val failure and reenable -O
This commit is contained in:
Jeff Bolz
GitHub
parent
fc727bcdd5
commit
79f26e9e12
13 changed files with 368 additions and 67 deletions
|
|
@ -63,7 +63,8 @@ const std::vector<std::string> type_names = {
|
|||
"iq3_xxs",
|
||||
"iq3_s",
|
||||
"iq4_xs",
|
||||
"iq4_nl"
|
||||
"iq4_nl",
|
||||
"bf16",
|
||||
};
|
||||
|
||||
namespace {
|
||||
|
|
@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
|
|||
std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
|
||||
|
||||
std::map<std::string, std::string> base_dict = {
|
||||
{"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"},
|
||||
{"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"},
|
||||
};
|
||||
std::string shader_name = "matmul";
|
||||
|
|
@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
|
|||
|
||||
const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
|
||||
|
||||
// Shaders with f16 B_TYPE
|
||||
string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string {
|
||||
if (t == "bf16") {
|
||||
// scalar path promotes to float
|
||||
if (!coopmat && !coopmat2) {
|
||||
return "float";
|
||||
}
|
||||
return "bfloat16_t";
|
||||
}
|
||||
if (coopmat2 || fp16) {
|
||||
return "float16_t";
|
||||
}
|
||||
return "float";
|
||||
};
|
||||
|
||||
string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
// Shaders with f16 B_TYPE
|
||||
string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
|
||||
string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
|
||||
// bf16
|
||||
{
|
||||
std::string load_vec_a_unaligned = "1";
|
||||
// For aligned matmul loads
|
||||
std::string load_vec_a = coopmat2 ? "1" : "4";
|
||||
|
||||
// scalar path promotes to float
|
||||
std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
|
||||
|
||||
// If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
|
||||
#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
|
||||
if (!(coopmat || coopmat2))
|
||||
#endif
|
||||
{
|
||||
string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_bf16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
}
|
||||
}
|
||||
|
||||
for (const auto& tname : type_names) {
|
||||
std::string load_vec_quant = "2";
|
||||
|
|
@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
|
|||
else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl"))
|
||||
load_vec_quant = "4";
|
||||
|
||||
if (tname == "bf16") {
|
||||
continue;
|
||||
}
|
||||
|
||||
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
|
||||
// For unaligned, load one at a time for f32/f16, or two at a time for quants
|
||||
std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant;
|
||||
std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
|
||||
// For aligned matmul loads
|
||||
std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant;
|
||||
std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
|
||||
|
||||
// don't generate f32 variants for coopmat2
|
||||
if (!coopmat2) {
|
||||
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
}
|
||||
|
||||
if (tname != "f16" && tname != "f32") {
|
||||
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
|
||||
}
|
||||
|
||||
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
|
||||
if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
|
||||
string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
|
||||
string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
|
@ -393,6 +430,7 @@ void process_shaders() {
|
|||
if (tname == "f32") {
|
||||
continue;
|
||||
}
|
||||
if (tname == "bf16") continue;
|
||||
|
||||
if (tname == "f16") {
|
||||
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
|
||||
|
|
@ -417,12 +455,12 @@ void process_shaders() {
|
|||
string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
|
||||
|
||||
// Dequant shaders
|
||||
if (tname != "f16") {
|
||||
if (tname != "f16" && tname != "bf16") {
|
||||
string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
|
||||
}
|
||||
|
||||
if (!string_ends_with(tname, "_k")) {
|
||||
shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
|
||||
shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";
|
||||
|
||||
if (tname == "f16") {
|
||||
string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
|
||||
|
|
@ -447,9 +485,11 @@ void process_shaders() {
|
|||
string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
|
||||
string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
|
||||
string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
|
||||
string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
|
||||
string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
|
||||
string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
|
||||
string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
|
||||
string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
|
||||
|
||||
for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
|
||||
string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue