From c747227a57203cbb1a6d83ed74ac6d08e4be9a5e Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Wed, 7 May 2025 15:02:11 -0500 Subject: [PATCH] vulkan: reduce register usage in scalar FA, but perf may be slightly worse --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 16 ++++++------ .../vulkan-shaders/flash_attn.comp | 26 ++++++++++++++----- 2 files changed, 27 insertions(+), 15 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 8158586d4..b285f43e9 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1921,14 +1921,14 @@ static void ggml_vk_load_shaders(vk_device& device) { }; #define CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, D) \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1]); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true); \ #define CREATE_FA(TYPE, NAMELC, SCALAR, SUFFIX) \ CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 64) \ diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp index abdadee28..cf3b72a4a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp @@ -106,6 +106,9 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x]; +shared float16_t masksh[Bc][Br]; +shared vec4 Qf[Br][D / 4]; + void main() { #ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); @@ -164,11 +167,11 @@ void main() { uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV; uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4; - vec4 Qf[Br][D_per_thread / 4]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { if (i * Br + r < N) { [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { - Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d * D_split + d_tid]) * p.scale; + Qf[r][d * D_split + d_tid] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d * D_split + d_tid]) * p.scale; } } } @@ -218,7 +221,7 @@ void main() { [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]); [[unroll]] for (uint32_t r = 0; r < Br; ++r) { - Sf[r][c] += dot(Qf[r][d], K_Tf); + Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf); } } } @@ -241,13 +244,22 @@ void main() { } if (p.mask != 0) { + + [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) { + uint32_t c = (idx + tid) % Bc; + uint32_t r = (idx + tid) / Bc; + masksh[c][r] = data_m[(i * Br + r) * m_stride + (j * Bc + c)]; + } + barrier(); + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { [[unroll]] for (uint32_t r = 0; r < Br; ++r) { - float mvf = data_m[(i * Br + r) * m_stride + (j * Bc + c * cols_per_iter + col_tid)]; + float mvf = float(masksh[c * cols_per_iter + col_tid][r]); Sf[r][c] += slope[r]*mvf; } } + barrier(); } float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br]; @@ -363,7 +375,7 @@ void main() { [[unroll]] for (uint32_t r = 0; r < Br; ++r) { if (r < N) { - for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); } @@ -398,7 +410,7 @@ void main() { if (p.gqa_ratio > 1) { [[unroll]] for (uint32_t r = 0; r < Br; ++r) { if (r < N) { - for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); } @@ -408,7 +420,7 @@ void main() { } else { [[unroll]] for (uint32_t r = 0; r < Br; ++r) { if (i * Br + r < N) { - for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]); }