diff --git a/include/ck_tile/core/arch/amd_buffer_addressing.hpp b/include/ck_tile/core/arch/amd_buffer_addressing.hpp index 037e86909d..7a9c017eb2 100644 --- a/include/ck_tile/core/arch/amd_buffer_addressing.hpp +++ b/include/ck_tile/core/arch/amd_buffer_addressing.hpp @@ -1833,14 +1833,17 @@ CK_TILE_DEVICE void amd_async_buffer_load(CK_TILE_LDS_ADDR T* smem, if constexpr(oob_conditional_check) v_offset = flag ? v_offset : src_wave_buffer_resource[2]; - llvm_amdgcn_raw_buffer_load_lds( - src_wave_buffer_resource, - reinterpret_cast(reinterpret_cast(smem)), - bytes, - v_offset, - src_wave_addr_offset, - /*src_immediate_addr_offset*/ 0, - static_cast(coherence)); +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wold-style-cast" + // Use C-style cast to change address space without dropping llvm noalias attribute + llvm_amdgcn_raw_buffer_load_lds(src_wave_buffer_resource, + (as3_uint32_ptr)(smem), + bytes, + v_offset, + src_wave_addr_offset, + /*src_immediate_addr_offset*/ 0, + static_cast(coherence)); +#pragma clang diagnostic pop } template & src_thread_ template __device__ auto amd_transpose_load_to_vgpr(const T* __restrict__ in_ptr) { +#define __LDS_ADDR __attribute__((address_space(3))) static_assert(__has_builtin(__builtin_amdgcn_raw_buffer_load_b32), "We need to have the compatible compiler version to build this instruction"); + +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wold-style-cast" + // Use C-style cast to change address space without dropping llvm noalias attribute + const auto in_ptr_ = (__LDS_ADDR T*)(const_cast(in_ptr)); +#pragma clang diagnostic pop if constexpr(std::is_same_v, ck_tile::half_t>) { typedef __attribute__((__vector_size__(4 * sizeof(__fp16)))) __fp16 llvm_fp16x4_t; - __attribute__((address_space(3))) llvm_fp16x4_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_fp16x4_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_fp16x4_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr16_b64_v4f16(lds_ptr)); } else if constexpr(std::is_same_v, ck_tile::bf16_t>) { typedef __attribute__((__vector_size__(4 * sizeof(__bf16)))) __bf16 llvm_bf16x4_t; - __attribute__((address_space(3))) llvm_bf16x4_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_bf16x4_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_bf16x4_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr16_b64_v4bf16(lds_ptr)); } else if constexpr(std::is_same_v, ck_tile::fp8_t> || @@ -2812,15 +2818,14 @@ __device__ auto amd_transpose_load_to_vgpr(const T* __restrict__ in_ptr) std::is_same_v, ck_tile::int8_t>) { typedef __attribute__((__vector_size__(2 * sizeof(index_t)))) index_t llvm_i32x2_t; - __attribute__((address_space(3))) llvm_i32x2_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_i32x2_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_i32x2_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr8_b64_v2i32(lds_ptr)); } else { static_assert(false, "not implemented"); } +#undef __LDS_ADDR } #endif diff --git a/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp b/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp index d1e4eb3da3..4013b51479 100644 --- a/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp +++ b/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp @@ -1603,14 +1603,17 @@ CK_TILE_DEVICE void amd_async_buffer_load(CK_TILE_LDS_ADDR T* smem, if constexpr(oob_conditional_check) v_offset = flag ? v_offset : src_wave_buffer_resource[2]; - llvm_amdgcn_raw_buffer_load_lds( - src_wave_buffer_resource, - reinterpret_cast(reinterpret_cast(smem)), - bytes, - v_offset, - src_wave_addr_offset, - /*src_immediate_addr_offset*/ 0, - static_cast(coherence)); +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wold-style-cast" + // Use C-style cast to change address space without dropping llvm noalias attribute + llvm_amdgcn_raw_buffer_load_lds(src_wave_buffer_resource, + (as3_uint32_ptr)(smem), + bytes, + v_offset, + src_wave_addr_offset, + /*src_immediate_addr_offset*/ 0, + static_cast(coherence)); +#pragma clang diagnostic pop } template __device__ auto amd_transpose_load_to_vgpr(const T* __restrict__ in_ptr) { +#define __LDS_ADDR __attribute__((address_space(3))) static_assert(__has_builtin(__builtin_amdgcn_raw_buffer_load_b32), "We need to have the compatible compiler version to build this instruction"); + +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wold-style-cast" + // Use C-style cast to change address space without dropping llvm noalias attribute + const auto in_ptr_ = (__LDS_ADDR T*)(const_cast(in_ptr)); +#pragma clang diagnostic pop if constexpr(std::is_same_v, ck_tile::half_t>) { typedef __attribute__((__vector_size__(4 * sizeof(__fp16)))) __fp16 llvm_fp16x4_t; - __attribute__((address_space(3))) llvm_fp16x4_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_fp16x4_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_fp16x4_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr16_b64_v4f16(lds_ptr)); } else if constexpr(std::is_same_v, ck_tile::bf16_t>) { typedef __attribute__((__vector_size__(4 * sizeof(__bf16)))) __bf16 llvm_bf16x4_t; - __attribute__((address_space(3))) llvm_bf16x4_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_bf16x4_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_bf16x4_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr16_b64_v4bf16(lds_ptr)); } else if constexpr(std::is_same_v, ck_tile::fp8_t> || @@ -2630,15 +2636,14 @@ __device__ auto amd_transpose_load_to_vgpr(const T* __restrict__ in_ptr) std::is_same_v, ck_tile::int8_t>) { typedef __attribute__((__vector_size__(2 * sizeof(index_t)))) index_t llvm_i32x2_t; - __attribute__((address_space(3))) llvm_i32x2_t* lds_ptr = - reinterpret_cast<__attribute__((address_space(3))) llvm_i32x2_t*>( - reinterpret_cast(in_ptr)); + auto lds_ptr = reinterpret_cast<__LDS_ADDR llvm_i32x2_t*>(in_ptr_); return bit_cast>(__builtin_amdgcn_ds_read_tr8_b64_v2i32(lds_ptr)); } else { static_assert(false, "not implemented"); } +#undef __LDS_ADDR } #endif diff --git a/include/ck_tile/core/tensor/buffer_view.hpp b/include/ck_tile/core/tensor/buffer_view.hpp index ca314a6abe..d1e770ef42 100644 --- a/include/ck_tile/core/tensor/buffer_view.hpp +++ b/include/ck_tile/core/tensor/buffer_view.hpp @@ -62,12 +62,12 @@ struct buffer_view -CK_TILE_HOST_DEVICE constexpr auto make_buffer_view(T* p, BufferSizeType buffer_size) +CK_TILE_HOST_DEVICE constexpr auto make_buffer_view(T* __restrict__ p, BufferSizeType buffer_size) { return buffer_view{p, buffer_size}; } @@ -1266,7 +1266,7 @@ template , remove_cvref_t>::value, bool>::type = false> CK_TILE_HOST_DEVICE constexpr auto -make_buffer_view(T* p, BufferSizeType buffer_size, X invalid_element_value) +make_buffer_view(T* __restrict__ p, BufferSizeType buffer_size, X invalid_element_value) { return buffer_view{ p, buffer_size, invalid_element_value}; diff --git a/include/ck_tile/core/tensor/tensor_view.hpp b/include/ck_tile/core/tensor/tensor_view.hpp index a85dbc6d00..6fa8f898e5 100644 --- a/include/ck_tile/core/tensor/tensor_view.hpp +++ b/include/ck_tile/core/tensor/tensor_view.hpp @@ -449,7 +449,7 @@ template -CK_TILE_HOST_DEVICE constexpr auto make_tensor_view(DataType* p, +CK_TILE_HOST_DEVICE constexpr auto make_tensor_view(DataType* __restrict__ p, const tensor_descriptor& desc) { auto buffer_view = @@ -468,7 +468,7 @@ template ::type = false> CK_TILE_HOST_DEVICE constexpr auto -make_naive_tensor_view(DataType* p, +make_naive_tensor_view(DataType* __restrict__ p, const tuple& lengths, const tuple& strides, number = number<-1>{}, @@ -491,7 +491,7 @@ template CK_TILE_HOST_DEVICE constexpr auto -make_naive_tensor_view_packed(DataType* p, +make_naive_tensor_view_packed(DataType* __restrict__ p, const tuple& lengths, number = number<-1>{}) { diff --git a/include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp b/include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp index 5e16fc563b..3f5bef366e 100644 --- a/include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp +++ b/include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp @@ -1115,7 +1115,8 @@ struct FmhaBwdDQDKDVKernel {i_n0, 0}); if constexpr(!kUseQrQtrDorPipeline) { - auto [dk_acc_tile, dv_acc_tile] = FmhaPipeline{}(q_dram_window, + auto [dk_acc_tile, dv_acc_tile] = FmhaPipeline{}(smem_ptr, + q_dram_window, k_dram_window, v_dram_window, bias_dram_window, @@ -1131,7 +1132,6 @@ struct FmhaBwdDQDKDVKernel kargs.scale, rp_undrop, scale_rp_undrop, - smem_ptr, dropout); KGradEpiloguePipeline{}(dk_dram_window, dk_acc_tile); @@ -1139,7 +1139,8 @@ struct FmhaBwdDQDKDVKernel } else { - FmhaPipeline{}(q_dram_window, + FmhaPipeline{}(smem_ptr, + q_dram_window, k_dram_window, v_dram_window, bias_dram_window, @@ -1160,7 +1161,6 @@ struct FmhaBwdDQDKDVKernel kargs.scale, rp_undrop, scale_rp_undrop, - smem_ptr, dropout); } } diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp index d36f8ad724..5e63fb714a 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp @@ -93,7 +93,8 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR typename BiasGradDramBlockWindowTmp, typename PositionEncoding> CK_TILE_HOST_DEVICE auto - operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, + operator()(void* smem_ptr, + const QDramBlockWindowTmp& q_dram_block_window_tmp, const KDramBlockWindowTmp& k_dram_block_window_tmp, const VDramBlockWindowTmp& v_dram_block_window_tmp, const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, @@ -109,7 +110,6 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR float scale, float rp_undrop, float scale_rp_undrop, - void* smem_ptr, FmhaDropout& dropout) const { static_assert( diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp.hpp index 88fb1281aa..b883aad155 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp.hpp @@ -93,7 +93,8 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP typename BiasGradDramBlockWindowTmp, typename PositionEncoding> CK_TILE_HOST_DEVICE auto - operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, + operator()(void* smem_ptr, + const QDramBlockWindowTmp& q_dram_block_window_tmp, const KDramBlockWindowTmp& k_dram_block_window_tmp, const VDramBlockWindowTmp& v_dram_block_window_tmp, const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, @@ -109,7 +110,6 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP float scale, float rp_undrop, float scale_rp_undrop, - void* smem_ptr, FmhaDropout& dropout) const { static_assert( diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_kr_ktr_vr.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_kr_ktr_vr.hpp index 9a31498dd1..9bd78b4077 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_kr_ktr_vr.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_kr_ktr_vr.hpp @@ -90,6 +90,53 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR else return raw_lse; }; + template + CK_TILE_DEVICE auto operator()(void* smem_ptr, Ts&&... args) const + { + // LDS allocation + // cast to char* to do pointer arithmetic + const auto smem_ptr_ = reinterpret_cast(smem_ptr); + const auto k_lds_ptr = reinterpret_cast(smem_ptr_); + const auto v_lds_ptr = + reinterpret_cast(smem_ptr_ + Policy::template GetSmemSizeK()); + + const auto do_lds_ptr0 = reinterpret_cast(smem_ptr_); + const auto do_lds_ptr1 = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeOGrad()); + const auto q_lds_ptr0 = reinterpret_cast( // + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeOGrad()); + const auto q_lds_ptr1 = reinterpret_cast( // + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ()); + const auto lse_lds_ptr = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ()); + const auto d_lds_ptr = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ() + + Policy::template GetSmemSizeLSE()); + const auto ds_lds_ptr = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ() + + Policy::template GetSmemSizeLSE() + Policy::template GetSmemSizeD()); + const auto bias_lds_ptr = reinterpret_cast(ds_lds_ptr); + return run(k_lds_ptr, + v_lds_ptr, + do_lds_ptr0, + do_lds_ptr1, + q_lds_ptr0, + q_lds_ptr1, + lse_lds_ptr, + d_lds_ptr, + ds_lds_ptr, + bias_lds_ptr, + std::forward(args)...); + } template - CK_TILE_DEVICE auto operator()( // + CK_TILE_DEVICE auto run( // + KDataType* __restrict__ k_lds_ptr, + VDataType* __restrict__ v_lds_ptr, + OGradDataType* __restrict__ do_lds_ptr0, + OGradDataType* __restrict__ do_lds_ptr1, + QDataType* __restrict__ q_lds_ptr0, + QDataType* __restrict__ q_lds_ptr1, + LSEDataType* __restrict__ lse_lds_ptr, + DDataType* __restrict__ d_lds_ptr, + GemmDataType* __restrict__ ds_lds_ptr, + BiasDataType* __restrict__ bias_lds_ptr, const QDramBlockWindowTmp& q_dram_block_window_tmp, const KDramBlockWindowTmp& k_dram_block_window_tmp, const VDramBlockWindowTmp& v_dram_block_window_tmp, @@ -119,7 +176,6 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR float scale, float rp_undrop, float scale_rp_undrop, - void* smem_ptr, FmhaDropout& dropout) const { static_assert( @@ -184,40 +240,6 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR } } - // LDS allocation - const auto smem_ptr_ = - reinterpret_cast(smem_ptr); // cast to char* to do pointer arithmetic - - const auto k_lds_ptr = reinterpret_cast(smem_ptr_); - const auto v_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeK()); - - const auto do_lds_ptr0 = reinterpret_cast(smem_ptr_); - const auto do_lds_ptr1 = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad()); - const auto q_lds_ptr0 = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeOGrad()); - const auto q_lds_ptr1 = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ()); - const auto lse_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ()); - const auto d_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ() + - Policy::template GetSmemSizeLSE()); - const auto ds_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeQ() + - Policy::template GetSmemSizeLSE() + Policy::template GetSmemSizeD()); - const auto bias_lds_ptr = reinterpret_cast(ds_lds_ptr); - auto k_lds = make_tensor_view( k_lds_ptr, Policy::template MakeKLdsWriteBlockDescriptor()); auto k_lds_write_window = @@ -453,13 +475,12 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR decltype(load_tile(d_dram_window)) d_block_tile; index_t i_total_bodys = 0; - auto main_body = [&](auto is_prologue_, auto is_epilogue_) mutable { - const bool is_even = (i_total_bodys % 2 == 0); - QDataType* const __restrict__ q_lds_ptr_curr = is_even ? q_lds_ptr1 : q_lds_ptr0; - QDataType* const __restrict__ q_lds_ptr_next = is_even ? q_lds_ptr0 : q_lds_ptr1; - OGradDataType* const __restrict__ do_lds_ptr_curr = is_even ? do_lds_ptr1 : do_lds_ptr0; - OGradDataType* const __restrict__ do_lds_ptr_next = is_even ? do_lds_ptr0 : do_lds_ptr1; - + auto main_body_impl = [&](auto is_prologue_, + auto is_epilogue_, + QDataType* const __restrict__ q_lds_ptr_curr, + QDataType* const __restrict__ q_lds_ptr_next, + OGradDataType* const __restrict__ do_lds_ptr_curr, + OGradDataType* const __restrict__ do_lds_ptr_next) mutable { constexpr bool is_prologue = is_prologue_.value; constexpr bool is_epilogue = is_epilogue_.value; static_assert(is_prologue || is_epilogue, "is_prologue or is_epilogue should be true"); @@ -467,19 +488,19 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR if constexpr(is_prologue) { + lse_block_tile = load_tile(lse_dram_window); + move_tile_window(lse_dram_window, {kM0}); + + d_block_tile = load_tile(d_dram_window); + move_tile_window(d_dram_window, {kM0}); + q_lds_write_window.set_bottom_tensor_view_data_ptr(q_lds_ptr_next); async_load_tile(q_lds_write_window, q_dram_window); move_tile_window(q_dram_window, {kM0, 0}); - lse_block_tile = load_tile(lse_dram_window); - move_tile_window(lse_dram_window, {kM0}); - do_lds_write_window.set_bottom_tensor_view_data_ptr(do_lds_ptr_next); async_load_tile(do_lds_write_window, do_dram_window); move_tile_window(do_dram_window, {kM0, 0}); - - d_block_tile = load_tile(d_dram_window); - move_tile_window(d_dram_window, {kM0}); } if constexpr(is_epilogue) { @@ -611,8 +632,8 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR constexpr auto i_j_idx = make_tuple(idx0, idx1); bool undrop_flag = p[i_j_idx] >= 0; ds(i_j_idx) = p[i_j_idx] * (!FmhaDropout::IsDropout || undrop_flag - ? (dp_acc[i_j_idx] - d[i_idx]) - : d[i_idx]); + ? (dp_acc[i_j_idx] - d[i_idx]) + : d[i_idx]); }); }); @@ -725,6 +746,20 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadKRKTRVR } move_tile_window(dq_dram_window, {kM0, 0}); } + }; + + auto main_body = [&](auto is_prologue_, auto is_epilogue_) mutable { + const bool is_even = (i_total_bodys % 2 == 0); + const auto q_lds_ptr_curr = is_even ? q_lds_ptr1 : q_lds_ptr0; + const auto q_lds_ptr_next = is_even ? q_lds_ptr0 : q_lds_ptr1; + const auto do_lds_ptr_curr = is_even ? do_lds_ptr1 : do_lds_ptr0; + const auto do_lds_ptr_next = is_even ? do_lds_ptr0 : do_lds_ptr1; + main_body_impl(is_prologue_, + is_epilogue_, + q_lds_ptr_curr, + q_lds_ptr_next, + do_lds_ptr_curr, + do_lds_ptr_next); i_total_bodys += 1; }; diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_qr_qtr_dor.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_qr_qtr_dor.hpp index 789cfb3ea4..5adb64564d 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_qr_qtr_dor.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_trload_qr_qtr_dor.hpp @@ -93,6 +93,42 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadQRQTRDOR return raw_lse; }; + template + CK_TILE_DEVICE auto operator()(void* smem_ptr, Ts&&... args) const + { + // LDS allocation + const auto smem_ptr_ = + reinterpret_cast(smem_ptr); // cast to char* to do pointer arithmetic + + const auto k_lds_ptr = reinterpret_cast(smem_ptr_); + const auto v_lds_ptr = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeK()); + + const auto do_lds_ptr = reinterpret_cast(smem_ptr_); + const auto q_lds_ptr = reinterpret_cast( // + smem_ptr_ + Policy::template GetSmemSizeOGrad()); + const auto lse_lds_ptr = reinterpret_cast( // + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ()); + const auto d_lds_ptr = reinterpret_cast( + smem_ptr_ + Policy::template GetSmemSizeOGrad() + + Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeLSE()); + + const auto ds_lds_ptr = + reinterpret_cast(smem_ptr_ + Policy::template GetSmemSizeK() + + Policy::template GetSmemSizeV()); + const auto bias_lds_ptr = reinterpret_cast(ds_lds_ptr); + return run(k_lds_ptr, + v_lds_ptr, + do_lds_ptr, + q_lds_ptr, + lse_lds_ptr, + d_lds_ptr, + ds_lds_ptr, + bias_lds_ptr, + std::forward(args)...); + } + template - CK_TILE_DEVICE auto operator()( // + CK_TILE_DEVICE auto run( // + KDataType* __restrict__ k_lds_ptr, + VDataType* __restrict__ v_lds_ptr, + OGradDataType* __restrict__ do_lds_ptr, + QDataType* __restrict__ q_lds_ptr, + LSEDataType* __restrict__ lse_lds_ptr, + DDataType* __restrict__ d_lds_ptr, + GemmDataType* __restrict__ ds_lds_ptr, + BiasDataType* __restrict__ bias_lds_ptr, const QDramBlockWindowTmp& q_dram_block_window_tmp, const KDramBlockWindowTmp& k_dram_block_window_tmp, const VDramBlockWindowTmp& v_dram_block_window_tmp, @@ -131,7 +175,6 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadQRQTRDOR float scale, float rp_undrop, float scale_rp_undrop, - void* smem_ptr, FmhaDropout& dropout) const { static_assert( @@ -181,29 +224,6 @@ struct BlockFmhaBwdDQDKDVPipelineTrLoadQRQTRDOR {seqlen_kv_start, 0}, Policy::template MakeKDramTileDistribution()); - // LDS allocation - const auto smem_ptr_ = - reinterpret_cast(smem_ptr); // cast to char* to do pointer arithmetic - - const auto k_lds_ptr = reinterpret_cast(smem_ptr_); - const auto v_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeK()); - - const auto do_lds_ptr = reinterpret_cast(smem_ptr_); - const auto q_lds_ptr = reinterpret_cast( // - smem_ptr_ + Policy::template GetSmemSizeOGrad()); - const auto lse_lds_ptr = reinterpret_cast( // - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ()); - const auto d_lds_ptr = reinterpret_cast( - smem_ptr_ + Policy::template GetSmemSizeOGrad() + - Policy::template GetSmemSizeQ() + Policy::template GetSmemSizeLSE()); - - const auto ds_lds_ptr = - reinterpret_cast(smem_ptr_ + Policy::template GetSmemSizeK() + - Policy::template GetSmemSizeV()); - const auto bias_lds_ptr = reinterpret_cast(ds_lds_ptr); - auto k_lds = make_tensor_view( k_lds_ptr, Policy::template MakeKLdsWriteBlockDescriptor()); auto k_lds_write_window = diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_trload.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_trload.hpp index 39d8814692..aafe481d2b 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_trload.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_trload.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -638,11 +638,11 @@ struct BlockFmhaPipelineQRKSVSAsyncTrload typename LSEaccDramBlockWindowTmp, typename PositionEncoding> CK_TILE_HOST_DEVICE auto - operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile - const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile - const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile - const BiasDramBlockWindowTmp& bias_dram_block_window_tmp, // M0*N0 tile - LSEaccDramBlockWindowTmp& lse_acc_dram_window_tmp, // M0*1 tile + operator()(const QDramBlockWindowTmp& __restrict__ q_dram_block_window_tmp, // M0*K0 tile + const KDramBlockWindowTmp& __restrict__ k_dram_block_window_tmp, // N0*K0 tile + const VDramBlockWindowTmp& __restrict__ v_dram_block_window_tmp, // N1*K1 tile + const BiasDramBlockWindowTmp& __restrict__ bias_dram_block_window_tmp, // M0*N0 tile + LSEaccDramBlockWindowTmp& __restrict__ lse_acc_dram_window_tmp, // M0*1 tile FmhaMask mask, PositionEncoding position_encoding, float scale_s, @@ -854,18 +854,10 @@ struct BlockFmhaPipelineQRKSVSAsyncTrload __builtin_amdgcn_sched_barrier(0); - auto mainloop = [&](index_t cur_loop) { - const bool is_even_loop = (cur_loop % 2 == 0); - - auto k_lds_write_ptr = is_even_loop ? static_cast(smem_ptrk0) - : static_cast(smem_ptrk1); - auto k_lds_read_ptr = is_even_loop ? static_cast(smem_ptrk1) - : static_cast(smem_ptrk0); - auto v_lds_write_ptr = is_even_loop ? static_cast(smem_ptrv1) - : static_cast(smem_ptrv0); - auto v_lds_read_ptr = is_even_loop ? static_cast(smem_ptrv0) - : static_cast(smem_ptrv1); - + auto mainloop = [&](KDataType* __restrict__ k_lds_write_ptr, + KDataType* __restrict__ k_lds_read_ptr, + KDataType* __restrict__ v_lds_write_ptr, + KDataType* __restrict__ v_lds_read_ptr) { // move V tile windows block_sync_lds(); move_tile_window(v_dram_window, {kN0, 0}); @@ -1110,11 +1102,20 @@ struct BlockFmhaPipelineQRKSVSAsyncTrload __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS_READ }); - }; + }; // mainloop do { - mainloop(i_total_loops); + bool is_even_loop = i_total_loops % 2 == 0; + auto k_lds_write_ptr = is_even_loop ? static_cast(smem_ptrk0) + : static_cast(smem_ptrk1); + auto k_lds_read_ptr = is_even_loop ? static_cast(smem_ptrk1) + : static_cast(smem_ptrk0); + auto v_lds_write_ptr = is_even_loop ? static_cast(smem_ptrv1) + : static_cast(smem_ptrv0); + auto v_lds_read_ptr = is_even_loop ? static_cast(smem_ptrv0) + : static_cast(smem_ptrv1); + mainloop(k_lds_write_ptr, k_lds_read_ptr, v_lds_write_ptr, v_lds_read_ptr); i_total_loops++; } while(i_total_loops < num_total_loop);