From 8146e471f14ad61b623c8a6410b9eb13787df91f Mon Sep 17 00:00:00 2001 From: Illia Silin <98187287+illsilin@users.noreply.github.com> Date: Fri, 23 May 2025 14:58:25 -0700 Subject: [PATCH 1/6] fix the buffer intrinsic names for clang >=20 (#2228) --- .../ck_tile/18_flatmm/run_flatmm_example.inc | 74 +++--- .../amd_buffer_addressing_builtins.hpp | 20 +- .../arch/amd_buffer_addressing_builtins.hpp | 218 +++++++++--------- 3 files changed, 160 insertions(+), 152 deletions(-) diff --git a/example/ck_tile/18_flatmm/run_flatmm_example.inc b/example/ck_tile/18_flatmm/run_flatmm_example.inc index 15a9df2c0c..c191fff7d0 100644 --- a/example/ck_tile/18_flatmm/run_flatmm_example.inc +++ b/example/ck_tile/18_flatmm/run_flatmm_example.inc @@ -4,14 +4,22 @@ #include template -constexpr const char* DataTypeToString() { - if constexpr (std::is_same_v) { +constexpr const char* DataTypeToString() +{ + if constexpr(std::is_same_v) + { return "fp16"; - } else if constexpr (std::is_same_v) { + } + else if constexpr(std::is_same_v) + { return "fp8"; - } else if constexpr (std::is_same_v) { + } + else if constexpr(std::is_same_v) + { return "bf8"; - } else { + } + else + { return "unknown"; } } @@ -112,8 +120,9 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf, args.stride_B = stride_B; args.stride_C = stride_C; - float ave_time = flatmm_calc( - args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat}); + float ave_time = + flatmm_calc( + args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat}); std::size_t flop = std::size_t(2) * M * N * K; std::size_t num_byte = @@ -121,18 +130,15 @@ float invoke_flatmm(ck_tile::DeviceMem& a_dev_buf, float tflops = static_cast(flop) / 1.E9 / ave_time; float gb_per_sec = num_byte / 1.E6 / ave_time; - std::cout << "Run Flatmm kernel with DataType = " << DataTypeToString() << " M =" << M << " N =" << N << " K =" << K - << " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C - << " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " - << std::endl; + std::cout << "Run Flatmm kernel with DataType = " << DataTypeToString() + << " M =" << M << " N =" << N << " K =" << K << " StrideA =" << stride_A + << " StrideB =" << stride_B << " StrideC =" << stride_C << " : " << ave_time + << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl; return ave_time; } -template +template int run_flatmm_example_with_layouts(int argc, char* argv[], const ALayout a_layout = ALayout{}, @@ -147,7 +153,7 @@ int run_flatmm_example_with_layouts(int argc, using BDataType = typename GemmBasicTypeConfig::BDataType; using CDataType = typename GemmBasicTypeConfig::CDataType; using AccDataType = typename GemmBasicTypeConfig::AccDataType; - + ck_tile::index_t M = arg_parser.get_int("m"); ck_tile::index_t N = arg_parser.get_int("n"); ck_tile::index_t K = arg_parser.get_int("k"); @@ -182,7 +188,7 @@ int run_flatmm_example_with_layouts(int argc, c_rslt_host.SetZero(); // do pre-shuffle - std::string mfma = arg_parser.get_str("prec"); + std::string mfma = arg_parser.get_str("prec"); #if defined(USING_MFMA_16x16x32) && defined(ENABLE_FP8) ck_tile::index_t mfma_type = 1; #else @@ -193,18 +199,18 @@ int run_flatmm_example_with_layouts(int argc, b_shuffle_dev_buf.ToDevice(b_shuffle_host.data()); invoke_flatmm( - a_dev_buf, - b_shuffle_dev_buf, - c_dev_buf, - M, - N, - K, - stride_A, - stride_B, - stride_C, - kbatch, - n_warmup, - n_repeat); + a_dev_buf, + b_shuffle_dev_buf, + c_dev_buf, + M, + N, + K, + stride_A, + stride_B, + stride_C, + kbatch, + n_warmup, + n_repeat); c_dev_buf.FromDevice(c_rslt_host.data()); bool pass = true; @@ -219,8 +225,9 @@ int run_flatmm_example_with_layouts(int argc, a_host, b_origin_host, c_ref_host); const float max_accumulated_value = *std::max_element(c_ref_host.mData.begin(), c_ref_host.mData.end()); - const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value); - pass = ck_tile::check_err(c_rslt_host, + const auto rtol_atol = calculate_rtol_atol( + K, kbatch, max_accumulated_value); + pass = ck_tile::check_err(c_rslt_host, c_ref_host, "Error: Incorrect results!", rtol_atol.at(ck_tile::number<0>{}), @@ -277,8 +284,9 @@ int run_flatmm_example_with_layouts(int argc, c_gpu_ref_dev_buf.FromDevice(c_gpu_ref_host.data()); const float max_accumulated_value = *std::max_element(c_gpu_ref_host.mData.begin(), c_gpu_ref_host.mData.end()); - const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value); - pass = ck_tile::check_err(c_rslt_host, + const auto rtol_atol = calculate_rtol_atol( + K, kbatch, max_accumulated_value); + pass = ck_tile::check_err(c_rslt_host, c_gpu_ref_host, "Error: Incorrect results!", rtol_atol.at(ck_tile::number<0>{}), diff --git a/include/ck/utility/amd_buffer_addressing_builtins.hpp b/include/ck/utility/amd_buffer_addressing_builtins.hpp index 19869906dc..296c1d44d7 100644 --- a/include/ck/utility/amd_buffer_addressing_builtins.hpp +++ b/include/ck/utility/amd_buffer_addressing_builtins.hpp @@ -80,7 +80,7 @@ __device__ half2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16"); // buffer atomic-add i32 __device__ int32_t llvm_amdgcn_raw_buffer_atomic_add_i32( @@ -88,7 +88,7 @@ __device__ int32_t llvm_amdgcn_raw_buffer_atomic_add_i32( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32"); // buffer atomic-add fp32 __device__ float llvm_amdgcn_raw_buffer_atomic_add_fp32( @@ -96,15 +96,15 @@ __device__ float llvm_amdgcn_raw_buffer_atomic_add_fp32( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32"); // buffer atomic-add fp32 -__device__ double llvm_amdgcn_raw_buffer_atomic_max_fp64( - double vdata, - int32x4_t rsrc, // dst_wave_buffer_resource - int voffset, // dst_thread_addr_offset - int soffset, // dst_wave_addr_offset - int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64.v4i32"); +__device__ double +llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata, + int32x4_t rsrc, // dst_wave_buffer_resource + int voffset, // dst_thread_addr_offset + int soffset, // dst_wave_addr_offset + int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64"); // memory coherency bit for buffer store/load instruction // check ISA manual for each GFX target @@ -827,7 +827,7 @@ llvm_amdgcn_raw_buffer_load_lds(int32x4_t rsrc, index_t voffset, index_t soffset, index_t offset, - index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds.v4i32"); + index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds"); #ifndef __HIPCC_RTC__ template 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 0b9956cd01..53a344c7b0 100644 --- a/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp +++ b/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp @@ -881,95 +881,95 @@ CK_TILE_DEVICE_EXTERN int8_t llvm_amdgcn_raw_buffer_load_i8(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i8"); CK_TILE_DEVICE_EXTERN int8x2_t llvm_amdgcn_raw_buffer_load_i8x2(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i8"); CK_TILE_DEVICE_EXTERN int8x4_t llvm_amdgcn_raw_buffer_load_i8x4(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8"); // buffer load i16 CK_TILE_DEVICE_EXTERN int16_t llvm_amdgcn_raw_buffer_load_i16(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16"); CK_TILE_DEVICE_EXTERN int16x2_t llvm_amdgcn_raw_buffer_load_i16x2(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16"); CK_TILE_DEVICE_EXTERN int16x4_t llvm_amdgcn_raw_buffer_load_i16x4(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i16"); // buffer load i32 CK_TILE_DEVICE_EXTERN int32_t llvm_amdgcn_raw_buffer_load_i32(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i32"); CK_TILE_DEVICE_EXTERN int32x2_t llvm_amdgcn_raw_buffer_load_i32x2(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i32"); CK_TILE_DEVICE_EXTERN int32x4_t llvm_amdgcn_raw_buffer_load_i32x4(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i32"); // buffer load fp16 CK_TILE_DEVICE_EXTERN _Float16 llvm_amdgcn_raw_buffer_load_fp16(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f16"); -CK_TILE_DEVICE_EXTERN fp16x2_t llvm_amdgcn_raw_buffer_load_fp16x2( - int32x4_t srsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f16.v4i32"); +CK_TILE_DEVICE_EXTERN fp16x2_t +llvm_amdgcn_raw_buffer_load_fp16x2(int32x4_t srsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f16"); -CK_TILE_DEVICE_EXTERN fp16x4_t llvm_amdgcn_raw_buffer_load_fp16x4( - int32x4_t srsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f16.v4i32"); +CK_TILE_DEVICE_EXTERN fp16x4_t +llvm_amdgcn_raw_buffer_load_fp16x4(int32x4_t srsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f16"); // buffer load fp32 CK_TILE_DEVICE_EXTERN float llvm_amdgcn_raw_buffer_load_fp32(int32x4_t srsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f32"); -CK_TILE_DEVICE_EXTERN fp32x2_t llvm_amdgcn_raw_buffer_load_fp32x2( - int32x4_t srsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f32.v4i32"); +CK_TILE_DEVICE_EXTERN fp32x2_t +llvm_amdgcn_raw_buffer_load_fp32x2(int32x4_t srsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f32"); -CK_TILE_DEVICE_EXTERN fp32x4_t llvm_amdgcn_raw_buffer_load_fp32x4( - int32x4_t srsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f32.v4i32"); +CK_TILE_DEVICE_EXTERN fp32x4_t +llvm_amdgcn_raw_buffer_load_fp32x4(int32x4_t srsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f32"); // buffer store i8 CK_TILE_DEVICE_EXTERN void @@ -977,21 +977,21 @@ llvm_amdgcn_raw_buffer_store_i8(int8_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i8"); CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i8x2(int8x2_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i8"); CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i8x4(int8x4_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i8.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i8"); // buffer store i16 CK_TILE_DEVICE_EXTERN void @@ -999,21 +999,21 @@ llvm_amdgcn_raw_buffer_store_i16(int16_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i16x2( - int16x2_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_i16x2(int16x2_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i16x4( - int16x4_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_i16x4(int16x4_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16"); // buffer store i32 CK_TILE_DEVICE_EXTERN void @@ -1021,7 +1021,7 @@ llvm_amdgcn_raw_buffer_store_i32(int32_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i32"); // buffer store ui16 CK_TILE_DEVICE_EXTERN void @@ -1029,35 +1029,35 @@ llvm_amdgcn_raw_buffer_store_ui16(uint16_t vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_ui16x2( - uint16x2_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_ui16x2(uint16x2_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_ui16x4( - uint16x4_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_ui16x4(uint16x4_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i32x2( - int32x2_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i32.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_i32x2(int32x2_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i32"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_i32x4( - int32x4_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i32.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_i32x4(int32x4_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i32"); // buffer store fp16 CK_TILE_DEVICE_EXTERN void @@ -1065,21 +1065,21 @@ llvm_amdgcn_raw_buffer_store_fp16(_Float16 vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_fp16x2( - fp16x2_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_fp16x2(fp16x2_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f16"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_fp16x4( - fp16x4_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f16.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_fp16x4(fp16x4_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f16"); // buffer store fp32 CK_TILE_DEVICE_EXTERN void @@ -1087,21 +1087,21 @@ llvm_amdgcn_raw_buffer_store_fp32(float vdata, int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f32"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_fp32x2( - fp32x2_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f32.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_fp32x2(fp32x2_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f32"); -CK_TILE_DEVICE_EXTERN void llvm_amdgcn_raw_buffer_store_fp32x4( - fp32x4_t vdata, - int32x4_t rsrc, - index_t voffset, - index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f32.v4i32"); +CK_TILE_DEVICE_EXTERN void +llvm_amdgcn_raw_buffer_store_fp32x4(fp32x4_t vdata, + int32x4_t rsrc, + index_t voffset, + index_t soffset, + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f32"); // buffer atomic-add fp16 CK_TILE_DEVICE_EXTERN fp16x2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2( @@ -1109,7 +1109,7 @@ CK_TILE_DEVICE_EXTERN fp16x2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16"); // buffer atomic-add i32 CK_TILE_DEVICE_EXTERN int32_t llvm_amdgcn_raw_buffer_atomic_add_i32( @@ -1117,7 +1117,7 @@ CK_TILE_DEVICE_EXTERN int32_t llvm_amdgcn_raw_buffer_atomic_add_i32( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32"); // buffer atomic-add fp32 CK_TILE_DEVICE_EXTERN float llvm_amdgcn_raw_buffer_atomic_add_fp32( @@ -1125,15 +1125,15 @@ CK_TILE_DEVICE_EXTERN float llvm_amdgcn_raw_buffer_atomic_add_fp32( int32x4_t rsrc, index_t voffset, index_t soffset, - index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32.v4i32"); + index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32"); // buffer atomic-max fp64 -CK_TILE_DEVICE_EXTERN double llvm_amdgcn_raw_buffer_atomic_max_fp64( - double vdata, - int32x4_t rsrc, // dst_wave_buffer_resource - int voffset, // dst_thread_addr_offset - int soffset, // dst_wave_addr_offset - int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64.v4i32"); +CK_TILE_DEVICE_EXTERN double +llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata, + int32x4_t rsrc, // dst_wave_buffer_resource + int voffset, // dst_thread_addr_offset + int soffset, // dst_wave_addr_offset + int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64"); // Direct loads from global to LDS. CK_TILE_DEVICE_EXTERN void @@ -1143,7 +1143,7 @@ llvm_amdgcn_raw_buffer_load_lds(int32x4_t rsrc, index_t voffset, index_t soffset, index_t offset, - index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds.v4i32"); + index_t aux) __asm("llvm.amdgcn.raw.buffer.load.lds"); template CK_TILE_DEVICE void async_buffer_load_dword_v(void* smem, From ece38b9d7a43f9fd54da164084121932e86dd873 Mon Sep 17 00:00:00 2001 From: Zzz9990 Date: Mon, 26 May 2025 19:17:18 +0800 Subject: [PATCH 2/6] [VLLM V1] Add chunked prefill for FA to pass seq with small seqlen_q (#2221) * fix splitkv compiler issue since lse is used to select kernel instances * bypass seqlen == 1 * add chunked prefill into mha varlen This reverts commit aa9847e42d258e48ec40e5d6469b5df103701e16. * skip compile when receipt 2-4 and add comments * fix --------- Co-authored-by: fsx950223 --- .../ck_tile/01_fmha/codegen/ops/fmha_fwd.py | 62 +++++++++++-------- example/ck_tile/01_fmha/fmha_fwd.hpp | 7 ++- include/ck_tile/core.hpp | 2 +- include/ck_tile/host.hpp | 1 + .../ops/fmha/kernel/fmha_fwd_kernel.hpp | 24 ++++++- .../pipeline/block_fmha_pipeline_problem.hpp | 1 + .../ops/fmha/pipeline/tile_fmha_traits.hpp | 4 +- 7 files changed, 72 insertions(+), 29 deletions(-) diff --git a/example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py b/example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py index 2f1287c87a..e70b896703 100644 --- a/example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py +++ b/example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py @@ -58,7 +58,8 @@ using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad}, {F_lse}, {F_dropout}, {F_squant}, - {F_occupancy}>; + {F_occupancy}, + {F_skip}>; using fmha_variant_{F_idx} = ck_tile::ComposedAttention<{F_logits} * ck_tile::LOGITS_SOFT_CAP, CK_TILE_FMHA_FWD_FAST_EXP2>; @@ -94,7 +95,7 @@ using fmha_kernel_{F_idx} = ck_tile::FmhaFwdKernel; using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, - {F_pipeline_enum}, {F_logits}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>; + {F_pipeline_enum}, {F_logits}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_skip}>; #include @@ -129,9 +130,9 @@ FMHA_FWD_API_PER_HDIM_CASE=""" {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v < }} """ -FMHA_FWD_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && (t.has_logits_soft_cap == {F_logits}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) && +FMHA_FWD_API_INNER_DISPATCH=""" {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && (t.has_logits_soft_cap == {F_logits}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse}) && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) && (t.skip_min_seqlen_q == {F_skip}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{ - using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_logits}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>; + using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_logits}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_skip}>; return fmha_fwd_(s, a); }} """ @@ -160,11 +161,12 @@ class FmhaFwdApiTrait: skpad : str dpad : str dvpad : str + skip : str @property def name(self) -> str: return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\ - f'{self.vlayout}-{self.logits}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}' + f'{self.vlayout}-{self.logits}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}-{self.skip}' @property def scheck(self) -> str: @@ -227,6 +229,7 @@ class FmhaFwdPipeline: F_dropout : str # F_squant : str # F_mask : str # value from MASK_MAP + F_skip : str # true/false @property def name(self) -> str: @@ -262,8 +265,12 @@ class FmhaFwdPipeline: if self.F_dropout == 't' : n += '_dropout' else: n += '_ndropout' + if self.F_skip == 't' : n += '_skip' + else: n += '_nskip' + if self.F_squant == 't' : n += '_squant' else: n += '_nsquant' + return n class FmhaFwdApiPool: @@ -293,7 +300,7 @@ class FmhaFwdApiPool: inners = inners + FMHA_FWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout], F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_logits=BOOL_MAP[trait.logits], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias], - F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] , + F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout], F_skip=BOOL_MAP[trait.skip], F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max, @@ -381,6 +388,7 @@ class FmhaFwdKernel: F_lse = BOOL_MAP[self.F_pipeline.F_lse], F_dropout = BOOL_MAP[self.F_pipeline.F_dropout], F_squant = BOOL_MAP[self.F_pipeline.F_squant], + F_skip = BOOL_MAP[self.F_pipeline.F_skip], F_occupancy = self.F_tile.F_occupancy, F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag], F_mask = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask], @@ -419,7 +427,8 @@ class FmhaFwdKernel: spad=self.F_pipeline.F_spad, skpad=self.F_pipeline.F_skpad, dpad=self.F_pipeline.F_dpad, - dvpad=self.F_pipeline.F_dvpad) + dvpad=self.F_pipeline.F_dvpad, + skip=self.F_pipeline.F_skip) # TODO: design a more practical way to do it # this is current supported tile size per hdim @@ -453,36 +462,36 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl squant = 't' if dtype == 'fp8' else 'f' pipelines = [] if dtype in ['fp16', 'bf16']: - for logits, mask, bias, lse, dropout in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]): + for logits, mask, bias, lse, dropout, skip in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"]): if hdim == 256: # if True: - pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) # the below two is used for hdim vectorize load - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) else: if bias == "bias": # TODO: rocm 6.2 compiler problem if using qr_async for bias case - pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) else: - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask)) - pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) + pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) if receipt == 1 and bias != "bias": - pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim - pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim + pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) # TODO: cover arbitraty hdim + pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', logits, bias, lse, dropout, squant, mask, skip)) # TODO: cover arbitraty hdim elif dtype in ['fp8', 'bf8']: # no need lse/dropout kernels for logits, mask, bias in itertools.product(["t", "f"], get_mask_map(mask_impl).keys(), BIAS_MAP.keys()): - pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, 'f', 'f', squant, mask)) + pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', logits, bias, 'f', 'f', squant, mask, 'f')) elif dtype in ['fp8fp16', 'fp8bf16']: # TODO None @@ -532,6 +541,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl cond &= pipeline.F_vlayout == 'row' cond &= pipeline.F_bias in ['no', 'alibi'] cond &= pipeline.F_squant == 'f' + cond &= pipeline.F_skip == 'f' if not cond: continue # PyTorch integration @@ -540,6 +550,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl cond &= pipeline.F_vlayout == 'row' cond &= pipeline.F_bias in ['no', 'bias'] cond &= pipeline.F_squant == 'f' + cond &= pipeline.F_skip == 'f' if not cond: continue # Aiter(mha_fwd) integration @@ -565,6 +576,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl cond &= pipeline.F_squant == 'f' if not cond: continue + api_pool.register_traits(k.api_trait()) gen.append(k) diff --git a/example/ck_tile/01_fmha/fmha_fwd.hpp b/example/ck_tile/01_fmha/fmha_fwd.hpp index 1838ee5bd9..5ce56d48b5 100644 --- a/example/ck_tile/01_fmha/fmha_fwd.hpp +++ b/example/ck_tile/01_fmha/fmha_fwd.hpp @@ -169,6 +169,7 @@ struct fmha_fwd_args ck_tile::index_t window_size_left; ck_tile::index_t window_size_right; ck_tile::index_t mask_type; + ck_tile::index_t min_seqlen_q; float p_drop; bool s_randval; @@ -433,6 +434,7 @@ auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args) args.window_size_left, args.window_size_right, args.mask_type, + args.min_seqlen_q, args.p_drop, args.s_randval, args.drop_seed_offset); @@ -837,7 +839,8 @@ template + bool kPadDv_, + bool kSkipMinSeqlenQ_ = false> struct fmha_fwd_traits_ { static constexpr ck_tile::index_t HDim = HDim_; @@ -861,6 +864,7 @@ struct fmha_fwd_traits_ static constexpr bool kPadSK = kPadSK_; static constexpr bool kPadD = kPadD_; static constexpr bool kPadDv = kPadDv_; + static constexpr bool kSkipMinSeqlenQ = kSkipMinSeqlenQ_; }; template @@ -995,6 +999,7 @@ struct fmha_fwd_traits bool has_lse; bool has_dropout; bool do_fp8_static_quant; + bool skip_min_seqlen_q = false; // TODO: padding check is inside this api }; float fmha_fwd(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&); diff --git a/include/ck_tile/core.hpp b/include/ck_tile/core.hpp index 27af59c192..be84842347 100644 --- a/include/ck_tile/core.hpp +++ b/include/ck_tile/core.hpp @@ -55,8 +55,8 @@ #include "ck_tile/core/tensor/tile_distribution_encoding.hpp" #include "ck_tile/core/tensor/tile_elementwise.hpp" #include "ck_tile/core/tensor/tile_scatter_gather.hpp" -#include "ck_tile/core/tensor/tile_window_base.hpp" #include "ck_tile/core/tensor/tile_window.hpp" +#include "ck_tile/core/tensor/tile_window_base.hpp" #include "ck_tile/core/tensor/tile_window_linear.hpp" #include "ck_tile/core/tensor/tile_window_utils.hpp" #include "ck_tile/core/tensor/transpose_tile.hpp" diff --git a/include/ck_tile/host.hpp b/include/ck_tile/host.hpp index 5a5e01460f..24feaf7c62 100644 --- a/include/ck_tile/host.hpp +++ b/include/ck_tile/host.hpp @@ -35,4 +35,5 @@ #include "ck_tile/host/reference/reference_softmax.hpp" #include "ck_tile/host/reference/reference_topk.hpp" #include "ck_tile/host/stream_config.hpp" +#include "ck_tile/host/stream_utils.hpp" #include "ck_tile/host/timer.hpp" diff --git a/include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp b/include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp index a4b3765455..bedf20626f 100644 --- a/include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp +++ b/include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp @@ -53,6 +53,8 @@ struct FmhaFwdKernel static constexpr bool kStoreLSE = FmhaPipeline::kStoreLSE; static constexpr bool kHasDropout = FmhaPipeline::kHasDropout; static constexpr bool kDoFp8StaticQuant = FmhaPipeline::Problem::kDoFp8StaticQuant; + static constexpr bool kSkipMinSeqlenQ = FmhaPipeline::Problem::kSkipMinSeqlenQ; + using AttentionVariant = ck_tile::remove_cvref_t; using FmhaMask = ck_tile::remove_cvref_t; static constexpr bool kHasMask = FmhaMask::IsMasking; @@ -257,6 +259,11 @@ struct FmhaFwdKernel ck_tile::index_t batch_stride_randval = 0; }; + struct FmhaFwdSkipMinSeqlenQKargs + { + ck_tile::index_t min_seqlen_q = 0; + }; + struct FmhaFwdBatchModeKargs : FmhaFwdCommonKargs, std::conditional_t>, std::conditional_t>, std::conditional_t>, - std::conditional_t> + std::conditional_t>, + std::conditional_t> { const int32_t* seqstart_q_ptr; const int32_t* seqstart_k_ptr; @@ -664,6 +672,7 @@ struct FmhaFwdKernel ck_tile::index_t window_size_left, ck_tile::index_t window_size_right, ck_tile::index_t mask_type, + ck_tile::index_t min_seqlen_q, float p_drop, bool s_randval, std::variant, std::pair> @@ -698,6 +707,7 @@ struct FmhaFwdKernel {}, // placeholder for fp8_static_quant args {}, // placeholder for dropout {}, // placeholder for logits_soft_cap + {}, // placeholder for min_seqlen_q reinterpret_cast(seqstart_q_ptr), reinterpret_cast(seqstart_k_ptr), reinterpret_cast(seqlen_k_ptr)}; @@ -753,6 +763,10 @@ struct FmhaFwdKernel { kargs.init_logits_soft_cap(logits_soft_cap); } + if constexpr(kSkipMinSeqlenQ) + { + kargs.min_seqlen_q = min_seqlen_q; + } return kargs; } @@ -1053,6 +1067,14 @@ struct FmhaFwdKernel const auto adjusted_seqstart_q_ptr = kargs.seqstart_q_ptr + i_batch; kargs.seqlen_q = adjusted_seqstart_q_ptr[1] - adjusted_seqstart_q_ptr[0]; + if constexpr(kSkipMinSeqlenQ) + { + if(kargs.seqlen_q <= kargs.min_seqlen_q) + { + return; + } + } + // # of required blocks is different in each groups, terminate unnecessary blocks // earlier if(kargs.seqlen_q <= i_m0) diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp index f35c00c268..21cc4950eb 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_problem.hpp @@ -53,6 +53,7 @@ struct BlockFmhaPipelineProblem static constexpr bool kPadHeadDimQ = Traits::kPadHeadDimQ; static constexpr bool kPadHeadDimV = Traits::kPadHeadDimV; static constexpr bool kHasLogitsSoftCap = Traits::kHasLogitsSoftCap; + static constexpr bool kSkipMinSeqlenQ = Traits::kSkipMinSeqlenQ; static constexpr auto BiasEnum = Traits::BiasEnum; static constexpr bool kStoreLSE = Traits::kStoreLSE; static constexpr bool kHasDropout = Traits::kHasDropout; diff --git a/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp b/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp index 4530b58d85..442619a3dc 100644 --- a/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp +++ b/include/ck_tile/ops/fmha/pipeline/tile_fmha_traits.hpp @@ -19,7 +19,8 @@ template + index_t kBlockPerCu_ = -1, /* overwrite occupancy if not -1 */ + bool kSkipMinSeqlenQ_ = false /* skip min seqlen q while chunked prefill */> struct TileFmhaTraits { static constexpr bool kPadSeqLenQ = kPadSeqLenQ_; @@ -33,6 +34,7 @@ struct TileFmhaTraits static constexpr bool kHasDropout = kHasDropout_; static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_; static constexpr index_t kBlockPerCu = kBlockPerCu_; + static constexpr bool kSkipMinSeqlenQ = kSkipMinSeqlenQ_; }; template Date: Mon, 26 May 2025 16:51:09 +0200 Subject: [PATCH 3/6] Fix grid size calc for bwd wei (#2226) --- ...onv_bwd_weight_multiple_d_xdl_cshuffle.hpp | 20 ++++++++++++------- ..._grouped_conv_bwd_weight_wmma_cshuffle.hpp | 6 +++--- .../transform_conv_bwd_weight_to_gemm.hpp | 14 ++++++------- .../transform_conv_bwd_weight_to_gemm_v2.hpp | 12 +++++------ 4 files changed, 29 insertions(+), 23 deletions(-) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp index 57c4b1a5cf..33b6d7c585 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -393,8 +393,10 @@ struct DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle { const index_t GemmM = K; const index_t GemmN = C * X; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = + GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = + GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; return transform_tensor_descriptor( wei_grid_desc, @@ -432,8 +434,10 @@ struct DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle { const index_t GemmM = K; const index_t GemmN = C * X * Y; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = + GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = + GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; return transform_tensor_descriptor( wei_grid_desc, @@ -472,8 +476,10 @@ struct DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle { const index_t GemmM = K; const index_t GemmN = C * X * Y * Z; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = + GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = + GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; return transform_tensor_descriptor( wei_grid_desc, diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp index 0831b754c8..e9e02eae81 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -208,8 +208,8 @@ struct DeviceGroupedConvBwdWeight_Wmma_CShuffle const index_t GemmM = K; const index_t GemmN = C * Z * X * Y; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmK0 = math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock) * K0PerBlock; diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp index c11bf845d0..bd3ab10802 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp +++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm.hpp @@ -1,6 +1,6 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -166,8 +166,8 @@ struct TransformConvBwdWeightToGemm const index_t GemmM = K; const index_t GemmN = C * X; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = @@ -365,8 +365,8 @@ struct TransformConvBwdWeightToGemm const index_t GemmM = K; const index_t GemmN = C * X * Y; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = @@ -558,8 +558,8 @@ struct TransformConvBwdWeightToGemm const index_t GemmM = K; const index_t GemmN = C * Z * X * Y; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp index f34e0e59b3..b72ddb8243 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp +++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp @@ -346,8 +346,8 @@ struct TransformConvBwdWeightToGemmV2 const index_t GemmM = K * NumGroupsToMerge; const index_t GemmN = C * X * NumGroupsToMerge; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = @@ -534,8 +534,8 @@ struct TransformConvBwdWeightToGemmV2 const index_t GemmM = K * NumGroupsToMerge; const index_t GemmN = C * X * Y * NumGroupsToMerge; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = @@ -737,8 +737,8 @@ struct TransformConvBwdWeightToGemmV2 const index_t GemmM = K * NumGroupsToMerge; const index_t GemmN = C * Z * X * Y * NumGroupsToMerge; - const auto PadGemmM = MPerBlock - GemmM % MPerBlock; - const auto PadGemmN = NPerBlock - GemmN % NPerBlock; + const auto PadGemmM = GemmM % MPerBlock == 0 ? 0 : MPerBlock - GemmM % MPerBlock; + const auto PadGemmN = GemmN % NPerBlock == 0 ? 0 : NPerBlock - GemmN % NPerBlock; const index_t GemmKBatch = batch_k; const index_t GemmK0 = From 4583aeffad950e7709ba9d0a26fcc9189789f5e7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= Date: Mon, 26 May 2025 22:46:18 +0200 Subject: [PATCH 4/6] Remove not needed bwd wei merged groups instances (#2218) * Grouped conv bwd wei add two stage instances for larger filter and Merge Groups * Fix * fix * Revert "Restore oddc instances (#2201)" This reverts commit 6342f6b5e8bbb9f2b4cefa33d2a863a8bb35329b. * fix --------- Co-authored-by: Rostyslav Geyyer <46627076+geyyer@users.noreply.github.com> --- ...conv_bwd_weight_two_stage_xdl_instance.hpp | 39 +----- .../gpu/grouped_convolution_forward.hpp | 8 -- .../gpu/grouped_convolution_forward_wmma.inc | 111 ------------------ .../gpu/grouped_conv2d_fwd/CMakeLists.txt | 4 - ...ma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp | 40 ------- ...mma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp | 40 ------- ...ma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp | 40 ------- ...mma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp | 40 ------- ...hwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp | 9 -- ...l_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp | 8 -- ...c_gkyxc_nhwgk_bf16_comp_part2_instance.cpp | 9 -- ...nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp | 9 -- ...dl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp | 8 -- ...gc_gkyxc_nhwgk_f16_comp_part2_instance.cpp | 9 -- ...dl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp | 10 +- ...l_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp | 28 +---- ...wd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp | 10 +- ...fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp | 10 +- ...fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp | 10 +- ...wd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 10 +- ...fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp | 10 +- ...fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp | 10 +- ...wd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp | 10 +- ...gc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_int8_mem_inter_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_int8_mem_intra_instance.cpp | 11 +- .../gpu/grouped_conv3d_fwd/CMakeLists.txt | 4 - ...gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp | 41 ------- ..._gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp | 41 ------- ...ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp | 41 ------- ..._ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp | 41 ------- 36 files changed, 20 insertions(+), 718 deletions(-) delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp index fbcda3ca57..87b3d4d8cd 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp @@ -158,11 +158,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -251,23 +247,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8> // clang-format on >; @@ -335,24 +317,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, - - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> - + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp index 545826650c..cf5dbaa323 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp @@ -613,7 +613,6 @@ struct DeviceOperationInstanceFactory>>& instances); -void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances( std::vector>>& instances); -void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances( - std::vector>>& instances); #endif #ifdef CK_ENABLE_FP16 @@ -291,20 +236,6 @@ void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instances( PassThrough, PassThrough>>>& instances); -void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances( std::vector>>& instances); - -void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances( - std::vector>>& instances); #endif } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt index 22e9d726b0..eba6fd789e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt @@ -93,8 +93,6 @@ add_instance_library(device_grouped_conv2d_fwd_instance wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp - wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp - wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp ## NHWGC, GKYXC, NHWGK wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp @@ -102,6 +100,4 @@ add_instance_library(device_grouped_conv2d_fwd_instance wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp - wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp - wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp ) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp deleted file mode 100644 index a8f723dfec..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp +++ /dev/null @@ -1,40 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k] -void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_f16_instances<2, - GNHWC, - GKYXC, - Empty_Tuple, - GNHWK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp deleted file mode 100644 index 784a118897..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp +++ /dev/null @@ -1,40 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k] -void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_i8_instances<2, - GNHWC, - GKYXC, - Empty_Tuple, - GNHWK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp deleted file mode 100644 index 8c621543a9..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp +++ /dev/null @@ -1,40 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] -void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_f16_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp deleted file mode 100644 index 5cb313b3ca..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp +++ /dev/null @@ -1,40 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] -void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_i8_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp index c078f8ed04..f5df7278d0 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp @@ -52,15 +52,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp index a67b11f1cf..db048679bd 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp @@ -49,14 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_bf16_comp_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp index 5c0391a25f..ee9507a80a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp @@ -52,15 +52,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instanc Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp index 726276c461..132d3c8411 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp @@ -52,15 +52,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_f16_comp_instances_2x<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp index 8b7bdec2a8..a7deb969ba 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp @@ -49,14 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f16_comp_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp index c66114b9a3..d2732547fa 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp @@ -52,15 +52,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_f16_comp_instances_part2<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp index 93e07e08fb..8a0caebc9f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" @@ -48,14 +48,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f32_comp_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp index 6acbb7475c..e45df1e107 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" @@ -50,14 +50,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( NHWGK, ConvFwd1x1S1P0>{}); - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_int8_comp_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); - if(ck::get_device_name() != "gfx950") { add_device_operation_instances( @@ -86,15 +78,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_int8_comp_instances_part2<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } if(ck::get_device_name() == "gfx950") @@ -125,15 +108,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances( - instances, - device_grouped_conv_fwd_xdl_int8_comp_instances_2x<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp index 2afbfdc386..078221f89f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_bf16_instances<2, - GNHWC, - GKYXC, - Empty_Tuple, - GNHWK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp index 822ef51e00..3a481dd204 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f16_instances<2, - GNHWC, - GKYXC, - Empty_Tuple, - GNHWK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp index 79a1fb99a8..5add0f8add 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f32_instances<2, - GNHWC, - GKYXC, - Empty_Tuple, - GNHWK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp index e567c0df75..0257c7d315 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_bf16_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp index 3e42184996..2715506fe2 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f16_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp index c035d4c3da..8d3e4d91b1 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f32_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp index 5c425effd8..465fa927a5 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,14 +46,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_int8_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp index e8a763c527..87423801cb 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance NHWGK, ConvFwd1x1S1P0, Interwave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_bf16_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp index 3ae3fb5186..ebb213461a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance NHWGK, ConvFwd1x1S1P0, Intrawave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_bf16_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp index cb7e912936..c2c8a099b2 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instances NHWGK, ConvFwd1x1S1P0, Interwave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f16_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp index d787f4b048..11cb853f0d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances NHWGK, ConvFwd1x1S1P0, Intrawave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f16_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp index 5644289790..1992d7f7c1 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instances NHWGK, ConvFwd1x1S1P0, Interwave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f32_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp index 5b12dad5a3..2b8fd3d9db 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances NHWGK, ConvFwd1x1S1P0, Intrawave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_f32_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp index f667481fa4..5579ec62cc 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance NHWGK, ConvFwd1x1S1P0, Interwave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_int8_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp index 2ff2c7f51f..77f3df2c11 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp @@ -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. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,15 +49,6 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance NHWGK, ConvFwd1x1S1P0, Intrawave>{}); - - add_device_operation_instances(instances, - device_grouped_conv_fwd_xdl_int8_mem_instances<2, - NHWGC, - GKYXC, - Empty_Tuple, - NHWGK, - ConvFwdOddC, - Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt index f8efa5a7c1..f55bdd45c9 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt @@ -66,10 +66,6 @@ set(GROUPED_CONV3D_FWD wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp - wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp - wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp - wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp - wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp ) if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp deleted file mode 100644 index fa378af1ee..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, -// wo, k] -void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_f16_instances<3, - GNDHWC, - GKZYXC, - Empty_Tuple, - GNDHWK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp deleted file mode 100644 index d41416fd4a..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, -// wo, k] -void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_i8_instances<3, - GNDHWC, - GKZYXC, - Empty_Tuple, - GNDHWK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp deleted file mode 100644 index 8a7bc26178..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = out[n, do, ho, wo, -// g, k] -void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_f16_instances<3, - NDHWGC, - GKZYXC, - Empty_Tuple, - NDHWGK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp deleted file mode 100644 index 7649f86971..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { -// Compilation parameters for in[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = out[n, do, ho, wo, -// g, k] -void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances( - std::vector>>& instances) -{ - add_device_operation_instances(instances, - device_grouped_conv_fwd_wmma_i8_instances<3, - NDHWGC, - GKZYXC, - Empty_Tuple, - NDHWGK, - Empty_Tuple, - PassThrough, - ConvFwdOddC>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck From b1ed92b13117d839b7fe9a68b47c442d5649265a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?= Date: Mon, 26 May 2025 23:26:04 +0200 Subject: [PATCH 5/6] Revert "Remove not needed bwd wei merged groups instances (#2218)" (#2235) This reverts commit 4583aeffad950e7709ba9d0a26fcc9189789f5e7. --- ...conv_bwd_weight_two_stage_xdl_instance.hpp | 39 +++++- .../gpu/grouped_convolution_forward.hpp | 8 ++ .../gpu/grouped_convolution_forward_wmma.inc | 111 ++++++++++++++++++ .../gpu/grouped_conv2d_fwd/CMakeLists.txt | 4 + ...ma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp | 40 +++++++ ...mma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp | 40 +++++++ ...ma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp | 40 +++++++ ...mma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp | 40 +++++++ ...hwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp | 9 ++ ...l_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp | 8 ++ ...c_gkyxc_nhwgk_bf16_comp_part2_instance.cpp | 9 ++ ...nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp | 9 ++ ...dl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp | 8 ++ ...gc_gkyxc_nhwgk_f16_comp_part2_instance.cpp | 9 ++ ...dl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp | 10 +- ...l_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp | 28 ++++- ...wd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp | 10 +- ...fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp | 10 +- ...fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp | 10 +- ...wd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 10 +- ...fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp | 10 +- ...fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp | 10 +- ...wd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp | 10 +- ...gc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp | 11 +- ...wgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_int8_mem_inter_instance.cpp | 11 +- ...gc_gkyxc_nhwgk_int8_mem_intra_instance.cpp | 11 +- .../gpu/grouped_conv3d_fwd/CMakeLists.txt | 4 + ...gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp | 41 +++++++ ..._gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp | 41 +++++++ ...ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp | 41 +++++++ ..._ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp | 41 +++++++ 36 files changed, 718 insertions(+), 20 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp index 87b3d4d8cd..fbcda3ca57 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp @@ -158,7 +158,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -247,9 +251,23 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> // clang-format on >; @@ -317,9 +335,24 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, + + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp index cf5dbaa323..545826650c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp @@ -613,6 +613,7 @@ struct DeviceOperationInstanceFactory>>& instances); +void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances( std::vector>>& instances); +void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances( std::vector>>& instances); +void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances( std::vector>>& instances); +void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances( + std::vector>>& instances); #endif #ifdef CK_ENABLE_FP16 @@ -236,6 +291,20 @@ void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instances( PassThrough, PassThrough>>>& instances); +void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances( std::vector>>& instances); +void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances( std::vector>>& instances); +void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances( + std::vector>>& instances); + void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances( std::vector>>& instances); + +void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances( + std::vector>>& instances); #endif } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt index eba6fd789e..22e9d726b0 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt @@ -93,6 +93,8 @@ add_instance_library(device_grouped_conv2d_fwd_instance wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp + wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp + wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp ## NHWGC, GKYXC, NHWGK wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp @@ -100,4 +102,6 @@ add_instance_library(device_grouped_conv2d_fwd_instance wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp + wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp + wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp ) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp new file mode 100644 index 0000000000..a8f723dfec --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp @@ -0,0 +1,40 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k] +void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_f16_instances<2, + GNHWC, + GKYXC, + Empty_Tuple, + GNHWK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp new file mode 100644 index 0000000000..784a118897 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp @@ -0,0 +1,40 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k] +void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_i8_instances<2, + GNHWC, + GKYXC, + Empty_Tuple, + GNHWK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp new file mode 100644 index 0000000000..8c621543a9 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp @@ -0,0 +1,40 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_f16_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp new file mode 100644 index 0000000000..5cb313b3ca --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp @@ -0,0 +1,40 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k] +void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_i8_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp index f5df7278d0..c078f8ed04 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instance.cpp @@ -52,6 +52,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_2x_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_2x<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp index db048679bd..a67b11f1cf 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instance.cpp @@ -49,6 +49,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp index ee9507a80a..5c0391a25f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instance.cpp @@ -52,6 +52,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_part2_instanc Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_bf16_comp_instances_part2<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp index 132d3c8411..726276c461 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instance.cpp @@ -52,6 +52,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_2x_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_f16_comp_instances_2x<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp index a7deb969ba..8b7bdec2a8 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instance.cpp @@ -49,6 +49,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f16_comp_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp index d2732547fa..c66114b9a3 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance.cpp @@ -52,6 +52,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_part2_instance Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_f16_comp_instances_part2<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp index 8a0caebc9f..93e07e08fb 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" @@ -48,6 +48,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f32_comp_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp index e45df1e107..6acbb7475c 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/comp/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp" @@ -50,6 +50,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( NHWGK, ConvFwd1x1S1P0>{}); + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_int8_comp_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); + if(ck::get_device_name() != "gfx950") { add_device_operation_instances( @@ -78,6 +86,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_int8_comp_instances_part2<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } if(ck::get_device_name() == "gfx950") @@ -108,6 +125,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_comp_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances( + instances, + device_grouped_conv_fwd_xdl_int8_comp_instances_2x<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp index 078221f89f..2afbfdc386 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<2, + GNHWC, + GKYXC, + Empty_Tuple, + GNHWK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp index 3a481dd204..822ef51e00 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f16_instances<2, + GNHWC, + GKYXC, + Empty_Tuple, + GNHWK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp index 5add0f8add..79a1fb99a8 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances( Empty_Tuple, GNHWK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f32_instances<2, + GNHWC, + GKYXC, + Empty_Tuple, + GNHWK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp index 0257c7d315..e567c0df75 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp index 2715506fe2..3e42184996 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f16_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp index 8d3e4d91b1..c035d4c3da 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f32_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp index 465fa927a5..5c425effd8 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp" @@ -46,6 +46,14 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_instances( Empty_Tuple, NHWGK, ConvFwd1x1S1P0>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_int8_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp index 87423801cb..e8a763c527 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instance NHWGK, ConvFwd1x1S1P0, Interwave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp index ebb213461a..3ae3fb5186 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance NHWGK, ConvFwd1x1S1P0, Intrawave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_bf16_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp index c2c8a099b2..cb7e912936 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_inter_instances NHWGK, ConvFwd1x1S1P0, Interwave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f16_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp index 11cb853f0d..d787f4b048 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances NHWGK, ConvFwd1x1S1P0, Intrawave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f16_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp index 1992d7f7c1..5644289790 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_inter_instances NHWGK, ConvFwd1x1S1P0, Interwave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f32_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp index 2b8fd3d9db..5b12dad5a3 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances NHWGK, ConvFwd1x1S1P0, Intrawave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_f32_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp index 5579ec62cc..f667481fa4 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_inter_instance NHWGK, ConvFwd1x1S1P0, Interwave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_int8_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Interwave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp index 77f3df2c11..2ff2c7f51f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_mem_instance.hpp" @@ -49,6 +49,15 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_int8_mem_intra_instance NHWGK, ConvFwd1x1S1P0, Intrawave>{}); + + add_device_operation_instances(instances, + device_grouped_conv_fwd_xdl_int8_mem_instances<2, + NHWGC, + GKYXC, + Empty_Tuple, + NHWGK, + ConvFwdOddC, + Intrawave>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt index f55bdd45c9..f8efa5a7c1 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt @@ -66,6 +66,10 @@ set(GROUPED_CONV3D_FWD wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp + wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp + wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp + wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp + wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp ) if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp new file mode 100644 index 0000000000..fa378af1ee --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, +// wo, k] +void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_f16_instances<3, + GNDHWC, + GKZYXC, + Empty_Tuple, + GNDHWK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp new file mode 100644 index 0000000000..d41416fd4a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = out[g, n, do, ho, +// wo, k] +void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_i8_instances<3, + GNDHWC, + GKZYXC, + Empty_Tuple, + GNDHWK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp new file mode 100644 index 0000000000..8a7bc26178 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = out[n, do, ho, wo, +// g, k] +void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_f16_instances<3, + NDHWGC, + GKZYXC, + Empty_Tuple, + NDHWGK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp new file mode 100644 index 0000000000..7649f86971 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp @@ -0,0 +1,41 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { +// Compilation parameters for in[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = out[n, do, ho, wo, +// g, k] +void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv_fwd_wmma_i8_instances<3, + NDHWGC, + GKZYXC, + Empty_Tuple, + NDHWGK, + Empty_Tuple, + PassThrough, + ConvFwdOddC>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck From 5727af98d1a70f9c4cc8aa7ac0cf6941690eeafb Mon Sep 17 00:00:00 2001 From: Yi DING Date: Tue, 27 May 2025 09:51:20 +0800 Subject: [PATCH 6/6] Add operator/instance filters to ckProfiler (#2233) --- profiler/src/CMakeLists.txt | 275 ++++++++++++++++++++---------------- 1 file changed, 156 insertions(+), 119 deletions(-) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 17c8c277eb..65dd704610 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -1,6 +1,16 @@ # ckProfiler -set(PROFILER_SOURCES - profiler.cpp +set(CK_PROFILER_OP_FILTER "" CACHE STRING "Filter for the operators to be profiled. Default is to include all") +set(CK_PROFILER_INSTANCE_FILTER "" CACHE STRING "Filter for the kernels instances to be profiled. Default is to be the same as the operator filter") +if (CK_PROFILER_OP_FILTER STREQUAL "") + set(CK_PROFILER_OP_FILTER ".+") +endif() +if (CK_PROFILER_INSTANCE_FILTER STREQUAL "") + set(CK_PROFILER_INSTANCE_FILTER ${CK_PROFILER_OP_FILTER}) +endif() +message(STATUS "CK_PROFILER_OP_FILTER: ${CK_PROFILER_OP_FILTER}") +message(STATUS "CK_PROFILER_INSTANCE_FILTER: ${CK_PROFILER_INSTANCE_FILTER}") + +set(PROFILER_OPS profile_gemm.cpp profile_reduce.cpp profile_groupnorm_bwd_data.cpp @@ -26,161 +36,188 @@ set(PROFILER_SOURCES if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp) - list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp) + list(APPEND PROFILER_OPS profile_contraction_bilinear.cpp) + list(APPEND PROFILER_OPS profile_contraction_scale.cpp) endif() if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp) + list(APPEND PROFILER_OPS profile_gemm_reduce.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm_gemm.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm_add_relu_gemm_add.cpp) + list(APPEND PROFILER_OPS profile_gemm_add.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_grouped_gemm.cpp) + list(APPEND PROFILER_OPS profile_gemm_streamk.cpp) + list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_relu.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_silu.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_relu_add_layernorm.cpp) + list(APPEND PROFILER_OPS profile_grouped_gemm_fixed_nk.cpp) + list(APPEND PROFILER_OPS profile_grouped_gemm_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_grouped_gemm_tile_loop.cpp) + list(APPEND PROFILER_OPS profile_grouped_gemm_multiply_tile_loop.cpp) endif() - list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp) - if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply_wp.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp) + list(APPEND PROFILER_OPS profile_gemm_multiply_add.cpp) + if(SUPPORTED_GPU_TARGETS MATCHES "gfx9[45]") + list(APPEND PROFILER_OPS profile_gemm_multiply_multiply.cpp) + list(APPEND PROFILER_OPS profile_gemm_multiply_multiply_wp.cpp) + list(APPEND PROFILER_OPS profile_gemm_ab_scale.cpp) endif() - list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_add_multiply.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp) - list(APPEND PROFILER_SOURCES profile_batched_gemm_b_scale.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp) - list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu_add.cpp) - list(APPEND PROFILER_SOURCES profile_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd_outelementop.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm_reduce.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_multiply.cpp) + list(APPEND PROFILER_OPS profile_gemm_bias_add_reduce.cpp) + list(APPEND PROFILER_OPS profile_gemm_splitk.cpp) + list(APPEND PROFILER_OPS profile_gemm_b_scale.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm_b_scale.cpp) + list(APPEND PROFILER_OPS profile_gemm_universal_batched.cpp) + list(APPEND PROFILER_OPS profile_gemm_universal_reduce.cpp) + list(APPEND PROFILER_OPS profile_gemm_universal_streamk.cpp) + list(APPEND PROFILER_OPS profile_conv_fwd_bias_relu.cpp) + list(APPEND PROFILER_OPS profile_conv_fwd_bias_relu_add.cpp) + list(APPEND PROFILER_OPS profile_conv_bwd_data.cpp) + list(APPEND PROFILER_OPS profile_conv_fwd.cpp) + list(APPEND PROFILER_OPS profile_grouped_conv_fwd_outelementop.cpp) endif() if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) + list(APPEND PROFILER_OPS profile_gemm_bilinear.cpp) endif() - list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) + list(APPEND PROFILER_OPS profile_gemm_universal.cpp) + list(APPEND PROFILER_OPS profile_grouped_conv_fwd.cpp) + list(APPEND PROFILER_OPS profile_grouped_conv_bwd_data.cpp) + list(APPEND PROFILER_OPS profile_grouped_conv_bwd_weight.cpp) endif() if(DL_KERNELS) - list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp) - list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_weight.cpp) + list(APPEND PROFILER_OPS profile_batched_gemm_multi_d.cpp) + list(APPEND PROFILER_OPS profile_grouped_conv_bwd_weight.cpp) endif() +set(PROFILER_SOURCES profiler.cpp) +foreach(SOURCE ${PROFILER_OPS}) + string(REGEX REPLACE "profile_(.+)\.cpp" "\\1" OP_NAME ${SOURCE}) + if (OP_NAME STREQUAL "") + message(FATAL_ERROR "Unexpected source file name: ${SOURCE}") + endif() + if("${OP_NAME}" MATCHES "${CK_PROFILER_OP_FILTER}") + list(APPEND PROFILER_SOURCES ${SOURCE}) + endif() +endforeach() +message(STATUS "ckProfiler sources: ${PROFILER_SOURCES}") + set(PROFILER_EXECUTABLE ckProfiler) add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES}) target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors) # flags to compress the library if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) - message("Adding --offload-compress flag for ${PROFILER_EXECUTABLE}") + message(STATUS "Adding --offload-compress flag for ${PROFILER_EXECUTABLE}") target_compile_options(${PROFILER_EXECUTABLE} PRIVATE --offload-compress) endif() -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool2d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool2d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance) -target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance) + +set(DEVICE_INSTANCES "") +list(APPEND DEVICE_INSTANCES device_gemm_instance) +list(APPEND DEVICE_INSTANCES device_normalization_fwd_instance) +list(APPEND DEVICE_INSTANCES device_normalization_bwd_data_instance) +list(APPEND DEVICE_INSTANCES device_normalization_bwd_gamma_beta_instance) +list(APPEND DEVICE_INSTANCES device_softmax_instance) +list(APPEND DEVICE_INSTANCES device_reduce_instance) +list(APPEND DEVICE_INSTANCES device_batchnorm_instance) +list(APPEND DEVICE_INSTANCES device_pool2d_fwd_instance) +list(APPEND DEVICE_INSTANCES device_pool3d_fwd_instance) +list(APPEND DEVICE_INSTANCES device_avg_pool2d_bwd_instance) +list(APPEND DEVICE_INSTANCES device_avg_pool3d_bwd_instance) +list(APPEND DEVICE_INSTANCES device_max_pool_bwd_instance) +list(APPEND DEVICE_INSTANCES device_image_to_column_instance) +list(APPEND DEVICE_INSTANCES device_column_to_image_instance) +list(APPEND DEVICE_INSTANCES device_transpose_instance) +list(APPEND DEVICE_INSTANCES device_permute_scale_instance) if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) + list(APPEND DEVICE_INSTANCES device_contraction_bilinear_instance) + list(APPEND DEVICE_INSTANCES device_contraction_scale_instance) endif() if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_tile_loop_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_gemm_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_add_relu_gemm_add_instance) + list(APPEND DEVICE_INSTANCES device_grouped_gemm_instance) + list(APPEND DEVICE_INSTANCES device_gemm_streamk_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_relu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_silu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_relu_add_layernorm_instance) + list(APPEND DEVICE_INSTANCES device_grouped_gemm_fixed_nk_instance) + list(APPEND DEVICE_INSTANCES device_grouped_gemm_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_grouped_gemm_tile_loop_instance) endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance) - if(SUPPORTED_GPU_TARGETS MATCHES "gfx94") - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_wp_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_reduce_instance) + list(APPEND DEVICE_INSTANCES device_gemm_multiply_add_instance) + if(SUPPORTED_GPU_TARGETS MATCHES "gfx9[45]") + list(APPEND DEVICE_INSTANCES device_gemm_multiply_multiply_instance) + list(APPEND DEVICE_INSTANCES device_gemm_multiply_multiply_wp_instance) + list(APPEND DEVICE_INSTANCES device_gemm_ab_scale_instance) endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_b_scale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convscale_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) + list(APPEND DEVICE_INSTANCES device_gemm_splitk_instance) + list(APPEND DEVICE_INSTANCES device_gemm_b_scale_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_b_scale_instance) + list(APPEND DEVICE_INSTANCES device_gemm_universal_batched_instance) + list(APPEND DEVICE_INSTANCES device_gemm_universal_reduce_instance) + list(APPEND DEVICE_INSTANCES device_gemm_universal_streamk_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_multiply_instance) + list(APPEND DEVICE_INSTANCES device_gemm_reduce_instance) + list(APPEND DEVICE_INSTANCES device_gemm_bias_add_reduce_instance) + list(APPEND DEVICE_INSTANCES device_conv2d_fwd_instance) + list(APPEND DEVICE_INSTANCES device_conv2d_fwd_bias_relu_instance) + list(APPEND DEVICE_INSTANCES device_conv2d_fwd_bias_relu_add_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv1d_fwd_instance) + list(APPEND DEVICE_INSTANCES device_conv1d_bwd_data_instance) + list(APPEND DEVICE_INSTANCES device_conv3d_bwd_data_instance) + list(APPEND DEVICE_INSTANCES device_conv2d_bwd_data_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv1d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv2d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_fwd_convscale_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_fwd_convinvscale_instance) endif() if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) + list(APPEND DEVICE_INSTANCES device_gemm_bilinear_instance) endif() - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_gemm_universal_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_fwd_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv2d_bwd_data_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_data_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv2d_fwd_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_weight_instance) endif() if(DL_KERNELS) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance) - target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_batched_gemm_multi_d_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv1d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv2d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_weight_instance) endif() + +set(PROFILER_LIBS utility getopt::getopt) +foreach(LIB ${DEVICE_INSTANCES}) + string(REGEX REPLACE "device_(.+)_instance" "\\1" INSTANCE_NAME ${LIB}) + if (INSTANCE_NAME STREQUAL "") + message(FATAL_ERROR "Unexpected kernel instance name: ${LIB}") + endif() + if("${INSTANCE_NAME}" MATCHES "${CK_PROFILER_INSTANCE_FILTER}") + list(APPEND PROFILER_LIBS ${LIB}) + endif() +endforeach() +message(STATUS "ckProfiler libs: ${PROFILER_LIBS}") +target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE ${PROFILER_LIBS}) + rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)