From e25c18aeb769ba4e6bbcfdb82e23f16ec45191c3 Mon Sep 17 00:00:00 2001 From: zjing14 Date: Fri, 17 Nov 2023 07:06:24 -0600 Subject: [PATCH] Improve 4k gemm perf (#1047) * improve 4k gemm perf * add f8 instances * format --------- Co-authored-by: Jing Zhang [ROCm/composable_kernel commit: e8cddfdc3bc7fbdec765ee0bfbb391ef7173b455] --- example/01_gemm/gemm_xdl_fp16.cpp | 6 +- include/ck/host_utility/kernel_launch.hpp | 9 +- include/ck/stream_config.hpp | 2 + .../tensor_operation_instance/gpu/gemm.hpp | 16 +- ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp | 36 ++--- ..._shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp | 91 ++++++++--- .../include/profiler/profile_gemm_impl.hpp | 147 +++++++++++------- 7 files changed, 198 insertions(+), 109 deletions(-) diff --git a/example/01_gemm/gemm_xdl_fp16.cpp b/example/01_gemm/gemm_xdl_fp16.cpp index 54fbd9cdd4..ef9e64cc69 100644 --- a/example/01_gemm/gemm_xdl_fp16.cpp +++ b/example/01_gemm/gemm_xdl_fp16.cpp @@ -9,13 +9,13 @@ using ADataType = ck::half_t; using BDataType = ck::half_t; using AccDataType = float; -using CShuffleDataType = float; +using CShuffleDataType = ck::half_t; using CDataType = ck::half_t; using F16 = ck::half_t; using ALayout = Row; -using BLayout = Col; +using BLayout = Row; using CLayout = Row; using AElementOp = PassThrough; @@ -39,7 +39,7 @@ using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffl // ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| // ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| // ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>; + < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 2, S<1, 16, 1, 16>, 8, ck::LoopScheduler::Interwave, ck::PipelineVersion::v1>; // clang-format on using DeviceGemmInstance = DeviceGemmInstance1; diff --git a/include/ck/host_utility/kernel_launch.hpp b/include/ck/host_utility/kernel_launch.hpp index df311620e1..cd00598b52 100644 --- a/include/ck/host_utility/kernel_launch.hpp +++ b/include/ck/host_utility/kernel_launch.hpp @@ -33,10 +33,13 @@ float launch_and_time_kernel(const StreamConfig& stream_config, printf("Warm up 1 time\n"); #endif // warm up - kernel<<>>(args...); - hip_check_error(hipGetLastError()); + for(int i = 0; i < stream_config.cold_niters_; ++i) + { + kernel<<>>(args...); + hip_check_error(hipGetLastError()); + } - const int nrepeat = 10; + const int nrepeat = stream_config.nrepeat_; #if DEBUG_LOG printf("Start running %d times...\n", nrepeat); #endif diff --git a/include/ck/stream_config.hpp b/include/ck/stream_config.hpp index 505a602b24..1681b223f3 100644 --- a/include/ck/stream_config.hpp +++ b/include/ck/stream_config.hpp @@ -11,4 +11,6 @@ struct StreamConfig hipStream_t stream_id_ = nullptr; bool time_kernel_ = false; int log_level_ = 0; + int cold_niters_ = 50; + int nrepeat_ = 200; }; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp index fa42f02449..ea7552d004 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp @@ -377,7 +377,7 @@ struct DeviceOperationInstanceFactory< if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs); #endif @@ -386,7 +386,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs); #endif @@ -395,7 +395,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs); #endif @@ -404,7 +404,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs); #endif @@ -418,7 +418,7 @@ struct DeviceOperationInstanceFactory< if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs); @@ -430,7 +430,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs); @@ -443,7 +443,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs); @@ -455,7 +455,7 @@ struct DeviceOperationInstanceFactory< else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs); + /// add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs); #ifdef DL_KERNELS add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs); add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs); diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp index 1880fadeaa..ad846e4c80 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp @@ -58,27 +58,27 @@ using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple< DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>, DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1>, DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v1> -#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES +#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES // pipeline v1, 2 waves , - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1> + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Interwave, PipelineVersion::v1> #endif -#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES +#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES // pipeline v2, 1 wave , DeviceGemm_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8, LoopScheduler::Default, PipelineVersion::v2>, diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp index 0b1bbe3af9..82eae9f0a2 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp @@ -31,31 +31,72 @@ static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecializat // Compilation parameters for a[m, k] * b[k, n] = c[m, n] template -using device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances = - std::tuple< - // clang-format off - //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| - //#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| - //#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| - //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16>, - DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16> - // clang-format on - >; +using device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances = std::tuple< + // clang-format off + //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler| Pipeline| + //#####################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| | | + //#####################| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | + //#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // pipeline v1, 1 wave + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v1> +#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES + // pipeline v1, 2 waves + , + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Interwave, PipelineVersion::v1> + +#endif +#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES + // pipeline v2, 1 wave + , + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 4, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 256, 64, 16, 16, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 128, 64, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 2>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 4, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 128, 64, 128, 64, 16, 16, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 16, 1, 1, 1, S<1, 32, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 4, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 128, 64, 64, 16, 16, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 4, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2>, + DeviceGemm_Xdl_CShuffle< Row, Row, Row, F8, F8, F8, F32, F8, PassThrough, PassThrough, PassThrough, GemmSpec, 1, 256, 64, 128, 64, 16, 16, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 16, 1, 1, 1, S<1, 64, 1, 4>, 16, LoopScheduler::Default, PipelineVersion::v2> +#endif + // clang-format on + >; void add_device_gemm_xdl_c_shuffle_f8_f8_f8_mk_kn_mn_instances( std::vector #include #include +#include #include "ck/ck.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" @@ -20,6 +21,7 @@ #include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/literals.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" +#include "ck/library/utility/fill.hpp" namespace ck { namespace profiler { @@ -69,14 +71,17 @@ int profile_gemm_impl(int do_verification, switch(init_method) { - case 0: break; + case 0: + ck::utils::FillConstant{static_cast(1.f)}(a_m_k); + ck::utils::FillConstant{static_cast(1.f)}(b_k_n); + break; case 1: - a_m_k.GenerateTensorValue(GeneratorTensor_2{-5, 5}); - b_k_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + ck::utils::FillUniformDistributionIntegerValue{-5.f, 5.f}(a_m_k); + ck::utils::FillUniformDistributionIntegerValue{-5.f, 5.f}(b_k_n); break; default: - a_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 0.1}); - b_k_n.GenerateTensorValue(GeneratorTensor_3{-0.01, 0.01}); + ck::utils::FillUniformDistribution{-1.f, 1.f}(a_m_k); + ck::utils::FillUniformDistribution{-1.f, 1.f}(b_k_n); } using AElementOp = ck::tensor_operation::element_wise::PassThrough; @@ -130,11 +135,10 @@ int profile_gemm_impl(int do_verification, ref_invoker.Run(ref_argument); } - std::string best_op_name; - float best_avg_time = 0; - float best_tflops = 0; - float best_gb_per_sec = 0; + float best_tflops = 0; + int best_instance_id = 0; + int instance_id = 0; // profile device op instances for(auto& op_ptr : op_ptrs) { @@ -178,10 +182,8 @@ int profile_gemm_impl(int do_verification, if(tflops > best_tflops) { - best_op_name = op_name; - best_tflops = tflops; - best_avg_time = avg_time; - best_gb_per_sec = gb_per_sec; + best_instance_id = instance_id; + best_tflops = tflops; } if(do_verification) @@ -205,53 +207,94 @@ int profile_gemm_impl(int do_verification, { std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl; } + + instance_id++; } - if constexpr(is_same::value) + sleep(2); + + // Run the best instance again { - std::cout << "Best Perf for datatype = f32"; - } - else if constexpr(is_same::value) - { - std::cout << "Best Perf for datatype = f16"; - } - else if constexpr(is_same::value) - { - std::cout << "Best Perf for datatype = bf16"; - } - else if constexpr(is_same::value) - { - std::cout << "Best Perf for datatype = int8"; - } + auto& op_ptr = op_ptrs[best_instance_id]; + auto argument_ptr = + op_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + static_cast(c_device_buf.GetDeviceBuffer()), + M, + N, + K, + StrideA, + StrideB, + StrideC, + a_element_op, + b_element_op, + c_element_op); + + auto invoker_ptr = op_ptr->MakeInvokerPointer(); + + if(op_ptr->IsSupportedArgument(argument_ptr.get())) + { + std::string op_name = op_ptr->GetTypeString(); + + float avg_time = invoker_ptr->Run(argument_ptr.get(), + StreamConfig{nullptr, time_kernel, 0, 50, 200}); + + std::size_t flop = std::size_t(2) * M * N * K; + + std::size_t num_btype = + sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N; + + float tflops = static_cast(flop) / 1.E9 / avg_time; + + float gb_per_sec = num_btype / 1.E6 / avg_time; + + if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f32"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = f16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = bf16"; + } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = int8"; + } #if defined CK_ENABLE_FP8 - else if constexpr(is_same::value) - { - std::cout << "Best Perf for datatype = fp8"; - } + else if constexpr(is_same::value) + { + std::cout << "Best Perf for datatype = fp8"; + } #endif - if constexpr(is_same::value) - { - std::cout << " ALayout = RowMajor"; - } - else if constexpr(is_same::value) - { - std::cout << " ALayout = ColumnMajor"; - } + if constexpr(is_same::value) + { + std::cout << " ALayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " ALayout = ColumnMajor"; + } - if constexpr(is_same::value) - { - std::cout << " BLayout = RowMajor"; - } - else if constexpr(is_same::value) - { - std::cout << " BLayout = ColumnMajor"; - } + if constexpr(is_same::value) + { + std::cout << " BLayout = RowMajor"; + } + else if constexpr(is_same::value) + { + std::cout << " BLayout = ColumnMajor"; + } - std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA - << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_avg_time - << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, " - << best_op_name << std::endl; + std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA + << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << avg_time + << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << op_name + << std::endl; + } + } return pass ? 0 : 1; }