Skip to content

GitLab

"include/ck/utility/reduction_common.hpp" did not exist on "b3e8d57d51300b88b591900621f71b6a1b3a7acc"
Unverified Commit 9e4429f9 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Gemm+Bilinear (#316) 

* refactor

* update example

* update example

* gemm bilinear

* clean

* update
parent 8e374781
Show whitespace changes
Inline Side-by-side
Showing with 515 additions and 1205 deletions
library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ 9e4429f9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using F16_TUPLE = ck::Tuple<F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
// no padding
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| DataType| Type| Type| 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|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, 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>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
// M/N/N padding
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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| DataType| Type| Type| 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|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 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<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_TUPLE, F16, PassThrough, PassThrough, Bilinear, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on
>;
void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
Col,
Row,
F16,
F16,
F16_TUPLE,
F16,
PassThrough,
PassThrough,
Bilinear>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 9e4429f9
......@@ -7,21 +7,19 @@ set(PROFILER_SOURCE
src/profiler.cpp
src/profile_gemm.cpp
src/profile_gemm_splitk.cpp
src/profile_gemm_bias_2d.cpp
src/profile_gemm_bias_relu.cpp
src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp
src/profile_gemm_bilinear.cpp
src/profile_gemm_bias_add_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
src/profile_gemm_reduce.cpp
src/profile_batched_gemm.cpp
src/profile_batched_gemm_reduce.cpp
src/profile_grouped_gemm.cpp
src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp
src/profile_convnd_fwd.cpp
src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
src/profile_reduce.cpp
src/profile_normalization.cpp
)
......@@ -31,12 +29,10 @@ target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE conv_util)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_splitk_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bilinear_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
......
profiler/include/profile_batched_gemm_impl.hpp
View file @ 9e4429f9
......@@ -159,10 +159,10 @@ bool profile_batched_gemm_impl(int do_verification,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchCount,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
BatchCount);
ck::tensor_operation::element_wise::PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
profiler/include/profile_gemm_bias_2d_impl.hpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceGemmAlphaBetaPtr = ck::tensor_operation::device::DeviceGemmBiasPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AlphaBetaAdd>;
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename C0DataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_2d_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
float alpha,
float beta)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<C0DataType> c0_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c0_m_n: " << c0_m_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c0_m_n.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5}, num_thread);
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
c0_m_n.GenerateTensorValue(GeneratorTensor_3<C0DataType>{-0.5, 0.5}, num_thread);
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AlphaBetaAdd;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{alpha, beta};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<ADataType,
BDataType,
C0DataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c0_m_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c0_device_buf(sizeof(C0DataType) * c0_m_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c0_device_buf.ToDevice(c0_m_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::instance::DeviceGemmAlphaBetaPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
is_same<CDataType, float>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<C0DataType*>(c0_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0 : ", c0_m_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_relu_add_impl.hpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation_add.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceGemmBiasReluAddPtr = ck::tensor_operation::device::DeviceGemmBiasActivationAddPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddReluAdd>;
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluAddPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_add_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideC1,
int KBatch = 1)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
// c1_m_n[m ,n]
Tensor<BDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{});
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddReluAdd;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivationAdd<ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
b_k_n,
c_m_n_host_result,
c0_n,
c1_m_n,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::instance::DeviceGemmBiasReluAddPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_add_f16_f16_f16_km_nk_mn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a: ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0: ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c1: ", c1_m_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_relu_impl.hpp → profiler/include/profile_gemm_bilinear_impl.hpp
View file @ 9e4429f9
......@@ -3,62 +3,45 @@
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using DeviceGemmBiasReluPtr = ck::tensor_operation::device::DeviceGemmBiasActivationPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmBiasReluPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename DDataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_relu_impl(int do_verification,
typename DELayout> // assume Ds and E have same layout
bool profile_gemm_bilinear_impl(int do_verification,
int init_method,
bool do_log,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch = 1)
int StrideD,
int StrideE,
float alpha,
float beta)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
......@@ -76,162 +59,151 @@ void profile_gemm_bias_relu_impl(int do_verification,
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, DELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::AddRelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Bilinear;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto cde_element_op = CDEElementOp{alpha, beta};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
DELayout,
ADataType,
BDataType,
ck::Tuple<DDataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op);
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::instance::DeviceGemmBiasReluPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
for(int m = 0; m < M; ++m)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
for(int n = 0; n < N; ++n)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::instance::
add_device_gemm_xdl_c_shuffle_bias_relu_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
std::string best_gemm_name;
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_m_n_device_buf.ToDevice(d_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
StrideC,
std::array<ck::index_t, 1>{StrideD},
StrideE,
a_element_op,
b_element_op,
c_element_op,
KBatch);
cde_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
......@@ -239,30 +211,22 @@ void profile_gemm_bias_relu_impl(int do_verification,
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c0 : ", c0_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
pass = pass &&
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
......
profiler/src/profile_batched_gemm.cpp
View file @ 9e4429f9
......@@ -27,8 +27,9 @@ enum struct GemmDataType
int profile_batched_gemm(int argc, char* argv[])
{
if(argc != 15)
if(argc != 18)
{
// clang-format off
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
......@@ -39,7 +40,8 @@ int profile_batched_gemm(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
// clang-format on
exit(1);
}
......@@ -58,7 +60,11 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
using F32 = float;
using F16 = ck::half_t;
......@@ -90,9 +96,13 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int BatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int BatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int BatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
......@@ -103,9 +113,9 @@ int profile_batched_gemm(int argc, char* argv[])
M,
N,
K,
BatchStrideA,
BatchStrideB,
BatchStrideC,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
......
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ 9e4429f9
......@@ -29,7 +29,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
......@@ -39,7 +39,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
......
profiler/src/profile_gemm_bias_2d.cpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_2d_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_2d(int argc, char* argv[])
{
if(!(argc == 16 || argc == 17))
{
printf("arg1: tensor operation (gemm: GEMM+Bias_2d)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: alpha\n");
printf("arg15: beta\n");
printf("arg16: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const float alpha = std::stof(argv[14]);
const float beta = std::stof(argv[15]);
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
float,
float,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu.cpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_relu_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_relu(int argc, char* argv[])
{
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu_add.cpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_relu_add_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_relu_add(int argc, char* argv[])
{
if(!(argc == 15 || argc == 16))
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU+Add)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
printf("arg15: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int StrideC1 = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bilinear.cpp 0 → 100644
View file @ 9e4429f9
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bilinear_impl.hpp"
int profile_gemm_bilinear(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F32_F32_F32_F32, // 0
F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8, // 3
};
if(argc != 17)
{
// clang-format off
printf("arg1: tensor operation (gemm_bilinear: GEMM+Bilinear)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
printf(" 1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
printf(" 2: E[m, n] = alpha * A[k, m] * B[k, n] + beta * D[m, n];\n");
printf(" 3: E[m, n] = alpha * A[k, m] * B[n, k] + beta * D[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
printf("arg15 to 16: alhpa, beta\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
const float alpha = std::stof(argv[15]);
const float beta = std::stof(argv[16]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d_type,
auto e_type,
auto a_layout,
auto b_layout,
auto de_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using DDataType = decltype(d_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using DELayout = decltype(de_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<DELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_bilinear_impl<ADataType,
BDataType,
AccDataType,
DDataType,
EDataType,
ALayout,
BLayout,
DELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD < 0) ? DefaultStrideD : StrideD,
(StrideE < 0) ? DefaultStrideE : StrideE,
alpha,
beta);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
profiler/src/profiler.cpp
View file @ 9e4429f9
......@@ -5,12 +5,10 @@
int profile_gemm(int, char*[]);
int profile_gemm_splitk(int, char*[]);
int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_gemm_bilinear(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
......@@ -28,12 +26,12 @@ static void print_helper_message()
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_splitk: Split-K GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_bilinear: GEMM+Bilinear\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" gemm_reduce: GEMM+Reduce\n"
" gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
" batched_gemm: Batched GEMM\n"
" batched_gemm_reduce: Batched GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
......@@ -63,17 +61,13 @@ int main(int argc, char* argv[])
{
return profile_gemm_splitk(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_2d") == 0)
{
return profile_gemm_bias_2d(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu") == 0)
else if(strcmp(argv[1], "gemm_bilinear") == 0)
{
return profile_gemm_bias_relu(argc, argv);
return profile_gemm_bilinear(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu_add") == 0)
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
return profile_gemm_bias_relu_add(argc, argv);
return profile_gemm_add_add_fastgelu(argc, argv);
}
else if(strcmp(argv[1], "gemm_reduce") == 0)
{
......@@ -119,17 +113,13 @@ int main(int argc, char* argv[])
{
return profile_convnd_bwd_data(argc, argv, 3);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "conv2d_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_gemm_add_add_fastgelu(argc, argv);
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 ||
strcmp(argv[1], "softmax") == 0)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment