Skip to content

GitLab

Commit 7e493730 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge branch 'develop' into wavelet_model

parents b89a88b5 40942b90
Hide whitespace changes
Inline Side-by-side
Showing with 989 additions and 217 deletions
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 7e493730
......@@ -6,63 +6,19 @@ function(add_instance_library INSTANCE_NAME)
clang_tidy_check(${INSTANCE_NAME})
endfunction(add_instance_library INSTANCE_NAME)
add_subdirectory(gemm)
add_subdirectory(gemm_splitk)
add_subdirectory(gemm_bilinear)
add_subdirectory(gemm_add_add_fastgelu)
add_subdirectory(gemm_reduce)
add_subdirectory(gemm_bias_add_reduce)
add_subdirectory(batched_gemm)
add_subdirectory(batched_gemm_reduce)
add_subdirectory(batched_gemm_gemm)
add_subdirectory(batched_gemm_softmax_gemm)
add_subdirectory(batched_gemm_add_relu_gemm_add)
add_subdirectory(grouped_gemm)
add_subdirectory(contraction_scale)
add_subdirectory(contraction_bilinear)
add_subdirectory(grouped_conv1d_fwd)
add_subdirectory(grouped_conv2d_fwd)
add_subdirectory(grouped_conv3d_fwd)
add_subdirectory(conv2d_fwd)
add_subdirectory(conv1d_bwd_data)
add_subdirectory(conv2d_bwd_data)
add_subdirectory(conv3d_bwd_data)
add_subdirectory(conv1d_bwd_weight)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(conv3d_bwd_weight)
add_subdirectory(conv2d_fwd_bias_relu)
add_subdirectory(conv2d_fwd_bias_relu_add)
add_subdirectory(reduce)
add_subdirectory(normalization)
add_subdirectory(elementwise)
file(GLOB dir_list LIST_DIRECTORIES true *)
set(CK_DEVICE_INSTANCES)
FOREACH(subdir_path ${dir_list})
set(target_dir)
IF(IS_DIRECTORY "${subdir_path}")
get_filename_component(target_dir ${subdir_path} NAME)
add_subdirectory(${target_dir})
list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
ENDIF()
ENDFOREACH()
add_library(device_operations STATIC
$<TARGET_OBJECTS:device_gemm_instance>
$<TARGET_OBJECTS:device_gemm_splitk_instance>
$<TARGET_OBJECTS:device_gemm_bilinear_instance>
$<TARGET_OBJECTS:device_gemm_add_add_fastgelu_instance>
$<TARGET_OBJECTS:device_gemm_bias_add_reduce_instance>
$<TARGET_OBJECTS:device_batched_gemm_instance>
$<TARGET_OBJECTS:device_batched_gemm_add_relu_gemm_add_instance>
$<TARGET_OBJECTS:device_batched_gemm_reduce_instance>
$<TARGET_OBJECTS:device_grouped_gemm_instance>
$<TARGET_OBJECTS:device_contraction_scale_instance>
$<TARGET_OBJECTS:device_contraction_bilinear_instance>
$<TARGET_OBJECTS:device_grouped_conv1d_fwd_instance>
$<TARGET_OBJECTS:device_grouped_conv2d_fwd_instance>
$<TARGET_OBJECTS:device_grouped_conv3d_fwd_instance>
$<TARGET_OBJECTS:device_conv1d_bwd_data_instance>
$<TARGET_OBJECTS:device_conv2d_bwd_data_instance>
$<TARGET_OBJECTS:device_conv3d_bwd_data_instance>
$<TARGET_OBJECTS:device_conv1d_bwd_weight_instance>
$<TARGET_OBJECTS:device_conv2d_bwd_weight_instance>
$<TARGET_OBJECTS:device_conv3d_bwd_weight_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_bias_relu_instance>
$<TARGET_OBJECTS:device_conv2d_fwd_bias_relu_add_instance>
$<TARGET_OBJECTS:device_reduce_instance>
$<TARGET_OBJECTS:device_normalization_instance>
$<TARGET_OBJECTS:device_elementwise_instance>
)
add_library(device_operations STATIC ${CK_DEVICE_INSTANCES})
add_library(composablekernels::device_operations ALIAS device_operations)
......
library/src/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute/CMakeLists.txt 0 → 100644
View file @ 7e493730
add_instance_library(device_batched_gemm_masking_scale_softmax_gemm_permute_instance
device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
)
library/src/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute/device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp 0 → 100644
View file @ 7e493730
// 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_batched_gemm_softmax_gemm_permute_xdl_cshuffle.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 Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
using device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
std::tuple<
// clang-format off
// 2 of them are commented out because they trigger the clang-13 issue.
//##############################################| ALayout| B0Layout| B1Layout| CPermuteNumDims_G_M_O| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//##############################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//##############################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
//DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
//DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, true>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>,
DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle< Row, Col, Row, CPermuteNumDims_G_M_O, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, Scale, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, true>
// clang-format on
>;
void add_device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<Row,
Col,
Row,
CPermuteNumDims_G_M_O,
F16,
F16,
F16,
F16,
PassThrough,
PassThrough,
Scale,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_batched_gemm_masking_scale_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ 7e493730
......@@ -26,32 +26,47 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmPadded =
ck::tensor_operation::device::GemmSpecialization::MNOPadding; // Padding K is currently flawed
static constexpr auto GemmPadded = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
std::tuple<
// clang-format off
//#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8>,
//#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 64, 32, 8, 8, 2, 32, 32, 2, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 256, 128, 32, 128, 32, 8, 8, 2, 32, 32, 2, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 64, 32, 8, 8, 2, 32, 32, 1, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 256, 32, 128, 32, 8, 8, 2, 32, 32, 1, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 32, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 128, 32, 8, 8, 2, 16, 16, 1, 16, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 8, S<1, 16, 1,16>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmDefault, 1, 256, 64, 256, 64, 64, 32, 8, 8, 2, 16, 16, 1, 16, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, 8, false>,
// Padded fallback kernel
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8>
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 64, 128, 32, 8, 8, 2, 32, 32, 1, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 64, 32, 128, 32, 8, 8, 2, 32, 32, 1, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>
// clang-format on
>;
using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances =
std::tuple<
// clang-format off
//#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle| A| B0| Acc0| B1| C| GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| MaskOut|
//#######################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Upper|
//#######################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| Triangle|
//#######################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 64, 32, 4, 4, 2, 32, 32, 2, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 256, 128, 40, 128, 32, 4, 4, 2, 32, 32, 2, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 64, 32, 4, 4, 2, 32, 32, 1, 8, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 256, 40, 128, 32, 4, 4, 2, 32, 32, 1, 8, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 64, 32, 4, 4, 2, 32, 32, 1, 4, 2, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>,
DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle< Row, Col, Row, Row, F16, F16, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough, GemmPadded, 1, 256, 128, 128, 40, 128, 32, 4, 4, 2, 32, 32, 1, 4, 4, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S<2,128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, false, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, false>
// clang-format on
>;
......@@ -73,6 +88,9 @@ void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_g
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
add_device_operation_instances(
instances,
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
View file @ 7e493730
......@@ -37,6 +37,7 @@ static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecial
using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
// clang-format off
// no padding
// N % 8 == 0 && K % 8 == 0
//##############################| A| B| Ds| E| 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|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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|
......@@ -55,7 +56,8 @@ using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances =
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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_Tuple, Row, F16, F16, F32, F16, 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
// M/N/K padding
// N % 8 == 0 && K % 8 == 0
//##############################| A| B| Ds| E| 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|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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|
......@@ -72,7 +74,48 @@ using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances =
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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_Tuple, Row, F16, F16, F32, F16, 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_Tuple, Row, F16, F16, F32, F16, 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_Tuple, Row, F16, F16, F32, F16, 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>
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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>,
// M/N/K padding
// N % 4 == 0 && K % 4 == 0
//##############################| A| B| Ds| E| 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|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 16>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 8, 1, 8>, 4>,
// M/N/K padding
// N % 8 == 0 && K % 1 == 0
//##############################| A| B| Ds| E| 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|
//##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| 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|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 32>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 32>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 4, 1, 32>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 64>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 32>, 1>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F16, F32, F16, 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, 2, 1, 32>, 1>
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f16_instance.cpp
View file @ 7e493730
......@@ -17,34 +17,42 @@ using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
template <index_t Rank, index_t Reduce>
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_layernorm_f16_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8, 8, 8>
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
DeviceLayernormImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
// clang-format on
>;
void add_device_layernorm_f16_rank2_instances(
std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 2, 1>>& instances)
void add_device_layernorm_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F16, F16, F16, F32, F16, Pass, 2, 1>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f16_instances<2, 1>{});
add_device_operation_instances(instances, device_layernorm_f16_instances<Pass, 2, 1>{});
}
void add_device_layernorm_f16_rank4_instances(
std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, Pass, 4, 3>>& instances)
void add_device_layernorm_rank_4_3_f16_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F16, F16, F16, F32, F16, Pass, 4, 3>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f16_instances<4, 3>{});
add_device_operation_instances(instances, device_layernorm_f16_instances<Pass, 4, 3>{});
}
void add_device_layernorm_rank_5_3_f16_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F16, F16, F16, F32, F16, Pass, 5, 3>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f16_instances<Pass, 5, 3>{});
}
} // namespace instance
......
library/src/tensor_operation_instance/gpu/normalization/device_layernorm_f32_instance.cpp
View file @ 7e493730
......@@ -16,33 +16,39 @@ using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
template <index_t Rank, index_t Reduce>
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_layernorm_f32_instances = std::tuple<
// clang-format off
// XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 2, 2, 2>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, Pass, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4, 4, 4>
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
DeviceLayernormImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>
// clang-format on
>;
void add_device_layernorm_f32_rank2_instances(
std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 2, 1>>& instances)
void add_device_layernorm_rank_2_1_f32_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F32, F32, F32, F32, F32, Pass, 2, 1>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f32_instances<2, 1>{});
add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 2, 1>{});
}
void add_device_layernorm_f32_rank4_instances(
std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, Pass, 4, 3>>& instances)
void add_device_layernorm_rank_4_3_f32_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F32, F32, F32, F32, F32, Pass, 4, 3>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f32_instances<4, 3>{});
add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 4, 3>{});
}
void add_device_layernorm_rank_5_3_f32_instances(
std::vector<std::unique_ptr<DeviceLayernorm<F32, F32, F32, F32, F32, Pass, 5, 3>>>& instances)
{
add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 5, 3>{});
}
} // namespace instance
......
profiler/CMakeLists.txt
View file @ 7e493730
......@@ -23,6 +23,7 @@ set(PROFILER_SOURCE
src/profile_conv_bwd_weight.cpp
src/profile_grouped_conv_fwd.cpp
src/profile_reduce.cpp
src/profile_groupnorm.cpp
src/profile_layernorm.cpp
src/profile_normalization.cpp
)
......
profiler/include/profile_batched_gemm_masking_scale_softmax_gemm_permute_impl.hpp 0 → 100644
View file @ 7e493730
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename B0DataType,
typename B1DataType,
typename CDataType,
typename ALayout,
typename B0Layout,
typename B1Layout,
typename CPermuteNumDims_G_M_O>
bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int O,
int G0,
int G1,
int StrideA = -1,
int StrideB0 = -1,
int StrideB1 = -1,
int BatchStrideA = -1,
int BatchStrideB0 = -1,
int BatchStrideB1 = -1,
float alpha = 1.f)
{
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
using PassThrough = tensor_operation::element_wise::PassThrough;
using Scale = tensor_operation::element_wise::Scale;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
using AccDataType = float;
// Ref Gemm0: various type in, fp32 out
using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, various type out
using ReferenceSoftmaxInstance =
tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: various type in, various type out
using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
bool pass = true;
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
const int BatchCount = G0 * G1;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_gs_ms_os_host_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
Tensor<CDataType> c_gs_ms_os_device_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
// Host verification: Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
std::vector<int>{M * O, O, 1});
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
std::srand(1); // work around test flakiness
switch(init_method)
{
case 0: break;
case 1:
// Still unsure whether this kind of deterministic floating point accurary issue is expected
// or not. May want to try exact same approach as the GPU kernel in the host reference
// GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
// shrink the input value range as it is less likely to produce errors of around ~1e-3.
// a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
// b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
// b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
using DeviceOp =
tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
ADataType,
B0DataType,
B1DataType,
CDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
if(do_verification)
{
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
ref_gemm0_invoker.Run(ref_gemm0_argument);
// mask out upper triangle
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(idx[1] < idx[2])
self(idx) = -ck::NumericLimits<float>::Infinity();
});
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device op instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
a_element_op,
b0_element_op,
acc0_element_op,
b1_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 ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
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, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
pass = pass & ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
c_gs_ms_os_device_result.mData,
",")
<< std::endl;
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " 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_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_groupnorm_impl.hpp 0 → 100644
View file @ 7e493730
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/layernorm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
namespace ck {
namespace profiler {
template <typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename AccDataType,
typename YDataType>
bool profile_groupnorm_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
if(length.size() != 5)
return false;
index_t G = length[3];
index_t C = length[4];
std::vector<index_t> reduce_dim = {1, 2, 4};
std::vector<index_t> gammaBetaLength = {G, C};
std::vector<index_t> gammaBetaStride = {0, 0, 0, C, 1};
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(gammaBetaLength);
Tensor<BetaDataType> beta(gammaBetaLength);
Tensor<YDataType> y(length);
Tensor<YDataType> host_y(length);
switch(init_method)
{
case 0:
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
break;
case 1:
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
break;
default:
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
}
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DeviceLayernorm<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
5,
3>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
PassThrough>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(
length,
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
gammaBetaStride,
gammaBetaStride,
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
reduce_dim,
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
PassThrough{});
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
continue;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
y.mDesc.GetElementSize() * sizeof(YDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
y_dev.FromDevice(y.mData.data());
bool pass =
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "x : ", x.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_y : ", host_y.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "y : ", y.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
std::cout << "num_kernel = " << num_kernel << ", best perf = " << best_avg_time << " ms, "
<< best_gb_per_sec << " GB/s, " << best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is tested" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_layernorm_impl.hpp
View file @ 7e493730
......@@ -6,8 +6,8 @@
#include <iomanip>
#include "ck/ck.hpp"
#include "profiler/include/data_type_enum.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/layernorm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
......@@ -15,26 +15,6 @@
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
void add_device_layernorm_f16_rank2_instances(
std::vector<DeviceLayernormPtr<F16, F16, F16, F32, F16, PassThrough, 2, 1>>&);
void add_device_layernorm_f32_rank2_instances(
std::vector<DeviceLayernormPtr<F32, F32, F32, F32, F32, PassThrough, 2, 1>>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
......@@ -53,8 +33,6 @@ void profile_layernorm_impl(int do_verification,
std::vector<index_t> strideGamma,
std::vector<index_t> strideBeta)
{
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
if(length.size() < 2)
......@@ -103,37 +81,24 @@ void profile_layernorm_impl(int do_verification,
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
// add device normalization instances
constexpr int NumReduceDim = Rank - 1;
std::vector<tensor_operation::device::DeviceLayernormPtr<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim>>
instances;
if constexpr(is_same<XDataType, F16>::value && is_same<GammaDataType, F16>::value &&
is_same<BetaDataType, F16>::value && is_same<YDataType, F16>::value &&
is_same<AccDataType, F32>::value)
{
if(length.size() == 2)
tensor_operation::device::instance::add_device_layernorm_f16_rank2_instances(instances);
}
else if constexpr(is_same<XDataType, F32>::value && is_same<GammaDataType, F32>::value &&
is_same<BetaDataType, F32>::value && is_same<YDataType, F32>::value &&
is_same<AccDataType, F32>::value)
{
if(length.size() == 2)
tensor_operation::device::instance::add_device_layernorm_f32_rank2_instances(instances);
}
if(instances.size() <= 0)
{
throw std::runtime_error("wrong! no device normalization instance found");
}
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DeviceLayernorm<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
......@@ -157,7 +122,7 @@ void profile_layernorm_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
for(auto& inst_ptr : instances)
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideXY,
......@@ -175,9 +140,9 @@ void profile_layernorm_impl(int do_verification,
if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = [", length, "], ") << std::endl;
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
return;
continue;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
......
profiler/src/profile_groupnorm.cpp 0 → 100644
View file @ 7e493730
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/include/data_type_enum.hpp"
#include "profiler/include/profile_groupnorm_impl.hpp"
using ck::index_t;
struct GroupnormArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_groupnorm()
{
std::cout << "arg1: tensor operation (groupnorm: Group normalization)\n"
<< "arg2: data type (0: fp16; 1: fp32)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg5: print tensor value (0: no; 1: yes)\n"
<< "arg6: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
<< std::endl;
}
int profile_groupnorm(int argc, char* argv[])
{
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
bool do_verification = false;
int init_method = 0;
bool do_log = 0;
bool time_kernel = 1;
std::vector<index_t> length = {64, 16, 16, 32, 40};
if(argc != 1 && argc != 13)
{
print_help_groupnorm();
return 0;
}
if(argc == 13)
{
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
do_verification = std::stoi(argv[3]);
init_method = std::stoi(argv[4]);
do_log = std::stoi(argv[5]);
time_kernel = std::stoi(argv[6]);
// parse the long options
GroupnormArgParser arg_parser;
arg_parser(argc, argv);
length = arg_parser.long_opts["length"];
}
using F16 = ck::half_t;
using F32 = float;
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
profiler/src/profile_layernorm.cpp
View file @ 7e493730
......@@ -5,6 +5,7 @@
#include <vector>
#include <unordered_map>
#include "profiler/include/data_type_enum.hpp"
#include "profiler/include/profile_layernorm_impl.hpp"
using ck::index_t;
......@@ -49,7 +50,7 @@ void print_help_layernorm()
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=n0, 1=yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1024 1024) \n"
<< "--strideXY: tensor strides (e.g, --strideXY 1024 1)\n"
<< "--strideGamma: tensor strides (e.g, --strideGamma 1)\n"
......@@ -114,10 +115,3 @@ int profile_layernorm(int argc, char* argv[])
return 0;
}
// hijack main() for quick debugging
// int main(int argc, char* argv[])
// {
// profile_layernorm(argc, argv);
// return 0;
// }
profiler/src/profiler.cpp
View file @ 7e493730
......@@ -22,6 +22,7 @@ int profile_conv_bwd_weight(int, char*[]);
int profile_grouped_conv_fwd(int, char*[]);
int profile_normalization(int, char*[]);
int profile_layernorm(int, char*[]);
int profile_groupnorm(int, char*[]);
int profile_reduce(int, char*[]);
static void print_helper_message()
......@@ -136,6 +137,10 @@ int main(int argc, char* argv[])
{
return profile_layernorm(argc, argv);
}
else if(strcmp(argv[1], "groupnorm") == 0)
{
return profile_groupnorm(argc, argv);
}
else
{
print_helper_message();
......
script/cmake-ck-dev.sh 0 → 100755
View file @ 7e493730
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=$1
cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=ON \
-D GPU_TARGETS=gfx908;gfx90a \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
#-D AMDGPU_TARGETS=gfx90a;gfx908
script/cmake-ck-release.sh 0 → 100755
View file @ 7e493730
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=$1
cmake \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_CXX_FLAGS="-O3" \
-D CMAKE_BUILD_TYPE=Release \
-D BUILD_DEV=OFF \
-D GPU_TARGETS=gfx908;gfx90a \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
${MY_PROJECT_SOURCE}
#-D AMDGPU_TARGETS=gfx90a;gfx908
script/cmake-rocm.sh deleted 100755 → 0
View file @ b89a88b5
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=../
MY_PROJECT_INSTALL=../install.dir
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS=" -O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE}
#-D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -ftemplate-backtrace-limit=0 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -gline-tables-only -save-temps=$PWD" \
#-D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -ftemplate-backtrace-limit=0 -gline-tables-only -save-temps=$PWD" \
script/process_perf_data.sh
View file @ 7e493730
......@@ -2,15 +2,14 @@
#
# in order to run this script you'd need the following python packages:
pip3 install --upgrade pip
pip3 install sqlalchemy pymysql pandas sshtunnel
#pip3 install --upgrade pip
#pip3 install sqlalchemy pymysql pandas sshtunnel
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#process results
gpu_arch=$1
python3 process_perf_data.py perf_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_resnet50_N256_"$gpu_arch".log
python3 process_perf_data.py perf_resnet50_N4_"$gpu_arch".log
python3 process_perf_data.py perf_gemm.log
python3 process_perf_data.py perf_resnet50_N256.log
python3 process_perf_data.py perf_resnet50_N4.log
script/process_qa_data.sh
View file @ 7e493730
......@@ -10,15 +10,14 @@
# please contact Illia.Silin@amd.com for more details
#process results
gpu_arch=$1
python3 process_perf_data.py perf_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_resnet50_N256_"$gpu_arch".log
python3 process_perf_data.py perf_resnet50_N4_"$gpu_arch".log
python3 process_perf_data.py perf_batched_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_grouped_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_conv_fwd_"$gpu_arch".log
python3 process_perf_data.py perf_conv_bwd_data_"$gpu_arch".log
python3 process_perf_data.py perf_gemm_bilinear_"$gpu_arch".log
python3 process_perf_data.py perf_reduction_"$gpu_arch".log
python3 process_perf_data.py perf_splitK_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_onnx_gemm_"$gpu_arch".log
python3 process_perf_data.py perf_gemm.log
python3 process_perf_data.py perf_resnet50_N256.log
python3 process_perf_data.py perf_resnet50_N4.log
python3 process_perf_data.py perf_batched_gemm.log
python3 process_perf_data.py perf_grouped_gemm.log
python3 process_perf_data.py perf_conv_fwd.log
python3 process_perf_data.py perf_conv_bwd_data.log
python3 process_perf_data.py perf_gemm_bilinear.log
python3 process_perf_data.py perf_reduction.log
python3 process_perf_data.py perf_splitK_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
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