ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 20:40:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add grouped conv bwd weight wmma (#985)

Browse Source 
* Add grouped conv bwd weight wmma

* Update README, changelog, profiler

* Minor fixes

* Fix grouped conv bwd wei dl kernel

* Minor fixes

* Minor stylistic fixes

[ROCm/composable_kernel commit: 16d7c4d2f7]
This commit is contained in:
Bartłomiej Kocot
2023-10-17 10:32:26 +02:00
committed by GitHub
parent 6fe4f27c05
commit 16f6f238ed
70 changed files with 1960 additions and 173 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
View File

@@ -6,8 +6,6 @@
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.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 {

119
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp Normal file
View File

@@ -0,0 +1,119 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, 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/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename CLayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_wmma_f16_instances =
std::tuple<
// clang-format off
//#####################################| NumDim| A| B| C| AData| BData| CData| AccData| A| B| C| ConvForward| Block| MPer| NPer| KPer| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################################| Spatial| Layout| Layout| Layout| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeatPerWave| NRepeatPerWave| _MBlock_MPerBlock| ScalarPerVector|
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 8, 8, 16, 16, 2, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 8, 8, 16, 16, 8, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 2>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 256, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 16>, 4>,
// blocksize=128
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 8, 8, 16, 16, 2, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 8, 8, 16, 16, 4, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 256, 8, 8, 16, 16, 1, 8, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 256, 32, 8, 8, 16, 16, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
// blocksize=64
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 8, 8, 16, 16, 4, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 8, 8, 16, 16, 1, 4, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 8, 8, 16, 16, 2, 4, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 8, 8, 16, 16, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 8, 8, 16, 16, 1, 8, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
// blocksize=32
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 32, 8, 8, 16, 16, 1, 2, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 64, 8, 8, 16, 16, 1, 4, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 32, 64, 8, 8, 16, 16, 2, 4, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 32, 32, 8, 8, 16, 16, 2, 2, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 64, 32, 8, 8, 16, 16, 4, 2, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 64, 16, 8, 8, 16, 16, 4, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 32, 16, 8, 8, 16, 16, 2, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename CLayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_wmma_i8_instances =
std::tuple<
// clang-format off
//#####################################| NumDim| A| B| C| AData| BData| CData| AccData| A| B| C| ConvForward| Block| MPer| NPer| KPer| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################################| Spatial| Layout| Layout| Layout| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeatPerWave| NRepeatPerWave| _MBlock_MPerBlock| ScalarPerVector|
//#####################################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 256, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 64, 1, 4>, 8>,
// blocksize=128
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 256, 8, 8, 16, 16, 4, 8, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 256, 8, 8, 16, 16, 2, 8, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 256, 8, 8, 16, 16, 1, 8, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 8, 8, 16, 16, 2, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 256, 32, 8, 8, 16, 16, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 256, 64, 8, 8, 16, 16, 8, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 2>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 256, 128, 8, 8, 16, 16, 8, 4, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
// blocksize=64
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 8, 8, 16, 16, 1, 8, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 128, 8, 8, 16, 16, 2, 8, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 64, 8, 8, 16, 16, 8, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 8, 8, 16, 16, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
// blocksize=32
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 64, 8, 8, 16, 16, 1, 4, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 64, 64, 8, 8, 16, 16, 4, 4, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 32, 32, 8, 8, 16, 16, 2, 2, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdWeight_Wmma_CShuffle<NDSpatial, ALayout, BLayout, CLayout, I8, I8, I8, I32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 64, 16, 8, 8, 16, 16, 4, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

126
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View File

@@ -163,6 +163,30 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC,
GKZYXC,
GNDHWK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC,
GKZYXC,
GNDHWK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
@@ -177,6 +201,31 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC,
GKZYXC,
GNDHWK,
int8_t,
int8_t,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC,
GKZYXC,
GNDHWK,
int8_t,
int8_t,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
@@ -202,6 +251,30 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
@@ -231,6 +304,31 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_
BF8,
F8>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
int8_t,
int8_t,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
int8_t,
int8_t,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef DL_KERNELS
// dl
@@ -694,6 +792,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
@@ -708,6 +810,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
}
else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
@@ -737,6 +849,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
@@ -752,6 +868,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> &&