ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-05 12:44:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit 'e77a7ca2bc65651b5e87a0127e0335733aca2f35' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2025-12-18 21:13:07 +00:00
parent cef729b554
commit 87b8b502e6
93 changed files with 8182 additions and 127 deletions
Download Patch File Download Diff File Expand all files Collapse all files

273
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_instance.hpp Normal file
View File

@@ -0,0 +1,273 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddClamp = ck::tensor_operation::element_wise::AddClamp;
using Clamp = ck::tensor_operation::element_wise::Clamp;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd1x1P0 = ConvolutionForwardSpecialization::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
static constexpr auto ConvFwdOddC =
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_instances_part1 = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 64, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_instances_part2 = std::tuple<
// clang-format off
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 32, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_instances_part3 = std::tuple<
// clang-format off
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 64, 8, 8, 16, 16, 4, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_instances_part4 = std::tuple<
// clang-format off
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 8, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_f16_instances_part1 = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 64, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_f16_instances_part2 = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 32, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_f16_instances_part3 = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 96, 64, 8, 8, 16, 16, 4, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_f16_instances_part4 = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 8, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, DsDataTypes, F16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

145
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_scaleadd_ab_instance.hpp Normal file
View File

@@ -0,0 +1,145 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_wmma_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using I8 = int8_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;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd1x1P0 = ConvolutionForwardSpecialization::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
static constexpr auto ConvFwdOddC =
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_wmma_cshufflev3_scaleadd_ab_bf16_instances =
std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 32, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 64, 8, 8, 16, 16, 4, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 8, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<BF16, BF16>, ck::Tuple<BF16, BF16>, F32, BF16, ck::Tuple<>, BF16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_wmma_cshufflev3_scaleadd_ab_f16_instances =
std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 48, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 32, 64, 8, 8, 16, 16, 2, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 32, 8, 8, 16, 16, 4, 3, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 96, 64, 8, 8, 16, 16, 4, 3, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 64, 32, 8, 8, 16, 16, 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, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 32, 32, 8, 8, 16, 16, 4, 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, 0, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 256, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 128, 128, 32, 8, 8, 16, 16, 8, 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, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 128, 64, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, ck::Tuple<>, ELayout, ck::Tuple<F16, F16>, ck::Tuple<F16, F16>, F32, F16, ck::Tuple<>, F16, ScaleAdd, ScaleAdd, PassThrough, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

97
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View File

@@ -24,8 +24,14 @@
#include "grouped_convolution_forward_mem_intra_xdl.inc"
#endif
#ifdef CK_USE_WMMA
#define CK_USE_WMMA_OLD
#ifdef CK_USE_WMMA_OLD
#include "grouped_convolution_forward_wmma.inc"
#endif
#include "grouped_convolution_forward_wmma_cshufflev3.inc"
#include "grouped_convolution_forward_mem_inter_wmma_cshufflev3.inc"
#include "grouped_convolution_forward_mem_intra_wmma_cshufflev3.inc"
#endif
namespace ck {
namespace tensor_operation {
@@ -652,7 +658,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif // CK_USE_XDL
#ifdef CK_USE_WMMA
// TODO: Put old Wmma instances back
#ifdef CK_USE_WMMA_OLD
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{
@@ -766,6 +774,93 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif
}
#endif // CK_USE_WMMA_OLD
#ifdef CK_USE_WMMA
// layout NHWGC/GKYXC/NHWGK
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
}
#endif
}
// 3D
// layout NDHWGC/GKZYXC/NDHWGK
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
}
#endif
}
#endif // CK_USE_WMMA
return op_ptrs;

89
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp.hpp
View File

@@ -16,6 +16,10 @@
#include "grouped_convolution_forward_bias_clamp_xdl.inc"
#endif
#ifdef CK_USE_WMMA
#include "grouped_convolution_forward_bias_clamp_wmma_cshufflev3.inc"
#endif
namespace ck {
namespace tensor_operation {
namespace device {
@@ -269,6 +273,91 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif // CK_USE_XDL
#ifdef CK_USE_WMMA
// layout NHWGC/GKYXC/NHWGK
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// op_ptrs);
}
#endif
}
// layout NDHWGC/GKZYXC/NDHWGK
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
// op_ptrs);
}
#endif
}
#endif // CK_USE_WMMA
return op_ptrs;
}
};

294
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp_wmma_cshufflev3.inc Normal file
View File

@@ -0,0 +1,294 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Tuple<NHWGK>,
// NHWGK,
// BF16,
// BF16,
// Tuple<BF16>,
// BF16,
// PassThrough,
// PassThrough,
// AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NDHWGC,
// GKZYXC,
// Tuple<NDHWGK>,
// NDHWGK,
// BF16,
// BF16,
// Tuple<BF16>,
// BF16,
// PassThrough,
// PassThrough,
// AddClamp>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Tuple<NHWGK>,
// NHWGK,
// F16,
// F16,
// Tuple<F16>,
// F16,
// PassThrough,
// PassThrough,
// AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NDHWGC,
// GKZYXC,
// Tuple<NDHWGK>,
// NDHWGK,
// F16,
// F16,
// Tuple<F16>,
// F16,
// PassThrough,
// PassThrough,
// AddClamp>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

89
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_clamp.hpp
View File

@@ -16,6 +16,10 @@
#include "grouped_convolution_forward_clamp_xdl.inc"
#endif
#ifdef CK_USE_WMMA
#include "grouped_convolution_forward_clamp_wmma_cshufflev3.inc"
#endif
namespace ck {
namespace tensor_operation {
namespace device {
@@ -266,6 +270,91 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif // CK_USE_XDL
#ifdef CK_USE_WMMA
// layout NHWGC/GKYXC/NHWGK
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// op_ptrs);
}
#endif
}
// layout NDHWGC/GKZYXC/NDHWGK
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t> &&
is_same_v<AComputeType, ck::bhalf_t> &&
is_same_v<BComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<AComputeType, half_t> &&
is_same_v<BComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
op_ptrs);
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
// op_ptrs);
}
#endif
}
#endif // CK_USE_WMMA
return op_ptrs;
}
};

294
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_clamp_wmma_cshufflev3.inc Normal file
View File

@@ -0,0 +1,294 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Tuple<>,
// NHWGK,
// BF16,
// BF16,
// Tuple<>,
// BF16,
// PassThrough,
// PassThrough,
// Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NDHWGC,
// GKZYXC,
// Tuple<>,
// NDHWGK,
// BF16,
// BF16,
// Tuple<>,
// BF16,
// PassThrough,
// PassThrough,
// Clamp>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
F16,
F16,
Tuple<>,
F16,
PassThrough,
PassThrough,
Clamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
// NHWGC,
// GKYXC,
// Tuple<>,
// NHWGK,
// F16,
// F16,
// Tuple<>,
// F16,
// PassThrough,
// PassThrough,
// Clamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
// NDHWGC,
// GKZYXC,
// Tuple<>,
// NDHWGK,
// F16,
// F16,
// Tuple<>,
// F16,
// PassThrough,
// PassThrough,
// Clamp>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

30
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_inter_wmma_cshufflev3.inc Normal file
View File

@@ -0,0 +1,30 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_int8_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

30
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_intra_wmma_cshufflev3.inc Normal file
View File

@@ -0,0 +1,30 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_int8_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

64
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp
View File

@@ -21,6 +21,7 @@ namespace instance {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ScaleAdd = ck::tensor_operation::element_wise::ScaleAdd;
#ifdef CK_USE_XDL
#ifdef CK_ENABLE_BF16
// grouped conv3d forward multi AB scaleadd, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
@@ -85,6 +86,42 @@ void add_device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_ins
ScaleAdd,
PassThrough>>>& instances);
#endif
#endif // CK_USE_XDL
#ifdef CK_USE_WMMA
#ifdef CK_ENABLE_BF16
// grouped conv3d forward multi AB scaleadd, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
ck::Tuple<BF16, BF16>,
ck::Tuple<BF16, BF16>,
ck::Tuple<>,
BF16,
ScaleAdd,
ScaleAdd,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<>,
NDHWGK,
ck::Tuple<F16, F16>,
ck::Tuple<F16, F16>,
ck::Tuple<>,
F16,
ScaleAdd,
ScaleAdd,
PassThrough>>>& instances);
#endif
#endif // CK_USE_WMMA
template <ck::index_t NumDimSpatial,
typename InLayout,
@@ -129,6 +166,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_USE_XDL
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
@@ -169,6 +207,32 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
}
#endif
}
#endif // CK_USE_XDL
#ifdef CK_USE_WMMA
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, ck::Tuple<half_t, half_t>> &&
is_same_v<WeiDataType, ck::Tuple<half_t, half_t>> &&
is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_wmma_cshufflev3_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::Tuple<ck::bhalf_t, ck::bhalf_t>> &&
is_same_v<WeiDataType, ck::Tuple<ck::bhalf_t, ck::bhalf_t>> &&
is_same_v<OutDataType, ck::bhalf_t> && is_same_v<ComputeType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_wmma_cshufflev3_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
}
#endif
}
#endif // CK_USE_WMMA
return op_ptrs;
}

236
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_wmma_cshufflev3.inc Normal file
View File

@@ -0,0 +1,236 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part1(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part2(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part3(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_instances_part4(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck