Skip to content

GitLab

Commit 930b2872 authored by Harisankar Sadasivan's avatar Harisankar Sadasivan
Browse files

best performing kernel for GEMV codex problem with M=1 with inverted B matrix

parents a1e17d18 a4f72a31
Hide whitespace changes
Inline Side-by-side
Showing with 1861 additions and 314 deletions
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp 0 → 100755
View file @ 930b2872
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
using Empty_Tuple = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
static constexpr auto ConvBwdData1x1S1P0 =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename DsDatatype,
typename CDEElementOp,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_wmma_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 256, 128, 256, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 256, 64, 256, 8, 8, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 256, 128, 256, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 256, 128, 64, 8, 8, 16, 16, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
// blocksize=128
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 128, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 128, 8, 8, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 128, 128, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 32, 256, 8, 8, 16, 16, 1, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
// blocksize=64
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 64, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 64, 64, 8, 8, 16, 16, 2, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 64, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 128, 8, 8, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
// blocksize=32
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 16, 64, 8, 8, 16, 16, 1, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 64, 32, 8, 8, 16, 16, 4, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 32, 32, 8, 8, 16, 16, 2, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F16, Empty_Tuple, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 16, 32, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename DsDatatype,
typename CDEElementOp,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_wmma_i8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 64, 4, 16, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 16, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 16, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 256, 64, 256, 8, 16, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 16, 1, 1, 1, S<1, 32, 1, 8>, 8>,
// blocksize=128
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 256, 8, 16, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 64, 128, 8, 16, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 128, 256, 8, 16, 16, 16, 4, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 32, 256, 8, 16, 16, 16, 1, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 128, 256, 128, 8, 16, 16, 16, 8, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
// blocksize=64
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 128, 8, 16, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 64, 128, 8, 16, 16, 16, 2, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 128, 8, 16, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 64, 32, 64, 8, 16, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
// blocksize=32
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 16, 64, 8, 16, 16, 16, 1, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 64, 64, 8, 16, 16, 16, 4, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 32, 32, 8, 16, 16, 16, 2, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, I32, I8, Empty_Tuple, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, 32, 16, 64, 8, 16, 16, 16, 1, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
View file @ 930b2872
...@@ -18,6 +18,8 @@ namespace instance { ...@@ -18,6 +18,8 @@ namespace instance {
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
using BF8 = ck::bf8_t;
using F8 = ck::f8_t;
using Empty_Tuple = ck::Tuple<>; using Empty_Tuple = ck::Tuple<>;
...@@ -143,6 +145,43 @@ using device_grouped_conv_bwd_data_xdl_f32_instances = ...@@ -143,6 +145,43 @@ using device_grouped_conv_bwd_data_xdl_f32_instances =
// clang-format on // clang-format on
>; >;
// f16_f16_f16_comp_f8
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionBackwardDataSpecialization ConvSpec>
using device_grouped_conv_bwd_data_xdl_input_fp16_comp_bf8f8_instances =
std::tuple<
// clang-format off
// ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer|
// ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock|
// ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, LoopScheduler::Default, BF8, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 32, 1, 4>, 1, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4, LoopScheduler::Default, BF8, F8>,
DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, F32, F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, LoopScheduler::Default, BF8, F8>
// clang-format on
>;
} // namespace instance } // namespace instance
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
View file @ 930b2872
...@@ -19,6 +19,14 @@ using BF16 = ck::bhalf_t; ...@@ -19,6 +19,14 @@ using BF16 = ck::bhalf_t;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using Empty_Tuple = ck::Tuple<>; using Empty_Tuple = ck::Tuple<>;
template <ck::index_t... Is> template <ck::index_t... Is>
...@@ -133,6 +141,43 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple< ...@@ -133,6 +141,43 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec>
using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| Compute| Compute|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| TypeA| TypeB|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
#if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
// generic instance
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>,
// instance for small conv.K
// for fp16 conv.K and conv.C must be divisible by 2
// since half_t atomic_add require scalar_per_x_vector % 2 == 0
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>,
DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>
#endif
// clang-format on
>;
} // namespace instance } // namespace instance
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp 0 → 100755
View file @ 930b2872
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.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;
using I32 = int32_t;
using Empty_Tuple = ck::Tuple<>;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd1x1P0 =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
static constexpr auto ConvFwd1x1S1P0 =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
static constexpr auto ConvFwdOddC =
ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename DsDatatype,
typename CDEElementOp,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_wmma_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| Ds| EData| AccData| CShuffle| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| DataType| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 4, 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>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 64, 256, 4, 8, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 256, 64, 4, 8, 16, 16, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 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>,
// blocksize=128
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 4, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 4, 8, 16, 16, 2, 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>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 8, 8, 16, 16, 2, 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>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 4, 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>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 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>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 32, 256, 4, 8, 16, 16, 1, 8, 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>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 256, 32, 4, 8, 16, 16, 8, 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>,
// blocksize=64
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 64, 4, 8, 16, 16, 1, 4, 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, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 64, 32, 4, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 32, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 128, 4, 8, 16, 16, 1, 8, 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, 32, 1, 2>, 8>,
// blocksize=32
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 16, 64, 4, 8, 16, 16, 1, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 64, 16, 4, 8, 16, 16, 4, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 32, 32, 4, 8, 16, 16, 2, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, F16, F16, DsDatatype, F16, F32, F16, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 16, 16, 4, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
template <index_t NDSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename DsDatatype,
typename CDEElementOp,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_wmma_i8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| Ds| EData| AccData| CShuffle| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| DataType| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | WMMA| WMMA| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
//generic instance
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 4, 16, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 16, 1, 1, 1, S<1, 32, 1, 4>, 1>,
// blocksize=256
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 4, 16, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 64, 256, 4, 16, 16, 16, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 256, 64, 4, 16, 16, 16, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 8, 16, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 8>, 8>,
// blocksize=128
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 4, 16, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 64, 8, 16, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 4, 16, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 64, 128, 8, 16, 16, 16, 2, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 4, 16, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 128, 64, 8, 16, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 32, 256, 4, 16, 16, 16, 1, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 128, 256, 32, 4, 16, 16, 16, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 4>, 8>,
// blocksize=64
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 64, 4, 16, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 64, 32, 4, 16, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 32, 8, 16, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 64, 32, 128, 4, 16, 16, 16, 1, 8, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 32, 1, 2>, 8>,
// blocksize=32
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 16, 64, 4, 16, 16, 16, 1, 4, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 64, 16, 4, 16, 16, 16, 4, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 32, 32, 4, 16, 16, 16, 2, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>,
DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<NDSpatial, ALayout, BLayout, DsLayout, ELayout, I8, I8, DsDatatype, I8, I32, I8, PassThrough, PassThrough, CDEElementOp, ConvSpec, GemmMNKPadding, 32, 16, 16, 4, 16, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 16, 1, 2>, 8>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
View file @ 930b2872
...@@ -13,6 +13,10 @@ namespace tensor_operation { ...@@ -13,6 +13,10 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
...@@ -174,6 +178,42 @@ using device_grouped_conv_fwd_xdl_int8_instances = std::tuple< ...@@ -174,6 +178,42 @@ using device_grouped_conv_fwd_xdl_int8_instances = std::tuple<
// clang-format on // clang-format on
>; >;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_f16_comp_f8_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| ComputeType|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#ifdef CK_ENABLE_FP8
// generic instance
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 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, F8>,
// instances for small conv.K and conv.C
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 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, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>,
DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, F8>
#endif
// clang-format on
>;
} // namespace instance } // namespace instance
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
View file @ 930b2872
...@@ -16,6 +16,7 @@ namespace device { ...@@ -16,6 +16,7 @@ namespace device {
namespace instance { namespace instance {
// conv2d backward data // conv2d backward data
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
...@@ -30,6 +31,35 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances( ...@@ -30,6 +31,35 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
...@@ -43,7 +73,8 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances( ...@@ -43,7 +73,8 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances( void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK, GNHWK,
...@@ -57,7 +88,37 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances( ...@@ -57,7 +88,37 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
...@@ -72,6 +133,36 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances( ...@@ -72,6 +133,36 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
...@@ -85,7 +176,8 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances( ...@@ -85,7 +176,8 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances( void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK, NHWGK,
...@@ -99,8 +191,38 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances( ...@@ -99,8 +191,38 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// conv3d backward data // conv3d backward data
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances( void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK, GNDHWK,
...@@ -115,6 +237,35 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances( ...@@ -115,6 +237,35 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances( void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK, GNDHWK,
...@@ -128,7 +279,8 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances( ...@@ -128,7 +279,8 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances( void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK, GNDHWK,
...@@ -142,7 +294,37 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances( ...@@ -142,7 +294,37 @@ void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances( void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK, NDHWGK,
...@@ -157,6 +339,35 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances( ...@@ -157,6 +339,35 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances( void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK, NDHWGK,
...@@ -170,7 +381,8 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances( ...@@ -170,7 +381,8 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances( void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK, NDHWGK,
...@@ -184,14 +396,62 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances( ...@@ -184,14 +396,62 @@ void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_input_f16_comp_bf8f8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough,
BF8,
F8>>>& instances);
#endif
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename OutLayout, typename OutLayout,
typename WeiLayout, typename WeiLayout,
typename InLayout, typename InLayout,
typename OutDataType, typename OutDataType,
typename WeiDataType, typename WeiDataType,
typename InDataType> typename InDataType,
typename ComputeTypeA,
typename ComputeTypeB>
struct DeviceOperationInstanceFactory< struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD< ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<
NumDimSpatial, NumDimSpatial,
...@@ -205,7 +465,9 @@ struct DeviceOperationInstanceFactory< ...@@ -205,7 +465,9 @@ struct DeviceOperationInstanceFactory<
InDataType, InDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>> ck::tensor_operation::element_wise::PassThrough,
ComputeTypeA,
ComputeTypeB>>
{ {
using DeviceOp = using DeviceOp =
DeviceGroupedConvBwdDataMultipleD<NumDimSpatial, DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
...@@ -219,7 +481,9 @@ struct DeviceOperationInstanceFactory< ...@@ -219,7 +481,9 @@ struct DeviceOperationInstanceFactory<
InDataType, InDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>; ck::tensor_operation::element_wise::PassThrough,
ComputeTypeA,
ComputeTypeB>;
static auto GetInstances() static auto GetInstances()
{ {
...@@ -230,42 +494,80 @@ struct DeviceOperationInstanceFactory< ...@@ -230,42 +494,80 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> && if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutLayout, GNHWK>) is_same_v<OutLayout, GNHWK>)
{ {
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_f16_1x1s1p0_instances(
op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances( add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_gnhwk_gkyxc_gnhwc_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
} }
else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> && else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutLayout, NHWGK>) is_same_v<OutLayout, NHWGK>)
{ {
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_f16_1x1s1p0_instances(
op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16>)
{ {
add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances( add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_instances(op_ptrs);
add_device_grouped_conv2d_bwd_data_wmma_nhwgk_gkyxc_nhwgc_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
} }
} }
else if constexpr(NumDimSpatial == 3) else if constexpr(NumDimSpatial == 3)
...@@ -274,46 +576,96 @@ struct DeviceOperationInstanceFactory< ...@@ -274,46 +576,96 @@ struct DeviceOperationInstanceFactory<
if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> && if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, GNDHWK>) is_same_v<OutLayout, GNDHWK>)
{ {
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_f16_1x1s1p0_instances(
op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances( add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
} }
else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> && else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, NDHWGK>) is_same_v<OutLayout, NDHWGK>)
{ {
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> && if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>) is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, F16> &&
is_same_v<ComputeTypeB, F16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances( add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_1x1s1p0_instances(
op_ptrs);
} }
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
else if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, bf8_t> &&
is_same_v<ComputeTypeB, f8_t>)
{
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_input_f16_comp_bf8f8_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> && else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>) is_same_v<OutDataType, F32>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances( add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> && else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>) is_same_v<OutDataType, BF16>)
{ {
add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances( add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_1x1s1p0_instances(
op_ptrs);
}
#endif
} }
} }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ 930b2872
...@@ -19,6 +19,7 @@ namespace instance { ...@@ -19,6 +19,7 @@ namespace instance {
// xdl // xdl
// conv1d backward weight // conv1d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -30,7 +31,8 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_insta ...@@ -30,7 +31,8 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_insta
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances( void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -42,7 +44,8 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances( ...@@ -42,7 +44,8 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances( void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -54,8 +57,9 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances( ...@@ -54,8 +57,9 @@ void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv2d backward weight // conv2d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -67,7 +71,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_in ...@@ -67,7 +71,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_in
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances( void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -79,7 +84,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -79,7 +84,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances( void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -91,7 +97,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances( ...@@ -91,7 +97,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -103,7 +110,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_in ...@@ -103,7 +110,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_in
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances( void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -115,7 +123,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances( ...@@ -115,7 +123,8 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances( void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -127,8 +136,9 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances( ...@@ -127,8 +136,9 @@ void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv3d backward weight // conv3d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -140,7 +150,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16 ...@@ -140,7 +150,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -152,7 +163,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances ...@@ -152,7 +163,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances( void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -164,7 +176,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances ...@@ -164,7 +176,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -176,7 +189,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16 ...@@ -176,7 +189,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -188,7 +202,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances ...@@ -188,7 +202,8 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances( void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -200,10 +215,27 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances ...@@ -200,10 +215,27 @@ void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_f8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough,
BF8,
F8>>>& instances);
#endif
#ifdef DL_KERNELS #ifdef DL_KERNELS
// dl // dl
// conv1d backward weight // conv1d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -215,7 +247,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instan ...@@ -215,7 +247,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_f32_bf16_instan
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances( void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -227,7 +260,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances( ...@@ -227,7 +260,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances( void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
GNWC, GNWC,
...@@ -239,7 +273,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances( ...@@ -239,7 +273,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances( void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
NWGC, NWGC,
...@@ -251,7 +286,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instan ...@@ -251,7 +286,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instan
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances( void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
NWGC, NWGC,
...@@ -263,7 +299,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances( ...@@ -263,7 +299,8 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances( void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
NWGC, NWGC,
...@@ -275,8 +312,9 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances( ...@@ -275,8 +312,9 @@ void add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv2d backward weight // conv2d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -288,7 +326,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_ins ...@@ -288,7 +326,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_ins
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances( void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -300,7 +339,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -300,7 +339,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances( void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC, GNHWC,
...@@ -312,7 +352,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances( ...@@ -312,7 +352,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -324,7 +365,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_ins ...@@ -324,7 +365,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_ins
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances( void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -336,7 +378,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances( ...@@ -336,7 +378,8 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances( void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC, NHWGC,
...@@ -348,8 +391,9 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances( ...@@ -348,8 +391,9 @@ void add_device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// conv3d backward weight // conv3d backward weight
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -361,7 +405,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_ ...@@ -361,7 +405,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances( void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -373,7 +418,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances( ...@@ -373,7 +418,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances( void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
GNDHWC, GNDHWC,
...@@ -385,7 +431,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances( ...@@ -385,7 +431,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -397,7 +444,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_ ...@@ -397,7 +444,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances( void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -409,7 +457,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances( ...@@ -409,7 +457,8 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances( void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3, std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC, NDHWGC,
...@@ -422,6 +471,7 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances( ...@@ -422,6 +471,7 @@ void add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif #endif
#endif
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename InLayout, typename InLayout,
...@@ -429,7 +479,9 @@ template <ck::index_t NumDimSpatial, ...@@ -429,7 +479,9 @@ template <ck::index_t NumDimSpatial,
typename OutLayout, typename OutLayout,
typename InDataType, typename InDataType,
typename WeiDataType, typename WeiDataType,
typename OutDataType> typename OutDataType,
typename ComputeTypeA,
typename ComputeTypeB>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvBwdWeight< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvBwdWeight<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
...@@ -440,7 +492,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -440,7 +492,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
OutDataType, OutDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>> ck::tensor_operation::element_wise::PassThrough,
ComputeTypeA,
ComputeTypeB>>
{ {
using DeviceOp = DeviceGroupedConvBwdWeight<NumDimSpatial, using DeviceOp = DeviceGroupedConvBwdWeight<NumDimSpatial,
InLayout, InLayout,
...@@ -451,7 +505,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -451,7 +505,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
OutDataType, OutDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>; ck::tensor_operation::element_wise::PassThrough,
ComputeTypeA,
ComputeTypeB>;
static auto GetInstances() static auto GetInstances()
{ {
...@@ -462,6 +518,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -462,6 +518,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> && if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> &&
is_same_v<OutLayout, GNWK>) is_same_v<OutLayout, GNWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
...@@ -470,6 +527,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -470,6 +527,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
...@@ -478,6 +537,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -478,6 +537,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
#endif #endif
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -489,21 +550,27 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -489,21 +550,27 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
} }
else if constexpr(is_same_v<InLayout, NWGC> && is_same_v<WeiLayout, GKXC> && else if constexpr(is_same_v<InLayout, NWGC> && is_same_v<WeiLayout, GKXC> &&
is_same_v<OutLayout, NWGK>) is_same_v<OutLayout, NWGK>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(op_ptrs); add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -511,6 +578,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -511,6 +578,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances( add_device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#endif #endif
} }
} }
...@@ -519,6 +587,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -519,6 +587,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> && if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutLayout, GNHWK>) is_same_v<OutLayout, GNHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
...@@ -529,6 +598,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -529,6 +598,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances( add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
...@@ -539,6 +610,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -539,6 +610,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances( add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -550,10 +623,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -550,10 +623,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
} }
else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> && else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
is_same_v<OutLayout, NHWGK>) is_same_v<OutLayout, NHWGK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
...@@ -564,6 +639,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -564,6 +639,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances( add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
...@@ -574,6 +651,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -574,6 +651,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances( add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -585,6 +664,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -585,6 +664,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
} }
} }
else if constexpr(NumDimSpatial == 3) else if constexpr(NumDimSpatial == 3)
...@@ -592,6 +672,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -592,6 +672,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> && if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, GNDHWK>) is_same_v<OutLayout, GNDHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
...@@ -602,6 +683,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -602,6 +683,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances( add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t>)
{ {
...@@ -612,6 +695,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -612,6 +695,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances( add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -623,10 +708,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -623,10 +708,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
} }
else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> && else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, NDHWGK>) is_same_v<OutLayout, NDHWGK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
...@@ -637,8 +724,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -637,8 +724,12 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances( add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>) is_same_v<OutDataType, half_t> &&
is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{ {
#ifdef DL_KERNELS #ifdef DL_KERNELS
add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances( add_device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
...@@ -647,6 +738,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -647,6 +738,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs); op_ptrs);
} }
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, ck::bhalf_t>) is_same_v<OutDataType, ck::bhalf_t>)
...@@ -658,6 +751,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -658,6 +751,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
op_ptrs); op_ptrs);
} }
#endif
#if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> &&
is_same_v<ComputeTypeA, bf8_t> && is_same_v<ComputeTypeB, f8_t>)
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_f8_instances(
op_ptrs);
}
#endif
} }
} }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ 930b2872
...@@ -17,6 +17,7 @@ namespace tensor_operation { ...@@ -17,6 +17,7 @@ namespace tensor_operation {
namespace device { namespace device {
namespace instance { namespace instance {
#ifdef CK_ENABLE_BF16
// grouped conv1d forward, GNWC/GKXC/GNWK // grouped conv1d forward, GNWC/GKXC/GNWK
void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances( void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
...@@ -31,7 +32,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances( ...@@ -31,7 +32,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances( void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
GNWC, GNWC,
...@@ -45,7 +48,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances( ...@@ -45,7 +48,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances( void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
GNWC, GNWC,
...@@ -59,7 +64,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances( ...@@ -59,7 +64,9 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances( void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
GNWC, GNWC,
...@@ -73,7 +80,73 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances( ...@@ -73,7 +80,73 @@ void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv2d forward, GNHWC/GKYXC/GNHWK // grouped conv2d forward, GNHWC/GKYXC/GNHWK
void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
...@@ -88,13 +161,436 @@ void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances( ...@@ -88,13 +161,436 @@ void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances( void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC, GNHWC,
GKYXC, GKYXC,
Empty_Tuple, Empty_Tuple,
GNHWK, GNHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<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_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC,
GKZYXC,
Empty_Tuple,
GNDHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<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_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16, F16,
F16, F16,
Empty_Tuple, Empty_Tuple,
...@@ -102,27 +598,31 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -102,27 +598,31 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances( #ifdef CK_ENABLE_FP16
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instances(
GNHWC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GKYXC, NDHWGC,
GKZYXC,
Empty_Tuple, Empty_Tuple,
GNHWK, NDHWGK,
F32, F16,
F32, F16,
Empty_Tuple, Empty_Tuple,
F32, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances( #ifdef CK_ENABLE_FP16
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instances(
GNHWC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GKYXC, NDHWGC,
GKZYXC,
Empty_Tuple, Empty_Tuple,
GNHWK, NDHWGK,
F16, F16,
F16, F16,
Empty_Tuple, Empty_Tuple,
...@@ -130,21 +630,25 @@ void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -130,21 +630,25 @@ void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances( #ifdef CK_ENABLE_FP16
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances(
GNHWC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GKYXC, NDHWGC,
GKZYXC,
Empty_Tuple, Empty_Tuple,
GNHWK, NDHWGK,
F32, F16,
F32, F16,
Empty_Tuple, Empty_Tuple,
F32, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances( void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC, GNHWC,
...@@ -158,12 +662,15 @@ void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -158,12 +662,15 @@ void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances( #endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC, GNHWC,
GKYXC, GKYXC,
Empty_Tuple, Empty_Tuple,
NHWGK, GNHWK,
F16, F16,
F16, F16,
Empty_Tuple, Empty_Tuple,
...@@ -171,69 +678,64 @@ void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances( ...@@ -171,69 +678,64 @@ void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances( #ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC, GNHWC,
GKYXC, GKYXC,
Empty_Tuple, Empty_Tuple,
GNHWK, GNHWK,
int8_t, F16,
int8_t, F16,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple, Empty_Tuple,
F32, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// grouped conv2d forward, NHWGC/GKYXC/NHWGK #ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances( void add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC, GNHWC,
GKYXC, GKYXC,
Empty_Tuple, Empty_Tuple,
NHWGK, GNHWK,
BF16, F16,
BF16, F16,
Empty_Tuple, Empty_Tuple,
BF16, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances( #ifdef CK_ENABLE_FP8
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_f8_instances(
NHWGC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GKYXC, NDHWGC,
GKZYXC,
Empty_Tuple, Empty_Tuple,
NHWGK, NDHWGK,
F16, F16,
F16, F16,
Empty_Tuple, Empty_Tuple,
F16, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough,
F8>>>& instances);
#endif
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances( #ifdef CK_ENABLE_FP32
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2, void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
NHWGC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GKYXC, NDHWGC,
GKZYXC,
Empty_Tuple, Empty_Tuple,
NHWGK, NDHWGK,
F32, F32,
F32, F32,
Empty_Tuple, Empty_Tuple,
...@@ -241,56 +743,57 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances( ...@@ -241,56 +743,57 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK #ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances( void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC, NDHWGC,
GKZYXC, GKZYXC,
Empty_Tuple, Empty_Tuple,
GNDHWK, NDHWGK,
BF16, int8_t,
BF16, int8_t,
Empty_Tuple, Empty_Tuple,
BF16, int8_t,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances( void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC, NDHWGC,
GKZYXC, GKZYXC,
Empty_Tuple, Empty_Tuple,
GNDHWK, NDHWGK,
F16, int8_t,
F16, int8_t,
Empty_Tuple, Empty_Tuple,
F16, int8_t,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances( void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC, NDHWGC,
GKZYXC, GKZYXC,
Empty_Tuple, Empty_Tuple,
GNDHWK, NDHWGK,
F32, int8_t,
F32, int8_t,
Empty_Tuple, Empty_Tuple,
F32, int8_t,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances( void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
GNDHWC, NDHWGC,
GKZYXC, GKZYXC,
Empty_Tuple, Empty_Tuple,
GNDHWK, NDHWGK,
int8_t, int8_t,
int8_t, int8_t,
Empty_Tuple, Empty_Tuple,
...@@ -299,27 +802,45 @@ void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances( ...@@ -299,27 +802,45 @@ void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK void add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances(
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
NDHWGC, NDHWGC,
GKZYXC, GKZYXC,
Empty_Tuple, Empty_Tuple,
NDHWGK, NDHWGK,
BF16, int8_t,
BF16, int8_t,
Empty_Tuple, Empty_Tuple,
BF16, int8_t,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( #if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
NDHWGC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GKZYXC, NHWGC,
GKYXC,
Empty_Tuple, Empty_Tuple,
NDHWGK, NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16, F16,
F16, F16,
Empty_Tuple, Empty_Tuple,
...@@ -327,34 +848,39 @@ void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances( ...@@ -327,34 +848,39 @@ void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances( #if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
NDHWGC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GKZYXC, GNHWC,
GKYXC,
Empty_Tuple, Empty_Tuple,
NDHWGK, GNHWK,
F32, F16,
F32, F16,
Empty_Tuple, Empty_Tuple,
F32, F16,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances( #if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3, void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
NDHWGC, std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GKZYXC, GNHWC,
GKYXC,
Empty_Tuple, Empty_Tuple,
NDHWGK, GNHWK,
int8_t, F32,
int8_t, F32,
Empty_Tuple, Empty_Tuple,
int8_t, F32,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
#endif
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename InLayout, typename InLayout,
...@@ -362,7 +888,8 @@ template <ck::index_t NumDimSpatial, ...@@ -362,7 +888,8 @@ template <ck::index_t NumDimSpatial,
typename OutLayout, typename OutLayout,
typename InDataType, typename InDataType,
typename WeiDataType, typename WeiDataType,
typename OutDataType> typename OutDataType,
typename ComputeType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD< struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
NumDimSpatial, NumDimSpatial,
InLayout, InLayout,
...@@ -375,7 +902,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -375,7 +902,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
OutDataType, OutDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>> ck::tensor_operation::element_wise::PassThrough,
ComputeType>>
{ {
using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial, using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout, InLayout,
...@@ -388,7 +916,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -388,7 +916,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
OutDataType, OutDataType,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>; ck::tensor_operation::element_wise::PassThrough,
ComputeType>;
static auto GetInstances() static auto GetInstances()
{ {
...@@ -397,127 +926,245 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -397,127 +926,245 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> && if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>) is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #endif
is_same_v<OutDataType, half_t>) #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #endif
is_same_v<OutDataType, int8_t>) #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs);
} }
#endif
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>) if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
}
#endif
#if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #endif
is_same_v<OutDataType, half_t>)
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instances(op_ptrs);
}
#endif
#if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>) #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>)
{ {
add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs); add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #endif
is_same_v<OutDataType, int8_t>)
#ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{ {
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances(op_ptrs); add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instances(op_ptrs);
} }
#endif
} }
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>) if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
}
#endif
#if(defined(CK_ENABLE_FP32) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #endif
is_same_v<OutDataType, half_t>)
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
}
#endif
#if(defined(CK_ENABLE_FP16) && defined(DL_KERNELS))
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{
add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>) #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>)
{ {
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
} }
#endif
#ifdef CK_ENABLE_INT8
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instances(op_ptrs);
add_device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instances(op_ptrs);
}
#endif
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>) if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #endif
is_same_v<OutDataType, half_t>) #ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #endif
is_same_v<OutDataType, int8_t>) #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instances(op_ptrs);
} }
#endif
} }
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>) if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
{ {
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> && if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>) is_same_v<OutDataType, float>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> && #endif
is_same_v<OutDataType, half_t>)
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<ComputeType, ck::f8_t>)
{
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_f8_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<ComputeType, half_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, ck::bhalf_t> && #endif
is_same_v<WeiDataType, ck::bhalf_t> && #ifdef CK_ENABLE_BF16
is_same_v<OutDataType, ck::bhalf_t>) if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> && #endif
is_same_v<OutDataType, int8_t>) #ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{ {
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instances(op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instances(op_ptrs);
} }
#endif
} }
return op_ptrs; return op_ptrs;
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
View file @ 930b2872
...@@ -2,13 +2,20 @@ ...@@ -2,13 +2,20 @@
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp"
...@@ -18,39 +25,10 @@ ...@@ -18,39 +25,10 @@
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp"
...@@ -60,17 +38,38 @@ ...@@ -60,17 +38,38 @@
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp"
......
library/include/ck/library/utility/host_tensor_generator.hpp
View file @ 930b2872
...@@ -95,7 +95,7 @@ struct GeneratorTensor_2<int8_t> ...@@ -95,7 +95,7 @@ struct GeneratorTensor_2<int8_t>
} }
}; };
#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8 #if defined CK_ENABLE_FP8
template <> template <>
struct GeneratorTensor_2<ck::f8_t> struct GeneratorTensor_2<ck::f8_t>
{ {
...@@ -111,6 +111,22 @@ struct GeneratorTensor_2<ck::f8_t> ...@@ -111,6 +111,22 @@ struct GeneratorTensor_2<ck::f8_t>
}; };
#endif #endif
#if defined CK_ENABLE_BF8
template <>
struct GeneratorTensor_2<ck::bf8_t>
{
int min_value = 0;
int max_value = 1;
template <typename... Is>
ck::bf8_t operator()(Is...)
{
float tmp = (std::rand() % (max_value - min_value)) + min_value;
return ck::type_convert<ck::bf8_t>(tmp);
}
};
#endif
template <typename T> template <typename T>
struct GeneratorTensor_3 struct GeneratorTensor_3
{ {
...@@ -143,7 +159,7 @@ struct GeneratorTensor_3<ck::bhalf_t> ...@@ -143,7 +159,7 @@ struct GeneratorTensor_3<ck::bhalf_t>
} }
}; };
#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8 #if defined CK_ENABLE_FP8
template <> template <>
struct GeneratorTensor_3<ck::f8_t> struct GeneratorTensor_3<ck::f8_t>
{ {
...@@ -162,6 +178,25 @@ struct GeneratorTensor_3<ck::f8_t> ...@@ -162,6 +178,25 @@ struct GeneratorTensor_3<ck::f8_t>
}; };
#endif #endif
#if defined CK_ENABLE_BF8
template <>
struct GeneratorTensor_3<ck::bf8_t>
{
float min_value = 0;
float max_value = 1;
template <typename... Is>
ck::bf8_t operator()(Is...)
{
float tmp = float(std::rand()) / float(RAND_MAX);
float fp32_tmp = min_value + tmp * (max_value - min_value);
return ck::type_convert<ck::bf8_t>(fp32_tmp);
}
};
#endif
template <typename T> template <typename T>
struct GeneratorTensor_4 struct GeneratorTensor_4
{ {
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 930b2872
function(add_instance_library INSTANCE_NAME) function(add_instance_library INSTANCE_NAME)
message("adding instance ${INSTANCE_NAME}") message("adding instance ${INSTANCE_NAME}")
add_library(${INSTANCE_NAME} OBJECT ${ARGN}) set(result 1)
target_compile_features(${INSTANCE_NAME} PUBLIC) if(DEFINED DTYPES)
set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON) foreach(source IN LISTS ARGN)
clang_tidy_check(${INSTANCE_NAME}) set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
#make an exception for reduction kernels
if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}" OR "${source}" MATCHES "device_reduce_instance")
#if filename matches any selected type, exit type loop and do no exclude the file from the list
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
endif()
endforeach()
if(test EQUAL 1)
message("removing instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
endif()
foreach(source IN LISTS ARGN)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(ARGN)
add_library(${INSTANCE_NAME} OBJECT ${ARGN})
target_compile_features(${INSTANCE_NAME} PUBLIC)
set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(${INSTANCE_NAME})
set(result 0)
endif()
#message("add_instance_library returns ${result}")
set(result ${result} PARENT_SCOPE)
endfunction(add_instance_library INSTANCE_NAME) endfunction(add_instance_library INSTANCE_NAME)
...@@ -15,33 +63,49 @@ IF(IS_DIRECTORY "${subdir_path}") ...@@ -15,33 +63,49 @@ IF(IS_DIRECTORY "${subdir_path}")
set(cmake_instance) set(cmake_instance)
file(READ "${subdir_path}/CMakeLists.txt" cmake_instance) file(READ "${subdir_path}/CMakeLists.txt" cmake_instance)
set(add_inst 0) set(add_inst 0)
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8") if(("${cmake_instance}" MATCHES "_fp8" OR "${cmake_instance}" MATCHES "_f8") AND DTYPES MATCHES "fp8")
#message("fp8 instance found!") message("fp8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16") if(("${cmake_instance}" MATCHES "_fp16" OR "${cmake_instance}" MATCHES "_f16") AND DTYPES MATCHES "fp16")
#message("fp16 instance found!") message("fp16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32") if(("${cmake_instance}" MATCHES "_fp32" OR "${cmake_instance}" MATCHES "_f32") AND DTYPES MATCHES "fp32")
#message("fp32 instance found!") message("fp32 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64") if(("${cmake_instance}" MATCHES "_fp64" OR "${cmake_instance}" MATCHES "_f64") AND DTYPES MATCHES "fp64")
#message("fp64 instance found!") message("fp64 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16") if("${cmake_instance}" MATCHES "_bf16" AND DTYPES MATCHES "bf16")
#message("bf16 instance found!") message("bf16 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8") if(("${cmake_instance}" MATCHES "_int8" OR "${cmake_instance}" MATCHES "_i8") AND DTYPES MATCHES "int8")
#message("int8 instance found!") message("int8 instance found!")
set(add_inst 1) set(add_inst 1)
endif() endif()
if(NOT "${cmake_instance}" MATCHES "DTYPES" OR NOT DEFINED DTYPES) if(NOT "${cmake_instance}" MATCHES "_fp8" OR
#message("instance should be built for all types!") NOT "${cmake_instance}" MATCHES "_f8" OR
set(add_inst 1) NOT "${cmake_instance}" MATCHES "_fp16" OR
NOT "${cmake_instance}" MATCHES "_f16" OR
NOT "${cmake_instance}" MATCHES "_fp32" OR
NOT "${cmake_instance}" MATCHES "_f32" OR
NOT "${cmake_instance}" MATCHES "_fp64" OR
NOT "${cmake_instance}" MATCHES "_f64" OR
NOT "${cmake_instance}" MATCHES "_bf16" OR
NOT "${cmake_instance}" MATCHES "_int8" OR
NOT "${cmake_instance}" MATCHES "_i8" OR
NOT "${cmake_instance}" MATCHES "_int4" OR
NOT DEFINED DTYPES)
message("instance should be built for all types!")
set(add_inst 1)
endif()
if("${cmake_instance}" MATCHES "quantization" AND DEFINED DTYPES AND NOT DTYPES MATCHES "int8")
message("quantization instances will not be built!")
set(add_inst 0)
endif() endif()
if("${cmake_instance}" MATCHES "ONLY DL_KERNELS" AND NOT DEFINED DL_KERNELS) if("${cmake_instance}" MATCHES "ONLY DL_KERNELS" AND NOT DEFINED DL_KERNELS)
message("Found only dl instances, but DL_KERNELS is not set. Skipping.") message("Found only dl instances, but DL_KERNELS is not set. Skipping.")
......
library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
View file @ 930b2872
set(DEVICE_AVGPOOL_BWD_INSTANCES) set(DEVICE_AVGPOOL_BWD_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f16_instance.cpp) device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
endif() device_avg_pool3d_bwd_ndhwc_f32_instance.cpp)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp)
endif()
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f32_instance.cpp)
endif()
add_instance_library(device_avg_pool3d_bwd_instance ${DEVICE_AVGPOOL_BWD_INSTANCES}) add_instance_library(device_avg_pool3d_bwd_instance ${DEVICE_AVGPOOL_BWD_INSTANCES})
library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
View file @ 930b2872
set(BATCHED_GEMM_INSTANCES) set(BATCHED_GEMM_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp) device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
endif() device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp) device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
endif() device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp) device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
endif() device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp) device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp)
endif()
add_instance_library(device_batched_gemm_instance ${BATCHED_GEMM_INSTANCES}) add_instance_library(device_batched_gemm_instance ${BATCHED_GEMM_INSTANCES})
library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_instance_library(device_batched_gemm_add_relu_gemm_add_instance add_instance_library(device_batched_gemm_add_relu_gemm_add_instance
device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
) )
endif()
\ No newline at end of file
library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_instance_library(device_batched_gemm_bias_permute_instance add_instance_library(device_batched_gemm_bias_permute_instance
device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
) )
endif()
library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_instance_library(device_batched_gemm_gemm_instance add_instance_library(device_batched_gemm_gemm_instance
device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
) )
endif()
library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
View file @ 930b2872
# ONLY DL_KERNELS # ONLY DL_KERNELS
if(DL_KERNELS) set(BATCHED_GEMM_MULTID_INSTANCES)
set(BATCHED_GEMM_MULTID_INSTANCES) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp)
endif() list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp) list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp)
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp) add_instance_library(device_batched_gemm_multi_d_instance ${BATCHED_GEMM_MULTID_INSTANCES})
list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp)
endif()
add_instance_library(device_batched_gemm_multi_d_instance ${BATCHED_GEMM_MULTID_INSTANCES})
endif()
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_instance_library(device_batched_gemm_reduce_instance add_instance_library(device_batched_gemm_reduce_instance
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
) )
endif()
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
View file @ 930b2872
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_instance_library(device_batched_gemm_softmax_gemm_instance add_instance_library(device_batched_gemm_softmax_gemm_instance
device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
) )
endif()
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
View file @ 930b2872
set(DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES) set(DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES
list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp) device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp) device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
endif() device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES) device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp)
list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp)
list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp)
endif()
add_instance_library(device_batched_gemm_softmax_gemm_permute_instance ${DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES}) add_instance_library(device_batched_gemm_softmax_gemm_permute_instance ${DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES})
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment