Skip to content

GitLab

Commit 14822d71 authored by Jing Zhang's avatar Jing Zhang
Browse files

merge

parents 5b02dfaf 80560ef2
Hide whitespace changes
Inline Side-by-side
Showing with 297 additions and 98 deletions
example/01_gemm/gemm_xdl_bf16_rtn.cpp 0 → 100644
View file @ 14822d71
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/utility/type_convert.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = float;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::ConvertBF16RTN;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
example/15_grouped_gemm/CMakeLists.txt
View file @ 14822d71
......@@ -9,16 +9,13 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
add_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp16
example_grouped_gemm_multiple_d_dl_fp16
example_grouped_gemm_xdl_splitk_fp16
example_grouped_gemm_xdl_fixed_nk_fp16
example_grouped_gemm_xdl_fixed_nk_fp8
example_grouped_gemm_xdl_fixed_nk_bias_fp16
)
example_grouped_gemm_xdl_fixed_nk_bias_fp16)
endif()
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
......@@ -28,6 +25,10 @@ if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
endif()
if(DTYPES MATCHES "f8" OR NOT DEFINED DTYPES)
add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
......
example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
View file @ 14822d71
......@@ -30,7 +30,7 @@ using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddBias = ck::tensor_operation::element_wise::AddBias;
using Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
......@@ -49,7 +49,7 @@ using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddBias;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MPadding;
......
example/35_splitK_gemm/CMakeLists.txt
View file @ 14822d71
......@@ -3,15 +3,15 @@ set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_splitK_gemm_xdl)
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
endif()
if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bfp16)
endif()
......
example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
View file @ 14822d71
......@@ -33,6 +33,7 @@ using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CDataType = F32;
using ComputeType = BF16;
using ALayout = Row;
using BLayout = Col;
......@@ -46,11 +47,11 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 4>;
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 4, ComputeType>;
// clang-format on
#include "run_splitK_gemm_example.inc"
......
example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
View file @ 14822d71
......@@ -30,6 +30,7 @@ using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CDataType = int32_t;
using ComputeType = int8_t;
using ALayout = Row;
using BLayout = Col;
......@@ -43,11 +44,11 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 16, 16, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 16, 16, true, 1, 1, S<1, 32, 1, 8>, 4>;
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type|
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 16, 16, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 16, 16, true, 1, 1, S<1, 32, 1, 8>, 4, ComputeType>;
// clang-format on
#include "run_splitK_gemm_example.inc"
......
include/ck/ck.hpp
View file @ 14822d71
......@@ -3,6 +3,8 @@
#pragma once
#include "ck/config.h"
#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
#include "hip/hip_runtime.h"
#include "hip/hip_fp16.h"
......@@ -200,9 +202,6 @@
// workaround: compiler issue on gfx908
#define CK_WORKAROUND_SWDEV_388832 1
// workaround: Grouped Conv2d_bwd_data fails for already implemented instance
#define CK_WORKAROUND_GITHUB_ISSUE_824 1
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0
......
include/ck/config.h.in 0 → 100644
View file @ 14822d71
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2023 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#ifndef CK_CONFIG_H_IN
#define CK_CONFIG_H_IN
// clang-format off
//
// DataType supports in the current CK build
//
#ifndef DTYPES
#cmakedefine DTYPES "@DTYPES@"
#endif
// if DTYPES is not defined, enable all datatypes in headerfiles
#ifndef CK_ENABLE_ALL_DTYPES
#cmakedefine CK_ENABLE_ALL_DTYPES @CK_ENABLE_ALL_DTYPES@
#if defined(CK_ENABLE_ALL_DTYPES)
#ifndef CK_ENABLE_INT8
#define CK_ENABLE_INT8 "ON"
#endif
#ifndef CK_ENABLE_FP8
#define CK_ENABLE_FP8 "ON"
#endif
#ifndef CK_ENABLE_FP16
#define CK_ENABLE_FP16 "ON"
#endif
#ifndef CK_ENABLE_BF16
#define CK_ENABLE_BF16 "ON"
#endif
#ifndef CK_ENABLE_FP32
#define CK_ENABLE_FP32 "ON"
#endif
#ifndef CK_ENABLE_FP64
#define CK_ENABLE_FP64 "ON"
#endif
#endif
#endif
// if DTYPES are selectively enabled
#ifndef CK_ENABLE_INT8
#cmakedefine CK_ENABLE_INT8 @CK_ENABLE_INT8@
#endif
#ifndef CK_ENABLE_FP8
#cmakedefine CK_ENABLE_FP8 @CK_ENABLE_FP8@
#endif
#ifndef CK_ENABLE_FP16
#cmakedefine CK_ENABLE_FP16 @CK_ENABLE_FP16@
#endif
#ifndef CK_ENABLE_BF16
#cmakedefine CK_ENABLE_BF16 @CK_ENABLE_BF16@
#endif
#ifndef CK_ENABLE_FP32
#cmakedefine CK_ENABLE_FP32 @CK_ENABLE_FP32@
#endif
#ifndef CK_ENABLE_FP64
#cmakedefine CK_ENABLE_FP64 @CK_ENABLE_FP64@
#endif
//
// Legacy DL kernel supports in the current CK build
// by default DL kernels are turned OFF
//
#ifndef CK_ENABLE_DL_KERNELS
#cmakedefine CK_ENABLE_DL_KERNELS @CK_ENABLE_DL_KERNELS@
#endif
//
// Instances supports in the current CK build
//
#ifndef CK_ENABLE_INSTANCES_ONLY
#cmakedefine CK_ENABLE_INSTANCES_ONLY @CK_ENABLE_INSTANCES_ONLY@
#endif
// clang-format on
#endif // CK_CONFIG_H_IN
include/ck/tensor_description/multi_index_transform.hpp
View file @ 14822d71
......@@ -1042,13 +1042,13 @@ struct Merge_v2_magic_division
using UpLengths =
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
using LowLengthsMagicDivisorMultipiler = decltype(
generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengths>{},
Number<NDimLow>{}));
using LowLengthsMagicDivisorMultipiler = decltype(generate_tuple(
lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengths>{},
Number<NDimLow>{}));
using LowLengthsMagicDivisorShift = decltype(
generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengths>{},
Number<NDimLow>{}));
using LowLengthsMagicDivisorShift = decltype(generate_tuple(
lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengths>{},
Number<NDimLow>{}));
LowLengths low_lengths_;
LowLengthsMagicDivisorMultipiler low_lengths_magic_divisor_multiplier_;
......@@ -1201,9 +1201,9 @@ struct Merge_v2r2_magic_division
lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengthsScan>{},
Number<NDimLow>{}));
using LowLengthsScanMagicDivisorShift = decltype(
generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengthsScan>{},
Number<NDimLow>{}));
using LowLengthsScanMagicDivisorShift = decltype(generate_tuple(
lambda_merge_generate_MagicDivision_calculate_magic_shift<LowLengthsScan>{},
Number<NDimLow>{}));
LowLengths low_lengths_;
LowLengthsScan low_lengths_scan_;
......
include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
View file @ 14822d71
......@@ -35,8 +35,8 @@ struct BlockwiseSoftmax
static constexpr index_t MRepeat = ThreadSliceDesc_M_K{}.GetLength(I0);
static constexpr index_t KRepeat = ThreadSliceDesc_M_K{}.GetLength(I1);
using ThreadSliceDesc_M = decltype(
make_naive_tensor_descriptor_packed(make_tuple(ThreadSliceDesc_M_K{}.GetLength(I0))));
using ThreadSliceDesc_M = decltype(make_naive_tensor_descriptor_packed(
make_tuple(ThreadSliceDesc_M_K{}.GetLength(I0))));
using ThreadwiseMaxReduce = typename conditional<
IgnoreNaN,
......
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -543,9 +543,13 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
EGridDesc_G_M_N e_grid_desc_g_m_n_;
};
using ComputeDataType = ADataType;
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
ADataType,
BDataType,
ComputeDataType,
AccDataType,
CShuffleDataType,
DsDataType,
......@@ -588,14 +592,18 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
LoopSched>;
// desc for blockwise copy
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}))>;
// block-to-e-tile map
using Block2ETileMap =
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
View file @ 14822d71
......@@ -331,8 +331,13 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
};
using ComputeDataType = ADataType;
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
ADataType,
BDataType,
ComputeDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<>, // DsDataType,
......@@ -378,13 +383,16 @@ struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
CDEBlockTransferScalarPerVector_NPerBlock,
LoopSched>;
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}));
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}));
using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
// Argument
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
View file @ 14822d71
......@@ -324,8 +324,12 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
index_t BatchStrideE_;
};
using ComputeDataType = ADataType;
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
BDataType,
ComputeDataType,
AccDataType,
CShuffleDataType,
DsDataType,
......@@ -368,14 +372,18 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
LoopSched>;
// desc for blockwise copy
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}))>;
// block-to-e-tile map
using Block2ETileMap =
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -510,12 +510,15 @@ struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched>;
using A0GridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(A0GridDesc_M_K{}))>;
using B0GridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(B0GridDesc_N_K{}))>;
using B1GridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(B1GridDesc_N_K{}))>;
using A0GridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(
A0GridDesc_M_K{}))>;
using B0GridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(
B0GridDesc_N_K{}))>;
using B1GridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(
B1GridDesc_N_K{}))>;
// Argument
struct Argument : public BaseArgument
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -310,9 +310,13 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N({}, {}));
using ComputeDataType = ADataType;
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
BDataType,
ComputeDataType,
AccDataType,
CShuffleDataType,
DsDataType,
......@@ -355,14 +359,18 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
LoopSched>;
// desc for blockwise copy
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}))>;
// block-to-e-tile map
using Block2ETileMap =
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ 14822d71
......@@ -532,11 +532,12 @@ struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_
float ave_time = 0;
const auto Run = [&](const auto& kernel) {
hipGetErrorString(hipMemset(
hipGetErrorString(hipMemsetAsync(
arg.p_c_grid_,
0,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
sizeof(CDataType)));
sizeof(CDataType),
stream_config.stream_id_));
ave_time =
launch_and_time_kernel(stream_config,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -364,11 +364,13 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
// We have to separate mean var descriptor for gemm and layernorm bacause of different grid
// layout(different padding)
using GemmMeanVarGridDesc_M_NBlock = decltype(
MakeMeanVarDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1, 1));
using GemmMeanVarGridDesc_M_NBlock =
decltype(MakeMeanVarDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1,
1));
using GemmCountGridDesc_M_NBlock = decltype(
MakeCountDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1, 1));
using GemmCountGridDesc_M_NBlock =
decltype(MakeCountDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1,
1));
using LayernormMeanVarGridDesc_M_NBlock =
decltype(MakeMeanVarDescriptor_M_N<Sequence<true, true>,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -337,10 +337,12 @@ struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
RThreadTransferDstScalarPerVector_MPerBlock,
LoopSched>;
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
View file @ 14822d71
......@@ -20,7 +20,8 @@
namespace ck {
template <typename GridwiseGemm,
typename ABDataType,
typename ADataType,
typename BDataType,
typename DsPointer,
typename EDataType,
typename AElementwiseOperation,
......@@ -36,8 +37,8 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_multiple_d_xdl_cshuffle(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
kernel_gemm_multiple_d_xdl_cshuffle(const ADataType* __restrict__ p_a_grid,
const BDataType* __restrict__ p_b_grid,
DsPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
const AElementwiseOperation a_element_op,
......@@ -242,9 +243,13 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
using ComputeDataType = EDataType;
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
BDataType,
ComputeDataType,
AccDataType,
CShuffleDataType,
DsDataType,
......@@ -288,14 +293,18 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
PipelineVer>;
// desc for blockwise copy
using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 = remove_cvref_t<decltype(
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using AGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
AGridDesc_M_K{}))>;
using BGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
BGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
DsGridDesc_M_N{}))>;
using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}))>;
// block-to-e-tile map
using Block2ETileMap =
......@@ -438,6 +447,7 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
const auto kernel = kernel_gemm_multiple_d_xdl_cshuffle<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
BDataType, // TODO: distiguish A/B datatype
typename GridwiseGemm::DsGridPointer,
EDataType,
AElementwiseOperation,
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ 14822d71
......@@ -57,7 +57,10 @@ template <typename ADataType,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
typename ComputeType = CDataType,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
......@@ -76,11 +79,11 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
// TODO: should be exposed as Tparams.
static constexpr index_t NumGemmKPrefetchStage = 1;
static constexpr LoopScheduler LoopSched = make_default_loop_scheduler();
static constexpr PipelineVersion PipelineVer = PipelineVersion::v1;
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
......@@ -120,7 +123,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CBlockTransferScalarPerVector_NWaveNPerXDL,
CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVer>;
PipelineVer,
ComputeType>;
using Argument = typename GridwiseGemm::Argument;
using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
......@@ -158,8 +162,10 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
const auto Run = [&](const auto& kernel) {
if(kbatch > 1)
hipGetErrorString(
hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType)));
hipGetErrorString(hipMemsetAsync(karg.p_c_grid,
0,
karg.M * karg.N * sizeof(CDataType),
stream_config.stream_id_));
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
......
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