Skip to content

GitLab

Commit 2724c519 authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents 1fb4a474 2eb74a9c
Hide whitespace changes
Inline Side-by-side
Showing with 2086 additions and 39 deletions
example/62_convnd_activ/unary/convnd_fwd_xdl_sigmoid_fp16.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::Sigmoid;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_softrelu_fp16.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::SoftRelu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_tanh_fp16.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::TanH;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/63_layernorm4d_fwd/CMakeLists.txt 0 → 100644
View file @ 2724c519
add_example_executable(example_layernorm4d_fwd_fp16 layernorm4d_fwd_fp16.cpp)
add_example_executable(example_layernorm4d_fwd_splitk_fp16 layernorm4d_fwd_splitk_fp16.cpp)
example/63_layernorm4d_fwd/common.hpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationFwdImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm4d_fwd_example.inc"
int main() { return run_layernorm4d_fwd_example<DeviceInstance>(); }
example/63_layernorm4d_fwd/layernorm4d_fwd_splitk_fp16.cpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
using DeviceInstance = ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl<
XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // XScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm4d_fwd_example.inc"
int main() { return run_layernorm4d_fwd_example<DeviceInstance>(); }
example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename DeviceInstance>
int run_layernorm4d_fwd_example()
{
bool time_kernel = false;
ck::index_t N = 256;
ck::index_t H = 16;
ck::index_t W = 16;
ck::index_t C = 8;
Tensor<XDataType> x({N, H, W, C});
Tensor<GammaDataType> gamma({H, W, C});
Tensor<BetaDataType> beta({H, W, C});
Tensor<YDataType> y({N, H, W, C});
Tensor<SaveMeanInvStdDataType> save_mean({N});
Tensor<SaveMeanInvStdDataType> save_inv_std({N});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
#ifdef SAVE_MEAN_INV_STD
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
#endif
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
auto device_instance = DeviceInstance{};
auto argument_ptr = device_instance.MakeArgumentPointer(
{N, H, W, C},
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
{0, W * C, C, 1},
{0, W * C, C, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
{1, 2, 3},
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters are not supported" << std::endl;
return 1;
};
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
{
Tensor<YDataType> host_y({N, H, W, C});
Tensor<SaveMeanInvStdDataType> host_save_mean({N});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({N});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
{N, H, W, C},
{1, 2, 3},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results (y)", 1e-3, 1e-3);
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.FromDevice(save_mean.mData.data());
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(
save_mean, host_save_mean, "Error: Incorrect results (mean)", 1e-3, 1e-3);
pass &= ck::utils::check_err(
save_inv_std, host_save_inv_std, "Error: Incorrect results (inv_std)", 1e-3, 1e-3);
#endif
}
return (pass ? 0 : 1);
}
example/CMakeLists.txt
View file @ 2724c519
......@@ -7,20 +7,112 @@ add_custom_target(examples)
function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
add_dependencies(examples ${EXAMPLE_NAME})
add_dependencies(check ${EXAMPLE_NAME})
rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS FILE_NAME)
set(test 0)
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing example source file ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
endif()
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
add_dependencies(examples ${EXAMPLE_NAME})
add_dependencies(check ${EXAMPLE_NAME})
rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
set(result 0)
endif()
#message("add_example returns ${result}")
set(result ${result} PARENT_SCOPE)
endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_dependencies EXAMPLE_NAME FILE_NAME)
if(result EQUAL 0)
add_dependencies(${EXAMPLE_NAME} ${FILE_NAME})
endif()
endfunction(add_example_dependencies EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_dependencies(examples ${EXAMPLE_NAME})
rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
set(result 1)
if(DEFINED DTYPES)
foreach(source IN LISTS FILE_NAME)
set(test 0)
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
endif()
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
message("removing dl example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
endif()
endforeach()
#only continue if there are some source files left on the list
if(FILE_NAME)
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
add_dependencies(examples ${EXAMPLE_NAME})
rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
set(result 0)
endif()
#message("add_example returns ${result}")
set(result ${result} PARENT_SCOPE)
endfunction(add_example_executable_no_testing EXAMPLE_NAME)
# add all example subdir
......
include/ck/ck.hpp
View file @ 2724c519
......@@ -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"
......@@ -27,15 +29,45 @@
#define CK_WAVELET_MIN_BLOCK_PER_CU 2
#endif
// kernel attribute: amdgpu_waves_per_eu()
#ifdef CK_USE_WAVES_PER_EU
// for 1-wave kernels, control arguments of amdgpu_waves_per_eu() attribute
#ifndef CK_MIN_WAVES_PER_EU
#define CK_MIN_WAVES_PER_EU 0
#endif
#ifndef CK_MAX_WAVES_PER_EU
#define CK_MAX_WAVES_PER_EU 0
#endif
#else
#define CK_USE_WAVES_PER_EU 0
#endif
// define general macros for various architectures
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#define __gfx94__
#endif
#if defined(__gfx1010__) || defined(__gfx1011__) || defined(__gfx1012__)
#define __gfx101__
#endif
#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__)
#define __gfx103__
#endif
#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
#define __gfx11__
#endif
// buffer resource
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx940__) // for GPU code
defined(__gfx90a__) || defined(__gfx94__)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx1030__) // for GPU code
#elif defined(__gfx103__)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
#elif defined(__gfx11__)
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
#endif
......@@ -43,32 +75,36 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code, define nothing
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
#define CK_USE_AMD_V_MAC_F32
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
defined(__gfx940__) // for GPU code
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx103__) || \
defined(__gfx94__) // for GPU code
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
#elif defined(__gfx11__)
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8_GFX11
#endif
// MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
#define CK_USE_AMD_MFMA
#endif
#if(defined(__gfx90a__) || defined(__gfx940__))
#if(defined(__gfx90a__) || defined(__gfx94__))
#define CK_USE_AMD_MFMA_BF16_1K_OP
#endif
#if defined(__gfx940__)
#if defined(__gfx94__)
#define CK_USE_AMD_MFMA_GFX940
#endif
// WMMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_WMMA
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
#elif defined(__gfx11__) // for GPU code
#define CK_USE_AMD_WMMA
#endif
......@@ -85,13 +121,13 @@
// buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
#endif
#if(defined(__gfx90a__) || defined(__gfx940__)) // for GPU code
#if(defined(__gfx90a__) || defined(__gfx94__)) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
#else
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
......@@ -100,8 +136,21 @@
// inline asm
#define CK_USE_AMD_INLINE_ASM 1
// inner product (DLOP)
#define CK_USE_AMD_INNER_PRODUCT_INLINE_ASM 1
// inner product (V_MAC/V_FMAC)
#define CK_USE_AMD_V_MAC_INLINE_ASM 1
// V_DOT inline instructions, less efficient since they require adding
// `s_nop`s to avoid hazard
#define CK_USE_AMD_V_DOT_INLINE_ASM 0
// inner product using V_DOT with DPP8 modifiers
#define CK_USE_AMD_V_DOT_DPP8_INLINE_ASM 1
// LDS direct loads using inline assembly
#define CK_USE_AMD_LDS_DIRECT_LOAD_INLINE_ASM 1
// set stochastic rounding as default for f8 conversions
#define CK_USE_SR_F8_CONVERSION 1
// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
#define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
......@@ -145,6 +194,10 @@
#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
// experimental feature: add instances using pipeline v2
#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
// experimental feature: optimize pipeline v2 by IGLP strategy (value=ID of strategy)
#ifndef CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
#define CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT 0
#endif
// hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
......@@ -170,13 +223,14 @@
// workaround: compiler issue on gfx908
#define CK_WORKAROUND_SWDEV_388832 1
// flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0
// denorm test fix, required to work around dissue
#ifndef CK_WORKAROUND_DENORM_FIX
#define CK_WORKAROUND_DENORM_FIX 0
#elif
#else
// enable only on MI200
#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#endif // CK_WORKAROUND_DENORM_FIX
......
include/ck/config.h.in 0 → 100644
View file @ 2724c519
/*******************************************************************************
*
* 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_BF8
#define CK_ENABLE_BF8 "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_BF8
#cmakedefine CK_ENABLE_BF8 @CK_ENABLE_BF8@
#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/host_utility/device_prop.hpp
View file @ 2724c519
......@@ -26,7 +26,7 @@ inline std::string get_device_name()
}
const std::string raw_name(props.gcnArchName);
// https://github.com/ROCmSoftwarePlatform/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
// https://github.com/ROCm/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
static std::map<std::string, std::string> device_name_map = {
{"Ellesmere", "gfx803"},
{"Baffin", "gfx803"},
......@@ -51,4 +51,37 @@ inline std::string get_device_name()
return name;
}
inline bool is_xdl_supported()
{
return ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942";
}
inline bool is_lds_direct_load_supported()
{
// Check if direct loads from global memory to LDS are supported.
return ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940" ||
ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942";
}
inline bool is_navi1_supported()
{
return ck::get_device_name() == "gfx1010" || ck::get_device_name() == "gfx1011" ||
ck::get_device_name() == "gfx1012";
}
inline bool is_navi2_supported()
{
return ck::get_device_name() == "gfx1030" || ck::get_device_name() == "gfx1031" ||
ck::get_device_name() == "gfx1032" || ck::get_device_name() == "gfx1034" ||
ck::get_device_name() == "gfx1035" || ck::get_device_name() == "gfx1036";
}
inline bool is_navi3_supported()
{
return ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102" || ck::get_device_name() == "gfx1103";
}
} // namespace ck
include/ck/host_utility/hip_check_error.hpp
View file @ 2724c519
......@@ -3,15 +3,32 @@
#pragma once
#include <sstream>
#include <hip/hip_runtime.h>
// To be removed, which really does not tell the location of failed HIP functional call
inline void hip_check_error(hipError_t x)
{
if(x != hipSuccess)
{
std::ostringstream ss;
ss << "HIP runtime error: " << hipGetErrorString(x) << ". " << __FILE__ << ": " << __LINE__
<< "in function: " << __func__;
ss << "HIP runtime error: " << hipGetErrorString(x) << ". "
<< "hip_check_error.hpp"
<< ": " << __LINE__ << "in function: " << __func__;
throw std::runtime_error(ss.str());
}
}
#define HIP_CHECK_ERROR(retval_or_funcall) \
do \
{ \
hipError_t _tmpVal = retval_or_funcall; \
if(_tmpVal != hipSuccess) \
{ \
std::ostringstream ostr; \
ostr << "HIP Function Failed (" \
<< "hip_check_error.hpp" \
<< "," << __LINE__ << ") " << hipGetErrorString(_tmpVal); \
throw std::runtime_error(ostr.str()); \
} \
} while(0)
include/ck/host_utility/kernel_launch.hpp
View file @ 2724c519
......@@ -30,15 +30,90 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
block_dim.y,
block_dim.z);
printf("Warm up 1 time\n");
printf("Warm up %d times\n", stream_config.cold_niters_);
#endif
const int nrepeat = 50;
// warm up
for(int i = 0; i < stream_config.cold_niters_; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
const int nrepeat = stream_config.nrepeat_;
#if DEBUG_LOG
printf("Start running %d times...\n", nrepeat);
#endif
hipEvent_t start, stop;
hip_check_error(hipEventCreate(&start));
hip_check_error(hipEventCreate(&stop));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(start, stream_config.stream_id_));
for(int i = 0; i < nrepeat; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
hip_check_error(hipEventSynchronize(stop));
float total_time = 0;
hip_check_error(hipEventElapsedTime(&total_time, start, stop));
return total_time / nrepeat;
}
else
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
}
#else
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
#endif
}
template <typename... Args, typename F, typename PreProcessFunc>
float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
PreProcessFunc preprocess,
F kernel,
dim3 grid_dim,
dim3 block_dim,
std::size_t lds_byte,
Args... args)
{
#if CK_TIME_KERNEL
if(stream_config.time_kernel_)
{
#if DEBUG_LOG
printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
__func__,
grid_dim.x,
grid_dim.y,
grid_dim.z,
block_dim.x,
block_dim.y,
block_dim.z);
printf("Warm up %d times\n", stream_config.cold_niters_);
#endif
// warm up
preprocess();
for(int i = 0; i < stream_config.cold_niters_; ++i)
{
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
const int nrepeat = stream_config.nrepeat_;
#if DEBUG_LOG
printf("Start running %d times...\n", nrepeat);
#endif
......@@ -52,7 +127,9 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
for(int i = 0; i < nrepeat; ++i)
{
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
}
hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
......@@ -66,12 +143,15 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
}
else
{
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
}
#else
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
hip_check_error(hipGetLastError());
return 0;
#endif
......
include/ck/host_utility/stream_utility.hpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <hip/hip_runtime.h>
#include "ck/stream_config.hpp"
#include "ck/host_utility/hip_check_error.hpp"
static inline int getAvailableComputeUnitCount(const StreamConfig& stream_config)
{
constexpr int MAX_MASK_DWORDS = 64;
// assume at most 64*32 = 2048 CUs
uint32_t cuMask[MAX_MASK_DWORDS];
for(int i = 0; i < MAX_MASK_DWORDS; i++)
cuMask[i] = 0;
auto countSetBits = [](uint32_t dword) {
int count = 0;
while(dword != 0)
{
if(dword & 0x1)
count++;
dword = dword >> 1;
};
return (count);
};
hip_check_error(hipExtStreamGetCUMask(stream_config.stream_id_, MAX_MASK_DWORDS, &cuMask[0]));
int ret = 0;
for(int i = 0; i < MAX_MASK_DWORDS; i++)
ret += countSetBits(cuMask[i]);
return (ret);
};
include/ck/stream_config.hpp
View file @ 2724c519
......@@ -11,4 +11,6 @@ struct StreamConfig
hipStream_t stream_id_ = nullptr;
bool time_kernel_ = false;
int log_level_ = 0;
int cold_niters_ = 5;
int nrepeat_ = 50;
};
include/ck/tensor_description/multi_index_transform.hpp
View file @ 2724c519
......@@ -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_gemm_dl_v2r3.hpp
View file @ 2724c519
......@@ -11,7 +11,7 @@
namespace ck {
// C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
// A and B are visable to the whole block, C is distributed among each thread
// A and B are visible to the whole block, C is distributed among each thread
// Assume:
// 1. A:
// 1. ABlockDesc_BK0_BM_BK1 is known at compile-time
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/warp/dpp_gemm.hpp"
namespace ck {
/**
* Blockwise GEMM that uses DPP instruction modifier to limit the amount of data loaded for each
* thread by sharing the data between threads in a lanegroup.
*
* In every iteration, each wave calculates a C tile of size `MPerDpp` * `NPerDpp`, there are
* `MRepeat` iterations for `M` dimension and `NRepeat` for `N` one.
* In total, the algorithm runs using
* `MPerBlock / (MRepeat * MPerDpp) * NPerBlock / (NRepeat * NPerDpp)` waves.
*/
template <index_t BlockSize,
typename ABDataType,
typename AccDataType,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
index_t MPerDpp,
index_t NPerDpp,
index_t MRepeat,
index_t NRepeat,
index_t KPack>
struct BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
static constexpr index_t WaveSize = get_warp_size();
static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
static constexpr index_t KPerBlock =
BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
static constexpr auto dpp_gemm = DppGemm<ABDataType, MPerDpp, NPerDpp, KPack>{};
static constexpr index_t KPerThread = KPerBlock / dpp_gemm.K0PerDpp;
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerDpp);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerDpp);
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
AccDataType,
MRepeat * NRepeat,
dpp_gemm.GetRegSizePerDpp(),
true>
c_thread_buf_;
__host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
__device__ static auto GetWaveIdx()
{
const index_t thread_id = ThisThreadBlock::GetThreadId();
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
__device__ static auto CalculateAThreadOriginDataIndex_M0_M1_M2_K()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto dpp_a_idx = dpp_gemm.CalculateAThreadOriginDataIndex_K_M();
const auto dpp_a_idx_k = dpp_a_idx[I0];
const auto dpp_a_idx_m = dpp_a_idx[I1];
return make_tuple(0, waveId_m, dpp_a_idx_m, KPerThread * dpp_a_idx_k);
}
__device__ static auto CalculateBThreadOriginDataIndex_N0_N1_N2_K()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_n = wave_idx[I1];
const auto dpp_b_idx = dpp_gemm.CalculateBThreadOriginDataIndex_K_N();
const auto dpp_b_idx_k = dpp_b_idx[I0];
const auto dpp_b_idx_n = dpp_b_idx[I1];
return make_tuple(0, waveId_n, dpp_b_idx_n, KPerThread * dpp_b_idx_k);
}
template <index_t m0, index_t n0>
__device__ static auto CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = dpp_gemm.GetBeginOfThreadBlk();
const auto blk_m_offset = blk_idx[I0];
const auto blk_n_offset = blk_idx[I1];
constexpr auto mrepeat_mwave_MPerDpp_to_m_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerDpp))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
constexpr auto nrepeat_nwave_NPerDpp_to_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerDpp))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
const index_t c_thread_m = mrepeat_mwave_MPerDpp_to_m_adaptor.CalculateBottomIndex(
make_tuple(m0, waveId_m, blk_m_offset))[I0];
const index_t c_thread_n = nrepeat_nwave_NPerDpp_to_n_adaptor.CalculateBottomIndex(
make_tuple(n0, waveId_n, blk_n_offset))[I0];
return make_tuple(c_thread_m, c_thread_n);
}
__host__ __device__ BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2()
{
static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
BK0NK1BlockDesc::IsKnownAtCompileTime(),
"Wrong! Block descriptors should be known at the time of compilation.");
#if defined(__HIP_DEVICE_COMPILE__)
// Host wave size can be different than the device one and this assert could fail for host,
// but it does matter only for device.
static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
"ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
#endif
static_assert(MPerBlock % (MPerDpp * MRepeat) == 0,
"Invalid parameters. MPerBlock must be divisible by MPerDpp * MRepeat.");
static_assert(NPerBlock % (NPerDpp * NRepeat) == 0,
"Invalid parameters. NPerBlock must be divisible by NPerDpp * NRepeat.");
}
__host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2()
{
constexpr auto c_m_n_tblk_lens = dpp_gemm.GetCMNThreadBlkLengths();
constexpr auto M = c_m_n_tblk_lens[I0];
constexpr auto N = c_m_n_tblk_lens[I1];
return make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M, N));
}
__host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_N2()
{
constexpr auto c_m_n_tblk_lens = dpp_gemm.GetCMNThreadBlkLengths();
constexpr auto M = c_m_n_tblk_lens[I0];
constexpr auto N = c_m_n_tblk_lens[I1];
return make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M, N));
}
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2()
{
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerDpp>{},
Number<NPerDpp>{}));
return c_block_desc_m0_n0_m1_n1_m2_n2;
}
__host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_N2()
{
constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(I1,
Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerDpp>{},
Number<NPerDpp>{}));
return c_block_desc_g_m0_n0_m1_n1_m2_n2;
}
template <typename CGridDesc_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2(const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerDpp), MWaves, MPerDpp)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerDpp), NWaves, NPerDpp))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
return c_grid_desc_m0_n0_m1_n1_m2_n2;
}
template <typename CGridDesc_G_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_G_M0_N0_M1_N1_M2_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
{
const auto G = c_grid_desc_g_m_n.GetLength(I0);
const auto M = c_grid_desc_g_m_n.GetLength(I1);
const auto N = c_grid_desc_g_m_n.GetLength(I2);
const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_grid_desc_g_m_n,
make_tuple(make_pass_through_transform(G),
make_unmerge_transform(make_tuple(M / (MWaves * MPerDpp), MWaves, MPerDpp)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerDpp), NWaves, NPerDpp))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
return c_grid_desc_g_m0_n0_m1_n1_m2_n2;
}
__host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
{
return transform_tensor_descriptor(
AK0MK1BlockDesc{},
make_tuple(
make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
make_unmerge_transform(
make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerDpp>{}))),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
}
__host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
{
return transform_tensor_descriptor(
BK0NK1BlockDesc{},
make_tuple(
make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
make_unmerge_transform(
make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerDpp>{}))),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
}
static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
__device__ void Run(const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ABDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ABDataType>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
static_for<0, KPerThread, KPack>{}([&](auto k) {
vector_type<ABDataType, KPack> a_thread_vec;
vector_type<ABDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<ABDataType>()(i) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
b_thread_vec.template AsType<ABDataType>()(i) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
});
using dpp_input_type =
typename vector_type<ABDataType, dpp_gemm.K1PerDpp>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
dpp_gemm.template Run(a_thread_vec.template AsType<dpp_input_type>(),
b_thread_vec.template AsType<dpp_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
protected:
// A[M0, M1, M2, KPerThread]
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
// B[N0, N1, N2, KPerThread]
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
// C[M, N, NumRegDpp]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, dpp_gemm.GetRegSizePerDpp()));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ABDataType,
ABDataType,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerThread>,
Sequence<0, 1, 2, 3>,
3,
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<ABDataType,
ABDataType,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerThread>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex_M0_M1_M2_K()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex_N0_N1_N2_K()};
};
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp 0 → 100644
View file @ 2724c519
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/utility/loop_scheduler.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
// Double LDS buffer
// Prefetech 2 stage
// Local prefetch 1 stage
namespace ck {
template <index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t ABufferLoadWidth,
index_t BBufferLoadWidth,
index_t ALDSWriteWidth,
index_t BLDSWriteWidth,
index_t ALDSReadWidth,
index_t BLDSReadWidth,
index_t MRepeat,
index_t NRepeat,
index_t MPerXDL,
index_t NPerXDL,
index_t KPerXDL>
struct BlockwiseGemmXdlops_pipeline_hotloop_inst
{
static constexpr index_t WaveSize = 64;
static constexpr index_t WaveNumM = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t WaveNumN = NPerBlock / (NRepeat * NPerXDL);
static constexpr index_t A_Buffer_Load_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * ABufferLoadWidth);
static constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * BBufferLoadWidth);
static constexpr index_t A_LDS_Write_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * ALDSWriteWidth);
static constexpr index_t B_LDS_Write_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * BLDSWriteWidth);
static constexpr index_t A_LDS_Read_Inst_Num =
WaveNumN * MPerBlock * KPerBlock / (BlockSize * ALDSReadWidth);
static constexpr index_t B_LDS_Read_Inst_Num =
WaveNumM * MPerBlock * KPerBlock / (BlockSize * BLDSReadWidth);
static constexpr index_t C_MFMA_Inst_Num =
MPerBlock * NPerBlock * KPerBlock / (BlockSize / WaveSize) / (MPerXDL * NPerXDL * KPerXDL);
static constexpr auto Print()
{
printf(" Blk/Wave Size: %d, %d, M/N/K PerBlk: %d, %d, %d, M/N/K PerXdl: %d, %d, %d\n",
BlockSize,
WaveSize,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
KPerXDL);
printf(" A/B buffer load inst: %d, %d\n A/B LDS write inst: %d, %d\n A/B LDS read inst: "
"%d, %d\n C MFMA inst: %d\n",
A_Buffer_Load_Inst_Num,
B_Buffer_Load_Inst_Num,
A_LDS_Write_Inst_Num,
B_LDS_Write_Inst_Num,
A_LDS_Read_Inst_Num,
B_LDS_Read_Inst_Num,
C_MFMA_Inst_Num);
}
};
template <
index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
bool TransposeC = false,
index_t AMmaKStride =
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops,
index_t BMmaKStride =
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{}.K0PerXdlops>
struct BlockwiseGemmXdlops_pipeline_v4
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
static constexpr index_t WaveSize = get_warp_size();
static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm =
XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
static constexpr index_t KRepeat = KPerThread / KPack;
static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
using HotLoopInstList = BlockwiseGemmXdlops_pipeline_hotloop_inst<BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
A_K1,
B_K1,
A_K1,
B_K1,
KPack,
KPack,
MRepeat,
NRepeat,
MPerXDL,
NPerXDL,
xdlops_gemm.KPerXdlops>;
static_assert(KPerThread % KPack == 0,
"Wrong KPack setting; try increasing KPerThread or decreasing KPack");
StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
FloatAcc,
MRepeat * NRepeat,
xdlops_gemm.GetRegSizePerXdlops(),
true>
c_thread_buf_;
__host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
__device__ static auto GetWaveIdx()
{
const index_t thread_id = ThisThreadBlock::GetThreadId();
constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
}
__device__ static auto CalculateAThreadOriginDataIndex()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
}
__device__ static auto CalculateBThreadOriginDataIndex()
{
const auto wave_idx = GetWaveIdx();
const auto waveId_n = wave_idx[I1];
const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
}
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
__device__ static auto
CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1, 2>{}));
const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
return make_tuple(c_thread_m, c_thread_n);
}
template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
__device__ static auto
CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
{
const auto wave_idx = GetWaveIdx();
const auto waveId_m = wave_idx[I0];
const auto waveId_n = wave_idx[I1];
const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
return make_tuple(
m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
}
using Tuple4 = decltype(CalculateAThreadOriginDataIndex());
__host__ __device__
BlockwiseGemmXdlops_pipeline_v4(Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
Tuple4 b_origin = CalculateBThreadOriginDataIndex())
: a_thread_copy_(a_origin), b_thread_copy_(b_origin)
{
static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
"ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
"wrong!");
// HotLoopInstList::Print();
}
// transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
__host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
{
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
return make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
}
// XDL output supporting C_xdl = A_xdl * B_xdl
__host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
return make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
}
__host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
constexpr auto N = c_m0_m1_m2_n_tblk_lens[I3];
return make_naive_tensor_descriptor_packed(
make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
}
// transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
{
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerXDL>{},
Number<NPerXDL>{}));
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
}
// XDL output supporting C_xdl = A_xdl * B_xdl
__host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerXDL>{},
Number<NPerXDL>{}));
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
}
__host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
{
constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
make_naive_tensor_descriptor_packed(make_tuple(I1,
Number<MRepeat>{},
Number<NRepeat>{},
Number<MWaves>{},
Number<NWaves>{},
Number<MPerXDL>{},
Number<NPerXDL>{}));
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
c_block_desc_g_m0_n0_m1_n1_m2_n2);
}
template <typename CGridDesc_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
const auto N = c_grid_desc_m_n.GetLength(I1);
const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
}
template <typename CGridDesc_G_M_N>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
{
const auto G = c_grid_desc_g_m_n.GetLength(I0);
const auto M = c_grid_desc_g_m_n.GetLength(I1);
const auto N = c_grid_desc_g_m_n.GetLength(I2);
const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
c_grid_desc_g_m_n,
make_tuple(make_pass_through_transform(G),
make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
c_grid_desc_g_m0_n0_m1_n1_m2_n2);
}
__device__ static constexpr auto HotLoopScheduler()
{
// schedule
constexpr auto num_ds_read_inst =
HotLoopInstList::A_LDS_Read_Inst_Num + HotLoopInstList::B_LDS_Read_Inst_Num;
constexpr auto num_ds_write_inst =
HotLoopInstList::A_LDS_Write_Inst_Num + HotLoopInstList::B_LDS_Write_Inst_Num;
;
constexpr auto num_buffer_load_inst =
HotLoopInstList::A_Buffer_Load_Inst_Num + HotLoopInstList::B_Buffer_Load_Inst_Num;
;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto num_issue = num_buffer_load_inst;
static_for<0, num_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(
0x100, num_ds_read_inst / num_buffer_load_inst, 0); // DS read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(
0x200, num_ds_write_inst / num_buffer_load_inst, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_inst / num_buffer_load_inst - 3, 0); // MFMA
});
}
template <index_t stage>
__device__ static constexpr auto TailScheduler()
{
}
template <>
__device__ static constexpr auto TailScheduler<1>()
{
// schedule
constexpr auto num_ds_read_inst =
HotLoopInstList::A_LDS_Read_Inst_Num + HotLoopInstList::B_LDS_Read_Inst_Num;
constexpr auto num_ds_write_inst =
HotLoopInstList::A_LDS_Write_Inst_Num + HotLoopInstList::B_LDS_Write_Inst_Num;
;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto num_issue = num_ds_write_inst;
static_for<0, num_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(
0x100, num_ds_read_inst / num_ds_write_inst - 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_inst / num_ds_write_inst - 3, 0); // MFMA
});
}
template <>
__device__ static constexpr auto TailScheduler<2>()
{
// schedule
constexpr auto num_ds_read_inst =
HotLoopInstList::A_LDS_Read_Inst_Num + HotLoopInstList::B_LDS_Read_Inst_Num;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto num_issue = num_ds_read_inst;
static_for<0, num_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_inst / num_ds_read_inst, 0); // MFMA
});
}
static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
template <bool HasMainLoop,
index_t TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
index_t num_loop) const
{
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
// Inst List:
// ds_read_b128: 16
// ds_write_b128: 8
// buffer_load_dwordx4: 16
// v_mfma: 0
// -------------------------------------------------------------------------------------------
// Global prefetch 1th, Fill Ping LDS
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
// Local prefetch 1th, Fill Ping Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(I0));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(I0),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(I0));
});
});
});
// Global prefetch 2th, Fill Pong LDS
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
// Global prefetch 3rd
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
// This hot loop has two legacy loopover, to implement the double local buffer strategy
do
{
// -------------------------------------------------------------------------------------------
using PingP1 = Number<0>;
using PongP1 = Number<1>;
// MFMA: Ping Reg
// DS_WRITE: To Ping LDS
// DS_READ: Pong LDS to Pong Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(PongP1{}),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(PongP1{}));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(PongP1{}),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(PongP1{}));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP1{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP1{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
// -------------------------------------------------------------------------------------------
using PingP2 = Number<1>;
using PongP2 = Number<0>;
// MFMA: Pong Reg
// DS_WRITE: To Pong LDS
// DS_READ: Ping LDS to Ping Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(PongP2{}),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(PongP2{}));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(PongP2{}),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(PongP2{}));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP2{}));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP2{}));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP2{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP2{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 2;
} while(i < (num_loop - 3));
}
// tail
if constexpr(TailNum == 3)
{
using PingP1 = Number<0>;
using PongP1 = Number<1>;
// MFMA: Ping Reg
// DS_WRITE: To Ping LDS
// DS_READ: Pong LDS to Pong Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(PongP1{}),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(PongP1{}));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(PongP1{}),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(PongP1{}));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(PingP1{}));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(PingP1{}));
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP1{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP1{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
TailScheduler<1>();
__builtin_amdgcn_sched_barrier(0);
// -------------------------------------------------------------------------------------------
using PingP2 = Number<1>;
using PongP2 = Number<0>;
// MFMA: Pong Reg
// DS_WRITE: To Pong LDS
// DS_READ: Ping LDS to Ping Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(PongP2{}),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(PongP2{}));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(PongP2{}),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(PongP2{}));
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP2{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP2{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
TailScheduler<2>();
__builtin_amdgcn_sched_barrier(0);
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PongP2{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PongP2{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
// 64 v_mfma
__builtin_amdgcn_sched_group_barrier(0x008, 64, 0); // MFMA
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(TailNum == 2)
{
using PingP1 = Number<0>;
using PongP1 = Number<1>;
// MFMA: Ping Reg
// DS_WRITE: To Ping LDS
// DS_READ: Pong LDS to Pong Reg
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(PongP1{}),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(PongP1{}));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(PongP1{}),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(PongP1{}));
});
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP1{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP1{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
TailScheduler<2>();
__builtin_amdgcn_sched_barrier(0);
// -------------------------------------------------------------------------------------------
using PingP2 = Number<1>;
// MFMA: Pong Reg
// DS_WRITE: To Pong LDS
// DS_READ: Ping LDS to Ping Reg
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<FloatAB>()(ik) =
a_thread_bufs[PingP2{}][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<FloatAB>()(ik) =
b_thread_bufs[PingP2{}][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
// 64 v_mfma
__builtin_amdgcn_sched_group_barrier(0x008, 64, 0); // MFMA
__builtin_amdgcn_sched_barrier(0);
}
}
protected:
// M1, N1 as double buffer index
// Read buffer + Compute buffer
// A[M0, M1, M2, KPack]
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
make_tuple(
Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
// B[N0, N1, N2, KPack]
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}),
make_tuple(
Number<KPack>{}, Number<KPack * MRepeat * KPack>{}, Number<MRepeat * KPack>{}, I1));
// C[M, N, NumRegXdlops]
static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
};
} // namespace ck
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