Skip to content

GitLab

Commit e5ebcc41 authored by Artur Wojcik's avatar Artur Wojcik
Browse files

Merge branch 'develop' into uif2-migraphx

parents 57cdd70b abac8b07
Hide whitespace changes
Inline Side-by-side
Showing with 2585 additions and 525 deletions
example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp 0 → 100644
View file @ e5ebcc41
// 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 @ e5ebcc41
// 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 @ e5ebcc41
// 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 @ e5ebcc41
......@@ -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 @ e5ebcc41
......@@ -45,16 +45,30 @@
#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__) || defined(__gfx941__) || \
defined(__gfx942__) // 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
......@@ -62,33 +76,36 @@
#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__) || defined(__gfx941__) || defined(__gfx942__) // 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__) || defined(__gfx941__) || \
defined(__gfx942__) // for GPU code
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) // for GPU code
#define CK_USE_AMD_MFMA
#endif
#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
#if(defined(__gfx90a__) || defined(__gfx94__))
#define CK_USE_AMD_MFMA_BF16_1K_OP
#endif
#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#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
......@@ -104,15 +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__) || defined(__gfx941__) || \
defined(__gfx942__) // 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__) || defined(__gfx941__) || \
defined(__gfx942__)) // 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
......@@ -131,6 +146,12 @@
// 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
......@@ -209,7 +230,7 @@
// 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
View file @ e5ebcc41
......@@ -43,6 +43,9 @@
#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
......@@ -66,6 +69,10 @@
#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
......
include/ck/host_utility/device_prop.hpp
View file @ e5ebcc41
......@@ -27,7 +27,7 @@ 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"},
......@@ -59,5 +59,31 @@ inline bool is_xdl_supported()
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
#endif
include/ck/host_utility/hip_check_error.hpp
View file @ e5ebcc41
......@@ -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 @ e5ebcc41
......@@ -30,12 +30,16 @@ 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
// warm up
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
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 = 10;
const int nrepeat = stream_config.nrepeat_;
#if DEBUG_LOG
printf("Start running %d times...\n", nrepeat);
#endif
......@@ -50,6 +54,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
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_));
......@@ -64,11 +69,13 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
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
......@@ -96,13 +103,17 @@ float launch_and_time_kernel_with_preprocess(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
// warm up
preprocess();
kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
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 = 10;
const int nrepeat = stream_config.nrepeat_;
#if DEBUG_LOG
printf("Start running %d times...\n", nrepeat);
#endif
......@@ -118,6 +129,7 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
{
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_));
......@@ -133,11 +145,13 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
{
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/stream_config.hpp
View file @ e5ebcc41
......@@ -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_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp deleted 100644 → 0
View file @ 57cdd70b
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/amd_gemm_dpp.hpp"
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_adaptor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp"
namespace ck {
/**
* DPP8 version of blockwise GEMM algorithm. It uses DPP8 instruction modifier to limit
* the data loaded from LDS to registers.
*
* The algorithm groups threads into groups of size `dpp8::lane_group_size` and splits the matrix C
* between them in such a way that threads from the same group need the same chunk of either
* matrix A (or B, respectively). Without the usage of DPP8, each thread would need to load the
* whole chunk from LDS to its own register space.
* Usage of DPP8 modifiers allow each thread to load less data, exactly `1 / dpp8::lane_group_size`
* of the chunk, and then share that data with other threads from the same lane group.
*
* Assumptions coming from the usage of DPP8:
* 1. `BM10BN10ThreadClusterBM10Xs[1] == dpp8::lane_group_size` or
* `BM10BN10ThreadClusterBN10Xs[1] == dpp8::lane_group_size` -
* - it makes consecutive `dpp8::lane_group_size` threads use the same chunk of either
* matrix A or B;
* - based on these values we determine which matrix to share.
* 2. `BM1PerThreadBM11 % dpp8::lane_group_size == 0` (if sharing A) or
* `BN1PerThreadBN11 % dpp8::lane_group_size == 0` (if sharing B) -
* - we have to make sure that the data to split is divisible by the number of
* threads in the group.
*
* General algorithm:
* C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
* 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
* 2. ABlockBuffer is DynamicBuffer
* 2. B:
* 1. BBlockDesc_BK0_BN_BK1 is known at compile-time
* 2. BBlockBuffer is DynamicBuffer
* 3. C:
* 1. CThreadDesc_BM0_BM11_BN0_BN11 is known at compile-time
* 2. CThreadBuffer is StaticBuffer
* 4. BM10BN10ThreadClusterBM10Xs::Size() = BM10BN10ThreadClusterBN10Xs::Size() == 2
*/
template <index_t BlockSize,
typename FloatA,
typename FloatB,
typename FloatC,
typename ABlockDesc_BK0_BM_BK1,
typename BBlockDesc_BK0_BN_BK1,
index_t BM1PerThreadBM11,
index_t BN1PerThreadBN11,
index_t BK0PerThread,
typename BM10BN10ThreadClusterBM10Xs, // Sequence<BM10BN10ThreadClusterBM100,
// BM10BN10ThreadClusterBM101, ...>
typename BM10BN10ThreadClusterBN10Xs, // Sequence<BM10BN10ThreadClusterBN100,
// BM10BN10ThreadClusterBN101, ...>
index_t AThreadCopyScalarPerVector_BM11,
index_t BThreadCopyScalarPerVector_BN11,
typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
bool>::type = false>
struct BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0
{
using AIndex = MultiIndex<4>;
using BIndex = MultiIndex<4>;
using CIndex = MultiIndex<4>;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr index_t BK0 = ABlockDesc_BK0_BM_BK1{}.GetLength(I0);
static constexpr index_t BK1 = ABlockDesc_BK0_BM_BK1{}.GetLength(I2);
static constexpr index_t BM = ABlockDesc_BK0_BM_BK1{}.GetLength(I1);
static constexpr index_t BN = BBlockDesc_BK0_BN_BK1{}.GetLength(I1);
static constexpr index_t BM100 = BM10BN10ThreadClusterBM10Xs{}[I0];
static constexpr index_t BN100 = BM10BN10ThreadClusterBN10Xs{}[I0];
static constexpr index_t BM101 = BM10BN10ThreadClusterBM10Xs{}[I1];
static constexpr index_t BN101 = BM10BN10ThreadClusterBN10Xs{}[I1];
static constexpr index_t BM11 = BM1PerThreadBM11;
static constexpr index_t BN11 = BN1PerThreadBN11;
static constexpr index_t BM1 = BM100 * BM101 * BM11;
static constexpr index_t BN1 = BN100 * BN101 * BN11;
static constexpr index_t BM0 = BM / BM1;
static constexpr index_t BN0 = BN / BN1;
// We assume that either `BM101` or `BN101` is equal to `dpp8::lane_group_size`. It makes all
// threads in a lane group need the same chunk of B or A matrices and we can share them using
// DPP.
static_assert(BM101 == dpp8::lane_group_size || BN101 == dpp8::lane_group_size);
static constexpr bool ShareB = BM101 == dpp8::lane_group_size ? true : false;
static constexpr bool ShareA = !ShareB;
// If DPP shares A (B, respectively), lane group gets `BM1PerThreadBM11` (`BN1PerThreadBN11`,
// respectively) elements, so we split them between threads in lane group so each thread loads
// less data from LDS.
static constexpr index_t BM1PerThread =
ShareA ? BM1PerThreadBM11 / dpp8::lane_group_size : BM1PerThreadBM11;
static constexpr index_t BN1PerThread =
ShareB ? BN1PerThreadBN11 / dpp8::lane_group_size : BN1PerThreadBN11;
__host__ __device__ static constexpr auto
MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
{
const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
a_block_desc_bk0_bm_bk1,
make_tuple(make_pass_through_transform(Number<BK0>{}),
make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
make_pass_through_transform(Number<BK1>{})),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
return a_block_bk0_bm0_bm1_bk1;
}
__host__ __device__ static constexpr auto
MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
{
const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
b_block_desc_bk0_bn_bk1,
make_tuple(make_pass_through_transform(Number<BK0>{}),
make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
make_pass_through_transform(Number<BK1>{})),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
return b_block_desc_bk0_bn0_bn1_bk1;
}
__host__ __device__ static constexpr auto
MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM_BN()
{
// upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
// lower: [BM, BN]
constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n =
make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(
Number<BM0>{}, Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
make_unmerge_transform(make_tuple(
Number<BN0>{}, Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n;
}
__host__ __device__ static constexpr auto
MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1()
{
// upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
// lower: [BM0, BM1, BN0, BN1]
constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1 =
make_single_stage_tensor_adaptor(
make_tuple(make_pass_through_transform(Number<BM0>{}),
make_unmerge_transform(
make_tuple(Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
make_pass_through_transform(Number<BN0>{}),
make_unmerge_transform(
make_tuple(Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1;
}
__host__ __device__ static constexpr auto GetCThreadTensorLengths_BM0_BM1_BN0_BN1()
{
return Sequence<BM0, BM11, BN0, BN11>{};
}
static constexpr auto a_block_desc_bk0_bm0_bm1_bk1_ =
MakeABlockDescriptor_BK0_BM0_BM1_BK1(ABlockDesc_BK0_BM_BK1{});
static constexpr auto b_block_desc_bk0_bn0_bn1_bk1_ =
MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
public:
__device__ BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0()
: c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
get_thread_local_1d_id())},
a_thread_copy_{CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()},
b_thread_copy_{CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()}
{
static_assert(ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
static_assert(BM % BM1 == 0 && BN % BN1 == 0, "wrong!");
static_assert(ABlockDesc_BK0_BM_BK1{}.GetLength(I0) ==
BBlockDesc_BK0_BN_BK1{}.GetLength(I0),
"wrong! K dimension not consistent");
static_assert(BM10BN10ThreadClusterBM10Xs::Size() == 2 &&
BM10BN10ThreadClusterBN10Xs::Size() == 2,
"wrong!");
}
__device__ static CIndex CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(index_t thread_id)
{
// lower: [BM0, BM1, BN0, BN1]
// upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
constexpr auto adaptor0 =
MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1();
// lower: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
// upper: [Tid, BM0, BM11, BN0, BN11]
constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(BM100, BN100, BM101, BN101)),
make_pass_through_transform(BM0),
make_pass_through_transform(BM11),
make_pass_through_transform(BN0),
make_pass_through_transform(BN11)),
make_tuple(
Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
}
__device__ AIndex CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()
{
const auto offsetBM0 = c_thread_origin_data_idx_[I0];
// If sharing matrix A, we need a separate BM1 offset for each thread in lane group.
const auto offsetBM1 = ShareA ? c_thread_origin_data_idx_[I1] +
dpp8::get_thread_idx_in_lane_group() * BM1PerThread
: c_thread_origin_data_idx_[I1];
return make_tuple(0, offsetBM0, offsetBM1, 0);
}
__device__ BIndex CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()
{
const auto offsetBN0 = c_thread_origin_data_idx_[I2];
// If sharing matrix B, we need a separate BN1 offset for each thread in lane group.
const auto offsetBN1 = ShareB ? c_thread_origin_data_idx_[I3] +
dpp8::get_thread_idx_in_lane_group() * BN1PerThread
: c_thread_origin_data_idx_[I3];
return make_tuple(0, offsetBN0, offsetBN1, 0);
}
template <typename CThreadDesc_BM0_BM11_BN0_BN11,
typename ABlockBuffer,
typename BBlockBuffer,
typename CThreadBuffer>
__device__ void Run(const CThreadDesc_BM0_BM11_BN0_BN11&,
const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
static_assert(CThreadDesc_BM0_BM11_BN0_BN11::IsKnownAtCompileTime(),
"wrong! Desc should be known at compile-time");
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
constexpr auto threadwise_contraction =
ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
FloatA,
FloatB,
FloatC,
decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
CThreadDesc_BM0_BM11_BN0_BN11,
Sequence<BK0PerThread, BK1>,
Sequence<1, BM1PerThreadBM11>,
Sequence<1, BN1PerThreadBN11>,
ShareA>{};
static_for<0, BN0, 1>{}([&](auto bn0) {
static_for<0, BM0, 1>{}([&](auto bm0) {
a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
make_tuple(I0, bm0, I0, I0),
a_block_buf,
a_thread_desc_bk0_bm0_bm1_bk1_,
make_tuple(I0, I0, I0, I0),
a_thread_buf);
b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
make_tuple(I0, bn0, I0, I0),
b_block_buf,
b_thread_desc_bk0_bn0_bn1_bk1_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
threadwise_contraction.Run(a_thread_buf,
make_tuple(I0, I0, I0, I0),
b_thread_buf,
make_tuple(I0, I0, I0, I0),
c_thread_buf,
make_tuple(bm0, I0, bn0, I0));
static_for<BK0PerThread, BK0, BK0PerThread>{}([&](auto bk0) {
a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
make_tuple(bk0, bm0, I0, I0),
a_block_buf,
a_thread_desc_bk0_bm0_bm1_bk1_,
make_tuple(I0, I0, I0, I0),
a_thread_buf);
b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
make_tuple(bk0, bn0, I0, I0),
b_block_buf,
b_thread_desc_bk0_bn0_bn1_bk1_,
make_tuple(I0, I0, I0, I0),
b_thread_buf);
threadwise_contraction.Run(a_thread_buf,
make_tuple(I0, I0, I0, I0),
b_thread_buf,
make_tuple(I0, I0, I0, I0),
c_thread_buf,
make_tuple(bm0, I0, bn0, I0));
});
});
});
}
private:
// A[BK0, BM0, BM1, BK1]
static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThread>{}, Number<BK1>{}));
// B[BK0, BN0, BN1, BK1]
static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThread>{}, Number<BK1>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
FloatA,
FloatA,
decltype(a_block_desc_bk0_bm0_bm1_bk1_),
decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
Sequence<BK0PerThread, 1, BM1PerThread, BK1>, // SliceLengths
Sequence<0, 1, 2, 3>, // DimAccessOrder
Sequence<1, 1, BM1PerThread, BK1>, // SrcVectorTensorLengths
Sequence<0, 1, 2, 3>>; // SrcVectorTensorContiguousDimOrder
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
FloatB,
FloatB,
decltype(b_block_desc_bk0_bn0_bn1_bk1_),
decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
Sequence<BK0PerThread, 1, BN1PerThread, BK1>, // SliceLengths
Sequence<0, 1, 2, 3>, // DimAccessOrder
Sequence<1, 1, BN1PerThread, BK1>, // SrcVectorTensorLengths
Sequence<0, 1, 2, 3>>; // SrcVectorTensorContiguousDimOrder
CIndex c_thread_origin_data_idx_;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
};
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp 0 → 100644
View file @ e5ebcc41
// 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 @ e5ebcc41
// 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
include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
View file @ e5ebcc41
......@@ -221,49 +221,102 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, KPerBlock / WmmaK, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, m0, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, n0, I0, I0, I0),
b_thread_buf);
vector_type<FloatA, WmmaK> a_thread_vec;
vector_type<FloatB, WmmaK> b_thread_vec;
static_for<0, WmmaK, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
// basic intrinsic to determine loopover direction
if constexpr(MRepeat < NRepeat)
{
static_for<0, KPerBlock / WmmaK, 1>{}(
[&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, m0, I0, I0, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(
b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, n0, I0, I0, I0),
b_thread_buf);
vector_type<FloatA, WmmaK> a_thread_vec;
vector_type<FloatB, WmmaK> b_thread_vec;
static_for<0, WmmaK, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
});
using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
else
{
static_for<0, KPerBlock / WmmaK, 1>{}(
[&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
static_for<0, NRepeat, 1>{}([&](auto n0) {
// read B
b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
make_tuple(I0, n0, I0, I0, I0),
b_thread_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
// read A
a_thread_copy_.Run(
a_block_desc_k0_m0_m1_m2_k1,
make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
make_tuple(I0, m0, I0, I0, I0),
a_thread_buf);
vector_type<FloatA, WmmaK> a_thread_vec;
vector_type<FloatB, WmmaK> b_thread_vec;
static_for<0, WmmaK, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
b_thread_vec.template AsType<FloatB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
});
using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
wmma_gemm.template Run(
a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
}
}
protected:
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
View file @ e5ebcc41
......@@ -4,27 +4,13 @@
#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"
namespace ck {
enum struct LoopScheduler
{
Default,
Interwave,
};
constexpr LoopScheduler make_default_loop_scheduler()
{
#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
return LoopScheduler::Interwave;
#else
return LoopScheduler::Default;
#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
}
template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
__host__ __device__ static constexpr auto
MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
......@@ -42,7 +28,8 @@ MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
}
template <index_t BlockSize,
typename FloatAB,
typename FloatA,
typename FloatB,
typename FloatAcc,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
......@@ -50,7 +37,9 @@ template <index_t BlockSize,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack>
index_t KPack,
typename ComputeTypeA = FloatA,
typename ComputeTypeB = FloatB>
struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
{
static constexpr auto I0 = Number<0>{};
......@@ -72,7 +61,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
static constexpr auto xdlops_gemm =
XdlopsGemm<ComputeTypeA, MPerXDL, NPerXDL, KPack, ComputeTypeB>{};
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
......@@ -308,9 +298,9 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -332,25 +322,27 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
b_thread_buf);
static_for<0, KPerThread, KPack>{}([&](auto k) {
vector_type<FloatAB, KPack> a_thread_vec;
vector_type<FloatAB, KPack> b_thread_vec;
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
a_thread_vec.template AsType<ComputeTypeA>()(i) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
b_thread_vec.template AsType<ComputeTypeB>()(i) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_a =
typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB, 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>(),
a_thread_vec.template AsType<mfma_input_type_a>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
......@@ -370,8 +362,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
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,
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
ComputeTypeA,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerThread>,
......@@ -380,8 +372,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
ComputeTypeB,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerThread>,
......@@ -399,7 +391,8 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
// the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the
// default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0
template <index_t BlockSize,
typename FloatAB,
typename FloatA,
typename FloatB,
typename FloatAcc,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
......@@ -408,10 +401,13 @@ template <index_t BlockSize,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
typename ComputeTypeA = FloatA,
typename ComputeTypeB = FloatB,
index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
: public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatA,
FloatB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
......@@ -419,10 +415,13 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
NPerXDL,
MRepeat,
NRepeat,
KPack>
KPack,
ComputeTypeA,
ComputeTypeB>
{
using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatA,
FloatB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
......@@ -430,7 +429,9 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
NPerXDL,
MRepeat,
NRepeat,
KPack>;
KPack,
ComputeTypeA,
ComputeTypeB>;
#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
using Base::a_block_desc_m0_m1_m2_k;
......@@ -454,9 +455,9 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
b_thread_desc_.GetElementSpaceSize());
static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
......@@ -493,20 +494,22 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static_for<0, KPerInnerLoop, KPack>{}([&](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;
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto i) {
a_thread_vec.template AsType<FloatAB>()(i) =
a_thread_vec.template AsType<ComputeTypeA>()(i) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, 0, 0, k_ + i))>{}];
b_thread_vec.template AsType<FloatAB>()(i) =
b_thread_vec.template AsType<ComputeTypeB>()(i) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, 0, 0, k_ + i))>{}];
});
using mfma_input_type =
typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_a =
typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops>::type;
using mfma_input_type_b =
typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
......@@ -528,8 +531,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
// TODO: insert setprio in more precise manner since we
// could have more than >1 MFMA instructions in single call
xdlops_gemm.template Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
a_thread_vec.template AsType<mfma_input_type_a>(),
b_thread_vec.template AsType<mfma_input_type_b>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
......@@ -555,8 +558,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
ComputeTypeA,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
......@@ -565,8 +568,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
A_K1,
A_K1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
ComputeTypeB,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPerInnerLoop>,
......@@ -582,7 +585,8 @@ struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
};
template <index_t BlockSize,
typename FloatAB,
typename FloatA,
typename FloatB,
typename FloatAcc,
typename AK0MK1BlockDesc,
typename BK0NK1BlockDesc,
......@@ -591,13 +595,16 @@ template <index_t BlockSize,
index_t MRepeat,
index_t NRepeat,
index_t KPack,
LoopScheduler LoopSched>
LoopScheduler LoopSched,
typename ComputeTypeA = FloatA,
typename ComputeTypeB = FloatB>
constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
{
if constexpr(LoopSched == LoopScheduler::Default)
{
return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatA,
FloatB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
......@@ -605,12 +612,15 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
KPack,
ComputeTypeA,
ComputeTypeB>{};
}
else if constexpr(LoopSched == LoopScheduler::Interwave)
{
return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
FloatA,
FloatB,
FloatAcc,
AK0MK1BlockDesc,
BK0NK1BlockDesc,
......@@ -618,7 +628,9 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
KPack,
ComputeTypeA,
ComputeTypeB>{};
}
};
......@@ -632,26 +644,27 @@ constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
* 3. configurable k index starting position and step size after each FMA/XDL instruction
*/
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, TransposeC>{}.K0PerXdlops,
index_t BMmaKStride =
KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops>
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_v2
{
static constexpr auto I0 = Number<0>{};
......@@ -668,7 +681,8 @@ struct BlockwiseGemmXdlops_v2
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, TransposeC>{};
static constexpr auto xdlops_gemm =
XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, FloatAB, TransposeC>{};
static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
......
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.hpp 0 → 100644
View file @ e5ebcc41
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
/**
* Transfer that uses direct load instructions to copy data from global to LDS memory.
*
* Traditional loads first copy data from global to registers, and then from registers to LDS.
* Direct loads do not need an intermediate step, data is copied directly from global to LDS,
* without the use of additional registers.
*
* However, the instruction has limitations:
* - each thread must copy exactly a single DWORD - 4 bytes;
* - threads within a single wavefront must write consecutive DWORDS into LDS,
* (data in global do not need to be contiguous, each thread might have its own offset).
*
* To make sure that all the transfers finished, the `waitcnt` instruction must be used with
* `vmcnt` instead of `lgkmcnt`.
*
* Limitations of the transfer class:
* - `SrcData` must be the same as `DstData` - no possibility to convert the data type in flight;
* - `DstVectorDim` must be the last dimension;
* - `SrcVectorDim` must be the last dimension if `ScalarPerVector` is greater than 1;
* - `ScalarPerVector` times the number of bytes of `DstData` must be equal to a single DWORD = 4B
* (for examlpe if `DstData` is fp32, then `ScalarPerVector` must be 1; if `DstData` is fp16,
* `ScalarPerVector` must be 2);
* - if `ScalarPerVector` is greater than 1, the contiguous dimension in src and dst must be
* the same dimension;
* - threads in a wavefront must write contiguous data to LDS (when wavefront size is 64,
* they must write 64 contiguous DWORDs) - `ThreadClusterLengths` must be prepared in such a way
* to guarantee that.
*/
template <typename ThreadGroup,
typename BlockSliceLengths,
typename ThreadClusterLengths,
typename SrcData,
typename DstData,
typename SrcDesc,
typename DstDesc,
index_t SrcVectorDim,
index_t DstVectorDim,
index_t ScalarPerVector>
struct ThreadGroupTensorSliceTransfer_DirectLoad
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
static constexpr auto I0 = Number<0>{};
static constexpr auto block_slice_lengths = BlockSliceLengths{};
static constexpr auto thread_cluster_lengths = ThreadClusterLengths{};
static constexpr auto thread_single_load_size = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, ScalarPerVector>{}, Number<nDim>{});
// After a load, each thread moves by `thread_steps` instead of loading the next elements.
// It makes the whole wavefront load contiguous memory, what is required for direct loads.
static constexpr auto thread_steps = thread_cluster_lengths * thread_single_load_size;
static constexpr auto thread_slice_lengths = block_slice_lengths / thread_steps;
static __device__ constexpr bool AreThreadClusterLengthsValid()
{
// Make sure that ThreadClusterLengths are set in a way that allows for contiguous writes to
// LDS by the threads from a single wavefront.
// Examples (assuming 64 threads in a wavefront, 128 in a thread block):
// 1. BlockSliceLengths = [K0PerBlock, MPerBlock, K1PerBlock] = [4, 128, 8],
// data type = fp32 -> ScalarPerVector = 1
// INVALID: ThreadClusterLengths = [4, 4, 8] since in the first iteration, threads 0-31
// write [0, 0, 0] - [0, 3, 7] and thread 32 writes [1, 0, 0] instead of
// [0, 4, 0].
// VALID: ThreadClusterLengths = [2, 8, 8] or [1, 16, 8] since in the first iteration,
// threads 0-63 write [0, 0, 0] - [0, 7, 7] -> 64 consecutive elements (DWORDs).
// 2. BlockSliceLengths = [K0PerBlock, MPerBlock, K1PerBlock] = [4, 128, 8],
// data type = fp16 -> ScalarPerVector = 2
// NOTE: ThreadClusterLengths must take into account that each thread writes two
// elements (single DWORD) along the contiguous dimension.
// INVALID: ThreadClusterLengths = [4, 4, 8] since each 8 threads would try to write
// 8 * 2 elements of K1PerBlock and there are only 8;
// ThreadClusterLengths = [4, 8, 4] since in the first iteration, threads 0-31
// write [0, 0, 0] - [0, 7, 7] (7 since each writes 2 elements) and thread 32
// writes [1, 0, 0] instead of [0, 8, 0].
// VALID: ThreadClusterLengths = [4, 16, 4] or [2, 32, 4] or [1, 64, 4] since in the
// first iteration, threads 0-63 write [0, 0, 0] - [0, 15, 7] -> 128 consecutive
// elements = 64 consecutive DWORDs.
int num_contiguous_dwords = 1;
bool is_contiguous = true;
static_for<0, nDim, 1>{}([&](auto i) {
if(is_contiguous)
{
num_contiguous_dwords *= thread_cluster_lengths[nDim - i - 1];
}
if(thread_slice_lengths[nDim - i - 1] > 1)
{
is_contiguous = false;
}
});
constexpr index_t wavefront_size = get_warp_size();
const bool wave_contiguous = num_contiguous_dwords % wavefront_size == 0;
bool thread_slice_lengths_correct = true;
static_for<0, nDim, 1>{}([&](auto i) {
if(thread_slice_lengths[i] <= 0)
{
thread_slice_lengths_correct = false;
}
});
return wave_contiguous && thread_slice_lengths_correct;
}
__device__ constexpr ThreadGroupTensorSliceTransfer_DirectLoad(
const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
{
static_assert(ck::is_same_v<SrcData, DstData>,
"Direct load transfer does not support datatypes conversion. Source and "
"destination data types must be the same.");
static_assert(
DstVectorDim == nDim - 1,
"Direct load transfer requires the destination vector dimension to be the last one.");
static_assert(ScalarPerVector == 1 || SrcVectorDim == DstVectorDim,
"When loading more than one element per thread at once, the contiguous "
"dimension must be the same between source and destination.");
constexpr auto dword_bytes = 4;
constexpr auto bytes_per_thread_load = ScalarPerVector * sizeof(SrcData);
static_assert(bytes_per_thread_load == dword_bytes,
"Direct load transfer requires each thread to load exactly a single "
"DWORD of data.");
static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
nDim == ThreadClusterLengths::Size(),
"Inconsistent number of dimensions across lengths and descriptors.");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"The number of threads cannot be less than the number of elements in "
"thread cluster lengths.");
static_assert(
AreThreadClusterLengthsValid(),
"Thread cluster lengths are incorrect. They must be set in a way that allows a single "
"wavefront to write contiguous DWORDs into LDS memory. ");
const auto thread_cluster_idx =
thread_cluster_desc_.CalculateBottomIndex(make_multi_index(ThreadGroup::GetThreadId()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_single_load_size;
SetSrcSliceOrigin(src_desc, src_block_slice_origin + thread_data_idx_begin);
SetDstSliceOrigin(dst_desc, dst_block_slice_origin + thread_data_idx_begin);
}
__device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
{
src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
src_slice_origin_ = src_slice_origin_idx;
}
__device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
{
dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
dst_slice_origin_ = dst_slice_origin_idx;
}
__device__ void ResetDstSliceWindow(const DstDesc& dst_desc)
{
dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_);
}
template <typename SrcBuffer, typename DstBuffer>
__device__ void Run(const SrcDesc& src_desc,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf)
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global,
"Source data must come from a global memory buffer.");
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"Destination data must be stored in an LDS memory buffer.");
static_assert(
ck::is_same_v<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>,
"SrcBuffer and SrcData data types must be consistent.");
static_assert(
ck::is_same_v<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>,
"DstBuffer and DstData data types must be consistent.");
constexpr auto dst_access_lengths = thread_slice_lengths;
const auto dst_forward_steps = generate_steps(dst_desc, 1);
const auto dst_backward_steps = generate_steps(dst_desc, -1);
const auto src_forward_steps = generate_steps(src_desc, 1);
const auto src_backward_steps = generate_steps(src_desc, -1);
// Loop over the destination block and copy data.
static_ford<decltype(dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
const auto src_offset = src_coord_.GetOffset();
const auto dst_offset = dst_coord_.GetOffset();
// Check if src data is not in the logic padding area.
const bool is_src_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
src_buf.template DirectCopyToLds<remove_cvref_t<decltype(dst_buf)>, ScalarPerVector>(
dst_buf, src_offset, dst_offset, is_src_valid);
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_dst_access_idx[i] < dst_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_dst_access_idx[j] == dst_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// Decide whether to move forward or backward.
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_dst_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * dst_access_lengths[j] + ordered_dst_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(dst_desc, dst_coord_, dst_forward_steps[i]);
move_tensor_coordinate(src_desc, src_coord_, src_forward_steps[i]);
}
else
{
move_tensor_coordinate(dst_desc, dst_coord_, dst_backward_steps[i]);
move_tensor_coordinate(src_desc, src_coord_, src_backward_steps[i]);
}
}
});
});
// Reset the destination slice since the entire buffer has been already filled.
ResetDstSliceWindow(dst_desc);
}
__device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
{
src_slice_origin_ = src_slice_origin_ + step;
src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_);
}
template <typename DescType>
__device__ auto generate_steps(const DescType& desc, int sign)
{
return generate_tuple(
[&](auto i) {
Index step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
step_idx(j) = (i.value == j.value) ? sign * thread_steps[i] : 0;
});
return make_tensor_coordinate_step(desc, step_idx);
},
Number<nDim>{});
}
private:
static constexpr auto thread_cluster_desc_ = make_cluster_descriptor(ThreadClusterLengths{});
SrcCoord src_coord_;
DstCoord dst_coord_;
Index src_slice_origin_;
Index dst_slice_origin_;
};
} // namespace ck
include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp 0 → 100644
View file @ e5ebcc41
// 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_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck {
// Thread-group level multi-source, multi-destination tensor slice data movement
// Assume:
// 1. All sources and destinations are DynamicBuffer
// 2. Same VectorDim and ScalerPerVector for all sources and destinations
// 3. DstInMemOps are per destination tensor
// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
//
// Does following things to avoid scratch memory issue
// 1. Pass tensor descritpors by reference (or tuple of references)
// 2. Does not keep reference to tensor descriptor
// 3. Does not construct new tensor coordinate when call Run()
template <typename ThreadGroup,
typename SrcDatas,
typename DstDatas,
typename SrcDescs,
typename DstDescs,
typename ElementwiseOperation,
typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
typename SliceLengths,
typename ThreadClusterLengths,
typename ThreadClusterArrangeOrder,
typename SrcDimAccessOrder,
typename DstDimAccessOrder,
index_t SrcVectorDim,
index_t DstVectorDim,
index_t SrcScalarPerVector,
index_t DstScalarPerVector,
typename ThreadTransferSrcResetCoordinateAfterRunFlags,
typename ThreadTransferDstResetCoordinateAfterRunFlags>
struct ThreadGroupTensorSliceTransfer_v7r2
{
static constexpr index_t nDim =
remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
using Index = MultiIndex<nDim>;
static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
__device__ constexpr ThreadGroupTensorSliceTransfer_v7r2(
const SrcDescs& src_descs,
const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
const DstDescs& dst_descs,
const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
const ElementwiseOperation& element_op)
: threadwise_transfer_(src_descs,
StaticallyIndexedArray<Index, nSrc>{},
dst_descs,
StaticallyIndexedArray<Index, nDst>{},
element_op)
{
static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
"wrong!");
static_for<0, nSrc, 1>{}([&](auto i) {
static_assert(
nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
"wrong!");
});
static_for<0, nDst, 1>{}([&](auto i) {
static_assert(
nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
"wrong!");
});
static_assert(nDim == ThreadClusterLengths::Size() &&
nDim == ThreadClusterArrangeOrder::Size() &&
nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
"wrong! nDim not consistent");
static_assert(
is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
"wrong! threads should be mapped to cover entire slicing window");
static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
"wrong! ThreadGroup::GetNumOfThread() too small");
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
make_multi_index(get_thread_local_1d_id()));
const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
const auto src_thread_slice_origins = generate_tuple(
[&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
Number<nSrc>{});
const auto dst_thread_slice_origins = generate_tuple(
[&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
Number<nDst>{});
threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
}
}
template <typename SrcBuffers>
__device__ void RunRead(const SrcDescs& src_descs, const SrcBuffers& src_bufs)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_descs, src_bufs);
}
}
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
template <typename DstBuffers>
__device__ void RunWrite(const DstDescs& dst_descs, DstBuffers dst_bufs)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
if constexpr(is_detected<is_tuple, decltype(dst_bufs)>::value)
threadwise_transfer_.RunWrite(dst_descs, dst_bufs);
else
threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs));
}
}
template <typename SrcBuffers, typename DstBuffers>
__device__ void Run(const SrcDescs& src_descs,
const SrcBuffers& src_bufs,
const DstDescs& dst_descs,
DstBuffers dst_bufs)
{
RunRead(src_descs, src_bufs);
RunWrite(dst_descs, dst_bufs);
}
template <index_t ISrc>
__device__ void
MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
}
}
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step)
{
static_for<0, SrcDescs::Size(), 1>{}(
[&](auto i) { MoveSrcSliceWindow(src_descs, i, step); });
}
template <index_t IDst>
__device__ void
MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
{
if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
}
}
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step)
{
static_for<0, DstDescs::Size(), 1>{}(
[&](auto i) { MoveDstSliceWindow(dst_descs, i, step); });
}
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v7r2<SrcDatas,
DstDatas,
SrcDescs,
DstDescs,
ElementwiseOperation,
DstInMemOps,
decltype(thread_slice_lengths),
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcScalarPerVector,
DstScalarPerVector,
ThreadTransferSrcResetCoordinateAfterRunFlags,
ThreadTransferDstResetCoordinateAfterRunFlags>;
ThreadwiseTransfer threadwise_transfer_;
};
} // namespace ck
include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp 0 → 100644
View file @ e5ebcc41
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace conv_tensor_rearrange_op {
struct BaseConvTensorRearrangeOp
{
};
struct ImageToColumn : public BaseConvTensorRearrangeOp
{
static constexpr const char* name = "Image to Column";
};
struct ColumnToImage : public BaseConvTensorRearrangeOp
{
static constexpr const char* name = "Column to Image";
};
template <typename Op,
typename std::enable_if<std::is_base_of<BaseConvTensorRearrangeOp, Op>::value,
bool>::type = false>
std::ostream& operator<<(std::ostream& os, const BaseConvTensorRearrangeOp&)
{
os << Op::name;
return os;
}
} // namespace conv_tensor_rearrange_op
} // namespace ck
include/ck/tensor_operation/gpu/device/device_base.hpp
View file @ e5ebcc41
......@@ -62,7 +62,9 @@ struct BaseOperator
};
virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const
virtual void SetWorkSpacePointer(BaseArgument* p_arg,
void* p_workspace,
const StreamConfig& = StreamConfig{}) const
{
assert(p_arg);
p_arg->p_workspace_ = p_workspace;
......
include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp 0 → 100644
View file @ e5ebcc41
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// GEMM:
// input : A0[M0, M1, ... K0, K1, ...], ...
// input : B0[N0, N1, ... K0, K1, ...], ...
// input : D0[M0, M1, ... N0, N1, ...], D1[M0, M1, ... N0, N1, ...], ...
// output : E[M0, M1, ... N0, N1, ...]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename AsDataType,
typename BsDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceContractionMultipleABD : public BaseOperator
{
static constexpr index_t NumATensor = AsDataType::Size();
static constexpr index_t NumBTensor = BsDataType::Size();
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
std::array<const void*, NumBTensor> p_bs,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_lengths,
const std::array<std::vector<index_t>, NumATensor>& a_ms_ks_strides,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_lengths,
const std::array<std::vector<index_t>, NumBTensor>& b_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& d_ms_ns_strides,
const std::vector<index_t>& e_ms_ns_length,
const std::vector<index_t>& e_ms_ns_stride,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // 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