Skip to content

GitLab

Commit fbc576b5 authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into lds_bypass_spilling
parents d4adc71a 8f455615
Hide whitespace changes
Inline Side-by-side
Showing with 478 additions and 75 deletions
CHANGELOG.md
View file @ fbc576b5
...@@ -18,6 +18,7 @@ Full documentation for Composable Kernel is not yet available. ...@@ -18,6 +18,7 @@ Full documentation for Composable Kernel is not yet available.
- Added multi-D GEMM client APIs (#534). - Added multi-D GEMM client APIs (#534).
- Added multi-embeddings support (#542). - Added multi-embeddings support (#542).
- Added Navi3x blockwise GEMM and real GEMM support (#541). - Added Navi3x blockwise GEMM and real GEMM support (#541).
- Added Navi grouped ConvBwdWeight support (#505).
### Changed ### Changed
- Changed ... - Changed ...
Jenkinsfile
View file @ fbc576b5
...@@ -471,6 +471,12 @@ def Build_CK(Map conf=[:]){ ...@@ -471,6 +471,12 @@ def Build_CK(Map conf=[:]){
//we only need the ckProfiler to run the performance tests, so we pack and stash it //we only need the ckProfiler to run the performance tests, so we pack and stash it
sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler' sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
stash "ckProfiler.tar.gz" stash "ckProfiler.tar.gz"
if (params.RUN_FULL_QA){
// build deb packages
sh 'make -j package'
archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
}
} }
} }
} }
...@@ -651,8 +657,8 @@ pipeline { ...@@ -651,8 +657,8 @@ pipeline {
{ {
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 " """ }" setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 " """ }"
execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a" -DCMAKE_CXX_FLAGS="-O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }" execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908,gfx90a;gfx1030" -DCMAKE_CXX_FLAGS="-O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
} }
steps{ steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local') Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
...@@ -674,7 +680,7 @@ pipeline { ...@@ -674,7 +680,7 @@ pipeline {
options { retry(2) } options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")} agent{ label rocmnode("gfx908 || gfx90a")}
environment{ environment{
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx908;gfx90a" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}" setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS=" -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -DGPU_TARGETS="gfx908;gfx90a;gfx1030" -DCMAKE_CXX_FLAGS=" -O3 " -DBUILD_DEV=On """}"
} }
steps{ steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release') runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......
client_example/07_grouped_conv2d_fwd/CMakeLists.txt → client_example/07_grouped_convnd_fwd/CMakeLists.txt
View file @ fbc576b5
add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp) add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_operations)
client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp 0 → 100644
View file @ fbc576b5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::GNWC;
using WeiLayout = ck::tensor_layout::convolution::GKXC;
using OutLayout = ck::tensor_layout::convolution::GNWK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr ck::index_t NumDimSpatial = 1;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Wo = 28;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Wi, C};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, X, C};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 1};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Wo, K};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 1};
std::partial_sum(rbegin(in_lengths),
std::prev(rend(in_lengths)),
std::next(rbegin(in_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(wei_lengths),
std::prev(rend(wei_lengths)),
std::next(rbegin(wei_strides)),
std::multiplies<>{});
std::partial_sum(rbegin(out_lengths),
std::prev(rend(out_lengths)),
std::next(rbegin(out_strides)),
std::multiplies<>{});
// transpose GNWC/GKXC/GNWK to GNCW/GKCX/GNCW
std::rotate(rbegin(in_lengths),
std::next(rbegin(in_lengths)),
std::next(rbegin(in_lengths), NumDimSpatial + 1));
std::rotate(rbegin(in_strides),
std::next(rbegin(in_strides)),
std::next(rbegin(in_strides), NumDimSpatial + 1));
std::rotate(rbegin(wei_lengths),
std::next(rbegin(wei_lengths)),
std::next(rbegin(wei_lengths), NumDimSpatial + 1));
std::rotate(rbegin(wei_strides),
std::next(rbegin(wei_strides)),
std::next(rbegin(wei_strides), NumDimSpatial + 1));
std::rotate(rbegin(out_lengths),
std::next(rbegin(out_lengths)),
std::next(rbegin(out_lengths), NumDimSpatial + 1));
std::rotate(rbegin(out_strides),
std::next(rbegin(out_strides)),
std::next(rbegin(out_strides), NumDimSpatial + 1));
std::array<ck::index_t, NumDimSpatial> filter_strides{1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Wo * X;
std::size_t num_bytes = sizeof(InDataType) * G * N * Wi * C +
sizeof(WeiDataType) * G * K * X * C +
sizeof(OutDataType) * G * N * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
PassThrough{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
}
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ fbc576b5
add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp 0 → 100644
View file @ fbc576b5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::GNWC;
using WeiLayout = ck::tensor_layout::convolution::GKXC;
using OutLayout = ck::tensor_layout::convolution::GNWK;
static constexpr ck::index_t NumDimSpatial = 1;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 256;
static constexpr ck::index_t K = 192;
static constexpr ck::index_t C = 192;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Wi = 28;
static constexpr ck::index_t Wo = 28;
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout>(G, N, K, C, {Wi}, {X}, {Wo}, {1}, {1}, {1}, {1})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
example/04_gemm_add_add_fastgelu/common.hpp
View file @ fbc576b5
...@@ -62,7 +62,7 @@ struct ExecutionConfig final ...@@ -62,7 +62,7 @@ struct ExecutionConfig final
}; };
inline bool inline bool
parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig config) parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
{ {
if(argc == 1) if(argc == 1)
{ {
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
View file @ fbc576b5
...@@ -7,10 +7,11 @@ using ADataType = BF16; ...@@ -7,10 +7,11 @@ using ADataType = BF16;
using BDataType = BF16; using BDataType = BF16;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F32; using CShuffleDataType = F32;
using D0DataType = BF16; using CDataType = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
using D1DataType = BF16; using D0DataType = BF16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using D1DataType = BF16;
using EDataType = BF16; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = BF16;
using ALayout = Row; using ALayout = Row;
using BLayout = Col; using BLayout = Col;
...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
AccDataType, CDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
View file @ fbc576b5
...@@ -7,10 +7,11 @@ using ADataType = F16; ...@@ -7,10 +7,11 @@ using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F32; using CShuffleDataType = F32;
using D0DataType = F16; using CDataType = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
using D1DataType = F16; using D0DataType = F16;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using D1DataType = F16;
using EDataType = F16; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F16;
using ALayout = Row; using ALayout = Row;
using BLayout = Col; using BLayout = Col;
...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
AccDataType, CDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
View file @ fbc576b5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp" #include "common.hpp"
...@@ -7,10 +6,11 @@ using ADataType = F32; ...@@ -7,10 +6,11 @@ using ADataType = F32;
using BDataType = F32; using BDataType = F32;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F32; using CShuffleDataType = F32;
using D0DataType = F32; using CDataType = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
using D1DataType = F32; using D0DataType = F32;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using D1DataType = F32;
using EDataType = F32; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = F32;
using ALayout = Row; using ALayout = Row;
using BLayout = Col; using BLayout = Col;
...@@ -36,7 +36,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -36,7 +36,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
AccDataType, CDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
View file @ fbc576b5
...@@ -11,10 +11,11 @@ using ADataType = I4; ...@@ -11,10 +11,11 @@ using ADataType = I4;
using BDataType = I4; using BDataType = I4;
using AccDataType = I32; using AccDataType = I32;
using CShuffleDataType = I32; using CShuffleDataType = I32;
using D0DataType = I4; using CDataType = I32; // C matrix doesn't exsit in GPU memory, this is used for host verification
using D1DataType = I4; using D0DataType = I4;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using D1DataType = I4;
using EDataType = I4; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = I4;
using KernelADataType = I8; using KernelADataType = I8;
using KernelBDataType = I8; using KernelBDataType = I8;
...@@ -47,7 +48,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -47,7 +48,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
AccDataType, CDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
View file @ fbc576b5
...@@ -7,10 +7,11 @@ using ADataType = I8; ...@@ -7,10 +7,11 @@ using ADataType = I8;
using BDataType = I8; using BDataType = I8;
using AccDataType = I32; using AccDataType = I32;
using CShuffleDataType = I32; using CShuffleDataType = I32;
using D0DataType = I8; using CDataType = I32; // C matrix doesn't exsit in GPU memory, this is used for host verification
using D1DataType = I8; using D0DataType = I8;
using DsDataType = ck::Tuple<D0DataType, D1DataType>; using D1DataType = I8;
using EDataType = I8; using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = I8;
using ALayout = Row; using ALayout = Row;
using BLayout = Col; using BLayout = Col;
...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -36,7 +37,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType, BDataType,
AccDataType, CDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
......
example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
View file @ fbc576b5
...@@ -124,7 +124,7 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC ...@@ -124,7 +124,7 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
if(config.do_verification) if(config.do_verification)
{ {
Tensor<AccDataType> c_m_n({M, N}); Tensor<CDataType> c_m_n({M, N});
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
......
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ fbc576b5
...@@ -6,3 +6,9 @@ add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd ...@@ -6,3 +6,9 @@ add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16 add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16) example_grouped_conv_bwd_weight_xdl_bf16)
add_custom_target(example_grouped_conv_bwd_weight_dl)
add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
example/20_grouped_conv_bwd_weight/common.hpp
View file @ fbc576b5
...@@ -9,7 +9,6 @@ ...@@ -9,7 +9,6 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp" #include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp 0 → 100644
View file @ fbc576b5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
16, // K0PerBlock
2, // K1
4, // M1PerThread
4, // N1PerThread
1, // KPerThread
S<8, 2>, // M1N1ThreadClusterM1Xs
S<8, 2>, // M1N1ThreadClusterN1Xs
S<1, 8, 1, 1, 2>, // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
S<1, 2, 1, 128, 1>, // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
S<0, 2, 3, 1, 4>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 3, 1, 4>, // ABlockTransferSrcAccessOrder
S<1, 1, 1, 1, 1>, // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
S<0, 2, 3, 1, 4>, // ABlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 1, 1>, // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
S<1, 1, 1, 8, 2>, // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
S<1, 16, 1, 16, 1>, // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
S<0, 1, 4, 2, 3>, // BBlockTransferThreadClusterArrangeOrder
S<0, 1, 4, 2, 3>, // BBlockTransferSrcAccessOrder
S<1, 1, 1, 8, 1>, // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
S<0, 1, 4, 2, 3>, // BBlockTransferSrcVectorTensorContiguousDimOrder
S<1, 1, 1, 1, 2>, // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
5, // CThreadTransferSrcDstVectorDim
4>; // CThreadTransferDstScalarPerVector
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
View file @ fbc576b5
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
using InDataType = BF16; using InDataType = BF16;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory // bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = F32; using WeiDataType = F32;
...@@ -13,6 +15,46 @@ using InElementOp = PassThrough; ...@@ -13,6 +15,46 @@ using InElementOp = PassThrough;
using WeiElementOp = PassThrough; using WeiElementOp = PassThrough;
using OutElementOp = PassThrough; using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
#include "run_grouped_conv_bwd_weight_example.inc" #include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); } int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
View file @ fbc576b5
...@@ -3,6 +3,8 @@ ...@@ -3,6 +3,8 @@
#include "common.hpp" #include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
using InDataType = F16; using InDataType = F16;
using WeiDataType = F16; using WeiDataType = F16;
using OutDataType = F16; using OutDataType = F16;
...@@ -12,6 +14,46 @@ using InElementOp = PassThrough; ...@@ -12,6 +14,46 @@ using InElementOp = PassThrough;
using WeiElementOp = PassThrough; using WeiElementOp = PassThrough;
using OutElementOp = PassThrough; using OutElementOp = PassThrough;
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
#include "run_grouped_conv_bwd_weight_example.inc" #include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); } int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ fbc576b5
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
template <ck::index_t NDimSpatial>
using DeviceConvBwdWeightInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial, using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType, InDataType,
...@@ -54,8 +14,19 @@ template <ck::index_t NDimSpatial> ...@@ -54,8 +14,19 @@ template <ck::index_t NDimSpatial>
bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
const ck::utils::conv::ConvParam& conv_param) const ck::utils::conv::ConvParam& conv_param)
{ {
constexpr ck::index_t split_k = 2; ck::index_t split_k;
// Set split_k = 2 for xdl op, split_k = 1 for dl
// Dl op doesn't support split_k > 1
// TODO: Add Dl op split_k > 1 support
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030"))
{
split_k = 2;
}
else
{
split_k = 1;
}
const auto in_g_n_c_wis_desc = const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed< ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
InputLayout<NDimSpatial>>(conv_param); InputLayout<NDimSpatial>>(conv_param);
...@@ -144,7 +115,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config, ...@@ -144,7 +115,7 @@ bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
std::cerr << "wrong! device_conv with the specified compilation parameters does " std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem" "not support this Conv problem"
<< std::endl; << std::endl;
return false; return true;
} }
float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel}); float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
......
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