Unverified Commit a0f2770b authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-405

No related merge requests found
Showing with 1539 additions and 440 deletions
+1539 -440
......@@ -663,8 +663,8 @@ pipeline {
{
agent{ label rocmnode("gfx908 || gfx90a") }
environment{
setup_args = "${params.COMPILER_VERSION == "release" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 --offload-arch=gfx90a -O3 " """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ }"
execute_args = "${params.COMPILER_VERSION == "release" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 --offload-arch=gfx90a -O3 " """ }"
execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
}
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')
......@@ -689,9 +689,8 @@ pipeline {
{
agent{ label rocmnode("gfx908")}
environment{
setup_args = "${params.COMPILER_VERSION == "release" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ }"
execute_args = "${params.COMPILER_VERSION == "release" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" """ : """ -DBUILD_DEV=Off -DCMAKE_INSTALL_PREFIX=../install -D CMAKE_CXX_FLAGS="--offload-arch=gfx908 -O3 " """ }"
execute_args = "${params.COMPILER_VERSION == "ck-9110" ? """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ : """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ }"
}
steps{
buildHipClangJobAndReboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......@@ -713,7 +712,7 @@ pipeline {
options { retry(2) }
agent{ label rocmnode("gfx908 || gfx90a")}
environment{
setup_args = "${params.COMPILER_VERSION == "release" ? """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 " -DBUILD_DEV=On """ : """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """}"
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx908 -O3 " -DBUILD_DEV=On """}"
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......@@ -728,7 +727,7 @@ pipeline {
options { retry(2) }
agent{ label rocmnode("gfx90a")}
environment{
setup_args = "${params.COMPILER_VERSION == "release" ? """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx90a -O3 " -DBUILD_DEV=On """ : """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """}"
setup_args = "${params.COMPILER_VERSION == "ck-9110" ? """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx90a -O3 -Xclang -mlink-builtin-bitcode -Xclang /opt/rocm/amdgcn/bitcode/oclc_abi_version_400.bc" -DBUILD_DEV=On """ : """ -D CMAKE_CXX_FLAGS=" --offload-arch=gfx90a -O3 " -DBUILD_DEV=On """}"
}
steps{
runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
......
......@@ -81,8 +81,8 @@ int main(int argc, char* argv[])
auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
{Stride, 1}, // xStrides
{1}, // gammaStrides
{1}, // betaStrides
{0, 1}, // gammaStrides
{0, 1}, // betaStrides
{Stride, 1}, // yStrides
{1}, // reduceDims
1e-4,
......
......@@ -6,9 +6,10 @@ find_package(composable_kernel 1.0.0 COMPONENTS device_operations)
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
add_subdirectory(01_gemm)
add_subdirectory(02_gemm_add_add_fastgelu)
add_subdirectory(03_gemm_layernorm)
add_subdirectory(04_contraction)
add_subdirectory(05_layernorm)
add_subdirectory(06_softmax)
# add all example subdir
file(GLOB dir_list LIST_DIRECTORIES true *)
FOREACH(subdir ${dir_list})
IF(IS_DIRECTORY "${subdir}")
add_subdirectory(${subdir})
ENDIF()
ENDFOREACH()
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_e_permute_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmEPermuteXdl
// clang-format off
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| DataType| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
const int M = 256;
const int N = 128;
const int K = 64;
const int stride_A = K;
const int stride_B = K;
const int batch_stride_A = M * K;
const int batch_stride_B = K * N;
const int G0 = 16;
const int G1 = 8;
const int batch_count = G0 * G1;
// output layout - [G0, M, G1, N]
const int stride_G0 = M * G1 * N;
const int stride_G1 = N;
const int stride_M = G1 * N;
const int stride_N = 1;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
// GEMM shape
ck::tensor_operation::device::BatchedGemmEPermuteDesc batched_gemm_e_permute_desc{
G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N};
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
auto f_host_e_tensor_descriptor = [](std::size_t G0_,
std::size_t G1_,
std::size_t M_,
std::size_t N_,
std::size_t stride_G0_,
std::size_t stride_G1_,
std::size_t stride_M_,
std::size_t stride_N_) {
return HostTensorDescriptor(
std::vector<std::size_t>({G0_, G1_, M_, N_}),
std::vector<std::size_t>({stride_G0_, stride_G1_, stride_M_, stride_N_}));
};
Tensor<EDataType> e_g0_g1_m_n_host_result(
f_host_e_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
Tensor<EDataType> e_g0_g1_m_n_device_result(
f_host_e_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "e_g0_g1_m_n: " << e_g0_g1_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) *
e_g0_g1_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEM
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
M,
N,
K,
stride_A,
stride_B,
batch_stride_A,
batch_stride_B,
batched_gemm_e_permute_desc,
batch_count,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * batch_count * M * N * K;
std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * K * N +
sizeof(EDataType) * batch_count * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
e_device_buf.FromDevice(e_g0_g1_m_n_device_result.mData.data());
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
Tensor<EDataType> c_g_m_n_host_result = HostTensorDescriptor(
std::vector<std::size_t>({batch_count, M, N}), std::vector<std::size_t>({M * N, N, 1}));
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
for(int g0 = 0; g0 < G0; g0++)
{
for(int g1 = 0; g1 < G1; g1++)
{
for(int m = 0; m < M; m++)
{
for(int n = 0; n < N; n++)
{
int g = g0 * G1 + g1;
e_g0_g1_m_n_host_result(g0, g1, m, n) = c_g_m_n_host_result(g, m, n);
}
}
}
}
pass = ck::utils::check_err(e_g0_g1_m_n_host_result.mData,
e_g0_g1_m_n_device_result.mData,
"Error: Incorrect results c");
}
return pass ? 0 : 1;
}
......@@ -29,24 +29,27 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
using DeviceInstance = ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
8, // GammaScalarPerVector
8, // BetaScalarPerVector
8>; // OutScalarPerVector
using DeviceInstance =
ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
int main()
{
......@@ -88,8 +91,8 @@ int main()
auto argument_ptr = device_instance.MakeArgumentPointer(
{M, N},
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
std::vector<ck::index_t>{gamma.mDesc.GetStrides().begin(), gamma.mDesc.GetStrides().end()},
std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(), beta.mDesc.GetStrides().end()},
{0, 1},
{0, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
{1},
1e-4,
......
......@@ -137,7 +137,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
......@@ -220,7 +220,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
......@@ -332,7 +332,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
......
......@@ -137,7 +137,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
......@@ -220,7 +220,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
......@@ -332,7 +332,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
......
......@@ -137,7 +137,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
......@@ -220,7 +220,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
......@@ -332,7 +332,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
......
......@@ -137,7 +137,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
......@@ -220,7 +220,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
......@@ -332,7 +332,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
......
......@@ -137,7 +137,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
......@@ -220,7 +220,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
......@@ -332,7 +332,7 @@ int main(int argc, char* argv[])
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
......
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_padded_batched_gemm_scale_softmax_gemm_xdl_fp16 padded_batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
add_custom_target(example_batched_gemm_scale_softmax_gemm)
add_dependencies(example_batched_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_batched_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_batched_gemm_scale_softmax_gemm example_padded_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_custom_target(example_gemm_scale_softmax_gemm)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
......@@ -16,7 +16,8 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -47,7 +48,9 @@ using CDataType = F16;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CLayout = Row;
using CPermuteNumDims_G_M_O =
S<2, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_G_M_O
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
......@@ -55,65 +58,67 @@ using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
MNPadding,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CShuffleBlockTransferScalarPerVector_NPerBlock
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
true>; // MaskOutUpperTriangle
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -143,22 +148,26 @@ int main(int argc, char* argv[])
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1020;
ck::index_t N = 1020;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 512;
ck::index_t N = 512;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t BatchCount = 4;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t StrideC = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
ck::index_t BatchStrideC = -1;
float alpha = 1;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
if(argc == 1)
{
// use default case
......@@ -169,74 +178,51 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 9)
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
BatchCount = std::stoi(argv[8]);
}
else if(argc == 18)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
StrideA = std::stoi(argv[9]);
StrideB0 = std::stoi(argv[10]);
StrideB1 = std::stoi(argv[11]);
StrideC = std::stoi(argv[12]);
BatchStrideA = std::stoi(argv[13]);
BatchStrideB0 = std::stoi(argv[14]);
BatchStrideB1 = std::stoi(argv[15]);
BatchStrideC = std::stoi(argv[16]);
alpha = std::stof(argv[17]);
alpha = std::stof(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
"BatchStrideB0, BatchStrideB1, BatchStrideC\n");
printf("arg17: scale (alpha)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
exit(0);
}
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? O : M;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
StrideC = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
BatchStrideC = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
const int BatchCount = G0 * G1;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
......@@ -263,15 +249,17 @@ int main(int argc, char* argv[])
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_g_m_o_host_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_o_device_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_gs_ms_os_host_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
Tensor<CDataType> c_gs_ms_os_device_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
switch(init_method)
{
......@@ -300,8 +288,8 @@ int main(int argc, char* argv[])
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
c_g_m_o_device_result.mDesc.GetElementSpaceSize());
DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
......@@ -320,20 +308,20 @@ int main(int argc, char* argv[])
gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
StrideA,
StrideB0,
StrideB1,
StrideC,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
BatchStrideC,
a_element_op,
b0_element_op,
acc0_element_op,
......@@ -361,26 +349,37 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
if(do_verification)
{
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
std::vector<int>{M * O, O, 1});
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
// gemm 0
ref_gemm0_invoker.Run(ref_gemm0_argument);
// mask out upper triangle
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(idx[1] < idx[2])
self(idx) = -ck::NumericLimits<float>::Infinity();
});
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
// softmax
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
......@@ -388,9 +387,22 @@ int main(int argc, char* argv[])
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
// gemm1
ref_gemm1_invoker.Run(ref_gemm1_argument);
return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
return ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData)
? 0
: 1;
}
return 0;
......
......@@ -58,7 +58,7 @@ using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNOPadding;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
......@@ -117,7 +117,8 @@ using DeviceGemmInstance =
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CShuffleBlockTransferScalarPerVector_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>; // MaskOutUpperTriangle
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -149,8 +150,8 @@ int main(int argc, char* argv[])
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 128;
ck::index_t N = 1024;
ck::index_t M = 120;
ck::index_t N = 1000;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t StrideA = -1;
......
......@@ -55,7 +55,7 @@ using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
ALayout,
......@@ -73,7 +73,7 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftma
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmDefault,
GemmSpec,
1,
256,
128, // MPerBlock
......@@ -113,7 +113,8 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftma
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CShuffleBlockTransferScalarPerVector_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>;
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......@@ -144,8 +145,8 @@ int main(int argc, char* argv[])
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t M = 1020;
ck::index_t N = 1020;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t BatchCount = 4;
......
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CPermuteNumDims_G_M_O =
S<1, 1, 1>; // "using CLayout = Row" has been replaced by CPermuteNumDims_M_O
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CPermuteNumDims_G_M_O,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
false>;
// Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: fp32 in, fp16 out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: fp16 in, fp16 out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 13;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
for(std::size_t i = 0; i < group_count; i++)
{
int M = 128 * (rand() % 8 + 1);
int N = 128 * (rand() % 8 + 1);
int K = 40;
int O = 40 * (rand() % 2 + 1);
int Batch = rand() % 8 + 1;
const int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int StrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int StrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int BatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int BatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int BatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
std::vector<ck::index_t> c_gs_ms_os_lengths{Batch, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides{O, Batch * O, 1};
problem_descs.push_back({M,
N,
K,
O,
Batch,
StrideA,
StrideB0,
StrideB1,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
c_gs_ms_os_lengths,
c_gs_ms_os_strides});
}
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::size_t flop = 0, num_byte = 0;
std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& K = problem_descs[i].K;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
const auto& StrideA = problem_descs[i].StrideA;
const auto& StrideB0 = problem_descs[i].StrideB0;
const auto& StrideB1 = problem_descs[i].StrideB1;
const auto& BatchStrideA = problem_descs[i].BatchStrideA;
const auto& BatchStrideB0 = problem_descs[i].BatchStrideB0;
const auto& BatchStrideB1 = problem_descs[i].BatchStrideB1;
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(Batch, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(Batch, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(Batch, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_gs_ms_os_device_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
Batch;
if(i < 4)
{
std::cout << "a_g_m_k[" << i << "]: " << a_g_m_k.mDesc << ", "
<< "b0_g_k_n[" << i << "]: " << b0_g_k_n.mDesc << ", "
<< "b1_g_n_o[" << i << "]: " << b1_g_n_o.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
}
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
case 3:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
a_tensors.push_back(a_g_m_k);
b0_tensors.push_back(b0_g_k_n);
b1_tensors.push_back(b1_g_n_o);
c_tensors.push_back(c_gs_ms_os_device_result);
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_g_m_k.mData.data());
b0_tensors_device[i]->ToDevice(b0_g_k_n.mData.data());
b1_tensors_device[i]->ToDevice(b1_g_n_o.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(p_a,
p_b0,
p_b1,
p_c,
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < group_count; i++)
{
const auto& M = problem_descs[i].M;
const auto& N = problem_descs[i].N;
const auto& O = problem_descs[i].O;
const auto& Batch = problem_descs[i].Batch;
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
const auto& a_g_m_k = a_tensors[i];
const auto& b0_g_k_n = b0_tensors[i];
const auto& b1_g_n_o = b1_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
Tensor<CDataType> c_gs_ms_os_host_result(
std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
// Output of Gemm0 is input A of Gemm1
Tensor<AccDataType> acc0_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(Batch, M, N, N, M * N, Row{}));
Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{Batch, M, O},
std::vector<int>{M * O, O, 1});
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// Note: in this example, we merely permute the dimensions by changing underlying
// strides so we simply access data as-is
c_gs_ms_os_host_result.ForEach(
[&](auto& self, auto idx) { self(idx) = c_g_m_o_host_result(idx); });
bool pass_ =
ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
pass &= pass_;
}
}
return pass ? 0 : 1;
}
add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
template <ck::index_t NDimSpatial,
typename OutDataType,
typename WeiDataType,
typename BiasDataType,
typename InDataType,
typename OutElementOp,
typename WeiElementOp,
typename InElementOp,
typename DeviceInstance>
int run_conv_bwd_data_bias_relu(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& bias_g_n_c_wis_desc,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const OutElementOp& out_element_op,
const WeiElementOp& wei_element_op,
const InElementOp& in_element_op)
{
Tensor<OutDataType> out(out_g_n_k_wos_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<BiasDataType> bias(bias_g_n_c_wis_desc);
Tensor<InDataType> in_host(in_g_n_c_wis_desc);
Tensor<InDataType> in_device(in_g_n_c_wis_desc);
std::cout << "out: " << out.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "bias: " << bias.mDesc << std::endl;
std::cout << "in: " << in_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
break;
default:
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{0.0, 1.0});
}
DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
out_device_buf.ToDevice(out.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
bias_device_buf.ToDevice(bias.mData.data());
// reset input to zero
in_device_buf.SetZero();
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(bias_g_n_c_wis_desc.GetLengths(), d0_g_n_c_wis_lengths);
copy(bias_g_n_c_wis_desc.GetStrides(), d0_g_n_c_wis_strides);
copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
// do conv
auto conv = DeviceInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(
out_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
in_device_buf.GetDeviceBuffer(),
a_g_n_k_wos_lengths,
a_g_n_k_wos_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_lengths},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_strides},
e_g_n_c_wis_lengths,
e_g_n_c_wis_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
out_element_op,
wei_element_op,
in_element_op);
if(!conv.IsSupportedArgument(argument))
{
printf("wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem\n");
return 1;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// c doesn't physically exist, any layout is fine
Tensor<float> c_host(in_g_n_c_wis_desc);
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
float,
WeiDataType,
OutDataType,
PassThrough,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(c_host,
wei,
out,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
PassThrough{},
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
// TODO: implement elementwise operation for host
in_host.ForEach(
[&](auto&, auto idx) { in_element_op(in_host(idx), c_host(idx), bias(idx)); });
in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData) ? 0 : 1;
}
return 0;
}
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "grouped_conv_bwd_data_bias_relu_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using OutDataType = ck::half_t;
using WeiDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using BiasDataType = ck::half_t; // bias
using InDataType = ck::half_t;
using OutLayout = ck::tensor_layout::convolution::GNHWK;
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
using BiasLayout = ck::tensor_layout::convolution::G_C;
using InLayout = ck::tensor_layout::convolution::GNHWC;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using CBiasInElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto ConvBwdDataDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvNdBwdDataInstance =
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<
NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
InLayout,
OutDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType>,
InDataType,
OutElementOp,
WeiElementOp,
CBiasInElementOp,
ConvBwdDataDefault,
true, // DoPadGemmM
true, // DoPadGemmN
1,
256,
128,
256,
32,
8,
2,
32,
32,
2,
4,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
0,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 2, 128, 256, 256, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = CBiasInElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 2)
{
// output image: GNHWK
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
// weight: GKYXC
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
// input image bias: G_C
const auto bias_g_n_c_wis_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.C_,
conv_param.input_spatial_lengths_[0],
conv_param.input_spatial_lengths_[1]},
{
conv_param.C_, // g
0, // n
1, // c
0, // hi
0 // wi
});
// input image: GNHWC
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
using DeviceInstance = DeviceConvNdBwdDataInstance<2>;
run_conv_bwd_data_bias_relu<2,
OutDataType,
WeiDataType,
BiasDataType,
InDataType,
OutElementOp,
WeiElementOp,
CBiasInElementOp,
DeviceInstance>(do_verification,
init_method,
time_kernel,
conv_param,
out_g_n_k_wos_desc,
wei_g_k_c_xs_desc,
bias_g_n_c_wis_desc,
in_g_n_c_wis_desc,
wei_element_op,
out_element_op,
in_element_op);
}
return 0;
}
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