ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-20 04:49:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

External Interface (#304)

Browse Source 
* add client example

* clean

* clean

* reorg

* clean up profiler

* reorg

* clea

* fix profiler

* function for getinstances

* update client example

* update client example

* update client example

* update

* update example

* update Jenkins file

* update cmake

* update Jenkins

[ROCm/composable_kernel commit: aebd211c36]
This commit is contained in:
Chao Liu
2022-06-26 19:39:02 -05:00
committed by GitHub
parent 73af96b913
commit 9096feca63
73 changed files with 2199 additions and 1961 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
test/batched_gemm/batched_gemm_util.hpp
View File

@@ -1,109 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef BATCHED_GEMM_UTILS_HPP
#define BATCHED_GEMM_UTILS_HPP
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
namespace ck {
namespace batched_gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename BatchedGemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostBatchedGemm(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_batched_gemm = BatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument =
ref_batched_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceBatchedGemm(DeviceGemmPtr& batched_gemm_ptr,
const ck::batched_gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_g_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_g_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_g_m_k_device_buf.ToDevice(A.mData.data());
b_g_k_n_device_buf.ToDevice(B.mData.data());
const auto batch_count = A.mDesc.GetLengths()[0];
auto invoker_ptr = batched_gemm_ptr->MakeInvokerPointer();
auto argument_ptr = batched_gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op,
batch_count);
if(!batched_gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_g_m_n_device_buf.FromDevice(C.mData.data());
}
} // namespace batched_gemm_util
} // namespace ck
#endif

121
test/gemm/gemm_util.hpp
View File

@@ -214,6 +214,11 @@ struct TestGemm
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::bhalf_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
@@ -234,121 +239,5 @@ struct TestGemm
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck

77
test/gemm/gemm_xdl_bf16.cpp
View File

@@ -47,6 +47,11 @@ void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
int main()
{
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using AccDataType = float;
using RowMajor = ck::tensor_layout::gemm::RowMajor;
using ColumnMajor = ck::tensor_layout::gemm::ColumnMajor;
@@ -58,13 +63,17 @@ int main()
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
@@ -73,13 +82,17 @@ int main()
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
ColumnMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
@@ -88,13 +101,17 @@ int main()
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
RowMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
gemmPtrs.clear();
@@ -103,13 +120,17 @@ int main()
for(auto& gemmPtr : gemmPtrs)
{
res &= ck::gemm_util::TestGemmBF16<DeviceGemmNoOpPtr,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
res &= ck::gemm_util::TestGemm<DeviceGemmNoOpPtr,
ADataType,
BDataType,
CDataType,
AccDataType,
RowMajor,
ColumnMajor,
RowMajor,
PassThrough,
PassThrough,
PassThrough>{}(gemmPtr);
}
std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;

10
test/gemm/gemm_xdl_fp16.cpp
View File

@@ -38,10 +38,12 @@ void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#if 0
void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#endif
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
@@ -69,8 +71,10 @@ int main()
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(gemmPtrs);
@@ -92,8 +96,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemmPtrs);
@@ -115,8 +121,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(gemmPtrs);
@@ -138,8 +146,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemmPtrs);
ck::tensor_operation::device::device_gemm_instance::

10
test/gemm/gemm_xdl_fp32.cpp
View File

@@ -38,10 +38,12 @@ void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNo
void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#if 0
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
#endif
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
@@ -67,8 +69,10 @@ int main()
std::vector<DeviceGemmNoOpPtr> gemmPtrs;
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(gemmPtrs);
@@ -90,8 +94,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemmPtrs);
@@ -113,8 +119,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(gemmPtrs);
@@ -136,8 +144,10 @@ int main()
gemmPtrs.clear();
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
#if 0
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
#endif
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);

2
test/gemm_split_k/CMakeLists.txt
View File

@@ -1,3 +1,3 @@
add_test_executable(test_gemm_split_k gemm_split_k.cpp)
target_link_libraries(test_gemm_split_k PRIVATE host_tensor)
target_link_libraries(test_gemm_split_k PRIVATE device_gemm_instance)
target_link_libraries(test_gemm_split_k PRIVATE device_gemm_splitk_instance)

23
test/gemm_split_k/gemm_split_k.cpp
View File

@@ -15,7 +15,6 @@
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/host_tensor/host_gemm.hpp"
@@ -28,20 +27,24 @@ enum struct GemmMatrixLayout
KM_NK_MN, // 3
};
using DeviceGemmNoOpPtr =
ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
using DeviceGemmSplitKNoOpPtr = ck::tensor_operation::device::DeviceGemmSplitKPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmSplitKNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmSplitKNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmSplitKNoOpPtr>&);
void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmSplitKNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
@@ -150,7 +153,7 @@ int test_gemm(const gemmArgs& args)
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
// add device GEMM instances
std::vector<DeviceGemmNoOpPtr> gemm_ptrs;
std::vector<DeviceGemmSplitKNoOpPtr> gemm_ptrs;
if(args.layout == GemmMatrixLayout::MK_KN_MN)
{