ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Remove CK_USE_AMD_MFMA_GFX950 (#1935)

Browse Source 
* Add runtime check in example_gemm_xdl_streamk for gfx950

* Add runtime check in grouped conv fwd examples for gfx950

* Disable CK_USE_AMD_MFMA_GFX950

* Add new instances for gfx950

* Fix test_gemm_universal on gfx950
This commit is contained in:
jefyang1
2025-03-04 10:32:25 -08:00
committed by GitHub
parent 540a6da40b
commit c95bda93ba
186 changed files with 3272 additions and 883 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
example/01_gemm/gemm_xdl_streamk.cpp Executable file → Normal file
View File

@@ -27,22 +27,24 @@ using DeviceGemmStreamK = ck::tensor_operation::device::DeviceGemmXdlStreamK
// ######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| 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| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(CK_USE_AMD_MFMA_GFX950)
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
#else // defined(CK_USE_AMD_MFMA_GFX950)
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 256, 128, 128, 4, 8, 32, 32, 2, 2, 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, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 256, 256, 128, 4, 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<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>;
// < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 128, 32, 128, 4, 8, 32, 32, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 1, 1, 1, S<1, 32, 1, 4>, 8>;
#endif // defined(CK_USE_AMD_MFMA_GFX950)
// instance for double rate mfma instruction on gfx950
using DeviceGemmStreamK2 = ck::tensor_operation::device::DeviceGemmXdlStreamK
// ######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| K0Per| K1| 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| Type| | | | Elementwise| Elementwise| Elementwise| 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| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, 256, 256, 128, 4, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// // clang-format on
// clang-format on
using DeviceGemmInstance = DeviceGemmStreamK;
using DeviceGemmInstance = DeviceGemmStreamK;
using DeviceGemmInstance2 = DeviceGemmStreamK2;
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
@@ -58,6 +60,6 @@ using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALa
BElementOp,
CElementOp>;
#include "run_gemm_example.inc"
#include "run_gemm_example_streamk.inc"
int main(int argc, char* argv[]) { return !run_gemm_streamk_example(argc, argv); }

78
example/01_gemm/run_gemm_example.inc
View File

@@ -3,8 +3,6 @@
#pragma once
#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
@@ -124,23 +122,12 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
using BaseStreamK = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
float ave_time = 0;
if constexpr(std::is_same<ProblemType, ProblemSize>::value &&
!std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
if constexpr(std::is_same<ProblemType, ProblemSize>::value)
{
auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE
@@ -171,61 +158,6 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
}
else if constexpr(std::is_same<ProblemType, ProblemSizeStreamK>::value &&
std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
{
auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
problem_size.NumSKBlocks);
if(!gemm.IsSupportedArgument(argument))
{
std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
return true;
}
std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
if(workspace_size != 0)
{
workspace.Realloc(workspace_size);
gemm.SetWorkSpacePointer(&argument, workspace.GetDeviceBuffer());
}
ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
#if 0
// TODO!!!!!
if(workspace_size != 0){
float * ws_ptr = reinterpret_cast<float*>(malloc(workspace_size));
size_t ws_dwords = workspace_size / sizeof(float);
workspace.FromDevice(ws_ptr);
for(size_t i = 0; i < ws_dwords; i++) {
uint32_t rere = reinterpret_cast<uint32_t*>(ws_ptr)[i];
printf("%4lu : %f(0x%08x)\n", i, ws_ptr[i], rere);
}
free(ws_ptr);
}
#endif
}
else
{
// When the Problem Type and Problem Size does not fit.
@@ -319,11 +251,3 @@ bool run_gemm_example(int argc, char* argv[])
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
}
bool run_gemm_streamk_example(int argc, char* argv[])
{
ProblemSizeStreamK problem_size;
ExecutionConfig config;
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
}

270
example/01_gemm/run_gemm_example_streamk.inc Normal file
View File

@@ -0,0 +1,270 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/host_utility/device_prop.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
#endif
using namespace ck::literals;
auto M = problem_size.M;
auto N = problem_size.N;
auto K = problem_size.K;
auto StrideA = problem_size.StrideA;
auto StrideB = problem_size.StrideB;
auto StrideC = problem_size.StrideC;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
if(stride == -1)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<std::size_t>(col);
}
else
{
return static_cast<std::size_t>(row);
}
}
else
return static_cast<std::size_t>(stride);
};
StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
switch(config.init_method)
{
case 0:
ck::utils::FillConstant<ADataType>{ck::type_convert<ADataType>(1.f)}(a_m_k);
ck::utils::FillConstant<BDataType>{ck::type_convert<BDataType>(1.f)}(b_k_n);
break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
case 2:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break;
case 3:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
case 4:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{1.f, 1.f}(b_k_n);
break;
case 5:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-2.f, 2.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-2.f, 2.f}(b_k_n);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-0.1f, 0.1f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-0.1f, 0.1f}(b_k_n);
}
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_ref_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
#ifdef BUILD_INT4_EXAMPLE
DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
c_m_n_device_result.mDesc.GetElementSpaceSize());
const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
#else
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_ref_buf(sizeof(CDataType) *
c_m_n_device_ref_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
#endif
DeviceMem workspace;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
using BaseStreamK = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// do GEMM
static_assert(std::is_base_of<BaseStreamK, DeviceGemmInstance>::value &&
std::is_base_of<BaseStreamK, DeviceGemmInstance2>::value);
auto gemm = DeviceGemmInstance{};
auto gemm2 = DeviceGemmInstance2{}; // instance for double rate mfma instruction
BaseStreamK* op_ptr = (ck::get_device_name() == "gfx950") ? static_cast<BaseStreamK*>(&gemm2)
: static_cast<BaseStreamK*>(&gemm);
float ave_time = 0;
auto invoker_ptr = op_ptr->MakeInvokerPointer();
auto argument_ptr = op_ptr->MakeArgumentPointer(
#ifdef BUILD_INT4_EXAMPLE
static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
#endif
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
problem_size.NumSKBlocks);
if(!op_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cerr << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
return true;
}
auto argument = argument_ptr.get();
std::size_t workspace_size = op_ptr->GetWorkSpaceSize(argument);
if(workspace_size != 0)
{
workspace.Realloc(workspace_size);
op_ptr->SetWorkSpacePointer(argument, workspace.GetDeviceBuffer());
}
ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, config.time_kernel});
std::size_t flop = 2_uz * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, "
<< op_ptr->GetTypeString() << std::endl;
bool pass = true;
if((config.do_verification == 1) || (config.do_verification == 3))
{
// CPU verification
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
std::cout << "Running verification on CPU." << std::endl;
ref_invoker.Run(ref_argument);
#ifdef BUILD_INT4_EXAMPLE
Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
#endif
}
if((config.do_verification == 2) || (config.do_verification == 3))
{
// GPU verification
auto ref_gemm_gpu = ReferenceGemmInstanceGPU{};
auto ref_invoker_gpu = ref_gemm_gpu.MakeInvoker();
auto ref_argument_gpu = ref_gemm_gpu.MakeArgument(
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_ref_buf.GetDeviceBuffer()),
M,
N,
K,
a_element_op,
b_element_op,
c_element_op);
std::cout << "Running verification on GPU." << std::endl;
ref_invoker_gpu.Run(ref_argument_gpu, StreamConfig{});
c_m_n_device_ref_buf.FromDevice(c_m_n_device_ref_result.mData.data());
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_device_ref_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
}
return pass == true;
}
bool run_gemm_streamk_example(int argc, char* argv[])
{
ProblemSizeStreamK problem_size;
ExecutionConfig config;
return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
}