ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-29 11:16:59 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
develop
composable_kernel/experimental/gemm_benchmark/run_mx_gemm_example_v3.inc
Illia Silin 717f2efef7 [rocm-libraries] ROCm/rocm-libraries#6978 (commit e58096d) 
[CK] add composable kernel support on gfx1250 (#6978)

## Motivation

Add composable kernel support on gfx1250.

## Technical Details

<!-- Explain the changes along with any relevant GitHub links. -->

## Test Plan

<!-- Explain any relevant testing done to verify this PR. -->

## Test Result

<!-- Briefly summarize test outcomes. -->

## Submission Checklist

- [ ] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

---------

Co-authored-by: Qun Lin <qlin@amd.com>
Co-authored-by: jialuo12_amdeng <jia.luo@amd.com>
Co-authored-by: Andriy Roshchenko <andriy.roshchenko@amd.com>
Co-authored-by: hsivasun_amdeng <haresh.sivasuntharampillai@amd.com>
2026-05-15 06:46:51 -07:00

379 lines
17 KiB
C++
Raw Permalink Blame History

// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <bool BPreShuffle, typename GemmInstance, typename ProblemType>
bool run_mx_gemm(GemmInstance& op_ptrs,
const ProblemType& problem_size,
const ExecutionConfig& config)
{
using BRefLayout = ck::conditional_t<BPreShuffle, Col, BLayout>;
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 KBatch = problem_size.KBatch;
auto f_host_tensor_descriptor =
[](ck::index_t row, ck::index_t col, ck::index_t stride, auto layout) {
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
return HostTensorDescriptor({row, col}, {stride, 1});
else
return HostTensorDescriptor({row, col}, {1, stride});
};
auto f_get_default_stride =
[](ck::index_t row, ck::index_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<ck::index_t>(col);
else
return static_cast<ck::index_t>(row);
}
else
return static_cast<ck::index_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{});
if(K % ScaleBlockSize != 0)
{
throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
};
if(K % ck::packed_size_v<ADataType> != 0 || K % ck::packed_size_v<BDataType> != 0)
{
throw std::runtime_error("wrong! K must be multiple of packed size.");
};
// Hardcode scale layouts as per pipeline assumptions
// TODO: Allow user to specify scale layouts
using AScaleLayout = Row;
using BScaleLayout = Col;
auto Scale_Padded_M = ck::math::integer_least_multiple(M, ScaleBlockSize);
auto Scale_Stride_AM =
f_get_default_stride(Scale_Padded_M, K / ScaleBlockSize, -1, AScaleLayout{});
auto Scale_Stride_BN = f_get_default_stride(K / ScaleBlockSize, N, -1, BScaleLayout{});
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
auto b_k_n =
std::make_shared<Tensor<BDataType>>(f_host_tensor_descriptor(K, N, StrideB, BRefLayout{}));
auto b_input = b_k_n;
if constexpr(BPreShuffle)
b_input = std::make_shared<Tensor<BDataType>>(
f_host_tensor_descriptor(K, N, StrideB, BRefLayout{})); // use layout only for size
// scales for A and B
Tensor<XDataType> a_m_k_scale(f_host_tensor_descriptor(
Scale_Padded_M, K / ScaleBlockSize, Scale_Stride_AM, AScaleLayout{}));
Tensor<XDataType> b_k_n_scale(
f_host_tensor_descriptor(K / ScaleBlockSize, N, Scale_Stride_BN, BScaleLayout{}));
// shuffled scales for A and B
Tensor<XDataType> a_shuffled_scale(f_host_tensor_descriptor(
Scale_Padded_M, K / ScaleBlockSize, Scale_Stride_AM, AScaleLayout{}));
Tensor<XDataType> b_shuffled_scale(
f_host_tensor_descriptor(K / ScaleBlockSize, N, Scale_Stride_BN, BScaleLayout{}));
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{})); // host verification
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{})); // device result downloaded to host
if(config.verbosity >= 0)
{
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "a_m_k_scale: " << a_m_k_scale.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n->mDesc << std::endl;
std::cout << "b_k_n_scale: " << b_k_n_scale.mDesc << std::endl;
std::cout << "c_m_n_device_result: " << c_m_n_device_result.mDesc << std::endl;
}
auto a_data_element = [](float x) { return ck::type_convert<ADataType>(x); };
auto b_data_element = [](float x) { return ck::type_convert<BDataType>(x); };
using int_distr = std::uniform_int_distribution<int>;
using float_distr = std::uniform_real_distribution<float>;
switch(config.init_method)
{
case 0: // Initializations for development and debugging
ck::utils::FillConstant<ADataType>{a_data_element(0.5f)}(a_m_k);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(2.0f)}(a_m_k_scale);
ck::utils::FillConstant<BDataType>{b_data_element(2.0f)}(*b_k_n);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(0.5f)}(b_k_n_scale);
if(config.verbosity > 0)
{
std::cout << "Init A = {0.5}" << std::endl;
std::cout << "Init A scale = {2.0}" << std::endl;
std::cout << "Init B = {2.0}" << std::endl;
std::cout << "Init B scale = {0.5}" << std::endl;
std::cout << "Expect C = {K}" << std::endl;
}
break;
case 1:
a_m_k.GenerateTensorDistr(
int_distr{-5, 5}, ck::identity{}, std::minstd_rand(time(nullptr))); // Z[-5,5]
b_k_n->GenerateTensorDistr(int_distr{-5, 5}); // Z[-5,5]
static_assert(ck::is_same_v<XDataType, ck::e8m0_bexp_t>);
a_m_k_scale.GenerateTensorDistr(int_distr{125, 128}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorDistr(int_distr{125, 128}); // scales: {0.25, 0.5, 1, 2}
break;
case 2:
a_m_k.GenerateTensorDistr(
float_distr{-2.0, 2.0}, ck::identity{}, std::minstd_rand(time(nullptr))); // R[-2,2]
a_m_k_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
b_k_n->GenerateTensorDistr(float_distr{-2.0, 2.0});
b_k_n_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
break;
case 3:
a_m_k.GenerateTensorDistr(float_distr{-2.0, 2.0}); // R[-2,2]
a_m_k_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
ck::utils::FillConstant<BDataType>{b_data_element(2.0f)}(*b_k_n);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(0.5f)}(b_k_n_scale);
break;
case 4:
ck::utils::FillConstant<ADataType>{a_data_element(0.5f)}(a_m_k);
ck::utils::FillConstant<XDataType>{ck::type_convert<XDataType>(2.0f)}(a_m_k_scale);
b_k_n->GenerateTensorDistr(float_distr{-2.0, 2.0});
b_k_n_scale.GenerateTensorDistr(float_distr{powf(2.0f, -125.0f), 1.0f});
break;
default:
if(config.verbosity > 0)
{
std::cout << "NOTE: No input data initialization." << std::endl;
}
}
if(ck::get_warp_size() == 64)
{
preShuffleScaleBuffer_gfx950<ck::is_same_v<ALayout, Row>>(a_m_k_scale.mData.data(),
a_shuffled_scale.mData.data(),
Scale_Padded_M,
K / ScaleBlockSize);
preShuffleScaleBuffer_gfx950<ck::is_same_v<BRefLayout, Col>>(
b_k_n_scale.mData.data(), b_shuffled_scale.mData.data(), N, K / ScaleBlockSize);
}
else if(ck::get_warp_size() == 32)
{
preShuffleScaleBuffer_gfx1250<ck::e8m0_bexp_t, ScaleBlockSize, ck::is_same_v<ALayout, Row>>(
a_m_k_scale.mData.data(),
a_shuffled_scale.mData.data(),
Scale_Padded_M,
K / ScaleBlockSize);
preShuffleScaleBuffer_gfx1250<ck::e8m0_bexp_t,
ScaleBlockSize,
ck::is_same_v<BRefLayout, Col>>(
b_k_n_scale.mData.data(), b_shuffled_scale.mData.data(), N, K / ScaleBlockSize);
}
else
{
throw std::runtime_error("wrong! unsupported warp size");
}
if constexpr(BPreShuffle)
{
int NPerXdl = 16; // Fixed 16
preShuffleBuffer(b_k_n->mData.data(), b_input->mData.data(), N, K, NPerXdl, KPack);
}
if(config.verbosity > 0)
std::cout << "Device memory allocation..." << std::endl;
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.GetElementSpaceSize());
DeviceMem a_scale_device_buf(sizeof(XDataType) * a_m_k_scale.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n->GetElementSpaceSize());
DeviceMem b_scale_device_buf(sizeof(XDataType) * b_k_n_scale.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.GetElementSpaceSize());
if(config.verbosity > 0)
std::cout << "Upload data to device..." << std::endl;
a_device_buf.ToDevice(a_m_k.mData.data());
a_scale_device_buf.ToDevice(a_shuffled_scale.mData.data());
b_device_buf.ToDevice(b_input->mData.data());
b_scale_device_buf.ToDevice(b_shuffled_scale.mData.data());
if(config.verbosity > 0)
std::cout << "Done." << std::endl;
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
bool pass = true;
if((config.do_verification == 1) || (config.do_verification == 3))
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMXGemm<ADataType,
BDataType,
CDataType,
AccDataType,
XDataType,
PassThrough,
PassThrough,
PassThrough,
float,
float>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
a_m_k_scale,
*b_k_n,
b_k_n_scale,
c_m_n_host_result,
PassThrough{},
PassThrough{},
PassThrough{});
ref_invoker.Run(ref_argument);
}
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// do GEMM
std::string best_op_name;
std::optional<std::string> best_op_object_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
size_t best_index = 0;
for(size_t i = 0; i < op_ptrs.size(); i++)
{
auto& gemm = op_ptrs[i];
auto invoker_ptr = gemm->MakeInvokerPointer();
float ave_time = 0;
auto argument_ptr =
gemm->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
a_scale_device_buf.GetDeviceBuffer(),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
b_scale_device_buf.GetDeviceBuffer(),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
Scale_Stride_AM,
StrideB,
Scale_Stride_BN,
StrideC,
KBatch,
a_element_op,
b_element_op,
c_element_op);
if(config.instance_index != -1 && config.instance_index != static_cast<int>(i))
{
continue;
}
if(!gemm->IsSupportedArgument(argument_ptr.get()))
{
std::cerr << gemm->GetTypeString() << " does not support this problem" << std::endl;
continue;
}
if(config.do_verification)
{
ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false, 1});
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
bool ret = ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType>(),
get_atol<CDataType>());
pass &= ret;
if(!ret)
{
std::cout << "Error: [" << i << "]: " << gemm->GetTypeString()
<< " results incorrect!" << std::endl;
}
}
if(config.time_kernel)
{
ave_time = invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr,
config.time_kernel,
0,
config.cold_niters,
config.nrepeat,
true,
config.rotating_count});
// Output size(M*N) * [dot product(2K) + product of scales(K/ScaleBlockSize) + scaling
// of partial sums(K/ScaleBlockSize)] FLOPS = 2 * M * N * K + 2 * M * N * K /
// ScaleBlockSize
std::size_t flop =
std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / ScaleBlockSize;
std::size_t num_btype = sizeof(ADataType) * M * K / ck::packed_size_v<ADataType> +
sizeof(BDataType) * K * N / ck::packed_size_v<BDataType> +
sizeof(CDataType) * M * N +
sizeof(XDataType) * M * K / ScaleBlockSize +
sizeof(XDataType) * N * K / ScaleBlockSize;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = static_cast<float>(num_btype) / 1e6f / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << " [" << i << "] " << gemm->GetTypeString() << std::endl;
if(tflops > best_tflops && ave_time > 1e-10)
{
std::string op_name = gemm->GetTypeString();
std::optional<std::string> op_obj_name = gemm->GetObjectName();
best_op_name = op_name;
best_op_object_name = op_obj_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = KBatch;
best_index = i;
}
}
}
if(config.time_kernel)
{
std::cout << "Best Perf for M = " << M << " N = " << N << " K = " << K
<< " StrideA = " << StrideA << " StrideB = " << StrideB
<< " StrideC = " << StrideC << " KBatch = " << best_kbatch << " : "
<< best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << " [" << best_index << "] " << best_op_name << std::endl;
if(best_op_object_name)
std::cout << best_op_object_name.value() << std::endl;
}
return pass;
}
template <bool BPreshuffle, typename GemmInstance>
bool run_mx_gemm_splitk_example(GemmInstance& op_ptrs, int argc, char* argv[])
{
ProblemSizeSplitK problem_size;
ExecutionConfig config;
return !parse_cmd_args(argc, argv, problem_size, config) ||
run_mx_gemm<BPreshuffle>(op_ptrs, problem_size, config);
}
Reference in New Issue View Git Blame Copy Permalink