mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-19 12:30:16 +00:00
2d67076cc0
* Fixed cmake errors related to gemm_bilinear. Previously, if the above flags are set, cmake build fails: GPU_TARGETS="gfx1100;gfx1201" -D DTYPES="fp16;bf16;fp8"
* Fixed cmake build errors related to test_fp8
* Updates to support mixed precision
* Adding support for RRR, F8xF16xF16 gemm_universal_wmma - wip
* Added support for F8xF16xF16 to gemm_wmma_universal
* Added support for F16xF8xF16 to gemm_wmma_universal
* Added support for BF16xI4xBF16 to gemm_wmma_universal
* Added support for F16xI4xF16 to gemm_wmma_universal
* Fixed IsSupportedArgument to check ComputeTypeA, ComputeTypeB instead of ADataType, BDataType
* Added missing test class for FP16_KM_NK
* Pre-commit hooks fixes
* Added padding instances for f16xf16xf16
* Fixed cmake errors related to gemm_bilinear. Previously, if the above flags are set, cmake build fails: GPU_TARGETS="gfx1100;gfx1201" -D DTYPES="fp16;bf16;fp8"
* Fixed cmake build errors related to test_fp8
* Ammending changes for adding support for padding instances for f16xf16xf16
* Fixes for padding instances for f16xf16xf16
* Added padding instances for bf16xbf16, f8xf8
* Added packed instances for bf16xi4xbf16
* Added padding instances for f8xf16xf16
* Added padding instances for f16xf8xf16, f16xi4xf16
* Fixed typos for bf16xbf16xbf16 padding instances
* Fixed typos for padded instances
* Added tests for fp16, KM_KN and KM_NK
* Padding not supported for when BDataType is pk_i4_t. Added fix for correct check and removed padding instances.
* Fixed typos
* Updated the set of tests for FP16
* Updated the set of tests for FP16
* Fix typo
* Moved f16xi4 test under the correct data layout group
* example for gemm_universal_bf16
* Adding examples for gemm_wmma instances
* Added the missing parameters
* Fixed review comments and added executable to cmakeLists
* Fixing clang format
* Fixing build erros
* Fixed compilation failure.
* Modified some code as per gemm_universal_examples
* Fixed the gemm specialization error
* Fixed the build errors.
* Fix strides of a/b_thread_desc
The descriptors are larger than needed (even though the compiler don't alloc registers for unused values).
* Load in M/NRepeat dims with thread copy's slice instead of a loop
* Clone BlockwiseGemmXdlops_pipeline_v1 for WMMA implementation
* Implement Intrawave and Interwave variants of pipeline v1
* Add instances for Interwave and Intrawave v1
* Add instances with ABlockLdsExtraM and BBlockLdsExtraN = 0
* Remove instances that are too slow (mostly because of register spilling)
* Add a workaround for fp8/bf8->f32 packed conversion issue
* Add instances for Interwave and Intrawave v1
* Enable profiling of mixed precision with f8 and int4 on WMMA
* Fix segfault in profiler when B is pk_i4_t
b_device_buf's size in bytes is larger than b_k_n_permute so b_device_buf.ToDevice reads out-of-bounds.
* Remove instances that are too slow (mostly because of register spilling)
* Add missing add_device_gemm_wmma_universal_f8_f8_bf16 declarations
* Add test case for bf16_i4
* Add missing Regular tests
* Add test_gemm_universal_xdl/wmma_fp16 to REGRESSION_TESTS
They take more than 30 seconds
* Fix a bug that fp16_i4 validation passes only with PermuteB
A permutation required by conversion from pk_i4_t to half_t does not
depend on PermuteB, they can be used independently.
* Use PermuteB with f16_i4 in most instances (as xdl)
Some instances use PermuteB = false for checking correctness.
See also the previous commit.
* Fix cache flushing for pk_i4
* Add mixed precision examples
* Disable all tests and instances with f8 on gfx11
Even though f8_f16 and f16_f8 don't require f8 WMMA instructions,
gfx11 still lacks hardware instructions for fast f8->f32 conversion.
* Add FP16 KM_NK and KM_KN test suites for XDL
These tests were added to common .inc for better testing of WMMA instances
* Fix int8 DTYPES check for gemm_bilinear
---------
Co-authored-by: Anca Hamuraru <anca@streamhpc.com>
Co-authored-by: Apoorva Kalyani <apoorva@streamhpc.com>
[ROCm/composable_kernel commit: 52b4860a30]
414 lines
16 KiB
C++
414 lines
16 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#pragma once
|
|
|
|
#include <iomanip>
|
|
#include <iostream>
|
|
#include <typeinfo>
|
|
|
|
#include "ck/ck.hpp"
|
|
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
|
#include "ck/tensor_operation/gpu/device/device_gemm_v2.hpp"
|
|
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
|
|
|
#include "ck/library/tensor_operation_instance/gpu/gemm_universal.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/literals.hpp"
|
|
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
|
|
|
namespace ck {
|
|
namespace profiler {
|
|
|
|
template <typename ADataType,
|
|
typename BDataType,
|
|
typename ComputeDataType,
|
|
typename AccDataType,
|
|
typename CDataType,
|
|
typename ALayout,
|
|
typename BLayout,
|
|
typename CLayout>
|
|
bool profile_gemm_universal_impl(int do_verification,
|
|
int init_method,
|
|
bool do_log,
|
|
bool time_kernel,
|
|
int M,
|
|
int N,
|
|
int K,
|
|
int StrideA,
|
|
int StrideB,
|
|
int StrideC,
|
|
int KBatch,
|
|
int n_warmup,
|
|
int n_iter,
|
|
uint64_t rotating = 0)
|
|
{
|
|
bool pass = true;
|
|
|
|
auto f_host_tensor_descriptor =
|
|
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
|
using namespace ck::literals;
|
|
|
|
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
|
}
|
|
else
|
|
{
|
|
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
|
}
|
|
};
|
|
|
|
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{}));
|
|
Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
|
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{}));
|
|
|
|
std::size_t total_gemm_needed =
|
|
a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes();
|
|
int rotating_count = std::max(
|
|
1,
|
|
std::min(n_iter,
|
|
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
|
|
|
|
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_device_result.mDesc << std::endl;
|
|
std::cout << "rotating count: " << rotating_count << std::endl;
|
|
|
|
switch(init_method)
|
|
{
|
|
case 0: break;
|
|
case 1:
|
|
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
|
|
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
|
|
break;
|
|
case 2:
|
|
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
|
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
|
break;
|
|
default:
|
|
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
|
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
|
|
}
|
|
|
|
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
|
|
|
|
const auto a_element_op = AElementOp{};
|
|
const auto b_element_op = BElementOp{};
|
|
const auto c_element_op = CElementOp{};
|
|
|
|
DeviceMem a_device_buf(a_m_k.GetElementSpaceSizeInBytes());
|
|
DeviceMem b_device_buf(b_k_n_permute.GetElementSpaceSizeInBytes());
|
|
DeviceMem c_device_buf(c_m_n_device_result.GetElementSpaceSizeInBytes());
|
|
|
|
a_device_buf.ToDevice(a_m_k.mData.data());
|
|
|
|
using DeviceOp = ck::tensor_operation::device::DeviceGemmV2<ALayout,
|
|
BLayout,
|
|
CLayout,
|
|
ADataType,
|
|
BDataType,
|
|
CDataType,
|
|
AElementOp,
|
|
BElementOp,
|
|
CElementOp>;
|
|
|
|
// get device op instances
|
|
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
|
|
DeviceOp>::GetInstances();
|
|
|
|
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
|
|
|
|
// Run reference GEMM
|
|
if(do_verification)
|
|
{
|
|
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
|
BDataType,
|
|
CDataType,
|
|
AccDataType,
|
|
AElementOp,
|
|
BElementOp,
|
|
CElementOp,
|
|
ComputeDataType>;
|
|
|
|
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);
|
|
|
|
ref_invoker.Run(ref_argument);
|
|
}
|
|
|
|
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;
|
|
|
|
// profile device GEMM instances
|
|
for(auto& op_ptr : op_ptrs)
|
|
{
|
|
const int KPerBlock = op_ptr->GetKPerBlock();
|
|
|
|
if(op_ptr->GetPermuteB())
|
|
{
|
|
int K1 = KPerBlock;
|
|
int K0 = K / KPerBlock;
|
|
|
|
// int K0, N, K1
|
|
for(int j = 0; j < K0; j++)
|
|
{
|
|
for(int i = 0; i < N; i++)
|
|
{
|
|
for(int jj = 0; jj < K1; jj++)
|
|
{
|
|
b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
|
|
}
|
|
}
|
|
}
|
|
}
|
|
else
|
|
{
|
|
b_k_n_permute = b_k_n;
|
|
}
|
|
|
|
#if CK_USE_PK4_LAYOUT_SHUFFLE
|
|
// Conversion from pk_i4_t to half_t expects a particular permutation
|
|
if constexpr(is_same_v<BDataType, pk_i4_t> && is_same_v<ComputeDataType, half_t>)
|
|
{
|
|
// vector pk_i4x4 permute
|
|
for(int i = 0; i < N; i++)
|
|
{
|
|
for(int j = 0; j < K; j += 8)
|
|
{
|
|
int input[8];
|
|
|
|
for(int k = 0; k < 4; k++)
|
|
{
|
|
int i4x2 = b_k_n_permute(j + k * 2, i).data;
|
|
input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
|
|
input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
|
|
}
|
|
|
|
// permute 01234567->20643175
|
|
{
|
|
int hi = input[2];
|
|
int lo = input[0];
|
|
int i4x2 = (hi << 4) | lo;
|
|
|
|
b_k_n_permute(j + 0, i) = i4x2;
|
|
}
|
|
|
|
{
|
|
int hi = input[6];
|
|
int lo = input[4];
|
|
int i4x2 = (hi << 4) | lo;
|
|
|
|
b_k_n_permute(j + 2, i) = i4x2;
|
|
}
|
|
|
|
{
|
|
int hi = input[3];
|
|
int lo = input[1];
|
|
int i4x2 = (hi << 4) | lo;
|
|
|
|
b_k_n_permute(j + 4, i) = i4x2;
|
|
}
|
|
|
|
{
|
|
int hi = input[7];
|
|
int lo = input[5];
|
|
int i4x2 = (hi << 4) | lo;
|
|
|
|
b_k_n_permute(j + 6, i) = i4x2;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
#endif
|
|
|
|
b_device_buf.ToDevice(b_k_n_permute.mData.data());
|
|
|
|
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38};
|
|
|
|
if(KBatch > 0)
|
|
{
|
|
kbatch_list = {KBatch};
|
|
}
|
|
|
|
for(std::size_t i = 0; i < kbatch_list.size(); i++)
|
|
{
|
|
auto kbatch_curr = kbatch_list[i];
|
|
|
|
auto argument_ptr =
|
|
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
|
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
|
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
|
M,
|
|
N,
|
|
K,
|
|
StrideA,
|
|
StrideB,
|
|
StrideC,
|
|
kbatch_curr,
|
|
a_element_op,
|
|
b_element_op,
|
|
c_element_op);
|
|
|
|
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
|
|
|
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
|
{
|
|
|
|
// re-init C to zero before profiling next kernel
|
|
c_device_buf.SetZero();
|
|
|
|
invoker_ptr->Run(argument_ptr.get(),
|
|
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
|
|
|
|
if(do_verification)
|
|
{
|
|
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
|
|
|
#if defined CK_ENABLE_FP8
|
|
// set softer tolerances for fp8
|
|
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
|
|
is_same_v<CDataType, f8_t>)
|
|
{
|
|
std::string msg = "Error: Incorrect results!";
|
|
double rtol = 1e-1;
|
|
double atol = 1e-1;
|
|
pass = pass & ck::utils::check_err(
|
|
c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
|
|
}
|
|
else
|
|
{
|
|
#endif
|
|
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
|
|
#if defined CK_ENABLE_FP8
|
|
}
|
|
#endif
|
|
|
|
if(do_log)
|
|
{
|
|
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
|
|
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
|
|
LogRangeAsType<float>(
|
|
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
|
|
<< std::endl;
|
|
LogRangeAsType<float>(
|
|
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
|
|
<< std::endl;
|
|
}
|
|
}
|
|
|
|
std::string op_name = op_ptr->GetTypeString();
|
|
std::optional<std::string> op_obj_name = op_ptr->GetObjectName();
|
|
|
|
float ave_time = invoker_ptr->Run(argument_ptr.get(),
|
|
StreamConfig{nullptr,
|
|
time_kernel,
|
|
0,
|
|
n_warmup,
|
|
n_iter,
|
|
rotating_count > 1,
|
|
rotating_count});
|
|
|
|
std::size_t flop = std::size_t(2) * M * N * K;
|
|
|
|
static constexpr index_t BPackedSize = []() {
|
|
if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
|
|
return 2;
|
|
else
|
|
return 1;
|
|
}();
|
|
|
|
std::size_t num_btype = sizeof(ADataType) * M * K +
|
|
sizeof(BDataType) * K * N / BPackedSize +
|
|
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: " << std::setw(10) << ave_time << " ms, " << tflops
|
|
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
|
|
<< kbatch_curr << std::endl;
|
|
|
|
if(tflops > best_tflops && ave_time > 1e-10)
|
|
{
|
|
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_curr;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
std::cout << op_ptr->GetTypeString() << " does not support this problem"
|
|
<< std::endl;
|
|
}
|
|
}
|
|
}
|
|
|
|
if constexpr(is_same<CDataType, float>::value)
|
|
{
|
|
std::cout << "Best Perf for datatype = f32";
|
|
}
|
|
else if constexpr(is_same<CDataType, half_t>::value)
|
|
{
|
|
std::cout << "Best Perf for datatype = f16";
|
|
}
|
|
else if constexpr(is_same<CDataType, bhalf_t>::value)
|
|
{
|
|
std::cout << "Best Perf for datatype = bf16";
|
|
}
|
|
else if constexpr(is_same<CDataType, int8_t>::value)
|
|
{
|
|
std::cout << "Best Perf for datatype = int8";
|
|
}
|
|
|
|
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
std::cout << " ALayout = RowMajor";
|
|
}
|
|
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
|
|
{
|
|
std::cout << " ALayout = ColumnMajor";
|
|
}
|
|
|
|
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
|
|
{
|
|
std::cout << " BLayout = RowMajor";
|
|
}
|
|
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
|
|
{
|
|
std::cout << " BLayout = ColumnMajor";
|
|
}
|
|
|
|
std::cout << " 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_op_name << std::endl;
|
|
|
|
if(best_op_object_name)
|
|
std::cout << best_op_object_name.value() << std::endl;
|
|
|
|
return pass;
|
|
}
|
|
|
|
} // namespace profiler
|
|
} // namespace ck
|