mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
39dc25a9b8
* Refactor universal gemm policy. * Adapt example to refactor changes. * Introduce static encoding pattern * Adding shuffled encoding patterns. * Fix err in reverse tuple. * Add transpose_tile2d * Small refactoring + doc * Enable reading on contiguous dimension in all layouts. * Transpose A/B register tile if needed for comp v3 pipeline. * Take contiguous dim size when calculating dram vector load size. * A/B smem pack size taken from WarpGemm attributes * Update B LDS layout and setup tile distribution pattern at class level. * Fix static assert. * Fix errors in examples. * Formatting & fix IsTranspose * Fix VectorSize & refactor. * Add error loging messages. * Fix VecLoadSize and TranspseC for mem pipeline. * Update unit-tests & disable mem pipeline. * Clang format * Update include/ck_tile/core/tensor/tile_window.hpp Co-authored-by: jakpiase <jakub.piasecki@amd.com> * Fix compilation and reviewers comments. * Refactor unit-test. Fallback to non-universal gemm. Need to use GemmPipelineAGmemBGmemCRegV1 for now, since GemmKernel is now supporting also non-K major vector reads. --------- Co-authored-by: jakpiase <jakub.piasecki@amd.com>
220 lines
10 KiB
C++
220 lines
10 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
|
|
#pragma once
|
|
|
|
template <typename Layout>
|
|
static constexpr inline auto is_row_major(Layout layout_)
|
|
{
|
|
return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
|
|
ck_tile::tensor_layout::gemm::RowMajor>>{};
|
|
}
|
|
|
|
auto calculate_rtol_atol(const ck_tile::index_t K,
|
|
const ck_tile::index_t kbatch,
|
|
const float max_accumulated_value)
|
|
{
|
|
using ComputeType =
|
|
std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
|
|
// Calculate thresholds
|
|
const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
|
|
ck_tile::integer_divide_ceil(K, kbatch));
|
|
const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
|
|
max_accumulated_value / kbatch, ck_tile::integer_divide_ceil(K, kbatch));
|
|
// Calculate error due to split_k accumulation
|
|
const auto rtol_split_k =
|
|
ck_tile::get_relative_threshold<CDataType, CDataType, CDataType>(kbatch);
|
|
const auto atol_split_k = ck_tile::get_absolute_threshold<CDataType, CDataType, CDataType>(
|
|
max_accumulated_value, kbatch);
|
|
// Use higher threshold
|
|
return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
|
|
}
|
|
|
|
template <typename ALayout, typename BLayout, typename CLayout>
|
|
float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
|
|
ck_tile::DeviceMem& b_k_n_dev_buf,
|
|
ck_tile::DeviceMem& c_m_n_dev_buf,
|
|
ck_tile::index_t M,
|
|
ck_tile::index_t N,
|
|
ck_tile::index_t K,
|
|
ck_tile::index_t stride_A,
|
|
ck_tile::index_t stride_B,
|
|
ck_tile::index_t stride_C,
|
|
ck_tile::index_t kbatch,
|
|
int n_warmup,
|
|
int n_repeat)
|
|
{
|
|
ck_tile::GemmHostArgs args;
|
|
args.a_ptr = a_m_k_dev_buf.GetDeviceBuffer();
|
|
args.b_ptr = b_k_n_dev_buf.GetDeviceBuffer();
|
|
args.c_ptr = c_m_n_dev_buf.GetDeviceBuffer();
|
|
args.k_batch = kbatch;
|
|
args.M = M;
|
|
args.N = N;
|
|
args.K = K;
|
|
args.stride_A = stride_A;
|
|
args.stride_B = stride_B;
|
|
args.stride_C = stride_C;
|
|
|
|
float ave_time = gemm_calc<ALayout, BLayout, CLayout>(
|
|
args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
|
|
|
|
std::size_t flop = std::size_t(2) * M * N * K;
|
|
std::size_t num_byte =
|
|
sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
|
|
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
|
float gb_per_sec = num_byte / 1.E6 / ave_time;
|
|
|
|
std::cout << "Run Gemm kernel with M =" << M << " N =" << N << " K =" << K
|
|
<< " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideC =" << stride_C
|
|
<< " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
|
<< std::endl;
|
|
|
|
return ave_time;
|
|
}
|
|
|
|
template <typename ALayout, typename BLayout, typename CLayout>
|
|
int run_gemm_example_with_layouts(int argc,
|
|
char* argv[],
|
|
const ALayout a_layout = ALayout{},
|
|
const BLayout b_layout = BLayout{},
|
|
[[maybe_unused]] const CLayout c_layout = CLayout{})
|
|
{
|
|
auto [result, arg_parser] = create_args(argc, argv);
|
|
if(!result)
|
|
return -1;
|
|
|
|
ck_tile::index_t M = arg_parser.get_int("m");
|
|
ck_tile::index_t N = arg_parser.get_int("n");
|
|
ck_tile::index_t K = arg_parser.get_int("k");
|
|
|
|
ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
|
|
ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
|
|
ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
|
|
|
|
ck_tile::index_t kbatch = arg_parser.get_int("split_k");
|
|
int n_warmup = arg_parser.get_int("warmup");
|
|
int n_repeat = arg_parser.get_int("repeat");
|
|
|
|
stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
|
|
stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
|
|
stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
|
|
|
|
ck_tile::HostTensor<ADataType> a_m_k(
|
|
ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
|
|
ck_tile::HostTensor<BDataType> b_k_n(
|
|
ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
|
|
ck_tile::HostTensor<CDataType> c_m_n_dev_result(
|
|
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
|
|
|
// TODO: add different init types
|
|
ck_tile::FillUniformDistribution<ADataType>{-5.f, 5.f}(a_m_k);
|
|
ck_tile::FillUniformDistribution<BDataType>{-5.f, 5.f}(b_k_n);
|
|
|
|
ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
|
|
ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
|
|
ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
|
|
|
|
a_m_k_dev_buf.ToDevice(a_m_k.data());
|
|
b_k_n_dev_buf.ToDevice(b_k_n.data());
|
|
c_m_n_dev_buf.SetZero();
|
|
c_m_n_dev_result.SetZero();
|
|
|
|
invoke_gemm<ALayout, BLayout, CLayout>(a_m_k_dev_buf,
|
|
b_k_n_dev_buf,
|
|
c_m_n_dev_buf,
|
|
M,
|
|
N,
|
|
K,
|
|
stride_A,
|
|
stride_B,
|
|
stride_C,
|
|
kbatch,
|
|
n_warmup,
|
|
n_repeat);
|
|
|
|
c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
|
|
bool pass = true;
|
|
|
|
if(arg_parser.get_int("v") == 1)
|
|
{
|
|
ck_tile::HostTensor<CDataType> c_m_n_host_ref(
|
|
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
|
c_m_n_host_ref.SetZero();
|
|
|
|
ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
|
|
a_m_k, b_k_n, c_m_n_host_ref);
|
|
const float max_accumulated_value =
|
|
*std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
|
|
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
|
|
pass = ck_tile::check_err(c_m_n_dev_result,
|
|
c_m_n_host_ref,
|
|
"Error: Incorrect results!",
|
|
rtol_atol.at(ck_tile::number<0>{}),
|
|
rtol_atol.at(ck_tile::number<1>{}));
|
|
|
|
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
|
|
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
|
|
<< std::endl;
|
|
std::cout << "The CPU veification result is:" << (pass ? "correct" : "fail") << std::endl;
|
|
}
|
|
else if(arg_parser.get_int("v") == 2)
|
|
{
|
|
ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
|
|
ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
|
|
ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
|
|
c_m_n_gpu_ref.SetZero();
|
|
c_m_n_gpu_buf_ref.SetZero();
|
|
|
|
ADataType* d_A;
|
|
BDataType* d_B;
|
|
CDataType* d_C;
|
|
|
|
ck_tile::hip_check_error(hipMalloc(&d_A, M * K * sizeof(ADataType)));
|
|
ck_tile::hip_check_error(hipMalloc(&d_B, N * K * sizeof(BDataType)));
|
|
ck_tile::hip_check_error(hipMalloc(&d_C, M * N * sizeof(CDataType)));
|
|
|
|
ck_tile::hip_check_error(hipMemcpy(d_A,
|
|
a_m_k_dev_buf.GetDeviceBuffer(),
|
|
M * K * sizeof(ADataType),
|
|
hipMemcpyHostToDevice));
|
|
ck_tile::hip_check_error(hipMemcpy(d_B,
|
|
b_k_n_dev_buf.GetDeviceBuffer(),
|
|
N * K * sizeof(BDataType),
|
|
hipMemcpyHostToDevice));
|
|
|
|
ck_tile::reference_gemm_gpu<ADataType,
|
|
BDataType,
|
|
AccDataType,
|
|
CDataType,
|
|
ALayout,
|
|
BLayout,
|
|
CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
|
|
|
|
ck_tile::hip_check_error(hipMemcpy(c_m_n_gpu_buf_ref.GetDeviceBuffer(),
|
|
d_C,
|
|
M * N * sizeof(CDataType),
|
|
hipMemcpyDeviceToHost));
|
|
|
|
ck_tile::hip_check_error(hipFree(d_A));
|
|
ck_tile::hip_check_error(hipFree(d_B));
|
|
ck_tile::hip_check_error(hipFree(d_C));
|
|
|
|
c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
|
|
const float max_accumulated_value =
|
|
*std::max_element(c_m_n_gpu_ref.mData.begin(), c_m_n_gpu_ref.mData.end());
|
|
const auto rtol_atol = calculate_rtol_atol(K, kbatch, max_accumulated_value);
|
|
pass = ck_tile::check_err(c_m_n_dev_result,
|
|
c_m_n_gpu_ref,
|
|
"Error: Incorrect results!",
|
|
rtol_atol.at(ck_tile::number<0>{}),
|
|
rtol_atol.at(ck_tile::number<1>{}));
|
|
|
|
std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
|
|
<< " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
|
|
<< std::endl;
|
|
std::cout << "The GPU veification result is: " << (pass ? "correct" : "fail") << std::endl;
|
|
}
|
|
|
|
return pass;
|
|
}
|