mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Wmma support for multiple ABD GEMM (#2803)
* multi_abd wmma support: - Add multiple A and B support to multiple D implementation (gridwise level) - Add multi_abd GEMM (device level) - Add instances (xdl parity) - Add tests (both xdl and wmma) - Add examples - Add ckProfiler support (both xdl and wmma) * Fix bug in device print function * Fix unused template parameter * Fix batched gemm for multiABD gridwise implementation * Fix gemm_universal_reduce with multiABDs gridwise implementation --------- Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com>
This commit is contained in:
Enrico Degregori
424
profiler/include/profiler/profile_gemm_multi_abd_impl.hpp
Normal file
424
profiler/include/profiler/profile_gemm_multi_abd_impl.hpp
Normal file
@@ -0,0 +1,424 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iomanip>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/tensor_operation_instance/gpu/gemm_multi_abd.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 {
|
||||
|
||||
// this function is also defined in CK but because of the way we use it in
|
||||
// profile_gemm_multi_impl, it requires the arguments to not be const
|
||||
template <typename... X, typename... Y>
|
||||
auto concat_tuple_of_refs(ck::Tuple<X&...>& tx, ck::Tuple<Y&...>& ty)
|
||||
{
|
||||
return ck::unpack2(
|
||||
[&](auto&&... zs) { return ck::Tuple<decltype(zs)...>{ck::forward<decltype(zs)>(zs)...}; },
|
||||
tx,
|
||||
ty);
|
||||
}
|
||||
|
||||
template <typename AsDataType,
|
||||
typename BsDataType,
|
||||
typename AccDataType,
|
||||
typename DsDataType,
|
||||
typename EDataType,
|
||||
typename AsLayout,
|
||||
typename BsLayout,
|
||||
typename DsLayout,
|
||||
typename ELayout,
|
||||
typename AElementOp,
|
||||
typename BElementOp,
|
||||
typename CDEElementOp>
|
||||
bool profile_gemm_multi_abd_impl(int do_verification,
|
||||
int init_method,
|
||||
bool /*do_log*/,
|
||||
bool time_kernel,
|
||||
int M,
|
||||
int N,
|
||||
int K,
|
||||
int StrideA,
|
||||
int StrideB,
|
||||
int StrideD,
|
||||
int StrideE)
|
||||
{
|
||||
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});
|
||||
}
|
||||
};
|
||||
|
||||
static constexpr index_t NumATensor = AsDataType::Size();
|
||||
auto as_m_k = generate_tuple(
|
||||
[&](auto i) {
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
using ALayout = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
|
||||
|
||||
return Tensor<ADataType>(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
},
|
||||
Number<NumATensor>{});
|
||||
|
||||
static constexpr index_t NumBTensor = BsDataType::Size();
|
||||
auto bs_k_n = generate_tuple(
|
||||
[&](auto i) {
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
using BLayout = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
|
||||
|
||||
return Tensor<BDataType>(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
},
|
||||
Number<NumBTensor>{});
|
||||
|
||||
static constexpr index_t NumDTensor = DsDataType::Size();
|
||||
auto ds_m_n = generate_tuple(
|
||||
[&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
|
||||
|
||||
return Tensor<DDataType>(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
|
||||
},
|
||||
Number<NumDTensor>{});
|
||||
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
static_for<0, NumATensor, 1>{}(
|
||||
[&](auto i) { std::cout << "a" << i.value << "_m_k: " << as_m_k(i).mDesc << std::endl; });
|
||||
static_for<0, NumBTensor, 1>{}(
|
||||
[&](auto i) { std::cout << "b" << i.value << "_k_n: " << bs_k_n(i).mDesc << std::endl; });
|
||||
static_for<0, NumDTensor, 1>{}(
|
||||
[&](auto i) { std::cout << "d" << i.value << "_m_n: " << ds_m_n(i).mDesc << std::endl; });
|
||||
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
|
||||
as_m_k(i).GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
});
|
||||
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
|
||||
bs_k_n(i).GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
});
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
|
||||
ds_m_n(i).GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
|
||||
});
|
||||
|
||||
break;
|
||||
default:
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
|
||||
as_m_k(i).GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
});
|
||||
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
|
||||
bs_k_n(i).GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
});
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
|
||||
ds_m_n(i).GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
|
||||
});
|
||||
}
|
||||
|
||||
const auto a_element_op = AElementOp{};
|
||||
const auto b_element_op = BElementOp{};
|
||||
const auto cde_element_op = CDEElementOp{};
|
||||
|
||||
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleABD<AsLayout,
|
||||
BsLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
AsDataType,
|
||||
BsDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp>;
|
||||
|
||||
// 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
|
||||
if(do_verification)
|
||||
{
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
Tensor<AccDataType> c_m_n({M, N});
|
||||
|
||||
using AComputeType =
|
||||
typename std::conditional<(NumATensor > 1),
|
||||
EDataType,
|
||||
remove_cvref_t<tuple_element_t<0, AsDataType>>>::type;
|
||||
|
||||
auto get_a_matrix = [&]() -> auto {
|
||||
// in case of pass through we avoid allocating a new
|
||||
// tensor and copying values
|
||||
if constexpr(is_same_v<AElementOp, PassThrough>)
|
||||
{
|
||||
return as_m_k(Number<0>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
Tensor<AComputeType> a_m_k({M, K});
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
// result
|
||||
auto data_refs1 = ck::tie(a_m_k(m, k));
|
||||
// inputs
|
||||
auto data_refs2 =
|
||||
generate_tie([&](auto i) -> auto& { return as_m_k(Number<i>{})(m, k); },
|
||||
Number<NumATensor>{});
|
||||
auto data_refs = concat_tuple_of_refs(data_refs1, data_refs2);
|
||||
unpack(a_element_op, data_refs);
|
||||
}
|
||||
}
|
||||
return a_m_k;
|
||||
}
|
||||
};
|
||||
|
||||
using BComputeType =
|
||||
typename std::conditional<(NumBTensor > 1),
|
||||
EDataType,
|
||||
remove_cvref_t<tuple_element_t<0, BsDataType>>>::type;
|
||||
|
||||
auto get_b_matrix = [&]() -> auto {
|
||||
// in case of pass through we avoid allocating a new
|
||||
// tensor and copying values
|
||||
if constexpr(is_same_v<AElementOp, PassThrough>)
|
||||
{
|
||||
return bs_k_n(Number<0>{});
|
||||
}
|
||||
else
|
||||
{
|
||||
Tensor<BComputeType> b_k_n({K, N});
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
// result
|
||||
auto data_refs1 = ck::tie(b_k_n(k, n));
|
||||
// inputs
|
||||
auto data_refs2 =
|
||||
generate_tie([&](auto i) -> auto& { return bs_k_n(Number<i>{})(k, n); },
|
||||
Number<NumBTensor>{});
|
||||
auto data_refs = concat_tuple_of_refs(data_refs1, data_refs2);
|
||||
unpack(b_element_op, data_refs);
|
||||
}
|
||||
}
|
||||
return b_k_n;
|
||||
}
|
||||
};
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<AComputeType,
|
||||
BComputeType,
|
||||
AccDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
get_a_matrix(), get_b_matrix(), c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
// compulsory
|
||||
auto data_refs1 = ck::tie(e_m_n_host_result(m, n), c_m_n(m, n));
|
||||
// optional (if multiple Ds)
|
||||
auto data_refs2 =
|
||||
generate_tie([&](auto i) -> auto& { return ds_m_n(Number<i>{})(m, n); },
|
||||
Number<NumDTensor>{});
|
||||
auto data_refs = concat_tuple_of_refs(data_refs1, data_refs2);
|
||||
unpack(cde_element_op, data_refs);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
std::array<DeviceMem*, NumATensor> as_device_buf;
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
as_device_buf[i] = new DeviceMem(sizeof(ADataType) * as_m_k(i).mDesc.GetElementSpaceSize());
|
||||
});
|
||||
|
||||
std::array<DeviceMem*, NumBTensor> bs_device_buf;
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
bs_device_buf[i] = new DeviceMem(sizeof(BDataType) * bs_k_n(i).mDesc.GetElementSpaceSize());
|
||||
});
|
||||
|
||||
std::array<DeviceMem*, NumDTensor> ds_device_buf;
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
|
||||
ds_device_buf[i] = new DeviceMem(sizeof(DDataType) * ds_m_n(i).mDesc.GetElementSpaceSize());
|
||||
});
|
||||
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
static_for<0, NumATensor, 1>{}(
|
||||
[&](auto i) { as_device_buf[i]->ToDevice(as_m_k(i).mData.data()); });
|
||||
|
||||
static_for<0, NumBTensor, 1>{}(
|
||||
[&](auto i) { bs_device_buf[i]->ToDevice(bs_k_n(i).mData.data()); });
|
||||
|
||||
static_for<0, NumDTensor, 1>{}(
|
||||
[&](auto i) { ds_device_buf[i]->ToDevice(ds_m_n(i).mData.data()); });
|
||||
|
||||
std::string best_op_name;
|
||||
float best_ave_time = 0;
|
||||
float best_tflops = 0;
|
||||
float best_gb_per_sec = 0;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
// profile device operation instances
|
||||
for(auto& op_ptr : op_ptrs)
|
||||
{
|
||||
std::array<const void*, NumATensor> as_pointer;
|
||||
std::array<ck::index_t, NumATensor> as_stride;
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
as_pointer[i] = as_device_buf[i]->GetDeviceBuffer();
|
||||
as_stride[i] = StrideA;
|
||||
});
|
||||
|
||||
std::array<const void*, NumBTensor> bs_pointer;
|
||||
std::array<ck::index_t, NumBTensor> bs_stride;
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
bs_pointer[i] = bs_device_buf[i]->GetDeviceBuffer();
|
||||
bs_stride[i] = StrideB;
|
||||
});
|
||||
std::array<const void*, NumDTensor> ds_pointer;
|
||||
std::array<ck::index_t, NumDTensor> ds_stride;
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) {
|
||||
ds_pointer[i] = ds_device_buf[i]->GetDeviceBuffer();
|
||||
ds_stride[i] = StrideD;
|
||||
});
|
||||
|
||||
auto argument_ptr = op_ptr->MakeArgumentPointer(as_pointer,
|
||||
bs_pointer,
|
||||
ds_pointer,
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
as_stride,
|
||||
bs_stride,
|
||||
ds_stride,
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
auto invoker_ptr = op_ptr->MakeInvokerPointer();
|
||||
|
||||
std::string op_name = op_ptr->GetTypeString();
|
||||
|
||||
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
// re-init E to zero before profiling a kernel
|
||||
e_device_buf.SetZero();
|
||||
|
||||
float ave_time =
|
||||
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
|
||||
std::size_t sizeADataType = 0;
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) {
|
||||
using ADataType = remove_cvref_t<tuple_element_t<i.value, AsDataType>>;
|
||||
sizeADataType = std::max(sizeADataType, sizeof(ADataType));
|
||||
});
|
||||
std::size_t sizeBDataType = 0;
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) {
|
||||
using BDataType = remove_cvref_t<tuple_element_t<i.value, BsDataType>>;
|
||||
sizeBDataType = std::max(sizeBDataType, sizeof(BDataType));
|
||||
});
|
||||
|
||||
std::size_t num_btype =
|
||||
sizeADataType * M * K + sizeBDataType * K * N + sizeof(EDataType) * 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 << std::endl;
|
||||
if(tflops > best_tflops)
|
||||
{
|
||||
best_op_name = op_name;
|
||||
best_tflops = tflops;
|
||||
best_ave_time = ave_time;
|
||||
best_gb_per_sec = gb_per_sec;
|
||||
}
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
std::cout << op_name << " does not support this problem" << std::endl;
|
||||
}
|
||||
}
|
||||
|
||||
static_for<0, NumATensor, 1>{}([&](auto i) { delete as_device_buf[i]; });
|
||||
|
||||
static_for<0, NumBTensor, 1>{}([&](auto i) { delete bs_device_buf[i]; });
|
||||
|
||||
static_for<0, NumDTensor, 1>{}([&](auto i) { delete ds_device_buf[i]; });
|
||||
|
||||
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
|
||||
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
|
||||
|
||||
return pass;
|
||||
}
|
||||
|
||||
} // namespace profiler
|
||||
} // namespace ck
|
||||
Reference in New Issue
Block a user