Skip to content

GitLab

Commit f27c50a7 authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

working version

parent 0ea428c3
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 614 additions and 67 deletions
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_splitk_xdl_cshuffle_two_stage.hpp
View file @ f27c50a7
......@@ -153,6 +153,9 @@ struct DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage
PipelineVer,
ComputeDataType>;
// CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
// indexy 1,3 -> MPerBlock, NPerBlock || podzielone przez MPerBlock -> NPerThread
template <typename ELay>
static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
{
......@@ -216,10 +219,8 @@ template <typename ELay>
static constexpr auto MakeElementwiseInputSequence()
{
return generate_sequence_v2(
[&](auto i) constexpr { return Number<i+1-i>{}; },
[&]([[maybe_unused]] auto i) constexpr { return Number<CDEShuffleBlockTransferScalarPerVector_NPerBlock>{}; },
Number<NumDTensor+1>{});
//CShuffleNXdlPerWavePerShuffle
}
using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
......@@ -227,20 +228,21 @@ template <typename ELay>
using DsGridDesc_M_N = decltype(MakeDsGridDescriptor_M_N({}, {}, {}));
using DsGridPointer = decltype(MakeDsGridPointer());
using CDGridDesc_M_N = decltype(concat_tuple(ck::Tuple<CGridDesc_M_N>{}, DsGridDesc_M_N{}));
//using CDDataTypes = decltype(concat_tuple(ck::Tuple<WorkspaceDataType*>{}, DsDataType{}));
using CDDataTypes = decltype(concat_tuple(ck::Tuple<WorkspaceDataType*>{}, DsGridPointer{}));
using ElementwiseInputSequence = decltype(MakeElementwiseInputSequence());
using GridwiseElementwise = GridwiseElementwise_2D<CDGridDesc_M_N, // zmien na C, D_0, ..., D_n / tuple<C, D_0, ..., D_N>
static constexpr index_t ClusterLengthMPerBlock = CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(1);
static constexpr index_t ClusterLengthNPerBlock = CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock::At(3);
using GridwiseElementwise = GridwiseElementwise_2D<CDGridDesc_M_N,
ck::Tuple<EGridDesc_M_N>,
CDDataTypes, // zmien na C, D_0, ..., D_n / tuple<C, D_0, ..., D_N>
CDDataTypes,
ck::Tuple<EDataType*>,
CDEElementwiseOperation,
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // MPerThread
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // NPerThread
MPerBlock / ClusterLengthMPerBlock,
NPerBlock / ClusterLengthNPerBlock,
ElementwiseInputSequence,
ck::Sequence<8>>;
ck::Sequence<CDEShuffleBlockTransferScalarPerVector_NPerBlock>>;
using Block2ETileMapKSplit =
BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>;
......@@ -324,10 +326,9 @@ template <typename ELay>
gemm_kernel_args_.reserve(group_count_);
elementwise_c_grid_descs_m_n_.reserve(group_count_);
elementwise_d_grid_descs_m_n_.reserve(group_count_);
ds_grid_pointer_.reserve(group_count_);
group_grid_size_.reserve(group_count_);
for(std::size_t i = 0; i < gemm_descs.size(); ++i)
{
......@@ -371,6 +372,7 @@ template <typename ELay>
const index_t block_end = grid_size_ + grid_size_grp;
grid_size_ += grid_size_grp;
group_grid_size_[i] = grid_size_grp;
// block-to-e-tile map
auto grouped_block_2_ctile_map =
GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
......@@ -449,12 +451,25 @@ template <typename ELay>
auto grouped_block_2_ctile_map =
GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
group_grid_size_[i] = grid_size_grp;
karg.KPadded = k_padded;
karg.K0Padded = k0_padded;
karg.k_batch = K_BATCH;
gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
gemm_kernel_args_[i].block_start_ = block_start;
gemm_kernel_args_[i].block_end_ = block_end;
#if DEBUG_LOG
index_t tiles = (block_end - block_start) / K_BATCH;
std::cout << "block_start: " << block_start << "\n"
<< "block_end: " << block_end << "\n"
<< "tiles: " << tiles << std::endl << std::endl;
std::cout << "KPadded: " << karg.KPadded << std::endl
<< "K0Padded: " << karg.K0Padded << std::endl
<< "KBatch: " << karg.k_batch << std::endl
<< "grid_size_: " << karg.KPadded << std::endl;
#endif
}
}
......@@ -515,6 +530,7 @@ template <typename ELay>
std::vector<std::array<const void*, NumDTensor>>& p_Ds_;
std::vector<std::array<index_t, NumDTensor>> stride_Ds_;
std::vector<GemmTransKernelArg> gemm_kernel_args_;
std::vector<index_t> group_grid_size_;
std::vector<CGridDesc_M_N> elementwise_c_grid_descs_m_n_;
std::vector<DsGridDesc_M_N> elementwise_d_grid_descs_m_n_;
......@@ -755,7 +771,7 @@ template <typename ELay>
for(int i=0; i < arg.group_count_; ++i) {
time += launch_and_time_kernel(stream_config,
elementwise_kernel,
dim3(arg.grid_size_), // chyba group_grid_size <<< tak zmienic na group_grid_size[i]
dim3(arg.group_grid_size_[i]), // chyba group_grid_size <<< tak zmienic na group_grid_size[i]
dim3(BlockSize),
0,
concat_tuple(make_tuple(arg.elementwise_c_grid_descs_m_n_[i]), arg.elementwise_d_grid_descs_m_n_[i]),
......@@ -763,10 +779,9 @@ template <typename ELay>
concat_tuple(make_tuple(arg.gemm_kernel_args_[i].karg_.p_c_grid), arg.ds_grid_pointer_[i]),
type_convert<EDataType*>(arg.e_ptrs_[i]),
arg.cde_element_op_,
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // num_threads_m
CDEShuffleBlockTransferScalarPerVector_NPerBlock); // num_threads_n
ClusterLengthMPerBlock, // num_threads_m
ClusterLengthNPerBlock); // num_threads_n
}
return time;
}
};
......@@ -870,7 +885,7 @@ template <typename ELay>
auto str = std::stringstream();
// clang-format off
str << "DeviceGroupedGemm_XdlSplitKTileLoop"
str << "DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage"
<< "<"
<< std::string(ALayout::name)[0] << ","
<< std::string(BLayout::name)[0] << ","
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ f27c50a7
......@@ -37,9 +37,9 @@ __global__ void
kernel_grouped_gemm_xdl_splitk(
const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
const index_t group_count,
const AElementwiseOperation a_element_op = AElementwiseOperation{},
const BElementwiseOperation b_element_op = BElementwiseOperation{},
const CElementwiseOperation c_element_op = CElementwiseOperation{})
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
......@@ -206,7 +206,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
static constexpr index_t B2E_M01 = 8;
using GroupedGemmBlock2ETileMap = OffsettedBlockToCTileMap<Block2ETileMapKSplit>;
using KernelArgument = typename GridwiseGemm::Argument;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
struct GemmTransKernelArg
{
KernelArgument karg_;
......@@ -450,7 +450,10 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
dim3(BlockSize),
0,
cast_pointer_to_constant_address_space(arg.p_workspace_),
arg.gemm_kernel_args_.size());
arg.gemm_kernel_args_.size(),
PassThrough{},
PassThrough{},
PassThrough{});
};
if(all_have_main_k0_block_loop)
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
View file @ f27c50a7
......@@ -146,18 +146,18 @@ void add_device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instances(
PassThrough,
PassThrough>>>& instances);
// void add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(
// std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
// Row,
// Empty_Tuple,
// Row,
// F16,
// F16,
// Empty_Tuple,
// F16,
// PassThrough,
// PassThrough,
// PassThrough>>>& instances)
void add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <typename ALayout,
typename BLayout,
......@@ -203,7 +203,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
op_ptrs);
// add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_multiple_d_xdl_two_stage_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
......
library/include/ck/library/utility/check_err.hpp
View file @ f27c50a7
......@@ -55,7 +55,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
......@@ -106,7 +106,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
......@@ -156,7 +156,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
......@@ -211,7 +211,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << " out[" << i << "] != ref[" << i << "]: " << o << " != " << r
<< std::endl;
......@@ -260,7 +260,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
......@@ -305,7 +305,7 @@ check_err(const Range& out,
{
max_err = err > max_err ? err : max_err;
err_count++;
if(err_count < 5)
if(err_count < 50000)
{
std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
<< "] != ref[" << i << "]: " << o << " != " << r << std::endl;
......
library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
View file @ f27c50a7
......@@ -9,5 +9,5 @@ add_instance_library(device_grouped_gemm_instance
device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp
#device_grouped_gemm_multiple_d_splitk_xdl_two_stage_f16_f16_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_multiple_d_splitk_xdl_two_stage_f16_f16_f16_mk_kn_mn_instance.cpp
)
profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
View file @ f27c50a7
......@@ -270,15 +270,16 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
for(std::size_t i = 0; i < gemm_descs.size(); i++)
c_device_buf[i]->SetZero();
std::cout << "p1\n";
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
std::cout << "p2\n";
if(do_verification)
{
bool instance_pass = true;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
std::cout << "p3\n";
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
......@@ -316,10 +317,10 @@ bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
pass = pass && instance_pass;
}
std::cout << "p4\n";
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::cout << "p5\n";
if(time_kernel)
{
std::size_t flop = 0, num_btype = 0;
......
profiler/include/profiler/profile_grouped_gemm_two_stage_impl.hpp 0 → 100644
View file @ f27c50a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multiple_d_splitk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.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/utility/fill.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
bool profile_grouped_gemm_two_stage_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const std::vector<int>& Ms,
const std::vector<int>& Ns,
const std::vector<int>& Ks,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
{
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});
}
};
std::size_t group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
{
throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
}
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(
Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
c_m_n_host_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
#if DEBUG_LOG
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
#endif // DEBUG_LOG
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
}
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{};
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
a_device_buf.reserve(group_count);
b_device_buf.reserve(group_count);
c_device_buf.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_c.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(std::size_t i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
b_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
c_device_buf.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
b_k_n[i],
c_m_n_host_results[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
}
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(p_a,
p_b,
p_ds,
p_c,
gemm_descs,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string gemm_name = gemm_ptr->GetTypeString();
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmMultipleDSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// skip non-splitk grouped_gemm
if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
{
continue;
}
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(kbatch > 0)
{
kbatch_list = {kbatch};
}
for(std::size_t j = 0; j < kbatch_list.size(); j++)
{
auto kbatch_curr = kbatch_list[j];
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
DeviceMem gemm_arg_dev_mem(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())->GetDeviceKernelArgSize(argument_ptr.get()));
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())->SetDeviceKernelArgs(argument_ptr.get(), gemm_arg_dev_mem.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
gemm_desc_workspace.SetZero();
for(std::size_t i = 0; i < gemm_descs.size(); i++)
c_device_buf[i]->SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
bool instance_pass = true;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
0.06);
}
else
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i]);
}
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_results[i].mData, ",")
<< std::endl;
}
}
std::cout << "Instance: " << gemm_name << " verification "
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
pass = pass && instance_pass;
}
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
if(time_kernel)
{
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
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, " << gemm_name << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
}
else
{
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
<< std::endl;
}
}
}
if(time_kernel)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
<< std::endl;
}
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ f27c50a7
......@@ -54,6 +54,7 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
endif()
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
......
profiler/src/profile_grouped_gemm_two_stage.cpp 0 → 100644
View file @ f27c50a7
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_grouped_gemm_two_stage_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
F8_F16_F16, // 4
F16_F8_F16, // 5
};
#define OP_NAME "grouped_gemm_two_stage"
#define OP_DESC "Grouped GEMM TwoStage"
namespace {
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
{
out.push_back(std::stoi(item));
}
return out;
}
int profile_grouped_gemm_two_stage(int argc, char* argv[])
{
if(argc < 14)
{
std::cout
<< "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: fp8@fp6; 5: f16@f8)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
<< " 1: A[m, k] * B[n, k] = C[m, n];\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0=n0, 1=yes)\n"
<< "arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n"
<< "arg15: kbatch value (default 1)\n"
<< "optional:\n"
<< "arg16: number of warm-up cycles (default 1)\n"
<< "arg17: number of iterations (default 10)\n"
<< std::endl;
exit(1);
}
const auto data_type = static_cast<GemmDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const auto Ms = argToIntArray(argv[8]);
const auto Ns = argToIntArray(argv[9]);
const auto Ks = argToIntArray(argv[10]);
auto StrideAs = argToIntArray(argv[11]); //a: mk b: kn, c: mn: stride a =
auto StrideBs = argToIntArray(argv[12]);
auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
const int DefaultStrideA = Ks[0];
const int DefaultStrideB = Ns[0];
const int DefaultStrideC = Ns[0];
for(size_t i=0; i<Ms.size(); ++i) {
StrideAs[i] = StrideAs[i] == -1 ? DefaultStrideA : StrideAs[i];
StrideBs[i] = StrideBs[i] == -1 ? DefaultStrideB : StrideBs[i];
StrideCs[i] = StrideCs[i] == -1 ? DefaultStrideC : StrideCs[i];
}
int n_warmup = 1;
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
}
#ifdef CK_ENABLE_FP16
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatch,
n_warmup,
n_iter);
}
else
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
#endif
return 0;
}
} // anonymous namespace
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_two_stage);
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment