Skip to content

GitLab

Commit ef5e60f6 authored by illsilin's avatar illsilin
Browse files

Merge branch 'develop' into gfx950

parents 2cc0fa26 5e93fa9e
Hide whitespace changes
Inline Side-by-side
Showing with 932 additions and 685 deletions
library/src/tensor_operation_instance/gpu/mha/CMakeLists.txt
View file @ ef5e60f6
......@@ -6,7 +6,7 @@ set(CK_TILE_SRC_FOLDER ${CMAKE_SOURCE_DIR}/include/ck_tile/)
# CK Codegen requires dataclass which is added in Python 3.7
# Python version 3.8 is required for general good practice as it is default for Ubuntu 20.04
if(NOT CK_USE_ALTERNATIVE_PYTHON)
find_package(PythonInterp 3 REQUIRED)
find_package(Python3 COMPONENTS Interpreter Development)
else()
message("Using alternative python version")
set(EXTRA_PYTHON_PATH)
......@@ -33,7 +33,7 @@ set(FMHA_KNOWN_APIS "fwd,fwd_splitkv,fwd_appendkv,bwd")
# Note: The receipt 3 arg filters the generated backwards instances to reduce compilation time.
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
execute_process(
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--list_blobs ${FMHA_CPP_FOLDER}/blob_list.txt
--api ${FMHA_KNOWN_APIS}
--receipt 3
......@@ -50,7 +50,7 @@ endif()
# With receipt 3 set, we are generating instances for datatype == {fp16 || bfp16}, bias == {no || alibi}, deterministic == off, and dpad == dvpad.
add_custom_command(
OUTPUT ${FMHA_GEN_BLOBS}
COMMAND ${PYTHON_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
COMMAND ${Python3_EXECUTABLE} ${FMHA_SRC_FOLDER}/generate.py
--output_dir ${FMHA_CPP_FOLDER}
--api ${FMHA_KNOWN_APIS}
--receipt 3
......
profiler/README.md
View file @ ef5e60f6
[Back to the main page](../README.md)
# Composable Kernel profiler
## Profile GEMM kernels
```bash
#arg1: tensor operation (gemm=GEMM)
......@@ -180,3 +182,13 @@ Note: Column to image kernel adds to the output memory, this will cause output b
################ op datatype verify init log time dim0 dim1 dim2 in_stride0 in_stride1 in_stride2 out_stride0 out_stride1 out_stride2
./bin/ckProfiler permute_scale 0 1 1 0 1 64 64 64 4096 64 1 1 64 4096
```
## Convert MIOpen driver command to CKProfiler
```bash
python3 ../script/convert_miopen_driver_to_profiler.py
/opt/rocm/bin/MIOpenDriver conv -n 32 -c 64 -H 28 -W 28 -k 64 -y 3 -x 3
-p 1 -q 1 -u 2 -v 2 -l 1 -j 1 -m conv -g 32 -F 1 -t 1
```
Only convolution driver is supported.
profiler/include/profiler/profile_gemm_universal_batched_impl.hpp 0 → 100644
View file @ ef5e60f6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.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_batched_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementOp,
typename BElementOp,
typename CElementOp,
typename DeviceOp>
bool profile_gemm_universal_batched_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int BatchStrideA,
int BatchStrideB,
int BatchStrideC,
int StrideA,
int StrideB,
int StrideC,
int BatchCount,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
bool pass = true;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
}
else
{
return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB, BatchStrideB, BLayout{}));
Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
int total_gemm_needed =
a_g_m_k.GetElementSpaceSizeInBytes() + b_g_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_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
}
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_verification)
{
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
CDataType,
float,
AElementOp,
BElementOp,
CElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
// 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;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device op instances
for(auto& op_ptr : op_ptrs)
{
std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
// false branch for multi d dl kernel
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,
BatchCount,
StrideA,
StrideB,
{},
StrideC,
BatchStrideA,
BatchStrideB,
{},
BatchStrideC,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
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();
std::string op_name = op_ptr->GetTypeString();
float ave_time = invoker_ptr->Run(
argument_ptr.get(),
StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter, true, rotating_count});
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N) *
BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << 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)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
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 << " B = " << BatchCount << " M = " << M << " N = " << N << " K = " << K
<< " StrideA = " << StrideA << " StrideB = " << StrideB << " StrideC = " << StrideC
<< ": " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View file @ ef5e60f6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -17,7 +17,6 @@
#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"
......@@ -42,11 +41,14 @@ bool profile_grouped_gemm_impl(int do_verification,
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)
const std::vector<int>& kbatches = {},
int n_warmup = 1,
int n_iter = 10)
{
bool pass = true;
// TODO: Fixme - we do not pass compute data type here but need it
// to compute error thresholds.
using ComputeDataType = ADataType;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
......@@ -75,6 +77,7 @@ bool profile_grouped_gemm_impl(int do_verification,
std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results;
ComputeDataType max_abs_in_val = 0.f;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
......@@ -93,17 +96,18 @@ bool profile_grouped_gemm_impl(int do_verification,
<< i << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
}
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);
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-2.f, 2.f}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-2.f, 2.f}(b_k_n[i]);
max_abs_in_val = 2.f;
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);
ck::utils::FillUniformDistribution<ADataType>{-0.5f, 0.5f}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5f, 0.5f}(b_k_n[i]);
max_abs_in_val = 0.5f;
}
}
......@@ -164,7 +168,20 @@ bool profile_grouped_gemm_impl(int do_verification,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
// If kbatch would be bigger than 1, then we will use SplitK version.
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_ptrs.size() <= 0)
......@@ -205,7 +222,6 @@ bool profile_grouped_gemm_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
}
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
......@@ -221,43 +237,44 @@ bool profile_grouped_gemm_impl(int do_verification,
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
std::size_t workspace_size = gemm_ptr->GetWorkSpaceSize(argument_ptr.get());
std::size_t kargs_size = gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get());
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string gemm_name = gemm_ptr->GetTypeString();
DeviceMem gemm_workspace, gemm_kargs;
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<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)
// The following is necessary since TwoStage kernel is using additional memory both
// for Workspace and kernel arguments.
if(kargs_size > 0)
{
continue;
gemm_kargs.Realloc(kargs_size);
gemm_ptr->SetDeviceKernelArgs(argument_ptr.get(), gemm_kargs.GetDeviceBuffer());
}
if(workspace_size > 0 && workspace_size != kargs_size)
{
gemm_workspace.Realloc(workspace_size);
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_workspace.GetDeviceBuffer());
}
std::string gemm_name = gemm_ptr->GetTypeString();
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(kbatch > 0)
// If the user will provide not empty kbatches list, then we test predefined set of kbatch
// values.
if(!kbatches.empty())
{
kbatch_list = {kbatch};
kbatch_list = kbatches;
}
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);
if(kbatch_curr > 1 && dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) != nullptr)
{
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
}
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
......@@ -272,23 +289,18 @@ bool profile_grouped_gemm_impl(int 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]);
}
auto atol = ck::utils::get_absolute_threshold<ComputeDataType, CDataType>(
max_abs_in_val, gemm_descs[i].K_);
auto rtol = ck::utils::get_relative_threshold<ComputeDataType, CDataType>(
gemm_descs[i].K_);
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
rtol,
atol);
if(do_log)
{
......@@ -311,11 +323,12 @@ bool profile_grouped_gemm_impl(int do_verification,
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)
{
float ave_time =
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
......
profiler/include/profiler/profile_grouped_gemm_multiply_tile_loop_impl.hpp
View file @ ef5e60f6
......@@ -143,8 +143,7 @@ bool profile_grouped_gemm_multiply_tile_loop_impl(int do_verification,
p_ds.reserve(group_count);
p_e.reserve(group_count);
using KernelArguments =
ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<NumDTensor>;
using KernelArguments = ck::tensor_operation::device::GroupedGemmKernelArgument<NumDTensor>;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<KernelArguments> gemm_kargs;
......
profiler/include/profiler/profile_grouped_gemm_tile_loop_impl.hpp
View file @ ef5e60f6
......@@ -127,7 +127,7 @@ bool profile_grouped_gemm_tile_loop_impl(int do_verification,
p_b.reserve(group_count);
p_c.reserve(group_count);
using KernelArguments = ck::tensor_operation::device::GroupedGemmTileLoopKernelArguments<>;
using KernelArguments = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<KernelArguments> gemm_kargs;
......
profiler/include/profiler/profile_grouped_gemm_two_stage_impl.hpp deleted 100644 → 0
View file @ 2cc0fa26
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, 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(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
{
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;
}
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 @ ef5e60f6
......@@ -43,13 +43,12 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
endif()
list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94" OR SUPPORTED_GPU_TARGETS MATCHES "gfx95")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
endif()
......@@ -59,6 +58,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_batched.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
......@@ -135,12 +135,13 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94" OR SUPPORTED_GPU_TARGETS MATCHES "gfx95")
if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
endif()
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_batched_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
......
profiler/src/profile_gemm_universal_batched.cpp 0 → 100644
View file @ ef5e60f6
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_universal_batched_impl.hpp"
#include "profiler_operation_registry.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_universal_batched.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
BF16_BF16_BF16, // 0
F8_F8_BF16, // 1
};
#define OP_NAME "gemm_universal_batched"
#define OP_DESC "Batched GEMM Universal"
int profile_batched_gemm_universal(int argc, char* argv[])
{
if(argc != 18 && argc != 21)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: bf16, 1: fp8->bf16)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
printf(" 1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
printf(" 2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
printf(" 3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
printf("optional:\n");
printf("arg18: number of warm-up cycles (default 1)\n");
printf("arg19: number of iterations (default 10)\n");
printf("arg20: memory for rotating buffer (default 0, size in MB)\n");
// clang-format on
exit(1);
}
int n_warmup = 1;
int n_iter = 10;
uint64_t rotating = 0;
if(argc == 21)
{
n_warmup = std::stoi(argv[18]);
n_iter = std::stoi(argv[19]);
rotating = std::stoull(argv[20]) * 1024 * 1024;
}
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 int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
#endif
using BF16 = ck::bhalf_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile =
[&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using DsDataType = ck::Tuple<>;
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using DsLayout = ck::Tuple<>;
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmV2MultiD<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
bool pass = ck::profiler::profile_gemm_universal_batched_impl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp,
DeviceOp>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount,
n_warmup,
n_iter,
rotating);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F8_F8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F8{}, BF16{}, Row{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_universal);
profiler/src/profile_gemm_universal_streamk.cpp 100644 → 100755
View file @ ef5e60f6
......@@ -85,8 +85,10 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
// using BF16 = ck::bhalf_t;
// using F8 = ck::f8_t;
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
using F8 = ck::f8_t;
#endif
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -145,6 +147,24 @@ int profile_gemm_universal_streamk(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#if defined(CK_USE_FP8_ON_UNSUPPORTED_ARCH) || defined(CK_USE_GFX94)
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_grouped_gemm.cpp
View file @ ef5e60f6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
......@@ -39,16 +39,13 @@ 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;
}
......@@ -69,7 +66,7 @@ int profile_grouped_gemm(int argc, char* argv[])
<< "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"
<< "arg15: kbatch values (default 1)\n"
<< "optional:\n"
<< "arg16: number of warm-up cycles (default 1)\n"
<< "arg17: number of iterations (default 10)\n"
......@@ -92,7 +89,7 @@ int profile_grouped_gemm(int argc, char* argv[])
const auto StrideAs = argToIntArray(argv[11]);
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
const auto kbatches = argc >= 15 ? argToIntArray(argv[14]) : std::vector<int>{};
int n_warmup = 1;
int n_iter = 10;
......@@ -102,7 +99,6 @@ int profile_grouped_gemm(int argc, char* argv[])
n_iter = std::stoi(argv[16]);
}
#ifdef CK_ENABLE_FP16
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::half_t,
......@@ -121,7 +117,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -143,7 +139,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -165,7 +161,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -187,7 +183,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -209,7 +205,7 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
......@@ -231,7 +227,73 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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,
kbatches,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_grouped_gemm_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
kbatches,
n_warmup,
n_iter);
}
......@@ -239,7 +301,6 @@ int profile_grouped_gemm(int argc, char* argv[])
{
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
}
#endif
return 0;
}
......
profiler/src/profile_grouped_gemm_fixed_nk.cpp
View file @ ef5e60f6
......@@ -32,9 +32,7 @@ namespace {
std::vector<int> argToIntArray(char* input)
{
std::vector<int> out;
std::istringstream in(input);
std::string item;
while(std::getline(in, item, ','))
......@@ -83,7 +81,7 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
const auto StrideAs = argToIntArray(argv[11]);
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
const int kbatch = argc >= 15 ? std::stoi(argv[14]) : 1;
using F32 = float;
using F16 = ck::half_t;
......@@ -97,8 +95,8 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
n_warmup = std::stoi(argv[15]);
n_iter = std::stoi(argv[16]);
}
#if defined(CK_ENABLE_BF16) && defined(CK_ENABLE_INT8)
......
profiler/src/profile_grouped_gemm_two_stage.cpp deleted 100644 → 0
View file @ 2cc0fa26
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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
{
F16_F16_F16, // 0
BF16_INT8_BF16, // 1
BF16_BF16_BF16 // 2
};
#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: fp16; 1: bf16@int8; 2: bf16)\n"
<< "arg3: matrix layout (0: A[m, k] * B[k, n] = 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]);
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]);
}
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::BF16_INT8_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
int8_t,
ck::bhalf_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::BF16_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
int8_t,
ck::bhalf_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 if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_grouped_gemm_two_stage_impl<ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_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");
}
return 0;
}
} // anonymous namespace
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_two_stage);
python/ck4inductor/batched_universal_gemm/gen_instances.py 0 → 100644
View file @ ef5e60f6
# SPDX-License-Identifier: MIT
# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
import logging
import os
import subprocess
from dataclasses import replace
from functools import lru_cache
from typing import List
from ..util import library_path
from .op import CKBatchedGemmOperation
log = logging.getLogger(__name__)
def _ck_library_dir():
gemm_instances_path = os.path.join(
library_path(),
"src",
"tensor_operation_instance",
"gpu",
"gemm_universal_batched",
)
if not os.path.exists(gemm_instances_path):
log.error("CK library path %s does not exist", gemm_instances_path)
return None
return gemm_instances_path
def parse_instances(str_instances: List[str]) -> List[CKBatchedGemmOperation]:
"""
Parse the lines containing Universal Gemm template instances into `CKBatchedGemmOperation` instances
"""
def maybe_int(s):
try:
return int(s)
except ValueError:
return s
op_instances = []
for line in str_instances:
s_template_args = line.split("DeviceBatchedGemmMultiD_Xdl_CShuffle_V3")[
-1
].strip("<>, ")
template_args = []
i_current = 0
while i_current < len(s_template_args):
if s_template_args[i_current] == " ":
# skip whitespace
i_current += 1
continue
elif s_template_args[i_current : i_current + 2] == "S<":
# parse template S<Index...>
i_next = s_template_args.find(">", i_current)
template_args.append(
tuple(map(int, s_template_args[i_current + 2 : i_next].split(",")))
)
i_current = i_next + 2
else:
# all string attributes must be either type aliases or global constants in C++
i_next = s_template_args.find(",", i_current)
template_args.append(
maybe_int(
s_template_args[i_current : i_next if i_next != -1 else None]
)
)
if i_next != -1:
i_current = i_next + 1
if i_next == -1:
break
# ds layout and dtype are parsed as placeholder; reset value
template_args[2] = tuple() # ds layout
template_args[6] = tuple() # ds dtype
new_instance = CKBatchedGemmOperation(
*template_args, # type: ignore[arg-type]
)
op_instances.append(new_instance)
return op_instances
@lru_cache(None)
def gen_ops_library() -> List[CKBatchedGemmOperation]:
"""
Parse the Universal Gemm instances defined in the composable kernel library folder.
"""
ck_library_dir = _ck_library_dir()
if not ck_library_dir:
return []
grep_result = subprocess.run(
[
"grep",
"-inR",
"DeviceBatchedGemmMultiD_Xdl_CShuffle_V3",
_ck_library_dir(),
],
capture_output=True,
text=True,
)
op_instances = parse_instances(grep_result.stdout.strip().split("\n"))
log.debug("ck instances from library: %d", len(op_instances))
schedulers = [
"BlockGemmPipelineScheduler::Intrawave",
"BlockGemmPipelineScheduler::Interwave",
]
gemm_specs = [
"GemmSpecialization::Default",
"GemmSpecialization::MPadding",
"GemmSpecialization::NPadding",
"GemmSpecialization::KPadding",
"GemmSpecialization::MNPadding",
"GemmSpecialization::MKPadding",
"GemmSpecialization::NKPadding",
"GemmSpecialization::MNKPadding",
]
# substitute templated args by looping through their domains
substitute_instances = []
for instance in op_instances:
sub_scheduler = instance.block_gemm_pipeline_scheduler == "BlkGemmPipeSched"
sub_spec = instance.gemm_specialization == "GemmSpec"
schedulers_range = (
schedulers if sub_scheduler else [instance.block_gemm_pipeline_scheduler]
)
spec_range = gemm_specs if sub_spec else [instance.gemm_specialization]
for scheduler in schedulers_range:
for spec in spec_range:
substitute_instances.append(
replace(
instance,
block_gemm_pipeline_scheduler=scheduler,
gemm_specialization=spec,
)
)
return substitute_instances
if __name__ == "__main__":
print(gen_ops_library())
python/ck4inductor/batched_universal_gemm/op.py 0 → 100644
View file @ ef5e60f6
# SPDX-License-Identifier: MIT
# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
from dataclasses import asdict, dataclass
from typing import Optional, Tuple
@dataclass
class CKBatchedGemmOperation:
"""
A python dataclass storing the template parameters of a CK Universal Gemm template instance
"""
a_layout: str
b_layout: str
ds_layouts: Tuple[str] # addmm specific
c_layout: str
a_element_dtype: str
b_element_dtype: str
ds_element_dtypes: Tuple[str] # addmm specific
c_element_dtype: str
acc_dtype: str
c_shuffle_dtype: str
a_elementwise_op: str
b_elementwise_op: str
c_elementwise_op: str
gemm_specialization: str
block_size: int
m_per_block: int
n_per_block: int
k_per_block: int
a_k1: int
b_k1: int
m_per_xdl: int
n_per_xdl: int
m_xdl_per_wave: int
n_xdl_per_wave: int
a_block_transfer_thread_cluster_lengths_ak0_m_ak1: Tuple[int, int, int]
a_block_transfer_thread_cluster_arrange_order: Tuple[int, int, int]
a_block_transfer_src_access_order: Tuple[int, int, int]
a_block_transfer_src_vector_dim: int
a_block_transfer_src_scalar_per_vector: int
a_block_transfer_dst_scalar_per_vector_ak1: int
a_block_lds_extra_m: bool
b_block_transfer_thread_cluster_lengths_bk0_n_bk1: Tuple[int, int, int]
b_block_transfer_thread_cluster_arrange_order: Tuple[int, int, int]
b_block_transfer_src_access_order: Tuple[int, int, int]
b_block_transfer_src_vector_dim: int
b_block_transfer_src_scalar_per_vector: int
b_block_transfer_dst_scalar_per_vector_bk1: int
b_block_lds_extra_n: bool
c_shuffle_m_xdl_per_wave_per_shuffle: int
c_shuffle_n_xdl_per_wave_per_shuffle: int
c_shuffle_block_transfer_cluster_lengths_m_block_m_per_block_n_block_n_per_block: (
Tuple[int, int, int, int]
)
c_shuffle_block_transfer_scalar_per_vector_n_per_block: Tuple[int]
block_gemm_pipeline_scheduler: str
block_gemm_pipeline_version: str
a_compute_dtype: Optional[str] = None
b_compute_dtype: Optional[str] = None
def name(self):
# cpp alias for template instance
return f"ck_device_batched_gemm_multi_d_xdl_c_shuffle_v3_{self.key_name()}"
def key_name(self):
# TBD; must be unique per instance. Intended to use as dict key
return "_".join(
[
"K"
+ field_name.replace("_", "").lower()
+ "V"
+ (
"x".join(map(str, iter(field_value)))
if isinstance(field_value, tuple)
else str(field_value).replace(":", "")
)
for field_name, field_value in self.dict_items()
]
)
def dict_items(self):
return asdict(self).items()
python/ck4inductor/grouped_conv_fwd/gen_instances.py
View file @ ef5e60f6
......@@ -130,9 +130,7 @@ def gen_conv_ops_library() -> List[CKGroupedConvFwdOp]:
# substitute templated args by looping through their domains
substitute_instances = []
for instance in op_instances:
sub_scheduler = (
instance.block_gemm_pipeline_scheduler == "BlkGemmPipeSched"
)
sub_scheduler = instance.block_gemm_pipeline_scheduler == "BlkGemmPipeSched"
sub_spec = instance.conv_forward_specialization == "ConvSpec"
schedulers_range = (
schedulers if sub_scheduler else [instance.block_gemm_pipeline_scheduler]
......
script/process_perf_data.py
View file @ ef5e60f6
......@@ -82,7 +82,7 @@ def parse_logfile(logfile):
StrideA=[]
StrideB=[]
StrideC=[]
if 'perf_gemm.log' in logfile:
if 'perf_gemm' in logfile and 'gemm_bilinear' not in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
......@@ -260,7 +260,7 @@ def main():
conn = sqlEngine.connect()
#save gemm performance tests:
if 'perf_gemm.log' in filename:
if 'perf_gemm' in filename and 'gemm_bilinear' not in filename:
#write the ck_gemm_test_params table only needed once the test set changes
#post_test_params(test_list,conn)
for i in range(1,len(results)+1):
......
script/process_perf_data.sh
View file @ ef5e60f6
......@@ -11,9 +11,22 @@
#process results
python3 process_perf_data.py perf_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
python3 process_perf_data.py perf_resnet50_N256.log
python3 process_perf_data.py perf_resnet50_N4.log
file=./perf_onnx_gemm_gfx10.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx10.log
fi
file=./perf_onnx_gemm_gfx11.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx11.log
fi
file=./perf_onnx_gemm_gfx12.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx12.log
fi
file=./perf_fmha_fwd_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_fwd_gfx942.log
......
script/process_qa_data.sh
View file @ ef5e60f6
......@@ -24,6 +24,18 @@ python3 process_perf_data.py perf_splitK_gemm.log
python3 process_perf_data.py perf_onnx_gemm.log
python3 process_perf_data.py perf_mixed_gemm.log
file=./perf_onnx_gemm_gfx10.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx10.log
fi
file=./perf_onnx_gemm_gfx11.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx11.log
fi
file=./perf_onnx_gemm_gfx12.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_onnx_gemm_gfx12.log
fi
file=./perf_fmha_fwd_gfx942.log
if [ -e "$file" ]; then
python3 process_perf_data.py perf_fmha_fwd_gfx942.log
......
script/run_full_performance_tests.sh
View file @ ef5e60f6
......@@ -5,7 +5,7 @@
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> <node name>
# input arguments:
# verification = 0 : do not verify result correctness on CPU
# = 1 : verifuy correctness on CPU (may take a long time)
......
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