ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Refactor elementwise kernels (#1222)

Browse Source 
* Refactor elementwise kernels

* Instances fixes

* Fix cmake

* Fix max pool bwd test

* Update two stage gemm split k

* Restore elementwise scale for hiptensor backward compatiblity

* Fix Acc data type check in conv fwd multiple abd

* Disable conv fp64 fwd example

* Update grouped conv weight multi d
This commit is contained in:
Bartłomiej Kocot
2024-04-19 13:31:17 +02:00
committed by GitHub
parent e0f3f918f1
commit ad1597c499
38 changed files with 513 additions and 2502 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
example/44_elementwise_permute/CMakeLists.txt
View File

@@ -1,12 +1,7 @@
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_row elementwise_permute_4D_fp32_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_row elementwise_permute_4D_fp16_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permute_4D_fp32_col.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
add_example_executable(example_elementwise_binary_4D_fp16 elementwise_binary_4D_fp16.cpp)
add_example_executable(example_elementwise_trinary_4D_fp16 elementwise_trinary_4D_fp16.cpp)
add_example_executable(example_elementwise_permute elementwise_permute.cpp)
if((NOT GPU_TARGETS MATCHES "gfx940") AND (NOT GPU_TARGETS MATCHES "gfx941") AND (NOT GPU_TARGETS MATCHES "gfx942"))
add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
endif()

121
example/44_elementwise_permute/elementwise_permute.cpp
View File

@@ -1,121 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
5, // NumDim
8, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> ncdhw = {16, 8, 8, 8, 8};
std::vector<std::size_t> ndhwc = {16, 8, 8, 8, 8};
std::array<ck::index_t, 5> ab_lengths;
std::array<ck::index_t, 5> a_strides = {
static_cast<int>(ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[3] * ncdhw[4]),
static_cast<int>(ncdhw[4]),
1,
static_cast<int>(ncdhw[2] * ncdhw[3] * ncdhw[4])};
std::array<ck::index_t, 5> b_strides = {
static_cast<int>(ndhwc[1] * ndhwc[2] * ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[2] * ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[3] * ndhwc[4]),
static_cast<int>(ndhwc[4]),
1};
ck::ranges::copy(ncdhw, ab_lengths.begin());
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
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"
<< std::endl;
bool pass = true;
if(do_verification)
{
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance =
ck::tensor_operation::host::ReferenceElementwise<1, ADataType, BDataType, PassThrough>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(as, host_b, PassThrough{});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}

118
example/44_elementwise_permute/elementwise_permute_3d.cpp
View File

@@ -1,118 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
#include "ck/library/utility/algorithm.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"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise3dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
2, // NumDim_m, {N, C}
2, // NumDim_n, {H, W}
1, // NumDim_k, {D}
4, // MPerThread
4, // NPerThread
4, // KPerThread
ck::Sequence<4>, // InScalarPerVectorSeq
ck::Sequence<4>>; // OutScalarPerVectorSeq
int main()
{
bool do_verification = true;
bool time_kernel = true;
const int N = 4;
const int C = 16;
const int H = 32;
const int W = 5;
const int D = 16;
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (ncdhw): " << a.mDesc << std::endl;
std::cout << "B (ndhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * ab_lengths[0] * ab_lengths[1] * ab_lengths[2] *
ab_lengths[3] * ab_lengths[4];
std::size_t num_btype =
(sizeof(ADataType) + sizeof(BDataType)) *
(ab_lengths[0] * ab_lengths[1] * ab_lengths[2] * ab_lengths[3] * ab_lengths[4]);
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"
<< std::endl;
bool pass = true;
if(do_verification)
{
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance =
ck::tensor_operation::host::ReferenceElementwise<1, ADataType, BDataType, PassThrough>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(as, host_b, PassThrough{});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}

113
example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
View File

@@ -1,113 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.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"
using F16 = ck::half_t;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // Elementwise op
3, // NumDim_M
1, // NumDim_N
1, // MPerThread
1, // NPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
int main()
{
bool do_verification = true;
bool time_kernel = true;
const int N = 120;
const int C = 128;
const int H = 32;
const int W = 32;
std::array<ck::index_t, 4> ab_lengths{N, H, W, C};
std::array<ck::index_t, 4> a_strides = {C * H * W, W, 1, H * W};
std::array<ck::index_t, 4> b_strides = {H * W * C, W * C, C, 1};
std::array<Tensor<ADataType>, 1> as = {Tensor<ADataType>(ab_lengths, a_strides)};
Tensor<ADataType>& a = as[0];
Tensor<BDataType> b(ab_lengths, b_strides);
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop =
std::size_t(2) * ab_lengths[0] * ab_lengths[1] * ab_lengths[2] * ab_lengths[3];
std::size_t num_btype = (sizeof(ADataType) + sizeof(BDataType)) *
(ab_lengths[0] * ab_lengths[1] * ab_lengths[2] * ab_lengths[3]);
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"
<< std::endl;
bool pass = true;
if(do_verification)
{
Tensor<BDataType> host_b(ab_lengths, b_strides);
using ReferenceElementwiseInstance =
ck::tensor_operation::host::ReferenceElementwise<1, ADataType, BDataType, PassThrough>;
auto ref_elementwise = ReferenceElementwiseInstance{};
auto ref_invoker = ref_elementwise.MakeInvoker();
auto ref_argument = ref_elementwise.MakeArgument(as, host_b, PassThrough{});
ref_invoker.Run(ref_argument);
b_device_buf.FromDevice(b.mData.data());
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}