Skip to content

GitLab

Commit 3e4d0ff3 authored by Jakub Piasecki's avatar Jakub Piasecki
Browse files

Merge remote-tracking branch 'origin/develop' into ggemm_multid_two_stage

parents 1ad29336 9e011bcd
Show whitespace changes
Inline Side-by-side
Showing with 987 additions and 18 deletions
client_example/19_pool/avg_pool3d_bwd.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......
client_example/19_pool/avg_pool3d_fwd.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......
client_example/19_pool/max_pool2d_bwd.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......
client_example/19_pool/max_pool2d_fwd.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......
client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......@@ -88,7 +88,7 @@ int main(int argc, char* argv[])
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <iostream>
......@@ -79,7 +79,7 @@ int main()
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/22_grouped_gemm/CMakeLists.txt
View file @ 3e4d0ff3
......@@ -6,3 +6,6 @@ target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel
add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_gemm_operations)
add_executable(client_grouped_gemm_fixed_nk_bf16 grouped_gemm_fixed_nk_bf16.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_bf16 PRIVATE composable_kernel::device_gemm_operations)
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp 0 → 100644
View file @ 3e4d0ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <iostream>
#include <vector>
#include <random>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
using I8 = int8_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = I8;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Row;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int main()
{
std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
int sum_of_m = 0;
const int group_count = 16;
for(int i = 0; i < group_count; ++i)
{
Ms.push_back(256 + 256 * i);
Ns.push_back(128 + 128 * i);
Ks.push_back(128 + 64 * i);
StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
sum_of_m += Ms[i];
}
auto f_matrix_space_size =
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
e_dev_bufs.reserve(group_count);
std::vector<void*> p_e;
p_e.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
grouped_gemm_kernel_args_;
grouped_gemm_kernel_args_.reserve(group_count);
for(int i = 0; i < group_count; ++i)
{
a_dev_bufs.emplace_back(sizeof(ADataType) *
f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
b_dev_bufs.emplace_back(sizeof(BDataType) *
f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
e_dev_bufs.emplace_back(sizeof(EDataType) *
f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
b_dev_bufs[i].GetDeviceBuffer(),
{},
e_dev_bufs[i].GetDeviceBuffer(),
Ms[i],
Ns[i],
Ks[i],
StrideAs[i],
StrideBs[i],
{},
StrideEs[i]});
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
std::vector<const void*> p_a = {}, p_b = {};
std::vector<std::array<const void*, 0>> p_ds = {};
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
SimpleDeviceMem grouped_gemm_kernel_args_dev(
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
std::string op_name = op_ptr->GetTypeString();
hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
grouped_gemm_kernel_args_.data(),
op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
hipMemcpyHostToDevice));
op_ptr->SetWorkSpacePointer(argument_ptr.get(),
grouped_gemm_workspace_dev.GetDeviceBuffer());
op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
grouped_gemm_kernel_args_dev.GetDeviceBuffer());
op_ptr->SetKBatch(argument_ptr.get(), 1);
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = 0, num_btype = 0;
for(std::size_t j = 0; j < gemm_descs.size(); ++j)
{
flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
sizeof(EDataType) * Ms[j] * Ns[j];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return 0;
}
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <iostream>
......@@ -76,7 +76,7 @@ int main()
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <iostream>
......@@ -77,7 +77,7 @@ int main()
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <iostream>
......@@ -77,7 +77,7 @@ int main()
[](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
using Layout = decltype(layout);
if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
if constexpr(std::is_same<Layout, Row>::value)
{
return (nRow - 1) * stride + nCol;
}
......
client_example/22_im2col_col2im/image_to_column.cpp
View file @ 3e4d0ff3
// 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 <cstdlib>
#include <iomanip>
......
client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp
View file @ 3e4d0ff3
// 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 <iomanip>
#include <vector>
......
client_example/24_grouped_conv_activation/CMakeLists.txt
View file @ 3e4d0ff3
......@@ -38,3 +38,11 @@ target_link_libraries(client_grouped_convnd_fwd_bilinear_residual_fp16 PRIVATE c
add_executable(client_grouped_convnd_bwd_data_bilinear_residual_fp16
grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp)
target_link_libraries(client_grouped_convnd_bwd_data_bilinear_residual_fp16 PRIVATE composable_kernel::device_conv_operations)
# Fwd scale
add_executable(client_grouped_convnd_fwd_scale_fp16
grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp)
target_link_libraries(client_grouped_convnd_fwd_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
# Bwd data scale
add_executable(client_grouped_convnd_bwd_data_scale_fp16
grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp)
target_link_libraries(client_grouped_convnd_bwd_data_scale_fp16 PRIVATE composable_kernel::device_conv_operations)
client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp 0 → 100644
View file @ 3e4d0ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_scale.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
// Use std tuple instead of ck tuple to avoid clang
// implicit instantiation of undefined template error.
using DDataTypes = std::tuple<ck::half_t>;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 64; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Z = 3; // filter D
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Di = 14; // input D
static constexpr ck::index_t Hi = 14; // input H
static constexpr ck::index_t Wi = 14; // input W
static constexpr ck::index_t Do = 14; // output D
static constexpr ck::index_t Ho = 14; // output H
static constexpr ck::index_t Wo = 14; // output W
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int execute_conv_bwd_data_scale()
{
std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, NumDimSpatial + 3> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, NumDimSpatial + 3> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, NumDimSpatial + 3> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * G * N * Di * Hi * Wi * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * G * N * Do * Ho * Wo * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
PassThrough,
PassThrough,
Scale>;
// 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;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
Scale{2.f});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop = std::size_t(2) * G * N * K * C * Do * Ho * Wo * Y * X +
3 * G * N * Di * Hi * Wi * C;
std::size_t num_bytes = 2 * sizeof(InDataType) * G * N * Di * Hi * Wi * C +
sizeof(WeiDataType) * G * K * Z * Y * X * C +
sizeof(OutDataType) * G * N * Do * Ho * Wo * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
in.GetDeviceBuffer(),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
Scale{2.f});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
int main() { return execute_conv_bwd_data_scale(); }
client_example/24_grouped_conv_activation/grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp 0 → 100644
View file @ 3e4d0ff3
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scale.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
// Use std tuple instead of ck tuple to avoid clang
// implicit instantiation of undefined template error.
using DDataTypes = std::tuple<ck::half_t>;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Scale = ck::tensor_operation::element_wise::Scale;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 32;
static constexpr ck::index_t N = 64; // batch size
static constexpr ck::index_t K = 64; // output channel
static constexpr ck::index_t C = 32; // input channel (per group)
static constexpr ck::index_t Z = 3; // filter D
static constexpr ck::index_t Y = 3; // filter H
static constexpr ck::index_t X = 3; // filter W
static constexpr ck::index_t Di = 14; // input D
static constexpr ck::index_t Hi = 14; // input H
static constexpr ck::index_t Wi = 14; // input W
static constexpr ck::index_t Do = 14; // output D
static constexpr ck::index_t Ho = 14; // output H
static constexpr ck::index_t Wo = 14; // output W
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
};
int execute_conv_fwd_scale()
{
// We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space.
// However, CK's API only accepts lengths and strides with order of GNCDHW/GKCZYX/GNKDHW.
// Hence, we need to adjust the order of strides.
std::array<ck::index_t, 6> in_lengths{G, N, C, Di, Hi, Wi};
std::array<ck::index_t, 6> in_strides{
C, Di * Hi * Wi * G * C, 1, Hi * Wi * G * C, Wi * G * C, G * C};
std::array<ck::index_t, 6> wei_lengths{G, K, C, Z, Y, X};
std::array<ck::index_t, 6> wei_strides{
K * Z * Y * X * C, Z * Y * X * C, 1, Y * X * C, X * C, C};
std::array<ck::index_t, 6> out_lengths{G, N, K, Do, Ho, Wo};
std::array<ck::index_t, 6> out_strides{
K, Do * Ho * Wo * G * K, 1, Ho * Wo * G * K, Wo * G * K, G * K};
// Logical broadcast bias (we have to pass bias lengths in the same format as output - GNKDHW)
std::array<ck::index_t, 6> bias_lengths{G, 1, K, 1, 1, 1};
std::array<ck::index_t, 6> bias_strides{K, 0, 1, 0, 0, 0};
std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
SimpleDeviceMem in(sizeof(InDataType) * N * Di * Hi * Wi * G * C);
SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Z * Y * X * C);
SimpleDeviceMem out(sizeof(OutDataType) * N * Do * Ho * Wo * G * K);
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
Scale>;
// 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;
int best_op_id = -1;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
float best_tflops = 0;
// profile device operation instances
std::cout << "Run all instances and do timing" << std::endl;
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
Scale{2.f});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t flop =
std::size_t(2) * G * N * K * C * Ho * Wo * Y * X + 3 * N * Ho * Wo * G * K;
std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
sizeof(WeiDataType) * G * K * Y * X * C +
sizeof(OutDataType) * 2 * N * Ho * Wo * G * K;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_id = i;
best_op_name = op_name;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
best_tflops = tflops;
}
}
else
{
std::cerr << op_name << " does not support this problem" << std::endl;
}
}
if(best_op_id < 0)
{
std::cerr << "no suitable instance" << std::endl;
return EXIT_FAILURE;
}
std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
<< " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
// run the best intance
{
auto& op_ptr = op_ptrs[best_op_id];
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
wei.GetDeviceBuffer(),
{},
out.GetDeviceBuffer(),
in_lengths,
in_strides,
wei_lengths,
wei_strides,
{},
{},
out_lengths,
out_strides,
filter_strides,
filter_dilations,
input_left_pads,
input_right_pads,
PassThrough{},
PassThrough{},
Scale{2.f});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
int main() { return execute_conv_fwd_scale(); }
client_example/25_wrapper/wrapper_basic_gemm.cpp
View file @ 3e4d0ff3
......@@ -213,4 +213,3 @@ int main(int argc, char* argv[])
3840, 4096, 4096, tile_shape, thread_layout);
return 0;
}
// MI300X Perf: 0.471337 ms, 273.369 TFlops, 204.671 GB/s,
client_example/25_wrapper/wrapper_optimized_gemm.cpp
View file @ 3e4d0ff3
......@@ -305,4 +305,3 @@ int main(int argc, char* argv[])
3840, 4096, 4096, tile_shape, thread_layout);
return 0;
}
// MI300X Perf: 0.411552 ms, 313.081 TFlops, 234.403 GB/s,
cmake/Embed.cmake 0 → 100644
View file @ 3e4d0ff3
#####################################################################################
# The MIT License (MIT)
#
# Copyright (c) 2015-2024 Advanced Micro Devices, Inc. All rights reserved.
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
# THE SOFTWARE.
#####################################################################################
if(WIN32)
set(EMBED_USE RC CACHE STRING "Use RC or CArrays to embed data files")
set_property(CACHE EMBED_USE PROPERTY STRINGS "RC;CArrays")
else()
if(BUILD_SHARED_LIBS)
set(EMBED_USE LD CACHE STRING "Use LD or CArrays to embed data files")
else()
set(EMBED_USE CArrays CACHE STRING "Use LD or CArrays to embed data files")
endif()
set_property(CACHE EMBED_USE PROPERTY STRINGS "LD;CArrays")
endif()
if(EMBED_USE STREQUAL "LD")
find_program(EMBED_LD ld REQUIRED)
find_program(EMBED_OBJCOPY objcopy REQUIRED)
endif()
function(embed_wrap_string)
set(options)
set(oneValueArgs VARIABLE AT_COLUMN)
set(multiValueArgs)
cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
string(LENGTH ${${PARSE_VARIABLE}} string_length)
math(EXPR offset "0")
while(string_length GREATER 0)
if(string_length GREATER ${PARSE_AT_COLUMN})
math(EXPR length "${PARSE_AT_COLUMN}")
else()
math(EXPR length "${string_length}")
endif()
string(SUBSTRING ${${PARSE_VARIABLE}} ${offset} ${length} line)
set(lines "${lines}\n${line}")
math(EXPR string_length "${string_length} - ${length}")
math(EXPR offset "${offset} + ${length}")
endwhile()
set(${PARSE_VARIABLE} "${lines}" PARENT_SCOPE)
endfunction()
function(generate_embed_source EMBED_NAME EMBED_DIR BASE_DIRECTORY)
set(options)
set(oneValueArgs)
set(multiValueArgs SYMBOLS FILES)
cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(RESOURCE_ID 100)
list(LENGTH PARSE_SYMBOLS SYMBOLS_LEN)
list(LENGTH PARSE_FILES FILES_LEN)
if(NOT ${SYMBOLS_LEN} EQUAL ${FILES_LEN})
message(FATAL_ERROR "Symbols and objects dont match: ${SYMBOLS_LEN} != ${FILES_LEN}")
endif()
math(EXPR LEN "${SYMBOLS_LEN} - 1")
foreach(idx RANGE ${LEN})
list(GET PARSE_SYMBOLS ${idx} SYMBOL)
list(GET PARSE_FILES ${idx} FILE)
file(RELATIVE_PATH BASE_NAME "${BASE_DIRECTORY}" ${FILE})
if(EMBED_USE STREQUAL "RC")
string(TOUPPER "${SYMBOL}" SYMBOL)
string(APPEND FILE_IDS "#define IDR_${SYMBOL} ${RESOURCE_ID}\n")
file(TO_NATIVE_PATH "${FILE}" NATIVE_FILE)
string(REPLACE "\\" "\\\\" NATIVE_FILE "${NATIVE_FILE}")
string(APPEND RC_FILE_MAPPING "IDR_${SYMBOL} TEXTFILE \"${NATIVE_FILE}\"\n")
string(APPEND INIT_KERNELS "\n {\"${BASE_NAME}\", resource::read(IDR_${SYMBOL})},")
math(EXPR RESOURCE_ID "${RESOURCE_ID} + 1" OUTPUT_FORMAT DECIMAL)
else()
set(START_SYMBOL "_binary_${SYMBOL}_start")
set(LENGTH_SYMBOL "_binary_${SYMBOL}_length")
if(EMBED_USE STREQUAL "LD")
string(APPEND EXTERNS "
extern const char ${START_SYMBOL}[];
extern const size_t _binary_${SYMBOL}_size;
const auto ${LENGTH_SYMBOL} = reinterpret_cast<size_t>(&_binary_${SYMBOL}_size);
")
else()
string(APPEND EXTERNS "
extern const char ${START_SYMBOL}[];
extern const size_t ${LENGTH_SYMBOL};
")
endif()
string(APPEND INIT_KERNELS "
{ \"${BASE_NAME}\", { ${START_SYMBOL}, ${LENGTH_SYMBOL}} },")
endif()
endforeach()
if(EMBED_USE STREQUAL "RC")
file(WRITE "${EMBED_DIR}/include/resource.h" "
#define TEXTFILE 256
${FILE_IDS}
")
file(WRITE "${EMBED_DIR}/resource.rc" "
#include \"resource.h\"
${RC_FILE_MAPPING}
")
set(EXTERNS "
#include <Windows.h>
#include \"resource.h\"
namespace resource {
std::string_view read(int id)
{
HMODULE handle = GetModuleHandle(nullptr);
HRSRC rc = FindResource(handle, MAKEINTRESOURCE(id), MAKEINTRESOURCE(TEXTFILE));
HGLOBAL data = LoadResource(handle, rc);
return {static_cast<const char*>(LockResource(data)), SizeofResource(handle, rc)};
}
}
")
set(EMBED_FILES ${EMBED_DIR}/include/resource.h ${EMBED_DIR}/resource.rc)
endif()
file(WRITE "${EMBED_DIR}/include/${EMBED_NAME}.hpp" "
#include <string_view>
#include <unordered_map>
#include <utility>
std::unordered_map<std::string_view, std::string_view> ${EMBED_NAME}();
")
file(WRITE "${EMBED_DIR}/${EMBED_NAME}.cpp" "
#include <${EMBED_NAME}.hpp>
${EXTERNS}
std::unordered_map<std::string_view, std::string_view> ${EMBED_NAME}()
{
static std::unordered_map<std::string_view, std::string_view> result = {${INIT_KERNELS}
};
return result;
}
")
list(APPEND EMBED_FILES ${EMBED_DIR}/${EMBED_NAME}.cpp ${EMBED_DIR}/include/${EMBED_NAME}.hpp)
set(EMBED_FILES ${EMBED_FILES} PARENT_SCOPE)
endfunction()
function(embed_file FILE BASE_DIRECTORY)
message(STATUS " ${FILE}")
file(RELATIVE_PATH REL_FILE "${BASE_DIRECTORY}" ${FILE})
string(MAKE_C_IDENTIFIER "${REL_FILE}" OUTPUT_SYMBOL)
get_filename_component(OUTPUT_FILE_DIR "${REL_FILE}" DIRECTORY)
file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/${OUTPUT_FILE_DIR}")
if(EMBED_USE STREQUAL "LD")
set(OUTPUT_FILE "${CMAKE_CURRENT_BINARY_DIR}/${REL_FILE}.o")
add_custom_command(
OUTPUT "${OUTPUT_FILE}"
COMMAND ${EMBED_LD} -r -o "${OUTPUT_FILE}" -z noexecstack --format=binary "${REL_FILE}"
COMMAND ${EMBED_OBJCOPY} --rename-section .data=.rodata,alloc,load,readonly,data,contents "${OUTPUT_FILE}"
WORKING_DIRECTORY "${BASE_DIRECTORY}"
DEPENDS "${FILE}"
VERBATIM)
set(OUTPUT_FILE ${OUTPUT_FILE} PARENT_SCOPE)
elseif(EMBED_USE STREQUAL "CArrays")
set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS ${FILE})
set(OUTPUT_FILE "${CMAKE_CURRENT_BINARY_DIR}/${REL_FILE}.cpp")
# reads source file contents as hex string
file(READ ${FILE} HEX_STRING HEX)
# wraps the hex string into multiple lines
embed_wrap_string(VARIABLE HEX_STRING AT_COLUMN 80)
# adds '0x' prefix and comma suffix before and after every byte respectively
string(REGEX REPLACE "([0-9a-f][0-9a-f])" "0x\\1, " ARRAY_VALUES ${HEX_STRING})
# removes trailing comma
string(REGEX REPLACE ", $" "" ARRAY_VALUES ${ARRAY_VALUES})
file(WRITE "${OUTPUT_FILE}" "
#include <cstddef>
extern const char _binary_${OUTPUT_SYMBOL}_start[] = { ${ARRAY_VALUES} };
extern const size_t _binary_${OUTPUT_SYMBOL}_length = sizeof(_binary_${OUTPUT_SYMBOL}_start);
")
set(OUTPUT_FILE ${OUTPUT_FILE} PARENT_SCOPE)
endif()
set(OUTPUT_SYMBOL ${OUTPUT_SYMBOL} PARENT_SCOPE)
endfunction()
function(add_embed_library EMBED_NAME)
set(options)
set(oneValueArgs RELATIVE)
set(multiValueArgs)
cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(EMBED_DIR ${CMAKE_CURRENT_BINARY_DIR}/embed/${EMBED_NAME})
file(MAKE_DIRECTORY ${EMBED_DIR})
message(STATUS "Embedding kernel files:")
foreach(FILE ${PARSE_UNPARSED_ARGUMENTS})
embed_file(${FILE} ${PARSE_RELATIVE})
list(APPEND OUTPUT_FILES ${OUTPUT_FILE})
list(APPEND SYMBOLS ${OUTPUT_SYMBOL})
endforeach()
message(STATUS "Generating embedding library '${EMBED_NAME}'")
generate_embed_source(${EMBED_NAME} ${EMBED_DIR} "${PARSE_RELATIVE}" SYMBOLS ${SYMBOLS} FILES ${PARSE_UNPARSED_ARGUMENTS})
set(INTERNAL_EMBED_LIB embed_lib_${EMBED_NAME})
if(EMBED_USE STREQUAL "LD")
add_library(${INTERNAL_EMBED_LIB} STATIC ${EMBED_FILES} ${OUTPUT_FILES})
else()
add_library(${INTERNAL_EMBED_LIB} OBJECT ${EMBED_FILES})
endif()
if(EMBED_USE STREQUAL "CArrays")
target_sources(${INTERNAL_EMBED_LIB} PRIVATE ${OUTPUT_FILES})
endif()
target_include_directories(${INTERNAL_EMBED_LIB} PRIVATE "${EMBED_DIR}/include")
target_compile_options(${INTERNAL_EMBED_LIB} PRIVATE -Wno-reserved-identifier -Wno-extern-initializer -Wno-missing-variable-declarations)
set_target_properties(${INTERNAL_EMBED_LIB} PROPERTIES POSITION_INDEPENDENT_CODE On)
add_library(${EMBED_NAME} INTERFACE)
if(EMBED_USE STREQUAL "RC")
target_link_libraries(${EMBED_NAME} INTERFACE $<TARGET_OBJECTS:${INTERNAL_EMBED_LIB}>)
elseif(EMBED_USE STREQUAL "LD")
target_link_libraries(${EMBED_NAME} INTERFACE ${INTERNAL_EMBED_LIB})
else()
target_sources(${EMBED_NAME} INTERFACE $<TARGET_OBJECTS:${INTERNAL_EMBED_LIB}>)
endif()
target_include_directories(${EMBED_NAME} INTERFACE "${EMBED_DIR}/include")
endfunction()
codegen/CMakeLists.txt 0 → 100644
View file @ 3e4d0ff3
cmake_minimum_required(VERSION 3.16)
project(composable_kernel_host)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/..)
find_package(ROCM)
include(ROCMInstallTargets)
include(ROCMTest)
list(APPEND CMAKE_MODULE_PATH ${CK_ROOT}/cmake)
include(Embed)
file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
${CK_ROOT}/include/ck/*.hpp)
message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
message(STATUS "RELATIVE: ${CK_ROOT}/include")
add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
add_definitions(-std=c++17)
file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
# TODO: Use object library
add_library(ck_host STATIC ${SOURCES})
target_link_libraries(ck_host PRIVATE ck_headers)
set_target_properties(ck_host PROPERTIES
LINKER_LANGUAGE CXX
POSITION_INDEPENDENT_CODE ON)
target_include_directories(ck_host PUBLIC
$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
)
add_executable(ck-template-driver driver/main.cpp)
target_link_libraries(ck-template-driver ck_host)
rocm_install(
TARGETS ck_host ck_headers
EXPORT ck_hostTargets
)
rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
if(BUILD_TESTING)
add_subdirectory(test)
endif()
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