Skip to content

GitLab

Commit 3c4fb1dd authored by Umang Yadav's avatar Umang Yadav
Browse files

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

parents 57cdd70b e8cddfdc
Hide whitespace changes
Inline Side-by-side
Showing with 1874 additions and 76 deletions
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.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;
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;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
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::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,
PassThrough>;
// 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{},
PassThrough{});
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;
std::size_t num_bytes = 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{},
PassThrough{});
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;
}
}
client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <iterator>
#include <numeric>
#include <vector>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_conv_fwd.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;
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;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 2;
static constexpr ck::index_t N = 16;
static constexpr ck::index_t K = 16;
static constexpr ck::index_t C = 16;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 14;
static constexpr ck::index_t Hi = 14;
static constexpr ck::index_t Wi = 14;
static constexpr ck::index_t Do = 14;
static constexpr ck::index_t Ho = 14;
static constexpr ck::index_t Wo = 14;
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::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,
PassThrough,
ck::bf8_t,
ck::f8_t>;
// 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{},
PassThrough{});
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;
std::size_t num_bytes = 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{},
PassThrough{});
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;
}
}
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 3c4fb1dd
......@@ -2,8 +2,10 @@ add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_f
add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_operations)
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ 3c4fb1dd
......@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial,
typename OutDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
typename OutLayout,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_bwd_weight(
const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
......@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
PassThrough,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = 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 AComputeType = ck::bf8_t;
using BComputeType = ck::f8_t;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
AComputeType,
BComputeType>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/16_convnd_fwd/CMakeLists.txt
View file @ 3c4fb1dd
add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
if((DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations)
target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
endif()
if((DTYPES MATCHES "fp8") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp16_comp_fp8 conv3d_fwd_fp16_comp_fp8.cpp)
target_link_libraries(client_conv3d_fwd_fp16_comp_fp8 PRIVATE composable_kernel::device_operations)
endif()
if((DTYPES MATCHES "fp32") OR NOT DEFINED DTYPES)
add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
endif()
client_example/16_convnd_fwd/common.hpp
View file @ 3c4fb1dd
......@@ -11,7 +11,7 @@
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
......@@ -94,7 +94,8 @@ template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename OutLayout,
ck::index_t NumNonSpatialDim = 3>
ck::index_t NumNonSpatialDim = 3,
typename ComputeType = InDataType>
bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
......@@ -173,18 +174,19 @@ bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialD
std::size_t flop = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
PassThrough>;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
PassThrough,
PassThrough,
PassThrough,
ComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
......
client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 1;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 64;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
int main()
{
return run_grouped_conv_fwd<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
3,
ck::f8_t>(
{N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
? EXIT_SUCCESS
: EXIT_FAILURE;
}
client_example/18_groupnorm/groupnorm_swish.cpp
View file @ 3c4fb1dd
......@@ -7,17 +7,19 @@
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp"
#include "ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp"
using XDataType = ck::half_t;
using GammaDataType = float;
using BetaDataType = float;
using YDataType = ck::half_t;
using ComputeDataType = float;
using Swish = ck::tensor_operation::element_wise::Swish;
using XDataType = ck::half_t;
using GammaDataType = float;
using BetaDataType = float;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using Swish = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
......@@ -49,22 +51,27 @@ int main(int argc, char* argv[])
std::size_t xy_size = N * H * W * G * C;
std::size_t gamma_beta_size = G * C;
std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1};
std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
std::vector<ck::index_t> xy_strides = {H * W * G * C, W * G * C, G * C, C, 1};
std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
std::vector<ck::index_t> save_mean_inv_std_strides = {G, 1};
SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
Swish,
Rank,
NumReduceDim>;
#ifdef SAVE_MEAN_INV_STD
SimpleDeviceMem save_mean_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
SimpleDeviceMem save_inv_std_device_buf(sizeof(SaveMeanInvStdDataType) * N * G);
#endif
using DeviceOp = ck::tensor_operation::device::DeviceNormalizationFwd<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
Swish,
Rank,
NumReduceDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -75,19 +82,26 @@ int main(int argc, char* argv[])
const auto& generic_op_ptr = op_ptrs[0];
auto generic_argument_ptr =
generic_op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
{1, 2, 4}, // reduceDims
generic_op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
......@@ -107,21 +121,29 @@ int main(int argc, char* argv[])
for(int i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
nullptr,
nullptr,
Swish{});
auto& op_ptr = op_ptrs[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
......@@ -129,12 +151,20 @@ int main(int argc, char* argv[])
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte =
sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
#ifdef SAVE_MEAN_INV_STD
num_byte += sizeof(SaveMeanInvStdDataType) * N * G * 2;
#endif
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
......@@ -165,25 +195,37 @@ int main(int argc, char* argv[])
std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
<< std::endl;
auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
nullptr,
nullptr,
Swish{});
auto argument_ptr =
op_ptr->MakeArgumentPointer({N, H, W, G, C}, // lengths
xy_strides, // xStrides
gamma_beta_strides, // gammaStrides
gamma_beta_strides, // betaStrides
xy_strides, // yStrides
save_mean_inv_std_strides, // save_mean Strides
save_mean_inv_std_strides, // save_inv_std Strides
{1, 2, 4}, // reduceDims
1e-6,
x_device_buf.GetDeviceBuffer(),
gamma_device_buf.GetDeviceBuffer(),
beta_device_buf.GetDeviceBuffer(),
y_device_buf.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_device_buf.GetDeviceBuffer(),
save_inv_std_device_buf.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
Swish{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
......
client_example/19_pool_fwd/CMakeLists.txt → client_example/19_pool/CMakeLists.txt
View file @ 3c4fb1dd
add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
\ No newline at end of file
target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
client_example/19_pool/avg_pool3d_bwd.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
using DOutDataType = ck::half_t;
using DInDataType = ck::half_t;
using DOutLayout = ck::tensor_layout::convolution::NDHWC;
using DInLayout = ck::tensor_layout::convolution::NDHWC;
struct SimpleDeviceMem
{
SimpleDeviceMem() = delete;
SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
{
(void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
}
void* GetDeviceBuffer() { return p_mem_; }
void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
~SimpleDeviceMem() { (void)hipFree(p_mem_); }
void* p_mem_;
std::size_t mMemSize_;
};
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Z = 2;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Di = 30;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_d = 2;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t window_dilation_d = 1;
ck::index_t window_dilation_h = 1;
ck::index_t window_dilation_w = 1;
ck::index_t in_left_pad_d = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_d = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCDHW
std::vector<ck::index_t> in_length = {N, C, Di, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Do, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations{
window_dilation_d, window_dilation_h, window_dilation_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
std::size_t in_tensor_size = N * C * Di * Hi * Wi;
std::size_t out_tensor_size = N * C * Do * Ho * Wo;
// tensor layout = NDHWC
std::vector<ck::index_t> in_tensor_stride = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
using DeviceOp = ck::tensor_operation::device::
DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
// 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;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 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(
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_length,
in_length,
out_tensor_stride,
in_tensor_stride,
window_spatial_lengths,
window_strides,
window_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
din_device_buf.SetZero();
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes =
in_tensor_size * sizeof(DInDataType) + out_tensor_size * sizeof(DOutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
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(
static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_length,
in_length,
out_tensor_stride,
in_tensor_stride,
window_spatial_lengths,
window_strides,
window_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
din_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/19_pool/max_pool2d_bwd.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using DOutDataType = ck::half_t;
using DInDataType = ck::half_t;
using IndexDataType = int32_t;
// We use pool3d to implement pool2d in this example
using InLayout = ck::tensor_layout::convolution::NDHWC;
using OutLayout = ck::tensor_layout::convolution::NDHWC;
constexpr ck::index_t InOutRank = 5;
constexpr ck::index_t WindowRank = 3;
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_;
};
void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
std::vector<ck::index_t>& window_lengths,
std::vector<ck::index_t>& output_lengths,
std::vector<ck::index_t>& input_stride,
std::vector<ck::index_t>& output_stride,
std::vector<ck::index_t>& indices_stride,
std::vector<ck::index_t>& window_strides,
std::vector<ck::index_t>& window_dilations,
std::vector<ck::index_t>& input_left_pads,
std::vector<ck::index_t>& input_right_pads,
std::vector<ck::index_t>& pooling_dims)
{
// NCHW to NCDHW
input_lengths.insert(input_lengths.begin() + 2, 1);
output_lengths.insert(output_lengths.begin() + 2, 1);
input_stride.insert(input_stride.begin() + 2, 0);
output_stride.insert(output_stride.begin() + 2, 0);
indices_stride.insert(indices_stride.begin() + 2, 0);
// YX to ZYX
window_lengths.insert(window_lengths.begin(), 1);
window_strides.insert(window_strides.begin(), 0);
window_dilations.insert(window_dilations.begin(), 0);
input_left_pads.insert(input_left_pads.begin(), 0);
input_right_pads.insert(input_right_pads.begin(), 0);
pooling_dims = {2, 3, 4};
}
int main(int argc, char* argv[])
{
ck::index_t N = 2;
ck::index_t C = 32;
ck::index_t Y = 2;
ck::index_t X = 2;
ck::index_t Hi = 30;
ck::index_t Wi = 30;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t window_dilation_h = 1;
ck::index_t window_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
// Pool API only support the order of NCHW
std::vector<ck::index_t> in_length = {N, C, Hi, Wi};
std::vector<ck::index_t> out_length = {N, C, Ho, Wo};
std::vector<ck::index_t> window_spatial_lengths = {Y, X};
std::vector<ck::index_t> window_strides = {window_stride_h, window_stride_w};
std::vector<ck::index_t> window_dilations = {window_dilation_h, window_dilation_w};
std::vector<ck::index_t> input_left_pads = {in_left_pad_h, in_left_pad_w};
std::vector<ck::index_t> input_right_pads = {in_right_pad_h, in_right_pad_w};
std::vector<ck::index_t> pooling_dims = {2, 3};
std::size_t in_tensor_size = N * C * Hi * Wi;
std::size_t out_tensor_size = N * C * Ho * Wo;
// tensor layout = NHWC
std::vector<ck::index_t> in_tensor_stride = {C * Hi * Wi, 1, Wi * C, C};
std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
TransformPool2dparamToPool3d(in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
window_dilations,
input_left_pads,
input_right_pads,
pooling_dims);
SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
SimpleDeviceMem indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
// Generate index data from max pool forward
{
using MaxPoolFwdDeviceOp =
ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
InLayout,
OutLayout,
ck::ReduceTensorOp::MAX,
true>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
MaxPoolFwdDeviceOp>::GetInstances();
auto& op_ptr = op_ptrs[0];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
in_length,
window_spatial_lengths,
out_length,
in_tensor_stride,
out_tensor_stride,
out_tensor_stride,
window_strides,
window_dilations,
input_left_pads,
input_right_pads,
pooling_dims);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
}
// Run MaxPool bwd
using MaxPoolBwdDeviceOp =
ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
MaxPoolBwdDeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
bool found = false;
int best_op_id = -1;
float best_ave_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 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(
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_tensor_size,
in_tensor_size,
window_spatial_lengths,
window_strides,
window_dilations);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_bytes = in_tensor_size * sizeof(DInDataType) +
out_tensor_size * sizeof(IndexDataType) +
out_tensor_size * sizeof(DOutDataType);
float gb_per_sec = num_bytes / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << "GB / s,"
<< op_name << std::endl;
if(ave_time < best_ave_time)
{
found = true;
best_op_id = i;
best_op_name = op_name;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
// run the best intance
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
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(
static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
out_tensor_size,
in_tensor_size,
window_spatial_lengths,
window_strides,
window_dilations);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
SimpleDeviceMem workspace(workspace_sz);
op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
}
std::cout << "Done" << std::endl;
}
return 0;
}
client_example/20_splitk_gemm/CMakeLists.txt
View file @ 3c4fb1dd
add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
endif()
client_example/21_grouped_gemm_bias/CMakeLists.txt 0 → 100644
View file @ 3c4fb1dd
add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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_bias.hpp"
using F16 = ck::half_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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using D0DataType = F32;
using DsDataType = ck::Tuple<D0DataType>;
using EDataType = F32;
using ALayout = Row;
using BLayout = Row;
using D0Layout = Row;
using DsLayout = ck::Tuple<D0Layout>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
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, ck::tensor_layout::gemm::RowMajor>::value)
{
return (nRow - 1) * stride + nCol;
}
else
{
return (nCol - 1) * stride + nRow;
}
};
std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, d0_dev_bufs, e_dev_bufs;
a_dev_bufs.reserve(group_count);
b_dev_bufs.reserve(group_count);
d0_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{}));
d0_dev_bufs.emplace_back(sizeof(D0DataType) *
f_matrix_space_size(Ms[i], Ns[i], 0, D0Layout{}));
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(),
std::array<const void*, 1>{d0_dev_bufs[i].GetDeviceBuffer()},
e_dev_bufs[i].GetDeviceBuffer(),
Ms[i],
Ns[i],
Ks[i],
StrideAs[i],
StrideBs[i],
std::array<ck::index_t, 1>{0},
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*, 1>> 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(), 2);
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/CMakeLists.txt 0 → 100644
View file @ 3c4fb1dd
add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_operations)
add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_operations)
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_operations)
client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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 F16 = ck::half_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 = F16;
using BDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
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, ck::tensor_layout::gemm::RowMajor>::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(), 32);
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_fp8.cpp 0 → 100644
View file @ 3c4fb1dd
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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 F8 = ck::f8_t;
using F16 = ck::half_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 = F16;
using BDataType = F8;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
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, ck::tensor_layout::gemm::RowMajor>::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(), 16);
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;
}
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