Skip to content

GitLab

Commit 289f15de authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into wmma_gemm

parents 9bd44685 d58b7f51
Hide whitespace changes
Inline Side-by-side
Showing with 729 additions and 133 deletions
client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iomanip>
#include <vector>
#include <iostream>
#include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
using ADataType = ck::half_t; // Input 1
using BDataType = ck::half_t; // Input 2
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using AccDataType = float;
using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
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()
{
bool time_kernel = true;
ck::index_t M = 48 * 256;
ck::index_t N = 1024;
ck::index_t Stride = N;
auto mn_size = (M - 1) * Stride + N;
SimpleDeviceMem a_dev_buf(sizeof(ADataType) * mn_size);
SimpleDeviceMem b_dev_buf(sizeof(BDataType) * mn_size);
SimpleDeviceMem gamma_dev_buf(sizeof(GammaDataType) * N);
SimpleDeviceMem beta_dev_buf(sizeof(BetaDataType) * N);
SimpleDeviceMem y_dev_buf(sizeof(YDataType) * mn_size);
std::array<const void*, 2> ab_input = {a_dev_buf.GetDeviceBuffer(),
b_dev_buf.GetDeviceBuffer()};
std::vector<ck::index_t> abStride = {Stride, 1};
std::array<std::vector<ck::index_t>, 2> abStrides = {abStride, abStride};
using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
ck::Tuple<ADataType, BDataType>,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
XElementwiseOperation,
YElementwiseOperation,
Rank,
NumReduceDim>;
// 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({M, N}, // lengths
abStrides,
{0, 1}, // gammaStrides
{0, 1}, // betaStrides
{Stride, 1}, // yStrides
{1}, // reduceDims
1e-4,
ab_input,
gamma_dev_buf.GetDeviceBuffer(),
beta_dev_buf.GetDeviceBuffer(),
y_dev_buf.GetDeviceBuffer(),
XElementwiseOperation{},
YElementwiseOperation{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
std::size_t num_byte = sizeof(ADataType) * M * N + sizeof(BDataType) * M * N +
sizeof(GammaDataType) * N + sizeof(BetaDataType) * N +
sizeof(YDataType) * M * N;
float gb_per_sec = num_byte / 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;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << 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({M, N}, // lengths
abStrides,
{1}, // gammaStrides
{1}, // betaStrides
{Stride, 1}, // yStrides
{1}, // reduceDims
1e-4,
ab_input,
gamma_dev_buf.GetDeviceBuffer(),
beta_dev_buf.GetDeviceBuffer(),
y_dev_buf.GetDeviceBuffer(),
XElementwiseOperation{},
YElementwiseOperation{});
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;
}
client_example/13_batchnorm/CMakeLists.txt 0 → 100644
View file @ 289f15de
add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)
client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <functional>
#include <numeric>
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
using XDataType = ck::half_t;
using DxDataType = float;
using DyDataType = float;
using AccDataType = float;
using ScaleDataType = ck::half_t;
using DscaleDbiasDataType = float;
using MeanVarDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 4;
constexpr int NumBatchNormReduceDim = 3;
const double epsilon = std::numeric_limits<float>::epsilon();
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(int argc, char* argv[])
{
std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
ck::index_t numXYElement =
std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
scaleBiasMeanVarLengths.end(),
1,
std::multiplies<ck::index_t>());
SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
SimpleDeviceMem dy(sizeof(DyDataType) * numXYElement);
SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem dx(sizeof(DxDataType) * numXYElement);
SimpleDeviceMem dscale(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem dbias(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
DxDataType,
DyDataType,
AccDataType,
ScaleDataType,
DscaleDbiasDataType,
MeanVarDataType,
PassThrough,
Rank,
NumBatchNormReduceDim>;
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(xyLengths,
xyStrides,
xyStrides,
xyStrides,
reduceDims,
scaleBiasMeanVarLengths,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
x.GetDeviceBuffer(),
dy.GetDeviceBuffer(),
scale.GetDeviceBuffer(),
mean.GetDeviceBuffer(),
invVariance.GetDeviceBuffer(),
epsilon,
PassThrough{},
dx.GetDeviceBuffer(),
dscale.GetDeviceBuffer(),
dbias.GetDeviceBuffer());
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 =
numXYElement * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
numScaleBiasMeanVarElement *
(sizeof(ScaleDataType) + sizeof(DscaleDbiasDataType) * 2 +
sizeof(MeanVarDataType) * 2);
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;
}
}
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
xyStrides,
xyStrides,
xyStrides,
reduceDims,
scaleBiasMeanVarLengths,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
x.GetDeviceBuffer(),
dy.GetDeviceBuffer(),
scale.GetDeviceBuffer(),
mean.GetDeviceBuffer(),
invVariance.GetDeviceBuffer(),
epsilon,
PassThrough{},
dx.GetDeviceBuffer(),
dscale.GetDeviceBuffer(),
dbias.GetDeviceBuffer());
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;
}
client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp 0 → 100644
View file @ 289f15de
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <functional>
#include <numeric>
#include <iomanip>
#include <iostream>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
using XDataType = float;
using YDataType = float;
using AccDataType = float;
using ScaleDataType = AccDataType;
using BiasDataType = AccDataType;
using MeanVarDataType = AccDataType;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 4;
constexpr int NumBatchNormReduceDim = 3;
const double epsilon = std::numeric_limits<float>::epsilon();
const double averageFactor = 0.1;
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(int argc, char* argv[])
{
std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
ck::index_t numXYElement =
std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
scaleBiasMeanVarLengths.end(),
1,
std::multiplies<ck::index_t>());
SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
YDataType,
AccDataType,
ScaleDataType,
BiasDataType,
MeanVarDataType,
PassThrough,
Rank,
NumBatchNormReduceDim>;
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(xyLengths,
xyStrides,
xyStrides,
reduceDims,
scaleBiasMeanVarLengths,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
x.GetDeviceBuffer(),
scale.GetDeviceBuffer(),
bias.GetDeviceBuffer(),
epsilon,
PassThrough{},
y.GetDeviceBuffer(),
mean.GetDeviceBuffer(),
invVariance.GetDeviceBuffer(),
averageFactor,
nullptr,
nullptr);
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 =
numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
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;
}
}
if(found)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << 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(xyLengths,
xyStrides,
xyStrides,
reduceDims,
scaleBiasMeanVarLengths,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
scaleBiasMeanVarStrides,
x.GetDeviceBuffer(),
scale.GetDeviceBuffer(),
bias.GetDeviceBuffer(),
epsilon,
PassThrough{},
y.GetDeviceBuffer(),
mean.GetDeviceBuffer(),
invVariance.GetDeviceBuffer(),
averageFactor,
nullptr,
nullptr);
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;
}
example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
View file @ 289f15de
......@@ -6,6 +6,8 @@
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp"
#include "ck/library/utility/literals.hpp"
using F16 = ck::half_t;
using F32 = float;
......@@ -135,15 +137,15 @@ int main(int argc, char* argv[])
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -240,7 +242,7 @@ int main(int argc, char* argv[])
show_2d_matrix(std::cout << "c_host :", c_m_n_host_result) << std::endl;
}
#endif
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
return 0;
......
example/01_gemm/run_gemm_example.inc
View file @ 289f15de
......@@ -32,15 +32,12 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
case 0: break;
case 1:
// CONFIRMED
// ck::utils::FillMNID<ADataType>{}(a_m_k.begin(), a_m_k.end());
// ck::utils::FillMNID<BDataType>{}(b_k_n.begin(), b_k_n.end());
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), a_m_k.end());
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(), b_k_n.end());
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end());
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end());
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
}
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
......@@ -133,11 +130,11 @@ bool run_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
return ck::utils::check_err(c_m_n_device_result_converted.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
#else
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
#endif
}
......
example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
View file @ 289f15de
......@@ -14,6 +14,7 @@
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -177,15 +178,15 @@ int main(int argc, char* argv[])
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -271,8 +272,7 @@ int main(int argc, char* argv[])
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
Tensor<CShuffleDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
......@@ -299,7 +299,7 @@ int main(int argc, char* argv[])
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
......
example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
View file @ 289f15de
......@@ -15,6 +15,7 @@
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -155,15 +156,15 @@ int main(int argc, char* argv[])
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
......@@ -275,7 +276,7 @@ int main(int argc, char* argv[])
}
}
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
......
example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
View file @ 289f15de
......@@ -124,7 +124,7 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
if(config.do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n({M, N});
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -147,9 +147,9 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
#ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> e_m_n_device_result_converted(e_m_n_device_result);
return ck::utils::check_err(e_m_n_device_result_converted.mData, e_m_n_host_result.mData);
return ck::utils::check_err(e_m_n_device_result_converted, e_m_n_host_result);
#else
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
#endif
}
......
example/09_convnd_fwd/CMakeLists.txt
View file @ 289f15de
......@@ -8,3 +8,4 @@ add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp6
add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
example/09_convnd_fwd/convnd_fwd_common.hpp
View file @ 289f15de
......@@ -10,6 +10,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -84,7 +85,7 @@ bool run_grouped_conv_fwd(bool do_verification,
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
......@@ -164,7 +165,7 @@ bool run_grouped_conv_fwd(bool do_verification,
out_device_buf.FromDevice(out_device.mData.data());
return ck::utils::check_err(
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
}
return true;
......
example/09_convnd_fwd/convnd_fwd_dl_common.hpp
View file @ 289f15de
......@@ -10,6 +10,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -29,6 +30,7 @@ void print_helper_msg()
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename DsDataType,
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
......@@ -45,8 +47,10 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
using DDataType = ck::remove_cvref_t<ck::tuple_element_t<0, DsDataType>>;
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<DDataType> bias(out_g_n_k_wos_desc);
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
......@@ -58,44 +62,53 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 3});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 3});
bias.GenerateTensorValue(GeneratorTensor_2<DDataType>{-2, 3});
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{-1});
bias.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(DDataType) * bias.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
bias_device_buf.ToDevice(bias.mData.data());
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> c_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> c_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(out_g_n_k_wos_desc.GetLengths(), c_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), c_g_n_k_wos_strides);
copy(out_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
......@@ -104,25 +117,32 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
// do Conv
auto conv = DeviceConvNDFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
c_g_n_k_wos_lengths,
c_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto argument = conv.MakeArgument(
in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
std::cout << "wrong! device_conv with the specified compilation parameters does not "
"support this Conv problem"
<< std::endl;
return true;
}
......@@ -138,28 +158,34 @@ bool run_grouped_conv_fwd_dl(bool do_verification,
if(do_verification)
{
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei,
out_host,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
ck::tensor_operation::element_wise::PassThrough>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument =
ref_conv.MakeArgument(in,
wei,
out_host,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
in_element_op,
wei_element_op,
ck::tensor_operation::element_wise::PassThrough{});
ref_invoker.Run(ref_argument);
// cde_elementwise
out_host.ForEach(
[&](auto&, auto idx) { out_element_op(out_host(idx), out_host(idx), bias(idx)); });
out_device_buf.FromDevice(out_device.mData.data());
return ck::utils::check_err(
......
example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
View file @ 289f15de
......@@ -3,13 +3,14 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using AccDataType = float;
using DsDataType = ck::Tuple<ck::half_t>;
using OutDataType = ck::half_t;
template <ck::index_t... Is>
......@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
......@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
// clang-format off
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
#include "run_convnd_fwd_dl_example.inc"
......
example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
View file @ 289f15de
......@@ -3,13 +3,14 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = float;
using WeiDataType = float;
using AccDataType = float;
using DsDataType = ck::Tuple<float>;
using OutDataType = float;
template <ck::index_t... Is>
......@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
......@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
// clang-format off
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
#include "run_convnd_fwd_dl_example.inc"
......
example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
View file @ 289f15de
......@@ -3,13 +3,14 @@
#include "convnd_fwd_dl_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = int8_t;
using WeiDataType = int8_t;
using AccDataType = int32_t;
using DsDataType = ck::Tuple<int8_t>;
using OutDataType = int8_t;
template <ck::index_t... Is>
......@@ -17,7 +18,7 @@ using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
......@@ -26,12 +27,12 @@ static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecial
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
// clang-format off
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// ######| NDim| InData| WeiData| MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ######| Spatial| Type| Type| Type| Type| Type| | | Layout| | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ######| | | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
// clang-format on
#include "run_convnd_fwd_dl_example.inc"
......
example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
View file @ 289f15de
......@@ -61,6 +61,7 @@ bool run_convnd_fwd_dl_example(int argc, char* argv[])
ndim_spatial_value,
InDataType,
WeiDataType,
DsDataType,
OutDataType,
InElementOp,
WeiElementOp,
......
example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
View file @ 289f15de
......@@ -16,6 +16,7 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
......@@ -140,9 +141,7 @@ make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
{
std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
std::copy(begin(problem_size.output_spatial_lengths_),
end(problem_size.output_spatial_lengths_),
std::back_inserter(dimensions));
ck::ranges::copy(problem_size.output_spatial_lengths_, std::back_inserter(dimensions));
return HostTensorDescriptor(dimensions);
}
......@@ -158,10 +157,3 @@ void unpack_host_tensor_descriptor(const HostTensorDescriptor& descriptor,
assert(size(descriptor.GetStrides()) == size(strides));
std::copy_n(begin(descriptor.GetStrides()), size(descriptor.GetStrides()), begin(strides));
}
template <typename Range, typename OutputIterator>
auto copy(const Range& range, OutputIterator iter)
-> decltype(std::copy(std::begin(range), std::end(range), iter))
{
return std::copy(std::begin(range), std::end(range), iter);
}
example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
View file @ 289f15de
......@@ -77,15 +77,12 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input.begin(),
conv_input.end());
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight.begin(),
conv_weight.end());
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input.begin(), conv_input.end());
ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight.begin(),
conv_weight.end());
ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input);
ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight);
}
DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
......@@ -123,10 +120,10 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
conv_output_g_n_k_wos_desc, conv_output_g_n_k_wos_lengths, conv_output_g_n_k_wos_strides);
unpack_host_tensor_descriptor(r0_desc, r0_lengths, r0_strides);
copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
copy(problem_size.input_left_pads_, begin(input_left_pads));
copy(problem_size.input_right_pads_, begin(input_right_pads));
ck::ranges::copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
ck::ranges::copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
ck::ranges::copy(problem_size.input_left_pads_, begin(input_left_pads));
ck::ranges::copy(problem_size.input_right_pads_, begin(input_right_pads));
// run Conv + Reduction on device
auto conv = DeviceInstance<NDimSpatial>{};
......@@ -276,16 +273,13 @@ bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
conv_output_device_buf.FromDevice(conv_output_device.mData.data());
r0_device_buf.FromDevice(r0_device.mData.data());
return ck::utils::check_err(conv_output_device.mData,
conv_output_host.mData,
return ck::utils::check_err(conv_output_device,
conv_output_host,
"Error: incorrect results! (Matrix E)",
1e-5f,
1e-4f) &&
ck::utils::check_err(r0_device.mData,
r0_host.mData,
"Error: incorrect results! (Matrix R0)",
1e-5f,
1e-4f);
ck::utils::check_err(
r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-5f, 1e-4f);
}
return true;
......
example/12_reduce/reduce_blockwise.cpp
View file @ 289f15de
......@@ -142,7 +142,7 @@ bool reduce_blockwise_test(bool do_verification,
std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
ck::ranges::copy(reduceDims, arrReduceDims.begin());
result = reduce_blockwise_impl<InOutDataType,
AccDataType,
......
example/12_reduce/reduce_blockwise_impl.hpp
View file @ 289f15de
......@@ -10,6 +10,7 @@
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -263,10 +264,10 @@ int reduce_blockwise_impl(bool do_verification,
std::array<index_t, NumOutDim> arrOutLengths;
std::array<index_t, NumOutDim> arrOutStrides;
std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
ck::ranges::copy(inLengths, arrInLengths.begin());
ck::ranges::copy(inStrides, arrInStrides.begin());
ck::ranges::copy(outLengths, arrOutLengths.begin());
ck::ranges::copy(outStrides, arrOutStrides.begin());
auto reduce = DeviceReduceInstance{};
......@@ -324,12 +325,12 @@ int reduce_blockwise_impl(bool do_verification,
#endif
out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
pass = pass && ck::utils::check_err(out, out_ref);
if(OutputIndex)
{
out_index_dev.FromDevice(out_indices.mData.data());
pass = pass && ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
};
};
......
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