Skip to content

GitLab

Commit 6fc49f91 authored by Chao Liu's avatar Chao Liu
Browse files

remove deprecated tensor descriptor

parent 506a823a
Show whitespace changes
Inline Side-by-side
Showing with 33 additions and 612 deletions
driver/include/tensor_generator.hpp → driver/include/host_tensor_generator.hpp
View file @ 6fc49f91
#ifndef TENSOR_GENERATOR_HPP #ifndef HOST_TENSOR_GENERATOR_HPP
#define TENSOR_GENERATOR_HPP #define HOST_TENSOR_GENERATOR_HPP
#include "config.hpp" #include "config.hpp"
......
driver/src/conv_bwd_data_driver.cpp
View file @ 6fc49f91
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "print_array.hpp" #include "print_array.hpp"
#include "print_sequence.hpp" #include "print_sequence.hpp"
#include "device.hpp" #include "device.hpp"
#include "tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "conv_common.hpp" #include "conv_common.hpp"
#include "host_conv_bwd_data.hpp" #include "host_conv_bwd_data.hpp"
...@@ -209,19 +209,19 @@ int main(int argc, char* argv[]) ...@@ -209,19 +209,19 @@ int main(int argc, char* argv[])
constexpr auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor( constexpr auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{}); in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: "); ostream_tensor_descriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: "); ostream_tensor_descriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: "); ostream_tensor_descriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
print_sequence("LeftPads", LeftPads{}); print_sequence("LeftPads", LeftPads{});
print_sequence("LeftPads", LeftPads{}); print_sequence("LeftPads", LeftPads{});
print_sequence("RightPads", RightPads{}); print_sequence("RightPads", RightPads{});
print_sequence("ConvStrides", ConvStrides{}); print_sequence("ConvStrides", ConvStrides{});
print_sequence("ConvDilations", ConvDilations{}); print_sequence("ConvDilations", ConvDilations{});
Tensor<float> in_nchw_device(make_TensorDescriptor(in_nchw_desc)); Tensor<float> in_nchw_device(make_HostTensorDescriptor(in_nchw_desc));
Tensor<float> in_nchw_host(make_TensorDescriptor(in_nchw_desc)); Tensor<float> in_nchw_host(make_HostTensorDescriptor(in_nchw_desc));
Tensor<float> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc)); Tensor<float> wei_kcyx(make_HostTensorDescriptor(wei_kcyx_desc));
Tensor<float> out_nkhw(make_TensorDescriptor(out_nkhw_desc)); Tensor<float> out_nkhw(make_HostTensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = std::thread::hardware_concurrency();
......
driver/src/conv_bwd_data_driver.cu deleted 100644 → 0
View file @ 506a823a
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "config.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "print_array.hpp"
#include "print_sequence.hpp"
#include "device.hpp"
#include "tensor_generator.hpp"
#include "device_tensor.hpp"
#include "conv_common.hpp"
#include "host_conv_bwd_data.hpp"
#include "device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw.hpp"
#include "device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw.hpp"
#include "device_convolution_backward_data_implicit_gemm_v2r1_nchw_kcyx_nkhw.hpp"
#include "device_convolution_backward_data_implicit_gemm_v3r1_nchw_kcyx_nkhw.hpp"
#include "device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
int main(int argc, char* argv[])
{
using namespace launcher;
#if 0
constexpr index_t N = 64;
constexpr index_t C = 256;
constexpr index_t HI = 56;
constexpr index_t WI = 56;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 3x3, 34x34
constexpr index_t N = 64;
constexpr index_t C = 256;
constexpr index_t HI = 34;
constexpr index_t WI = 34;
constexpr index_t K = 256;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 3x3, 28x28
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 1024;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 0
// 1x1 filter, 8x8 image
constexpr index_t N = 256;
constexpr index_t C = 1024;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 1024;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1 filter, 7x7 image
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 1024;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1 filter, 14x14 image
constexpr index_t N = 128;
constexpr index_t C = 512;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1 filter, 28x28 image
constexpr index_t N = 128;
constexpr index_t C = 128;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1 filter, 17x17 input
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 1024;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 5x5 filter, 2x2 pad, 7x7 input
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 1024;
constexpr index_t Y = 5;
constexpr index_t X = 5;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<2, 2>;
using RightPads = Sequence<2, 2>;
#elif 0
// 1x7 filter, 0x3 pad, 17x17 input
constexpr index_t N = 128;
constexpr index_t C = 128;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 7;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 3>;
using RightPads = Sequence<0, 3>;
#elif 0
// 7x1 filter, 3x0 pad, 17x17 input
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 1024;
constexpr index_t Y = 7;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<3, 0>;
using RightPads = Sequence<3, 0>;
#elif 1
// 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
constexpr index_t N = 128;
constexpr index_t C = 128;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 1024;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<2, 2>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#endif
constexpr auto in_nchw_desc = make_native_tensor_descriptor_packed(Sequence<N, C, HI, WI>{});
constexpr auto wei_kcyx_desc = make_native_tensor_descriptor_packed(Sequence<K, C, Y, X>{});
constexpr auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
print_sequence("LeftPads", LeftPads{});
print_sequence("LeftPads", LeftPads{});
print_sequence("RightPads", RightPads{});
print_sequence("ConvStrides", ConvStrides{});
print_sequence("ConvDilations", ConvDilations{});
Tensor<float> in_nchw_device(make_TensorDescriptor(in_nchw_desc));
Tensor<float> in_nchw_host(make_TensorDescriptor(in_nchw_desc));
Tensor<float> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<float> out_nkhw(make_TensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency();
if(argc != 3)
{
printf("arg1: do_verification, arg2: nrepeat\n");
exit(1);
}
bool do_verification = atoi(argv[1]);
std::size_t nrepeat = atoi(argv[2]);
if(do_verification)
{
#if 0
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{1}, num_thread);
out_nkhw.GenerateTensorValue(GeneratorTensor_1{1}, num_thread);
#else
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
out_nkhw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#endif
}
#if 0
device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw
#elif 0
device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw
#elif 0
device_convolution_backward_data_implicit_gemm_v2r1_nchw_kcyx_nkhw
#elif 0
device_convolution_backward_data_implicit_gemm_v3r1_nchw_kcyx_nkhw
#elif 1
device_convolution_backward_data_implicit_gemm_v4r1_nchw_kcyx_nkhw
#endif
(in_nchw_desc,
in_nchw_device,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw,
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{},
nrepeat);
if(do_verification)
{
host_direct_convolution_backward_data(in_nchw_host,
wei_kcyx,
out_nkhw,
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{});
check_error(in_nchw_host, in_nchw_device);
#if 0
LogRange(std::cout << "out_nkhw : ", out_nkhw.mData, ",") << std::endl;
LogRange(std::cout << "wei_kcyx : ", wei_kcyx.mData, ",") << std::endl;
LogRange(std::cout << "in_nchw_host : ", in_nchw_host.mData, ",") << std::endl;
LogRange(std::cout << "in_nchw_device : ", in_nchw_device.mData, ",") << std::endl;
#endif
}
}
driver/src/conv_bwd_data_driver.cu 0 → 120000
View file @ 6fc49f91
conv_bwd_data_driver.cpp
\ No newline at end of file
driver/src/conv_driver.cpp
View file @ 6fc49f91
#include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h> #include <stdlib.h>
#include <half.hpp> #include <half.hpp>
#include "config.hpp" #include "config.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "print_array.hpp" #include "print_array.hpp"
#include "print_sequence.hpp" #include "print_sequence.hpp"
#include "device.hpp" #include "device.hpp"
#include "tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "conv_common.hpp" #include "conv_common.hpp"
#include "host_conv.hpp" #include "host_conv.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
...@@ -505,14 +504,14 @@ int main(int argc, char* argv[]) ...@@ -505,14 +504,14 @@ int main(int argc, char* argv[])
using RightPads = Sequence<1, 1>; using RightPads = Sequence<1, 1>;
#endif #endif
auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{}); auto in_nchw_desc = make_native_tensor_descriptor_packed(Sequence<N, C, HI, WI>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor_packed(Sequence<K, C, Y, X>{}); auto wei_kcyx_desc = make_native_tensor_descriptor_packed(Sequence<K, C, Y, X>{});
auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor_deprecated( auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor(
in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{}); in_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
ostream_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: "); ostream_tensor_descriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: "); ostream_tensor_descriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: "); ostream_tensor_descriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
print_sequence("LeftPads", LeftPads{}); print_sequence("LeftPads", LeftPads{});
print_sequence("RightPads", RightPads{}); print_sequence("RightPads", RightPads{});
print_sequence("ConvStrides", ConvStrides{}); print_sequence("ConvStrides", ConvStrides{});
...@@ -526,10 +525,10 @@ int main(int argc, char* argv[]) ...@@ -526,10 +525,10 @@ int main(int argc, char* argv[])
using out_data_t = half_float::half; using out_data_t = half_float::half;
#endif #endif
Tensor<in_data_t> in_nchw(make_TensorDescriptor(in_nchw_desc)); Tensor<in_data_t> in_nchw(make_HostTensorDescriptor(in_nchw_desc));
Tensor<in_data_t> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc)); Tensor<in_data_t> wei_kcyx(make_HostTensorDescriptor(wei_kcyx_desc));
Tensor<out_data_t> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc)); Tensor<out_data_t> out_nkhw_host(make_HostTensorDescriptor(out_nkhw_desc));
Tensor<out_data_t> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc)); Tensor<out_data_t> out_nkhw_device(make_HostTensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency(); std::size_t num_thread = std::thread::hardware_concurrency();
......
driver/src/tensor.cpp → driver/src/host_tensor.cpp
View file @ 6fc49f91
#include <boost/range/adaptor/transformed.hpp> #include <boost/range/adaptor/transformed.hpp>
#include <cassert> #include <cassert>
#include "tensor.hpp" #include "host_tensor.hpp"
template <typename X> template <typename X>
TensorDescriptor::TensorDescriptor(std::vector<X> lens) : mLens(lens) HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens) : mLens(lens)
{ {
this->CalculateStrides(); this->CalculateStrides();
} }
template <typename X, typename Y> template <typename X, typename Y>
TensorDescriptor::TensorDescriptor(std::vector<X> lens, std::vector<Y> strides) HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens, std::vector<Y> strides)
: mLens(lens), mStrides(strides) : mLens(lens), mStrides(strides)
{ {
} }
void TensorDescriptor::CalculateStrides() void HostTensorDescriptor::CalculateStrides()
{ {
mStrides.clear(); mStrides.clear();
mStrides.resize(mLens.size(), 0); mStrides.resize(mLens.size(), 0);
...@@ -27,21 +27,21 @@ void TensorDescriptor::CalculateStrides() ...@@ -27,21 +27,21 @@ void TensorDescriptor::CalculateStrides()
mLens.rbegin(), mLens.rend() - 1, mStrides.rbegin() + 1, std::multiplies<std::size_t>()); mLens.rbegin(), mLens.rend() - 1, mStrides.rbegin() + 1, std::multiplies<std::size_t>());
} }
std::size_t TensorDescriptor::GetNumOfDimension() const { return mLens.size(); } std::size_t HostTensorDescriptor::GetNumOfDimension() const { return mLens.size(); }
std::size_t TensorDescriptor::GetElementSize() const std::size_t HostTensorDescriptor::GetElementSize() const
{ {
assert(mLens.size() == mStrides.size()); assert(mLens.size() == mStrides.size());
return std::accumulate( return std::accumulate(
mLens.begin(), mLens.end(), std::size_t{1}, std::multiplies<std::size_t>()); mLens.begin(), mLens.end(), std::size_t{1}, std::multiplies<std::size_t>());
} }
std::size_t TensorDescriptor::GetElementSpace() const std::size_t HostTensorDescriptor::GetElementSpace() const
{ {
auto ls = mLens | boost::adaptors::transformed([](std::size_t v) { return v - 1; }); auto ls = mLens | boost::adaptors::transformed([](std::size_t v) { return v - 1; });
return std::inner_product(ls.begin(), ls.end(), mStrides.begin(), std::size_t{0}) + 1; return std::inner_product(ls.begin(), ls.end(), mStrides.begin(), std::size_t{0}) + 1;
} }
const std::vector<std::size_t>& TensorDescriptor::GetLengths() const { return mLens; } const std::vector<std::size_t>& HostTensorDescriptor::GetLengths() const { return mLens; }
const std::vector<std::size_t>& TensorDescriptor::GetStrides() const { return mStrides; } const std::vector<std::size_t>& HostTensorDescriptor::GetStrides() const { return mStrides; }
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