Skip to content

GitLab

Unverified Commit 8f5f6496 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

backward data (#7) 

* enabled atomic add in tensor copy
* added gridwise GEMM
* added backward data conv using GEMM + atomic
* added backward data conv using GEMM, no atomic
parent 31ded4ac
Hide whitespace changes
Inline Side-by-side
Showing with 786 additions and 79 deletions
driver/src/col2im_driver.cpp 0 → 100644
View file @ 8f5f6496
#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_col2im.hpp"
#include "device_col2im_eb_nchw.hpp"
int main(int argc, char* argv[])
{
using namespace ck;
#if 1
constexpr index_t N = 2;
constexpr index_t C = 8;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 128;
constexpr index_t Y = 4;
constexpr index_t X = 4;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<2, 2>;
#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 = 128;
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
// 1x1 filter, 8x8 image
// cudnn@V100 68%, ck@V100 72%, ck@P100 52%, ck@VII 42%
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
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
// 1x1 filter, 8x8 image
// cudnn@V100 77%, ck@V100 76%, ck@P100 79%, ck@VII 51%
constexpr index_t N = 128;
constexpr index_t C = 2048;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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
// cudnn@V100 82%, ck@V100 76%, ck@P100 67%, ck@VII 64%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 384;
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, 8x8 image
// cudnn@V100 83%, ck@V100 75%, ck@P100 78%, ck@VII 65%
constexpr index_t N = 128;
constexpr index_t C = 1280;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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
// cudnn@V100 62%, ck@V100 68%, ck@P100 70%, ck@VII 50%
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, 8x8 image
// cudnn@V100 74%, ck@V100 57%, ck@P100 78%, ck@VII 61%
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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
// cudnn@V100 86%, ck@V100 84%, ck@P100 80%, ck@VII 69%
constexpr index_t N = 128;
constexpr index_t C = 256;
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, 7x7 image
// cudnn@V100 71%, ck@V100 55%, ck@P100 70%, ck@VII 62%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
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
// 1x1 filter, 17x17 input
// cudnn@V100 81%, ck@V100 76%, ck@P100 70%, ck@VII 76%
constexpr index_t N = 128;
constexpr index_t C = 768;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
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, 14x14 image
// cudnn@V100 73%, ck@V100 71%, ck@P100 70%, ck@VII 64%
constexpr index_t N = 128;
constexpr index_t C = 528;
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, 14x14 image
// cudnn@V100 73%, ck@V100 72%, ck@P100 79%, ck@VII 75%
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
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
// 1x1 filter, 7x7 image
// cudnn@V100 49%, ck@V100 50%, ck@P100 61%, ck@VII 52%
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
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
// 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
// cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81%
constexpr index_t N = 128;
constexpr index_t C = 288;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 384;
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>;
#elif 0
// 5x5 filter, 2x2 pad, 7x7 input
constexpr index_t N = 128;
constexpr index_t C = 48;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 128;
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
// 7x1 filter, 3x0 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 = 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
// 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>;
#endif
constexpr auto img_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(
img_nchw_desc, wei_kcyx_desc, ConvStrides{}, ConvDilations{}, LeftPads{}, RightPads{});
constexpr index_t HO = out_nkhw_desc.GetLengths()[2];
constexpr index_t WO = out_nkhw_desc.GetLengths()[3];
constexpr auto col_eb_desc =
make_native_tensor_descriptor_packed(Sequence<C * Y * X, N * HO * WO>{});
using FilterSizes = Sequence<Y, X>;
using OutputSizes = Sequence<HO, WO>;
ostream_ConstantTensorDescriptor(col_eb_desc, std::cout << "col_eb_desc: ");
ostream_ConstantTensorDescriptor(img_nchw_desc, std::cout << "img_nchw_desc: ");
print_sequence("FilterSizes", FilterSizes{});
print_sequence("OutputSizes", OutputSizes{});
print_sequence("LeftPads", LeftPads{});
print_sequence("LeftPads", LeftPads{});
print_sequence("RightPads", RightPads{});
print_sequence("ConvStrides", ConvStrides{});
print_sequence("ConvDilations", ConvDilations{});
Tensor<float> col_eb(make_TensorDescriptor(col_eb_desc));
Tensor<float> img_nchw_host(make_TensorDescriptor(img_nchw_desc));
Tensor<float> img_nchw_device(make_TensorDescriptor(img_nchw_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
col_eb.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#else
col_eb.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#endif
}
device_col2im_eb_nchw(col_eb_desc,
col_eb,
img_nchw_desc,
img_nchw_device,
FilterSizes{},
OutputSizes{},
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{},
nrepeat);
if(do_verification)
{
host_col2im(col_eb,
img_nchw_host,
FilterSizes{},
OutputSizes{},
ConvStrides{},
ConvDilations{},
LeftPads{},
RightPads{});
check_error(img_nchw_host, img_nchw_device);
#if 0
LogRange(std::cout << "col_eb : ", col_eb.mData, ",") << std::endl;
LogRange(std::cout << "img_nchw_host : ", img_nchw_host.mData, ",") << std::endl;
LogRange(std::cout << "img_nchw_device : ", img_nchw_device.mData, ",") << std::endl;
#endif
}
}
driver/src/col2im_driver.cu 0 → 120000
View file @ 8f5f6496
col2im_driver.cpp
\ No newline at end of file
driver/src/conv_bwd_data_driver.cpp 0 → 100644
View file @ 8f5f6496
#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"
int main(int argc, char* argv[])
{
using namespace ck;
#if 0
constexpr index_t N = 8;
constexpr index_t C = 128;
constexpr index_t HI = 16;
constexpr index_t WI = 16;
constexpr index_t K = 8;
constexpr index_t Y = 2;
constexpr index_t X = 2;
using ConvStrides = Sequence<4, 4>;
using ConvDilations = Sequence<2, 2>;
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 = 128;
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
// 1x1 filter, 8x8 image
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
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
// 1x1 filter, 8x8 image
constexpr index_t N = 128;
constexpr index_t C = 2048;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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 = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 384;
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, 8x8 image
constexpr index_t N = 128;
constexpr index_t C = 1280;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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, 8x8 image
constexpr index_t N = 64;
constexpr index_t C = 1536;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
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 = 256;
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, 7x7 image
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
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
// 1x1 filter, 17x17 input
constexpr index_t N = 128;
constexpr index_t C = 768;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
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, 14x14 image
constexpr index_t N = 128;
constexpr index_t C = 528;
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, 14x14 image
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
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
// 1x1 filter, 7x7 image
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
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
// 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
constexpr index_t N = 128;
constexpr index_t C = 288;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 384;
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>;
#elif 0
// 5x5 filter, 2x2 pad, 7x7 input
constexpr index_t N = 128;
constexpr index_t C = 48;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 128;
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
// 7x1 filter, 3x0 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 = 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
// 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>;
#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
#else
device_convolution_backward_data_implicit_gemm_v2r1_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 @ 8f5f6496
conv_bwd_data_driver.cpp
\ No newline at end of file
driver/src/driver.cpp → driver/src/conv_driver.cpp
View file @ 8f5f6496
...@@ -8,8 +8,10 @@ ...@@ -8,8 +8,10 @@
#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 "conv_common.hpp" #include "conv_common.hpp"
#include "host_conv.hpp" #include "host_conv.hpp"
#include "device_tensor.hpp"
//#include "device_convolution_direct_v2_nchw_kcyx_nkhw.hpp" //#include "device_convolution_direct_v2_nchw_kcyx_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp" //#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn.hpp"
//#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded.hpp" //#include "device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded.hpp"
...@@ -23,73 +25,24 @@ ...@@ -23,73 +25,24 @@
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_deprecated.hpp" #include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_deprecated.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp" #include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
struct GeneratorTensor_1
{
template <class... Is>
double operator()(Is... is)
{
return 1;
}
};
struct GeneratorTensor_2
{
int min_value = 0;
int max_value = 1;
template <class... Is>
double operator()(Is...)
{
return (std::rand() % (max_value - min_value)) + min_value;
}
};
struct GeneratorTensor_3
{
template <class... Is>
double operator()(Is... is)
{
std::array<index_t, sizeof...(Is)> dims = {{static_cast<index_t>(is)...}};
auto f_acc = [](auto a, auto b) { return 10 * a + b; };
return std::accumulate(dims.begin(), dims.end(), index_t(0), f_acc);
}
};
struct GeneratorTensor_Checkboard
{
template <class... Ts>
double operator()(Ts... Xs) const
{
std::array<index_t, sizeof...(Ts)> dims = {{Xs...}};
return std::accumulate(dims.begin(),
dims.end(),
true,
[](bool init, index_t x) -> int { return init != (x % 2); })
? 1
: -1;
}
};
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
using namespace ck; using namespace ck;
#if 0 #if 0
constexpr index_t N = 128; constexpr index_t N = 8;
constexpr index_t C = 128; constexpr index_t C = 32;
constexpr index_t HI = 17; constexpr index_t HI = 28;
constexpr index_t WI = 17; constexpr index_t WI = 28;
constexpr index_t K = 128; constexpr index_t K = 32;
constexpr index_t Y = 1; constexpr index_t Y = 5;
constexpr index_t X = 7; constexpr index_t X = 5;
using ConvStrides = Sequence<1, 1>; using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>; using ConvDilations = Sequence<2, 2>;
using LeftPads = Sequence<0, 3>; using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 3>; using RightPads = Sequence<0, 0>;
#elif 0 #elif 0
// 3x3, 34x34 // 3x3, 34x34
constexpr index_t N = 64; constexpr index_t N = 64;
...@@ -297,7 +250,7 @@ int main(int argc, char* argv[]) ...@@ -297,7 +250,7 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<0, 0>; using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>; using RightPads = Sequence<0, 0>;
#elif 0 #elif 1
// 3x3 filter, 2x2 stride, 35x35 input, 17x17 output // 3x3 filter, 2x2 stride, 35x35 input, 17x17 output
// cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81% // cudnn@V100 90%, ck@V100 93%, ck@P100 83%, ck@VII 81%
constexpr index_t N = 128; constexpr index_t N = 128;
...@@ -343,7 +296,7 @@ int main(int argc, char* argv[]) ...@@ -343,7 +296,7 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<3, 0>; using LeftPads = Sequence<3, 0>;
using RightPads = Sequence<3, 0>; using RightPads = Sequence<3, 0>;
#elif 1 #elif 0
// 1x7 filter, 0x3 pad, 17x17 input // 1x7 filter, 0x3 pad, 17x17 input
constexpr index_t N = 128; constexpr index_t N = 128;
constexpr index_t C = 128; constexpr index_t C = 128;
...@@ -362,7 +315,7 @@ int main(int argc, char* argv[]) ...@@ -362,7 +315,7 @@ int main(int argc, char* argv[])
auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{}); auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{});
auto wei_kcyx_desc = make_ConstantTensorDescriptor_packed(Sequence<K, C, Y, X>{}); auto wei_kcyx_desc = make_ConstantTensorDescriptor_packed(Sequence<K, C, Y, X>{});
auto out_nkhw_desc = get_convolution_with_padding_output_default_4d_tensor_descriptor( auto out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor_deprecated(
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_ConstantTensorDescriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
...@@ -492,7 +445,7 @@ int main(int argc, char* argv[]) ...@@ -492,7 +445,7 @@ int main(int argc, char* argv[])
ConvStrides{}, ConvStrides{},
ConvDilations{}, ConvDilations{},
nrepeat); nrepeat);
#elif 0 #elif 1
device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc, device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw, in_nchw,
wei_kcyx_desc, wei_kcyx_desc,
......
driver/src/conv_driver.cu 0 → 120000
View file @ 8f5f6496
conv_driver.cpp
\ No newline at end of file
driver/src/driver.cu deleted 120000 → 0
View file @ 31ded4ac
driver.cpp
\ No newline at end of file
driver/src/tensor.cpp
View file @ 8f5f6496
...@@ -3,12 +3,14 @@ ...@@ -3,12 +3,14 @@
#include "tensor.hpp" #include "tensor.hpp"
TensorDescriptor::TensorDescriptor(std::initializer_list<std::size_t> lens) : mLens(lens) template <typename X>
TensorDescriptor::TensorDescriptor(std::vector<X> lens) : mLens(lens)
{ {
this->CalculateStrides(); this->CalculateStrides();
} }
TensorDescriptor::TensorDescriptor(std::vector<std::size_t> lens, std::vector<std::size_t> strides) template <typename X, typename Y>
TensorDescriptor::TensorDescriptor(std::vector<X> lens, std::vector<Y> strides)
: mLens(lens), mStrides(strides) : mLens(lens), mStrides(strides)
{ {
} }
......
script/compile-hip.sh
View file @ 8f5f6496
...@@ -4,5 +4,5 @@ ...@@ -4,5 +4,5 @@
export KMDUMPLLVM=1 export KMDUMPLLVM=1
export KMDUMPDIR=$PWD export KMDUMPDIR=$PWD
make -j driver make -j $1
#/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.asm #/opt/rocm/hcc/bin/llvm-objdump -mcpu=gfx906 -source -line-numbers driver/dump-gfx906.isabin > driver/dump-gfx906.isabin.asm
script/docker-cuda.sh 0 → 100755
View file @ 8f5f6496
WORKSPACE=$1
echo "workspace: " $WORKSPACE
sudo docker run -it -v $WORKSPACE:/root/workspace --group-add sudo --runtime=nvidia asroy/cuda:10.1-cudnn7-devel-ubuntu18.04-latest /bin/bash
script/ds_read_offset.sh deleted 100755 → 0
View file @ 31ded4ac
for((i=0;i<=4096;i=i+64))
do
OFFSET=$i
echo "if(offset == $OFFSET)"
echo "{"
echo " asm volatile(\"\\n \\"
echo " ds_read_b128 %0, %1 offset:$OFFSET\n \\"
echo " \""
echo " : \"=v\"(r)"
echo " : \"v\"(__to_local(lds)));"
echo "}"
done
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