Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "882e51a47a460b26ed065ae612918791a0e4ff0a"
Unverified Commit 5c7cec11 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Code clean up (#20) 



* tuning para,

* testing on v100

* add fp16

* remove deprecated tensor descriptor

* sync with miopen

* update build script
Co-authored-by: default avatarJing Zhang <jizhan@amd.com>
parent 7d09790a
Hide whitespace changes
Inline Side-by-side
Showing with 6044 additions and 296 deletions
driver/src/conv_driver.cpp
View file @ 5c7cec11
......@@ -3,38 +3,28 @@
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "ConstantTensorDescriptor_deprecated.hpp"
#include "print_array.hpp"
#include "print_sequence.hpp"
#include "device.hpp"
#include "tensor_generator.hpp"
#include "host_tensor_generator.hpp"
#include "conv_common.hpp"
#include "host_conv.hpp"
#include "device_tensor.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_padded.hpp"
//#include "device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v2_chwn_cyxk_khwn.hpp"
//#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_deprecated.hpp"
#include "device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw.hpp"
//#include "device_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_deprecated.hpp"
#include "device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp"
int main(int argc, char* argv[])
{
using namespace ck;
#if 1
// 1x1
constexpr index_t N = 64;
constexpr index_t C = 64;
constexpr index_t HI = 56;
constexpr index_t WI = 56;
#if 0
// 1x1, 17x17
constexpr index_t N = 128;
constexpr index_t C = 1024;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
......@@ -45,12 +35,87 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x7
// 1x1, 8x8
constexpr index_t N = 128;
constexpr index_t C = 1024;
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, 73x73
constexpr index_t N = 128;
constexpr index_t C = 160;
constexpr index_t HI = 73;
constexpr index_t WI = 73;
constexpr index_t K = 64;
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, 35x35
constexpr index_t N = 128;
constexpr index_t C = 96;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 96;
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
// 3x3, 71x71
constexpr index_t N = 128;
constexpr index_t C = 192;
constexpr index_t HI = 71;
constexpr index_t WI = 71;
constexpr index_t K = 192;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<2, 2>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 0
// 7x1, 17x17
constexpr index_t N = 128;
constexpr index_t C = 128;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 1024;
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, 17x17
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;
......@@ -60,59 +125,74 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<0, 3>;
using RightPads = Sequence<0, 3>;
#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;
// 3x3, 299x299 stride=2
constexpr index_t N = 128;
constexpr index_t C = 3;
constexpr index_t HI = 299;
constexpr index_t WI = 299;
constexpr index_t K = 32;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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
// 3x3, 147x147
// v4r4@v100 xx.xx%, cudnn@v100 xx.xx%
constexpr index_t N = 128;
constexpr index_t C = 128;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 128;
constexpr index_t C = 32;
constexpr index_t HI = 147;
constexpr index_t WI = 147;
constexpr index_t K = 64;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<2, 2>;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 0
// 3x3, 149x149
// v4r4@v100 xx.xx%, cudnn@v100 xx.xx%
constexpr index_t N = 128;
constexpr index_t C = 32;
constexpr index_t HI = 149;
constexpr index_t WI = 149;
constexpr index_t K = 32;
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;
// 3x3, 17x17, stride 2
constexpr index_t N = 128;
constexpr index_t C = 192;
constexpr index_t HI = 17;
constexpr index_t WI = 17;
constexpr index_t K = 192;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 1x1, 35x35
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 C = 384;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 96;
constexpr index_t Y = 1;
constexpr index_t X = 1;
......@@ -121,110 +201,134 @@ int main(int argc, char* argv[])
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%
#elif 1
// 3x3, 35x35, stride 2
constexpr index_t N = 128;
constexpr index_t C = 832;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t C = 288;
constexpr index_t HI = 35;
constexpr index_t WI = 35;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t X = 1;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 1x3, 8x8
constexpr index_t N = 128;
constexpr index_t C = 1280;
constexpr index_t C = 384;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
constexpr index_t K = 448;
constexpr index_t Y = 1;
constexpr index_t X = 1;
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>;
using LeftPads = Sequence<0, 1>;
using RightPads = Sequence<0, 1>;
#elif 0
// 1x1 filter, 14x14 image
// cudnn@V100 62%, ck@V100 68%, ck@P100 70%, ck@VII 50%
// 3x1, 8x8
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 C = 448;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 512;
constexpr index_t Y = 3;
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>;
using LeftPads = Sequence<1, 0>;
using RightPads = Sequence<1, 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;
// 3x1, 8x8
constexpr index_t N = 128;
constexpr index_t C = 448;
constexpr index_t HI = 8;
constexpr index_t WI = 8;
constexpr index_t K = 384;
constexpr index_t Y = 1;
constexpr index_t K = 512;
constexpr index_t Y = 3;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<1, 0>;
using RightPads = Sequence<1, 0>;
#elif 0
// 3x3, 147x147
constexpr index_t N = 128;
constexpr index_t C = 64;
constexpr index_t HI = 147;
constexpr index_t WI = 147;
constexpr index_t K = 96;
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
// 1x1 filter, 28x28 image
// cudnn@V100 86%, ck@V100 84%, ck@P100 80%, ck@VII 69%
// 7x1, 73x73
// v44@v100 xx.xx%, cudnn@v100 xx.xx%
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 C = 64;
constexpr index_t HI = 73;
constexpr index_t WI = 73;
constexpr index_t K = 64;
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 0
// 3x3, 73x73
constexpr index_t N = 128;
constexpr index_t C = 64;
constexpr index_t HI = 73;
constexpr index_t WI = 73;
constexpr index_t K = 96;
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, 7x7 image
// cudnn@V100 71%, ck@V100 55%, ck@P100 70%, ck@VII 62%
// 1x1, 14x14, stride 2
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 C = 1024;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 2048;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 1x1, 14x14
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 C = 1024;
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;
......@@ -234,63 +338,104 @@ int main(int argc, char* argv[])
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%
// 1x1, 14x14, stride 2
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t C = 1024;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
constexpr index_t K = 128;
constexpr index_t K = 512;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 3x3, 28x28
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 = 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 1
// 3x3, 14x14
constexpr index_t N = 128;
constexpr index_t C = 528;
constexpr index_t C = 256;
constexpr index_t HI = 14;
constexpr index_t WI = 14;
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<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 1
// 1x1, 56x56, stride 2
constexpr index_t N = 128;
constexpr index_t C = 256;
constexpr index_t HI = 56;
constexpr index_t WI = 56;
constexpr index_t K = 128;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 7x7, 230x230 stride=2
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 C = 3;
constexpr index_t HI = 230;
constexpr index_t WI = 230;
constexpr index_t K = 64;
constexpr index_t Y = 7;
constexpr index_t X = 7;
using ConvStrides = Sequence<2, 2>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x1, 28x28, stride = 2
constexpr index_t N = 128;
constexpr index_t C = 512;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 1024;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvStrides = Sequence<2, 2>;
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%
// 1x1, 28x28, stride 2
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;
constexpr index_t C = 512;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 256;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<2, 2>;
using ConvDilations = Sequence<1, 1>;
......@@ -298,71 +443,92 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 5x5 filter, 2x2 pad, 7x7 input
// 1x1, 7x7
constexpr index_t N = 128;
constexpr index_t C = 48;
constexpr index_t C = 512;
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;
constexpr index_t K = 2048;
constexpr index_t Y = 1;
constexpr index_t X = 1;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<2, 2>;
using RightPads = Sequence<2, 2>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 0
// 1x7 filter, 0x3 pad, 17x17 input
// 3x3, 7x7
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;
constexpr index_t C = 512;
constexpr index_t HI = 7;
constexpr index_t WI = 7;
constexpr index_t K = 512;
constexpr index_t Y = 3;
constexpr index_t X = 3;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 3>;
using RightPads = Sequence<0, 3>;
using LeftPads = Sequence<1, 1>;
using RightPads = Sequence<1, 1>;
#elif 1
// 7x1 filter, 3x0 pad, 17x17 input
// 1x1, 56x56
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 C = 64;
constexpr index_t HI = 56;
constexpr index_t WI = 56;
constexpr index_t K = 64;
constexpr index_t Y = 1;
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>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 1
// 3x3, 56x56
constexpr index_t N = 128;
constexpr index_t C = 64;
constexpr index_t HI = 56;
constexpr index_t WI = 56;
constexpr index_t K = 64;
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>;
#endif
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 out_nkhw_desc = get_convolution_output_default_4d_tensor_descriptor_deprecated(
auto in_nchw_desc = make_native_tensor_descriptor_packed(Sequence<N, C, HI, WI>{});
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(
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: ");
ostream_tensor_descriptor(in_nchw_desc, std::cout << "in_nchw_desc: ");
ostream_tensor_descriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_tensor_descriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
print_sequence("LeftPads", LeftPads{});
print_sequence("RightPads", RightPads{});
print_sequence("ConvStrides", ConvStrides{});
print_sequence("ConvDilations", ConvDilations{});
#if 1
using in_data_t = float;
using out_data_t = float;
Tensor<in_data_t> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<in_data_t> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<out_data_t> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<out_data_t> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
#else
using in_data_t = half_float::half;
using out_data_t = half_float::half;
#endif
Tensor<in_data_t> in_nchw(make_HostTensorDescriptor(in_nchw_desc));
Tensor<in_data_t> wei_kcyx(make_HostTensorDescriptor(wei_kcyx_desc));
Tensor<out_data_t> out_nkhw_host(make_HostTensorDescriptor(out_nkhw_desc));
Tensor<out_data_t> out_nkhw_device(make_HostTensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency();
......@@ -399,42 +565,7 @@ int main(int argc, char* argv[])
#endif
}
#if 0
device_convolution_direct_v2_nchw_kcyx_nkhw
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn_padded(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
LeftPads{},
RightPads{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v2_chwn_cyxk_khwn(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(
(in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw_deprecated(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
#if 1
device_convolution_implicit_gemm_v4r1_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
......@@ -446,36 +577,6 @@ int main(int argc, char* argv[])
LeftPads{},
RightPads{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r2_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r3_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw_deprecated(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 1
device_convolution_implicit_gemm_v4r4_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
......@@ -492,7 +593,7 @@ int main(int argc, char* argv[])
if(do_verification)
{
#if 1
#if 0
if(Y == 3 && X == 3 && ConvStrides{}[0] == 1 && ConvStrides{}[1] == 1 &&
ConvDilations{}[0] == 1 && ConvDilations{}[1] == 1)
{
......
driver/src/device.cpp
View file @ 5c7cec11
......@@ -6,7 +6,7 @@ DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
#if CK_DEVICE_BACKEND_AMD
hipGetErrorString(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
#elif CK_DEVICE_BACKEND_NVIDIA
checkCudaErrors(cudaMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
cudaMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize);
#endif
}
......@@ -18,8 +18,7 @@ void DeviceMem::ToDevice(const void* p)
hipGetErrorString(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
#elif CK_DEVICE_BACKEND_NVIDIA
checkCudaErrors(
cudaMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, cudaMemcpyHostToDevice));
cudaMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, cudaMemcpyHostToDevice);
#endif
}
......@@ -28,7 +27,7 @@ void DeviceMem::FromDevice(void* p)
#if CK_DEVICE_BACKEND_AMD
hipGetErrorString(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
#elif CK_DEVICE_BACKEND_NVIDIA
checkCudaErrors(cudaMemcpy(p, mpDeviceBuf, mMemSize, cudaMemcpyDeviceToHost));
cudaMemcpy(p, mpDeviceBuf, mMemSize, cudaMemcpyDeviceToHost);
#endif
}
......@@ -37,7 +36,7 @@ DeviceMem::~DeviceMem()
#if CK_DEVICE_BACKEND_AMD
hipGetErrorString(hipFree(mpDeviceBuf));
#elif CK_DEVICE_BACKEND_NVIDIA
checkCudaErrors(cudaFree(mpDeviceBuf));
cudaFree(mpDeviceBuf);
#endif
}
......@@ -68,8 +67,10 @@ struct KernelTimerImpl
void Start()
{
#if CK_DEVICE_BACKEND_AMD
hipDeviceSynchronize();
hipEventRecord(mStart, 0);
#elif CK_DEVICE_BACKEND_NVIDIA
cudaDeviceSynchronize();
cudaEventRecord(mStart, 0);
#endif
}
......
driver/src/tensor.cpp → driver/src/host_tensor.cpp
View file @ 5c7cec11
#include <boost/range/adaptor/transformed.hpp>
#include <cassert>
#include "tensor.hpp"
#include "host_tensor.hpp"
template <typename X>
TensorDescriptor::TensorDescriptor(std::vector<X> lens) : mLens(lens)
HostTensorDescriptor::HostTensorDescriptor(std::vector<X> lens) : mLens(lens)
{
this->CalculateStrides();
}
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)
{
}
void TensorDescriptor::CalculateStrides()
void HostTensorDescriptor::CalculateStrides()
{
mStrides.clear();
mStrides.resize(mLens.size(), 0);
......@@ -27,21 +27,21 @@ void TensorDescriptor::CalculateStrides()
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());
return std::accumulate(
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; });
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; }
external/half/include/half.hpp 0 → 100644
View file @ 5c7cec11
This source diff could not be displayed because it is too large. You can view the blob instead.
script/cmake-cuda.sh
View file @ 5c7cec11
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=/home/chao/code/modular_convolution
MY_PROJECT_INSTALL=../install.dir
MY_PROJECT_SOURCE=../../../
export CUDA_ROOT=/usr/local/cuda
export CPATH=$CPATH:$CUDA_ROOT/include
export LIBRARY_PATH=$LIBRARY_PATH:$CUDA_ROOT/lib64
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_ROOT/lib64
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_CXX_COMPILER=clang++ \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D DEVICE_BACKEND=NVIDIA \
-D CUDA_COMMON_INCLUDE_DIR="/package/install/cuda/10.1/NVIDIA_CUDA-10.1_Samples/common/inc" \
-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -maxrregcount=128" \
${MY_PROJECT_SOURCE}
#-D BOOST_ROOT="/package/install/boost_1.67.0" \
#-D CMAKE_CUDA_COMPILER="/package/install/cuda_10.0/bin/nvcc" \
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61" \
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -maxrregcount=128" \
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_70,code=sm_70" \
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_70,code=sm_70 -Xptxas -v -maxrregcount=128" \
script/cmake-cuda_docker.sh deleted 100755 → 0
View file @ 7d09790a
#!/bin/bash
MY_PROJECT_SOURCE=../../../
MY_PROJECT_INSTALL=../install.dir
export CUDA_ROOT=/usr/local/cuda
export CPATH=$CPATH:$CUDA_ROOT/include
export LIBRARY_PATH=$LIBRARY_PATH:$CUDA_ROOT/lib64
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_ROOT/lib64
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_CXX_COMPILER=clang++-6.0 \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D DEVICE_BACKEND=NVIDIA \
-D CUDA_COMMON_INCLUDE_DIR="/root/NVIDIA_CUDA-10.1_Samples/common/inc" \
-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -maxrregcount=128" \
${MY_PROJECT_SOURCE}
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61" \
#-D CMAKE_CUDA_FLAGS="-ccbin clang++ -m64 -Xcompiler -fopenmp -lineinfo --source-in-ptx -keep -Xptxas -v -gencode=arch=compute_61,code=sm_61 -Xptxas -v -maxrregcount=128" \
script/cmake-hip.sh → script/cmake-rocm3.1.sh
View file @ 5c7cec11
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
......@@ -11,9 +10,10 @@ cmake
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_BUILD_TYPE=Release \
-D DEVICE_BACKEND="AMD" \
-D HIP_HIPCC_FLAGS="${HIP_HIPCC_FLAGS} -gline-tables-only -v" \
-D CMAKE_CXX_FLAGS="-gline-tables-only --amdgpu-target=gfx906" \
-D CMAKE_CXX_FLAGS="--amdgpu-target=gfx906" \
-D CMAKE_CXX_COMPILER=/opt/rocm/hip/bin/hipcc \
-D CMAKE_PREFIX_PATH="/opt/rocm" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE}
#-D CMAKE_CXX_FLAGS="-gline-tables-only -v --amdgpu-target=gfx906" \
script/cmake-rocm3.5.sh 0 → 100755
View file @ 5c7cec11
#!/bin/bash
rm -f CMakeCache.txt
rm -f *.cmake
rm -rf CMakeFiles
MY_PROJECT_SOURCE=../../../
MY_PROJECT_INSTALL=../install.dir
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D CMAKE_BUILD_TYPE=Release \
-D DEVICE_BACKEND="AMD" \
-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH="/opt/rocm" \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE}
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -save-temps" \
#-D CMAKE_CXX_FLAGS="-O3 --amdgpu-target=gfx906 -mllvm --amdgpu-enable-global-sgpr-addr -mllvm --amdgpu-spill-vgpr-to-agpr=0 -v -gline-tables-only -save-temps" \
script/docker-cuda.sh deleted 100755 → 0
View file @ 7d09790a
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/hack_isa.sh deleted 100755 → 0
View file @ 7d09790a
# step 1: GET ISA DUMP
#cd /root/workspace/mlopen/modular_convolution/build/hipcc/build.dir/driver && KMDUMPISA=1 /opt/rocm/hip/bin/hipcc -I/root/workspace/mlopen/modular_convolution/build/hipcc/build.dir/composable_kernel/include/utility -I/root/workspace/mlopen/modular_convolution/driver/include -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/kernel_algorithm -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/tensor_operation -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/tensor_description -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/utility -I/root/workspace/mlopen/modular_convolution/composable_kernel/include -gline-tables-only --amdgpu-target=gfx906 -fopenmp=libomp -O3 -DNDEBUG -std=c++14 -o CMakeFiles/driver.dir/src/driver.cpp.o -c /root/workspace/mlopen/modular_convolution/driver/src/driver.cpp -fno-gpu-rdc
# step 2: HACK ISA
#cd /root/workspace/mlopen/modular_convolution/build/hipcc/build.dir/driver && KMHACKISA=1 /opt/rocm/hip/bin/hipcc -I/root/workspace/mlopen/modular_convolution/build/hipcc/build.dir/composable_kernel/include/utility -I/root/workspace/mlopen/modular_convolution/driver/include -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/kernel_algorithm -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/tensor_operation -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/tensor_description -I/root/workspace/mlopen/modular_convolution/composable_kernel/include/utility -I/root/workspace/mlopen/modular_convolution/composable_kernel/include -gline-tables-only --amdgpu-target=gfx906 -fopenmp=libomp -O3 -DNDEBUG -std=c++14 -o CMakeFiles/driver.dir/src/driver.cpp.o -c /root/workspace/mlopen/modular_convolution/driver/src/driver.cpp -fno-gpu-rdc
# step 3: LINK
#/opt/rocm/hip/bin/hipcc -gline-tables-only --amdgpu-target=gfx906 -fopenmp=libomp -O3 -DNDEBUG CMakeFiles/driver.dir/src/driver.cpp.o -o driver -rdynamic libhost.so -Wl,-rpath,/root/workspace/mlopen/modular_convolution/build/hipcc/build.dir/driver
script/trace.sh deleted 100755 → 0
View file @ 7d09790a
#!/bin/bash
/root/workspace/rocprofiler_pkg/bin/rpl_run.sh --timestamp on -i /root/workspace/rocprofiler_pkg/input.xml -d ./trace ./driver/driver 0 10
script/tracer-hip.sh deleted 100755 → 0
View file @ 7d09790a
#!/bin/bash
/root/workspace/rocprofiler_pkg/bin/rpl_run.sh --timestamp on -i /root/workspace/rocprofiler_pkg/input.xml -d ./trace ./driver/driver 0 10
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