ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 10:37:44 +00:00
Code Issues Packages Projects Releases Wiki Activity

Added bwd data v3r1 v4r1, tweaking v1 (#10)

Browse Source 
* Added bwd data v3r1: breaking down compute into a series of load balanced GEMM, and launch in a single kernel
* Added bwd data v4r1: like v3r1, but launch GEMMs in multiple kernels
* Tweaked v1r1  and v1r2 (atomic) on AMD GPU

[ROCm/composable_kernel commit: c5da0377fb]
This commit is contained in:
Chao Liu
2020-01-20 10:20:03 -06:00
committed by GitHub
parent 24f7d66609
commit 7c9100b53f
43 changed files with 2123 additions and 452 deletions
Download Patch File Download Diff File Expand all files Collapse all files

81
driver/src/conv_driver.cpp
View File

@@ -30,32 +30,63 @@ int main(int argc, char* argv[])
using namespace ck;
#if 0
constexpr index_t N = 8;
constexpr index_t C = 32;
constexpr index_t HI = 28;
constexpr index_t WI = 28;
constexpr index_t K = 32;
constexpr index_t Y = 5;
constexpr index_t X = 5;
// 1x1
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<2, 2>;
using ConvDilations = Sequence<1, 1>;
using LeftPads = Sequence<0, 0>;
using RightPads = Sequence<0, 0>;
#elif 1
#elif 0
// 1x7
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 = 7;
using ConvStrides = Sequence<1, 1>;
using ConvDilations = Sequence<1, 1>;
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 = 128;
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 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 = 128;
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
@@ -282,21 +313,6 @@ int main(int argc, char* argv[])
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;
@@ -311,6 +327,21 @@ int main(int argc, char* argv[])
using LeftPads = Sequence<0, 3>;
using RightPads = Sequence<0, 3>;
#elif 1
// 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>;
#endif
auto in_nchw_desc = make_ConstantTensorDescriptor_packed(Sequence<N, C, HI, WI>{});