ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add xdlops v4r4r4 into online compilation (#48)

Browse Source 
* init for v4r4 xdlops olc

* refactor wrap

* init impl of v4r4 nchw xdlops olc

* tuning

* test perf

* fixed v4r4 nhwc

* tuned v4r4 nhwc

* use gridwise_gemm_xdlops_v2r3

* swap a/b

* add pointer support into offline v2r3

* debugging v4r4r4 transform for olc

* change timer of olc

* refactor v4r4 xdlops nchw olc

* remove transform fun in v4r4 xdlops nhwc olc

Co-authored-by: Chao Liu <chao.liu2@amd.com>
This commit is contained in:
zjing14
2021-07-16 23:27:08 -05:00
committed by GitHub
parent 0a72e4df94
commit fbdf4332c7
28 changed files with 1851 additions and 3909 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
driver/conv_driver_v2.cpp
View File

@@ -19,8 +19,6 @@
#include "device_dynamic_convolution_forward_implicit_gemm_v4r5r2_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v5r1_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk.hpp"
#include "device_dynamic_convolution_forward_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp"
#define USE_DYNAMIC_MODE 1
@@ -30,9 +28,7 @@
#define USE_CONV_FWD_V4R5_NCHW 0
#define USE_CONV_FWD_V4R5R2_NCHW 0
#define USE_CONV_FWD_V5R1_NCHW 0
#define USE_CONV_FWD_V4R4_XDL_NCHW 1
#define USE_CONV_FWD_V4R4R2_XDL_NHWC 0
#define USE_CONV_FWD_V4R4R3_XDL_NHWC 0
#define USE_CONV_FWD_V4R4R2_XDL_NCHW 1
#define USE_CONV_FWD_V4R4R4_XDL_NHWC 1
enum ConvForwardAlgo
@@ -43,10 +39,8 @@ enum ConvForwardAlgo
V4R5NCHW, // 3
V4R5R2NCHW, // 4
V5R1NCHW, // 5
V4R4XDLNCHW, // 6
V4R4R2XDLNHWC, // 7
V4R4R3XDLNHWC, // 8
V4R4R4XDLNHWC // 9
V4R4R2XDLNCHW, // 6
V4R4R4XDLNHWC // 7
};
int main(int argc, char* argv[])
@@ -462,8 +456,8 @@ int main(int argc, char* argv[])
}
#endif
#if USE_CONV_FWD_V4R4_XDL_NCHW
if(algo == ConvForwardAlgo::V4R4XDLNCHW)
#if USE_CONV_FWD_V4R4R2_XDL_NCHW
if(algo == ConvForwardAlgo::V4R4R2XDLNCHW)
{
if(layout != ConvTensorLayout::NCHW)
{
@@ -489,60 +483,6 @@ int main(int argc, char* argv[])
}
#endif
#if USE_CONV_FWD_V4R4R2_XDL_NHWC
if(algo == ConvForwardAlgo::V4R4R2XDLNHWC)
{
if(layout != ConvTensorLayout::NHWC)
{
throw std::runtime_error("wrong! layout");
}
const auto tmp = f_make_for_device_nhwc();
device_dynamic_convolution_forward_implicit_gemm_v4r4r2_xdlops_nhwc_kyxc_nhwk<in_data_t,
acc_data_t,
out_data_t>(
tmp[I0],
tmp[I1],
tmp[I2],
tmp[I3],
tmp[I4],
tmp[I5],
tmp[I6],
in,
wei,
out_device,
nrepeat);
}
#endif
#if USE_CONV_FWD_V4R4R3_XDL_NHWC
if(algo == ConvForwardAlgo::V4R4R3XDLNHWC)
{
if(layout != ConvTensorLayout::NHWC)
{
throw std::runtime_error("wrong! layout");
}
const auto tmp = f_make_for_device_nhwc();
device_dynamic_convolution_forward_implicit_gemm_v4r4r3_xdlops_nhwc_kyxc_nhwk<in_data_t,
acc_data_t,
out_data_t>(
tmp[I0],
tmp[I1],
tmp[I2],
tmp[I3],
tmp[I4],
tmp[I5],
tmp[I6],
in,
wei,
out_device,
nrepeat);
}
#endif
#if USE_CONV_FWD_V4R4R4_XDL_NHWC
if(algo == ConvForwardAlgo::V4R4R4XDLNHWC)
{