Difference between revisions of "Porting/ncnn"
Jump to navigation
Jump to search
JeremyRand (talk | contribs) (Finished: Update Vulkan dependency docs) |
JeremyRand (talk | contribs) (VSX Targets) |
||
| Line 14: | Line 14: | ||
* CI missing for POWER9/Clang | * CI missing for POWER9/Clang | ||
* Replace SSE with native VSX | * Replace SSE with native VSX | ||
| + | |||
| + | = VSX Targets = | ||
| + | |||
| + | When running Real-ESRGAN in ncnn on POWER9, most CPU time (over 81%) is spent inside [https://github.com/Tencent/ncnn/blob/575098640c254be2208095254f3e5de412751447/src/layer/x86/convolution_3x3_winograd.h#L613 <code>gemm_transB_packed_tile</code> in <code>convolution_3x3_winograd.h</code>], which uses SSE2. This may be a good target for rewriting in native VSX. | ||
= See Also = | = See Also = | ||
Revision as of 05:11, 18 December 2023
Contents
Finished
- Translate x86_64 SSE to ppc64le VSX intrinsics
- VSX toolchains: check for SSE2 support
- Add POWER8 VSX toolchains
- load_param_mem pybind
- test_squeezenet failed under Lubuntu 16.04 PowerPC 32-bit
- support big endian platform, add powerpc ci
- Update POWER Clang version docs
- Update Vulkan dependency docs
In progress
- CI missing for POWER9/Clang
- Replace SSE with native VSX
VSX Targets
When running Real-ESRGAN in ncnn on POWER9, most CPU time (over 81%) is spent inside gemm_transB_packed_tile in convolution_3x3_winograd.h, which uses SSE2. This may be a good target for rewriting in native VSX.
