commit 90eca0a793f285249897d9420dcedf8f876ae22d
author Marat Dukhan <maratek@google.com> Wed Mar 11 00:52:23 2020 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Wed Mar 11 00:53:03 2020 -0700
tree e0c61c8e2731f74b40f60e5481d364beccdcd093
parent f32ae3437e3ac59e9bcab48872281a6a5cf20ca5

AVX and AVX512F versions of PReLU micro-kernel

Up to 2X faster PReLU operator on AVX512F-enabled CPU

PiperOrigin-RevId: 300268958

CMakeLists.txt

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0df92bb..90fb349 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1068,6 +1068,8 @@
   src/f32-igemm/gen/3x16-avx-broadcast.c
   src/f32-igemm/gen/4x16-avx-broadcast.c
   src/f32-igemm/gen/5x16-avx-broadcast.c
+  src/f32-prelu/gen/avx-2x8.c
+  src/f32-prelu/gen/avx-2x16.c
   src/f32-rmax/avx.c
   src/f32-vbinary/gen/vadd-avx-x8.c
   src/f32-vbinary/gen/vadd-avx-x16.c
@@ -1295,6 +1297,8 @@
   src/f32-igemm/gen/6x16-avx512f-broadcast.c
   src/f32-igemm/gen/7x16-avx512f-broadcast.c
   src/f32-igemm/gen/8x16-avx512f-broadcast.c
+  src/f32-prelu/gen/avx512f-2x16.c
+  src/f32-prelu/gen/avx512f-2x32.c
   src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c
   src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c
   src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c