Refactor PReLU microkernels
- Support processing of multiple tiles of rows to lower function call overhead.
- Optimized SSE2 and NEON versions of the micro-kernel
- Unit test generator
- Process multiple SIMD vector of channels in a each call
- Rename arguments to use more descriptive names
PiperOrigin-RevId: 278916987
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7ad7f4c..9f83dd7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -195,7 +195,8 @@
src/f32-ppmm/3x3-scalar.c
src/f32-ppmm/4x2-scalar.c
src/f32-ppmm/4x4-scalar.c
- src/f32-prelu/x4-scalar.c
+ src/f32-prelu/scalar-2x1.c
+ src/f32-prelu/scalar-2x4.c
src/f32-rmax/scalar.c
src/f32-spmm/1x1-scalar-pipelined.c
src/f32-spmm/1x1-scalar.c
@@ -283,7 +284,8 @@
src/f32-pavgpool/mp9p8q-psimd.c
src/f32-pavgpool/up9-psimd.c
src/f32-ppmm/4x8-psimd.c
- src/f32-prelu/x4-psimd.c
+ src/f32-prelu/psimd-2x4.c
+ src/f32-prelu/psimd-2x8.c
src/f32-vadd/psimd.c
src/f32-vmul/psimd.c
src/f32-vmulcaddc/c4-psimd-x2.c
@@ -326,6 +328,8 @@
src/f32-pavgpool/up9-neon.c
src/f32-ppmm/4x8-neon.c
src/f32-ppmm/8x8-neon.c
+ src/f32-prelu/neon-2x4.c
+ src/f32-prelu/neon-2x8.c
src/f32-rmax/neon.c
src/f32-vmulcaddc/c4-neon-x2.c
src/q8-avgpool/mp9p8q-neon.c
@@ -445,7 +449,6 @@
src/f32-dwconv-spchw/3x3p1-sse.c
src/f32-dwconv-spchw/3x3s2p1-sse.c
src/f32-ppmm/4x8-sse.c
- src/f32-prelu/x4-sse.c
src/f32-rmax/sse.c
src/f32-spmm/4x1-sse.c
src/f32-spmm/8x1-sse.c
@@ -459,6 +462,8 @@
src/f32-argmaxpool/mp9p8q-sse2.c
src/f32-argmaxpool/up4-sse2.c
src/f32-argmaxpool/up9-sse2.c
+ src/f32-prelu/sse2-2x4.c
+ src/f32-prelu/sse2-2x8.c
src/q8-avgpool/mp9p8q-sse2.c
src/q8-avgpool/up9-sse2.c
src/q8-igemm/4x4c2-sse2.c
@@ -481,6 +486,10 @@
src/x8-zip/x4-sse2.c
src/x8-zip/xm-sse2.c)
+SET(XNNPACK_SSE41_MICROKERNEL_SRCS
+ src/f32-prelu/sse41-2x4.c
+ src/f32-prelu/sse41-2x8.c)
+
SET(XNNPACK_AVX_MICROKERNEL_SRCS
src/f32-rmax/avx.c
src/f32-vscale/avx-unroll32.c)
@@ -580,6 +589,7 @@
IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|x86_64)$" OR IOS_ARCH MATCHES "^(i386|x86_64)$")
LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_SSE_MICROKERNEL_SRCS})
LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_SSE2_MICROKERNEL_SRCS})
+ LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_SSE41_MICROKERNEL_SRCS})
LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX_MICROKERNEL_SRCS})
LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX2_MICROKERNEL_SRCS})
LIST(APPEND XNNPACK_MICROKERNEL_SRCS ${XNNPACK_AVX512F_MICROKERNEL_SRCS})
@@ -615,6 +625,7 @@
IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|x86_64)$" OR IOS_ARCH MATCHES "^(i386|x86_64)$")
SET_PROPERTY(SOURCE ${XNNPACK_SSE_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse ")
SET_PROPERTY(SOURCE ${XNNPACK_SSE2_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse2 ")
+ SET_PROPERTY(SOURCE ${XNNPACK_SSE41_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -msse4.1 ")
SET_PROPERTY(SOURCE ${XNNPACK_AVX_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mavx ")
SET_PROPERTY(SOURCE ${XNNPACK_AVX2_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mfma -mavx2 ")
SET_PROPERTY(SOURCE ${XNNPACK_AVX512F_MICROKERNEL_SRCS} APPEND_STRING PROPERTY COMPILE_FLAGS " -mavx512f ")