commit f71823280e81f70e5ebfb3e82ac418515212f622
author Marat Dukhan <maratek@google.com> Thu Sep 09 18:53:46 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Thu Sep 09 18:54:43 2021 -0700
tree 6fc8a481264051b11d466e0a50b9c2393a219b93
parent 548542c13770c9af1cf08542bc2c0138098beb18

X8 LUT NEON microkernels

PiperOrigin-RevId: 395833468

CMakeLists.txt

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 161f3d8..87b0128 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1962,7 +1962,7 @@
   src/math/sqrt-neonfma-nr2fma1adj.c
   src/math/sqrt-neonfma-nr3fma.c)
 
-SET(PROD_AARCH64_NEONFMA_MICROKERNEL_SRCS
+SET(PROD_AARCH64_NEON_MICROKERNEL_SRCS
   src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c
   src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c
   src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c
@@ -1981,7 +1981,7 @@
   src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c
   src/f32-vsqrt/gen/neon-sqrt-x4.c)
 
-SET(ALL_AARCH64_NEONFMA_MICROKERNEL_SRCS
+SET(ALL_AARCH64_NEON_MICROKERNEL_SRCS
   src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c
   src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c
   src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c
@@ -2093,7 +2093,11 @@
   src/math/sigmoid-neonfma-rr1-p5-div.c
   src/math/sigmoid-neonfma-rr2-lut64-p2-div.c
   src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c
-  src/math/sigmoid-neonfma-rr2-p5-div.c)
+  src/math/sigmoid-neonfma-rr2-p5-div.c
+  src/x8-lut/gen/lut-neon-tbx128x4-x16.c
+  src/x8-lut/gen/lut-neon-tbx128x4-x32.c
+  src/x8-lut/gen/lut-neon-tbx128x4-x48.c
+  src/x8-lut/gen/lut-neon-tbx128x4-x64.c)
 
 SET(PROD_NEONV8_MICROKERNEL_SRCS
   src/f32-vrnd/gen/vrndd-neonv8-x8.c
@@ -4849,13 +4853,13 @@
   LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_NEON_MICROKERNEL_SRCS})
   LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_NEONFMA_MICROKERNEL_SRCS})
   LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_NEONV8_MICROKERNEL_SRCS})
-  LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_AARCH64_NEONFMA_MICROKERNEL_SRCS})
+  LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_AARCH64_NEON_MICROKERNEL_SRCS})
   LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_AARCH64_NEONFP16ARITH_MICROKERNEL_SRCS})
   LIST(APPEND PROD_MICROKERNEL_SRCS ${PROD_NEONDOT_MICROKERNEL_SRCS})
   LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_NEON_MICROKERNEL_SRCS})
   LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_NEONFMA_MICROKERNEL_SRCS})
   LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_NEONV8_MICROKERNEL_SRCS})
-  LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_AARCH64_NEONFMA_MICROKERNEL_SRCS})
+  LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_AARCH64_NEON_MICROKERNEL_SRCS})
   LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_AARCH64_NEONFP16ARITH_MICROKERNEL_SRCS})
   LIST(APPEND ALL_MICROKERNEL_SRCS ${ALL_NEONDOT_MICROKERNEL_SRCS})
   IF(XNNPACK_ENABLE_ASSEMBLY)