commit 5243bb00db49a6d3c4491f91b3303ca5823f5dfc
author Frank Barchard <fbarchard@google.com> Fri Nov 22 16:37:50 2019 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Fri Nov 22 16:48:27 2019 -0800
tree d153d43babe3eeb452b8f3d9218b718de917c9f9
parent 17ec5f3f7b460701a113e3f916da2a4bc56e22c1

DUP Neon GEMM kernels for Exynos

PiperOrigin-RevId: 282066945

scripts/generate-f32-igemm.sh

diff --git a/scripts/generate-f32-igemm.sh b/scripts/generate-f32-igemm.sh
index 12b1160..e90fd27 100755
--- a/scripts/generate-f32-igemm.sh
+++ b/scripts/generate-f32-igemm.sh
@@ -26,6 +26,16 @@
 ### MRx2 micro-kernels-
 tools/xngen src/f32-igemm/MRx2-neon-ld64.c.in -D MR=4 -D NR=2 -D FMA=0 -D DUP=0 -o src/f32-igemm/4x2-neon-lane-ld64.c
 tools/xngen src/f32-igemm/MRx2-neon-ld64.c.in -D MR=4 -D NR=2 -D FMA=1 -D DUP=0 -o src/f32-igemm/4x2-neonfma-lane-ld64.c
+### DUP LD64 micro-kernels
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=1 -D NR=8 -D FMA=0 -D DUP=1 -o src/f32-igemm/1x8-neon-dup-ld64.c
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=1 -D NR=8 -D FMA=1 -D DUP=1 -o src/f32-igemm/1x8-neonfma-dup-ld64.c
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=4 -D NR=8 -D FMA=0 -D DUP=1 -o src/f32-igemm/4x8-neon-dup-ld64.c
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=4 -D NR=8 -D FMA=1 -D DUP=1 -o src/f32-igemm/4x8-neonfma-dup-ld64.c
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=6 -D NR=8 -D FMA=0 -D DUP=1 -o src/f32-igemm/6x8-neon-dup-ld64.c
+tools/xngen src/f32-igemm/neon-ld64.c.in      -D MR=6 -D NR=8 -D FMA=1 -D DUP=1 -o src/f32-igemm/6x8-neonfma-dup-ld64.c
+### DUP LD128 micro-kernels
+tools/xngen src/f32-igemm/neon-ld128.c.in     -D MR=4 -D NR=8 -D FMA=0 -D DUP=1 -o src/f32-igemm/4x8-neon-dup-ld128.c
+tools/xngen src/f32-igemm/neon-ld128.c.in     -D MR=4 -D NR=8 -D FMA=1 -D DUP=1 -o src/f32-igemm/4x8-neonfma-dup-ld128.c
 ### LOAD4+PERMUTE micro-kernels
 tools/xngen src/f32-igemm/neon-shuffle.c.in   -D MR=1 -D NR=8 -D FMA=0 -o src/f32-igemm/1x8s4-neon.c
 tools/xngen src/f32-igemm/neon-shuffle.c.in   -D MR=1 -D NR=8 -D FMA=1 -o src/f32-igemm/1x8s4-neonfma.c