commit c9f9d67435a5a31543a76f62387429a86f23198a
author Frank Barchard <fbarchard@google.com> Mon Oct 18 12:51:59 2021 -0700
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Oct 18 12:52:56 2021 -0700
tree 6fd165ebba72eb584abe25eb01340a0d68e3c21b
parent dbe781beb8003d74c70952baa725e615f5a6d2da

Add Channel Tile of 16 for float and 32 for half float.

- includes fix for test generator to pass primary_tile to test

PiperOrigin-RevId: 404035415

scripts/generate-f16-dwconv.sh

diff --git a/scripts/generate-f16-dwconv.sh b/scripts/generate-f16-dwconv.sh
index 08fac77..b633771 100755
--- a/scripts/generate-f16-dwconv.sh
+++ b/scripts/generate-f16-dwconv.sh
@@ -10,16 +10,22 @@
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x4-minmax-neonfp16arith.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=4 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x4-minmax-neonfp16arith-acc2.c &
 
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x9-minmax-neonfp16arith.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=9 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x9-minmax-neonfp16arith-acc2.c &
 
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=8  -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c &
 tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=16 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=1 -o src/f16-dwconv/gen/up32x25-minmax-neonfp16arith.c &
+tools/xngen src/f16-dwconv/up-neonfp16arith.c.in -D CHANNEL_TILE=32 -D KERNEL_TILE=25 -D ACCUMULATORS=2 -o src/f16-dwconv/gen/up32x25-minmax-neonfp16arith-acc2.c &
 
 ################################## Unit tests #################################
 tools/generate-dwconv-test.py --spec test/f16-dwconv-minmax.yaml --output test/f16-dwconv-minmax.cc &