aboutsummaryrefslogtreecommitdiff
path: root/scripts/generate-f32-dwconv2d-chw.sh
diff options
context:
space:
mode:
authorMarat Dukhan <maratek@google.com>2020-10-25 19:14:03 -0700
committerXNNPACK Team <xnnpack-github-robot@google.com>2020-10-25 19:14:39 -0700
commit0ff9718d746f1770847a6b19e4c3e5ba9be38556 (patch)
treed82ea2e237a4acf5a4f450328aee090e72e306af /scripts/generate-f32-dwconv2d-chw.sh
parentb16785a48511bea528eebbccb6ebff3083318621 (diff)
downloadXNNPACK-0ff9718d746f1770847a6b19e4c3e5ba9be38556.tar.gz
Auto-generate SSE versions of DWCONV2D CHW 3x3s2p1 micro-kernels
PiperOrigin-RevId: 338962367
Diffstat (limited to 'scripts/generate-f32-dwconv2d-chw.sh')
-rwxr-xr-xscripts/generate-f32-dwconv2d-chw.sh10
1 files changed, 10 insertions, 0 deletions
diff --git a/scripts/generate-f32-dwconv2d-chw.sh b/scripts/generate-f32-dwconv2d-chw.sh
index a6623ece1..47f588792 100755
--- a/scripts/generate-f32-dwconv2d-chw.sh
+++ b/scripts/generate-f32-dwconv2d-chw.sh
@@ -54,6 +54,16 @@ tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=
tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=1 -D ACCUMULATORS=4 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c
tools/xngen src/f32-dwconv2d-chw/3x3p1-ssse3.c.in -D ROW_TILE=2 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=2 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=3 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=4 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c
+
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=3 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=1 -D ACCUMULATORS=4 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c
+tools/xngen src/f32-dwconv2d-chw/3x3s2p1-sse.c.in -D ROW_TILE=2 -D ACCUMULATORS=2 -o src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c
+
################################### Scalar ####################################
tools/xngen src/f32-dwconv2d-chw/3x3p1-scalar.c.in -D ROW_TILE=1 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c
tools/xngen src/f32-dwconv2d-chw/3x3p1-scalar.c.in -D ROW_TILE=2 -D ACCUMULATORS=1 -o src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c