diff options
| author | Frank Barchard <fbarchard@google.com> | 2020-05-06 01:12:04 -0700 |
|---|---|---|
| committer | XNNPACK Team <xnnpack-github-robot@google.com> | 2020-05-06 01:12:34 -0700 |
| commit | 3f9f99f372081793f1f3765af7fd51046209b8ee (patch) | |
| tree | c2a6d79bd0d4adfa2167c8d25fe4648e43d35be2 /scripts/generate-f16-gemm.sh | |
| parent | 5d761c426d3978a1c364dc391e187fe70e10bf13 (diff) | |
| download | XNNPACK-3f9f99f372081793f1f3765af7fd51046209b8ee.tar.gz | |
Nx16 FP16 intrinsic GEMM and IGEMM ukernels
IGEMM performance mobilenet_v2 on Pixel 4
f16_igemm_6x16__neonfp16arith_ld64 8986179
f16_igemm_4x16__neonfp16arith_ld64 9969704
f16_igemm_8x16__neonfp16arith_ld64 10405198
f16_igemm_1x16__neonfp16arith_ld64 11418367
f16_igemm_6x8__neonfp16arith_ld64 11488318
f16_igemm_4x8__neonfp16arith_ld64 11566898
f16_igemm_8x8__neonfp16arith_ld64 11992713
f16_igemm_1x8__neonfp16arith_ld64 24999989
GEMM performance mobilenet_v2 on Pixel 4
f16_gemm_6x8__aarch64_neonfp16arith_ld64 5026706
f16_gemm_6x16__aarch64_neonfp16arith_ld32 5067141
f16_gemm_4x16__aarch64_neonfp16arith_ld32 5083430
f16_gemm_8x8__aarch64_neonfp16arith_ld64 5157512
f16_gemm_4x8__aarch64_neonfp16arith_ld64 5215077
f16_gemm_6x16__neonfp16arith_ld64 8605063
f16_gemm_4x16__neonfp16arith_ld64 8856811
f16_gemm_1x16__aarch64_neonfp16arith_ld32 9828810
f16_gemm_1x16__neonfp16arith_ld64 10187374
f16_gemm_1x8__aarch64_neonfp16arith_ld64 10449552
f16_gemm_4x8__neonfp16arith_ld64 10537397
f16_gemm_6x8__neonfp16arith_ld64 10797656
f16_gemm_8x8__neonfp16arith_ld64 11174740
f16_gemm_8x16__neonfp16arith_ld64 12283543
f16_gemm_1x8__neonfp16arith_ld64 16208999
PiperOrigin-RevId: 310106462
Diffstat (limited to 'scripts/generate-f16-gemm.sh')
| -rwxr-xr-x | scripts/generate-f16-gemm.sh | 9 |
1 files changed, 9 insertions, 0 deletions
diff --git a/scripts/generate-f16-gemm.sh b/scripts/generate-f16-gemm.sh index b91daf8b0..443af5ad0 100755 --- a/scripts/generate-f16-gemm.sh +++ b/scripts/generate-f16-gemm.sh @@ -32,5 +32,14 @@ tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=4 -D NR=8 -D INC=1 -o src tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=6 -D NR=8 -D INC=1 -o src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=8 -D NR=8 -D INC=1 -o src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=1 -D NR=16 -D INC=0 -o src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=4 -D NR=16 -D INC=0 -o src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=6 -D NR=16 -D INC=0 -o src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=8 -D NR=16 -D INC=0 -o src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=1 -D NR=16 -D INC=1 -o src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=4 -D NR=16 -D INC=1 -o src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=6 -D NR=16 -D INC=1 -o src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c +tools/xngen src/f16-gemm/neonfp16arith-ld64.c.in -D MR=8 -D NR=16 -D INC=1 -o src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c + ################################## Unit tests ################################# tools/generate-gemm-test.py --spec test/f16-gemm-minmax.yaml --output test/f16-gemm-minmax.cc |