diff options
| author | Parag Salasakar <img.mips1@gmail.com> | 2015-07-02 08:02:19 +0530 |
|---|---|---|
| committer | Parag Salasakar <img.mips1@gmail.com> | 2015-07-02 08:02:19 +0530 |
| commit | 6abf1aea6323df0bcc27e3806f8efda27837ba78 (patch) | |
| tree | 5d513f2ea418a348f15020357dbcad0c439ced04 /vpx_dsp/vpx_dsp_rtcd_defs.pl | |
| parent | e7578084299b7cc24e987367b20180ad4a91d78f (diff) | |
| download | libvpx-6abf1aea6323df0bcc27e3806f8efda27837ba78.tar libvpx-6abf1aea6323df0bcc27e3806f8efda27837ba78.tar.gz libvpx-6abf1aea6323df0bcc27e3806f8efda27837ba78.tar.bz2 libvpx-6abf1aea6323df0bcc27e3806f8efda27837ba78.zip | |
mips msa vpx_dsp sadx3 sadx8 optimization
average improvement ~3x-5x
Change-Id: Ifdb4670d31ae83c4e22a4238293d1377b16c90db
Diffstat (limited to 'vpx_dsp/vpx_dsp_rtcd_defs.pl')
| -rw-r--r-- | vpx_dsp/vpx_dsp_rtcd_defs.pl | 26 |
1 files changed, 16 insertions, 10 deletions
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index f12270c1c..f78d9ab6a 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -125,47 +125,53 @@ specialize qw/vpx_sad4x4_avg msa/, "$sse_x86inc"; # # Blocks of 3 add_proto qw/void vpx_sad64x64x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad64x64x3 msa/; add_proto qw/void vpx_sad32x32x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad32x32x3 msa/; add_proto qw/void vpx_sad16x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad16x16x3 sse3 ssse3/; +specialize qw/vpx_sad16x16x3 sse3 ssse3 msa/; add_proto qw/void vpx_sad16x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad16x8x3 sse3 ssse3/; +specialize qw/vpx_sad16x8x3 sse3 ssse3 msa/; add_proto qw/void vpx_sad8x16x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad8x16x3 sse3/; +specialize qw/vpx_sad8x16x3 sse3 msa/; add_proto qw/void vpx_sad8x8x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad8x8x3 sse3/; +specialize qw/vpx_sad8x8x3 sse3 msa/; add_proto qw/void vpx_sad4x4x3/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad4x4x3 sse3/; +specialize qw/vpx_sad4x4x3 sse3 msa/; # Blocks of 8 add_proto qw/void vpx_sad64x64x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad64x64x8 msa/; add_proto qw/void vpx_sad32x32x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad32x32x8 msa/; add_proto qw/void vpx_sad16x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad16x16x8 sse4_1/; +specialize qw/vpx_sad16x16x8 sse4_1 msa/; add_proto qw/void vpx_sad16x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad16x8x8 sse4_1/; +specialize qw/vpx_sad16x8x8 sse4_1 msa/; add_proto qw/void vpx_sad8x16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad8x16x8 sse4_1/; +specialize qw/vpx_sad8x16x8 sse4_1 msa/; add_proto qw/void vpx_sad8x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad8x8x8 sse4_1/; +specialize qw/vpx_sad8x8x8 sse4_1 msa/; add_proto qw/void vpx_sad8x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad8x4x8 msa/; add_proto qw/void vpx_sad4x8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; +specialize qw/vpx_sad4x8x8 msa/; add_proto qw/void vpx_sad4x4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array"; -specialize qw/vpx_sad4x4x8 sse4_1/; +specialize qw/vpx_sad4x4x8 sse4_1 msa/; # # Multi-block SAD, comparing a reference to N independent blocks |