diff options
| author | Luc Trudeau <luc@trud.ca> | 2018-06-13 13:39:04 -0400 |
|---|---|---|
| committer | Luc Trudeau <luc@trud.ca> | 2018-06-13 19:21:06 +0000 |
| commit | e3ce12cfc1c2d2cc245e1a6d49eaf3ff18538547 (patch) | |
| tree | 76c392d2752ac15a5466f0878db5f3b11aac2db4 /vpx_dsp | |
| parent | f950248b9b357b21e974e3ace94359d7ee8c7b29 (diff) | |
| download | libvpx-e3ce12cfc1c2d2cc245e1a6d49eaf3ff18538547.tar libvpx-e3ce12cfc1c2d2cc245e1a6d49eaf3ff18538547.tar.gz libvpx-e3ce12cfc1c2d2cc245e1a6d49eaf3ff18538547.tar.bz2 libvpx-e3ce12cfc1c2d2cc245e1a6d49eaf3ff18538547.zip | |
VSX Version of SAD8xN
VSX versions of the SAD functions of width 8.
SADTest Speed Test (POWER8 Model 2.1)
8x4 C time = 68.7 ms (±0.3 ms), VSX time = 31.8 ms (±0.1 ms) [2.2x]
8x8 C time = 55.6 ms (±0.3 ms), VSX time = 18.3 ms (±0.1 ms) [3.0x]
8x16 C time = 46.5 ms (±0.1 ms), VSX time = 15.6 ms (±0.1 ms) [3.0x]
Change-Id: I843f3b34e103b72deeade4a939193d8b53cee460
Diffstat (limited to 'vpx_dsp')
| -rw-r--r-- | vpx_dsp/ppc/sad_vsx.c | 25 | ||||
| -rw-r--r-- | vpx_dsp/ppc/types_vsx.h | 1 | ||||
| -rw-r--r-- | vpx_dsp/vpx_dsp_rtcd_defs.pl | 6 |
3 files changed, 29 insertions, 3 deletions
diff --git a/vpx_dsp/ppc/sad_vsx.c b/vpx_dsp/ppc/sad_vsx.c index bb49addae..b41c05b48 100644 --- a/vpx_dsp/ppc/sad_vsx.c +++ b/vpx_dsp/ppc/sad_vsx.c @@ -31,6 +31,28 @@ v_sad = vec_sum4s(v_absh, v_sad); \ v_sad = vec_sum4s(v_absl, v_sad); +#define SAD8(height) \ + unsigned int vpx_sad8x##height##_vsx(const uint8_t *a, int a_stride, \ + const uint8_t *b, int b_stride) { \ + int y = 0; \ + uint8x16_t v_a, v_b, v_abs; \ + uint32x4_t v_sad = vec_zeros_u32; \ + \ + do { \ + v_a = vec_vsx_ld(0, a); \ + v_b = vec_vsx_ld(0, b); \ + \ + v_abs = vec_sub(vec_max(v_a, v_b), vec_min(v_a, v_b)); \ + v_sad = vec_sum4s(v_abs, v_sad); \ + \ + a += a_stride; \ + b += b_stride; \ + y++; \ + } while (y < height); \ + \ + return v_sad[1] + v_sad[0]; \ + } + #define SAD16(height) \ unsigned int vpx_sad16x##height##_vsx(const uint8_t *a, int a_stride, \ const uint8_t *b, int b_stride) { \ @@ -95,6 +117,9 @@ return sad[3] + sad[2] + sad[1] + sad[0]; \ } +SAD8(4); +SAD8(8); +SAD8(16); SAD16(8); SAD16(16); SAD16(32); diff --git a/vpx_dsp/ppc/types_vsx.h b/vpx_dsp/ppc/types_vsx.h index 803d0377a..81c7b970a 100644 --- a/vpx_dsp/ppc/types_vsx.h +++ b/vpx_dsp/ppc/types_vsx.h @@ -82,6 +82,7 @@ static const int16x8_t vec_ones_s16 = { 1, 1, 1, 1, 1, 1, 1, 1 }; static const uint16x8_t vec_ones_u16 = { 1, 1, 1, 1, 1, 1, 1, 1 }; static const uint32x4_t vec_ones_u32 = { 1, 1, 1, 1 }; static const int32x4_t vec_zeros_s32 = { 0, 0, 0, 0 }; +static const uint32x4_t vec_zeros_u32 = { 0, 0, 0, 0 }; static const uint16x8_t vec_shift_sign_s16 = { 15, 15, 15, 15, 15, 15, 15, 15 }; static const uint32x4_t vec_shift_sign_s32 = { 31, 31, 31, 31 }; static const uint8x16_t vec_perm64 = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index f237e5503..9f3e268cc 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -748,13 +748,13 @@ add_proto qw/unsigned int vpx_sad16x8/, "const uint8_t *src_ptr, int src_stride, specialize qw/vpx_sad16x8 neon msa sse2 vsx mmi/; add_proto qw/unsigned int vpx_sad8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; -specialize qw/vpx_sad8x16 neon msa sse2 mmi/; +specialize qw/vpx_sad8x16 neon msa sse2 vsx mmi/; add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; -specialize qw/vpx_sad8x8 neon msa sse2 mmi/; +specialize qw/vpx_sad8x8 neon msa sse2 vsx mmi/; add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; -specialize qw/vpx_sad8x4 neon msa sse2 mmi/; +specialize qw/vpx_sad8x4 neon msa sse2 vsx mmi/; add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride"; specialize qw/vpx_sad4x8 neon msa sse2 mmi/; |