diff options
| author | Luca Barbato <lu_zero@gentoo.org> | 2017-04-12 01:18:35 +0200 |
|---|---|---|
| committer | James Zern <jzern@google.com> | 2017-04-19 19:57:51 -0700 |
| commit | 0b9be9320518dfe08057ad0e091e604d6d7d67a3 (patch) | |
| tree | 5435e9d664030c385f8d14686c224d21bed4d058 | |
| parent | ee9325b0bd24ed0c9ed11f85a6d5aa1518b5e9c0 (diff) | |
| download | libvpx-0b9be9320518dfe08057ad0e091e604d6d7d67a3.tar libvpx-0b9be9320518dfe08057ad0e091e604d6d7d67a3.tar.gz libvpx-0b9be9320518dfe08057ad0e091e604d6d7d67a3.tar.bz2 libvpx-0b9be9320518dfe08057ad0e091e604d6d7d67a3.zip | |
ppc: d63 predictor 8x8
10x faster.
Change-Id: I7cedbf4df2ce7df5b6f1108b11815d088fdb9ba8
| -rw-r--r-- | test/test_intra_pred_speed.cc | 2 | ||||
| -rw-r--r-- | vpx_dsp/ppc/intrapred_vsx.c | 23 | ||||
| -rw-r--r-- | vpx_dsp/vpx_dsp_rtcd_defs.pl | 2 |
3 files changed, 25 insertions, 2 deletions
diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc index 34b20cbcf..539f7d4fb 100644 --- a/test/test_intra_pred_speed.cc +++ b/test/test_intra_pred_speed.cc @@ -319,7 +319,7 @@ INTRA_PRED_TEST(VSX, TestIntraPred4, NULL, NULL, NULL, NULL, NULL, INTRA_PRED_TEST(VSX, TestIntraPred8, vpx_dc_predictor_8x8_vsx, NULL, NULL, NULL, NULL, NULL, vpx_d45_predictor_8x8_vsx, NULL, NULL, NULL, NULL, - NULL, vpx_tm_predictor_8x8_vsx) + vpx_d63_predictor_8x8_vsx, vpx_tm_predictor_8x8_vsx) INTRA_PRED_TEST(VSX, TestIntraPred16, vpx_dc_predictor_16x16_vsx, vpx_dc_left_predictor_16x16_vsx, vpx_dc_top_predictor_16x16_vsx, diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c index 8216e3e9e..0ff0f935d 100644 --- a/vpx_dsp/ppc/intrapred_vsx.c +++ b/vpx_dsp/ppc/intrapred_vsx.c @@ -642,6 +642,29 @@ void vpx_d45_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride, } } +void vpx_d63_predictor_8x8_vsx(uint8_t *dst, ptrdiff_t stride, + const uint8_t *above, const uint8_t *left) { + const uint8x16_t af = vec_vsx_ld(0, above); + const uint8x16_t above_right = vec_splat(af, 9); + const uint8x16_t a = xxpermdi(af, above_right, 1); + const uint8x16_t b = vec_perm(a, above_right, sl1); + const uint8x16_t c = vec_perm(b, above_right, sl1); + uint8x16_t row0 = vec_avg(a, b); + uint8x16_t row1 = avg3(a, b, c); + int i; + (void)left; + + for (i = 0; i < 4; i++) { + const uint8x16_t d0 = vec_vsx_ld(0, dst); + const uint8x16_t d1 = vec_vsx_ld(0, dst + stride); + vec_vsx_st(xxpermdi(row0, d0, 1), 0, dst); + vec_vsx_st(xxpermdi(row1, d1, 1), 0, dst + stride); + dst += stride * 2; + row0 = vec_perm(row0, above_right, sl1); + row1 = vec_perm(row1, above_right, sl1); + } +} + void vpx_d63_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride, const uint8_t *above, const uint8_t *left) { const uint8x16_t a0 = vec_vsx_ld(0, above); diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl index ba458c872..ef4ce24e9 100644 --- a/vpx_dsp/vpx_dsp_rtcd_defs.pl +++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl @@ -78,7 +78,7 @@ add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, con specialize qw/vpx_d45_predictor_8x8 neon sse2 vsx/; add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; -specialize qw/vpx_d63_predictor_8x8 ssse3/; +specialize qw/vpx_d63_predictor_8x8 ssse3 vsx/; add_proto qw/void vpx_h_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left"; specialize qw/vpx_h_predictor_8x8 neon dspr2 msa sse2/; |