shithub: libvpx

Info  •  Files  •  Log  •  Branches

ppc: d63 predictor 8x8

10x faster.

Change-Id: I7cedbf4df2ce7df5b6f1108b11815d088fdb9ba8

--- a/test/test_intra_pred_speed.cc

+++ b/test/test_intra_pred_speed.cc

@@ -319,7 +319,7 @@

 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,

--- a/vpx_dsp/ppc/intrapred_vsx.c

+++ b/vpx_dsp/ppc/intrapred_vsx.c

@@ -642,6 +642,29 @@

   }

 }

+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);

--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl

+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -78,7 +78,7 @@

 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/;