shithub: libvpx

Download patch

ref: a7a2d1653bbb6d42d292a0a10acd4d4e28c43d6b
parent: 7ad1faa6f8ed5472864155dca5a7a01b787f0365
author: Luca Barbato <lu_zero@gentoo.org>
date: Fri Apr 7 10:49:00 EDT 2017 

ppc: horizontal predictor 16x16

About 10x faster.

Change-Id: Ie81077fa32ad214cdb46bdcb0be4e9e2c7df47c2


--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -314,8 +314,8 @@
 
 #if HAVE_VSX
 INTRA_PRED_TEST(VSX, TestIntraPred16, NULL, NULL, NULL, NULL,
-                vpx_v_predictor_16x16_vsx, NULL, NULL, NULL, NULL, NULL, NULL,
-                NULL, NULL)
+                vpx_v_predictor_16x16_vsx, vpx_h_predictor_16x16_vsx, NULL,
+                NULL, NULL, NULL, NULL, NULL, NULL)
 
 INTRA_PRED_TEST(VSX, TestIntraPred32, NULL, NULL, NULL, NULL,
                 vpx_v_predictor_32x32_vsx, NULL, NULL, NULL, NULL, NULL, NULL,
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -34,3 +34,61 @@
     vec_vsx_st(d1, 16, dst);
   }
 }
+
+void vpx_h_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t d = vec_vsx_ld(0, left);
+  const uint8x16_t v0 = vec_splat(d, 0);
+  const uint8x16_t v1 = vec_splat(d, 1);
+  const uint8x16_t v2 = vec_splat(d, 2);
+  const uint8x16_t v3 = vec_splat(d, 3);
+
+  const uint8x16_t v4 = vec_splat(d, 4);
+  const uint8x16_t v5 = vec_splat(d, 5);
+  const uint8x16_t v6 = vec_splat(d, 6);
+  const uint8x16_t v7 = vec_splat(d, 7);
+
+  const uint8x16_t v8 = vec_splat(d, 8);
+  const uint8x16_t v9 = vec_splat(d, 9);
+  const uint8x16_t v10 = vec_splat(d, 10);
+  const uint8x16_t v11 = vec_splat(d, 11);
+
+  const uint8x16_t v12 = vec_splat(d, 12);
+  const uint8x16_t v13 = vec_splat(d, 13);
+  const uint8x16_t v14 = vec_splat(d, 14);
+  const uint8x16_t v15 = vec_splat(d, 15);
+
+  (void)above;
+
+  vec_vsx_st(v0, 0, dst);
+  dst += stride;
+  vec_vsx_st(v1, 0, dst);
+  dst += stride;
+  vec_vsx_st(v2, 0, dst);
+  dst += stride;
+  vec_vsx_st(v3, 0, dst);
+  dst += stride;
+  vec_vsx_st(v4, 0, dst);
+  dst += stride;
+  vec_vsx_st(v5, 0, dst);
+  dst += stride;
+  vec_vsx_st(v6, 0, dst);
+  dst += stride;
+  vec_vsx_st(v7, 0, dst);
+  dst += stride;
+  vec_vsx_st(v8, 0, dst);
+  dst += stride;
+  vec_vsx_st(v9, 0, dst);
+  dst += stride;
+  vec_vsx_st(v10, 0, dst);
+  dst += stride;
+  vec_vsx_st(v11, 0, dst);
+  dst += stride;
+  vec_vsx_st(v12, 0, dst);
+  dst += stride;
+  vec_vsx_st(v13, 0, dst);
+  dst += stride;
+  vec_vsx_st(v14, 0, dst);
+  dst += stride;
+  vec_vsx_st(v15, 0, dst);
+}
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -119,7 +119,7 @@
 specialize qw/vpx_d63_predictor_16x16 ssse3/;
 
 add_proto qw/void vpx_h_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_h_predictor_16x16 neon dspr2 msa sse2/;
+specialize qw/vpx_h_predictor_16x16 neon dspr2 msa sse2 vsx/;
 
 add_proto qw/void vpx_d117_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";