ppc: d45 predictor 16x16

About 16x faster.

Change-Id: Ie5469fb32d5fd11bb6cb06318cea475d8a5b00b9

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index fd27d71a9b358c265a6ff3f41656f4ca43737034..a97f90a1f5b32c077f18b86a3414cdd3af5ccc64 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -319,8 +319,8 @@ INTRA_PRED_TEST(VSX, TestIntraPred8, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
 INTRA_PRED_TEST(VSX, TestIntraPred16, vpx_dc_predictor_16x16_vsx,
                 vpx_dc_left_predictor_16x16_vsx, vpx_dc_top_predictor_16x16_vsx,
                 vpx_dc_128_predictor_16x16_vsx, vpx_v_predictor_16x16_vsx,
-                vpx_h_predictor_16x16_vsx, NULL, NULL, NULL, NULL, NULL, NULL,
-                vpx_tm_predictor_16x16_vsx)
+                vpx_h_predictor_16x16_vsx, vpx_d45_predictor_16x16_vsx, NULL,
+                NULL, NULL, NULL, NULL, vpx_tm_predictor_16x16_vsx)
 
 INTRA_PRED_TEST(VSX, TestIntraPred32, vpx_dc_predictor_32x32_vsx,
                 vpx_dc_left_predictor_32x32_vsx, vpx_dc_top_predictor_32x32_vsx,

diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index 23a5657565c4a7d69ec7eb1b609a4ba1dc9129cc..6c7d7f94d7f27c9eb3311edb59f6fe437caae91f 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -489,3 +489,32 @@ void vpx_dc_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
                                 const uint8_t *above, const uint8_t *left) {
   dc_fill_predictor_32x32(dst, stride, dc_avg32(above, left));
 }
+
+static uint8x16_t avg3(const uint8x16_t a, const uint8x16_t b,
+                       const uint8x16_t c) {
+  const uint8x16_t ac =
+      vec_adds(vec_and(a, c), vec_sr(vec_xor(a, c), vec_splat_u8(1)));
+
+  return vec_avg(ac, b);
+}
+
+// Workaround vec_sld/vec_xxsldi/vec_lsdoi being missing or broken.
+static const uint8x16_t sl1 = { 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8,
+                                0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF, 0x10 };
+
+void vpx_d45_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t a = vec_vsx_ld(0, above);
+  const uint8x16_t above_right = vec_splat(a, 15);
+  const uint8x16_t b = vec_perm(a, above_right, sl1);
+  const uint8x16_t c = vec_perm(b, above_right, sl1);
+  uint8x16_t row = avg3(a, b, c);
+  int i;
+  (void)left;
+
+  for (i = 0; i < 16; i++) {
+    vec_vsx_st(row, 0, dst);
+    dst += stride;
+    row = vec_perm(row, above_right, sl1);
+  }
+}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 4b372c2a479727a5d5476a366377d02fade828c9..707c1c8c28d22560ad97dc774c35d0a32496cec7 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -113,7 +113,7 @@ add_proto qw/void vpx_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vpx_d207_predictor_16x16 ssse3/;
 
 add_proto qw/void vpx_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_d45_predictor_16x16 neon ssse3/;
+specialize qw/vpx_d45_predictor_16x16 neon ssse3 vsx/;
 
 add_proto qw/void vpx_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_d63_predictor_16x16 ssse3/;