ppc: d63 predictor 32x32

About 10x faster.

Change-Id: If7d0645f75c5d7deb9751edd0bf47e2f9068e9e7

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index e578d43..08dd314 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -327,7 +327,8 @@ INTRA_PRED_TEST(VSX, TestIntraPred32, vpx_dc_predictor_32x32_vsx,
                 vpx_dc_left_predictor_32x32_vsx, vpx_dc_top_predictor_32x32_vsx,
                 vpx_dc_128_predictor_32x32_vsx, vpx_v_predictor_32x32_vsx,
                 vpx_h_predictor_32x32_vsx, vpx_d45_predictor_32x32_vsx, NULL,
-                NULL, NULL, NULL, NULL, vpx_tm_predictor_32x32_vsx)
+                NULL, NULL, NULL, vpx_d63_predictor_32x32_vsx,
+                vpx_tm_predictor_32x32_vsx)
 #endif  // HAVE_VSX
 
 // -----------------------------------------------------------------------------

diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index 5a00c3d..2006539 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -562,3 +562,33 @@ void vpx_d63_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
     row1 = vec_perm(row1, above_right, sl1);
   }
 }
+
+void vpx_d63_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t a0 = vec_vsx_ld(0, above);
+  const uint8x16_t a1 = vec_vsx_ld(16, above);
+  const uint8x16_t a2 = vec_vsx_ld(32, above);
+  const uint8x16_t above_right = vec_splat(a2, 0);
+  const uint8x16_t b0 = vec_perm(a0, a1, sl1);
+  const uint8x16_t b1 = vec_perm(a1, above_right, sl1);
+  const uint8x16_t c0 = vec_perm(b0, b1, sl1);
+  const uint8x16_t c1 = vec_perm(b1, above_right, sl1);
+  uint8x16_t row0_0 = vec_avg(a0, b0);
+  uint8x16_t row0_1 = vec_avg(a1, b1);
+  uint8x16_t row1_0 = avg3(a0, b0, c0);
+  uint8x16_t row1_1 = avg3(a1, b1, c1);
+  int i;
+  (void)left;
+
+  for (i = 0; i < 16; i++) {
+    vec_vsx_st(row0_0, 0, dst);
+    vec_vsx_st(row0_1, 16, dst);
+    vec_vsx_st(row1_0, 0, dst + stride);
+    vec_vsx_st(row1_1, 16, dst + stride);
+    dst += stride * 2;
+    row0_0 = vec_perm(row0_0, row0_1, sl1);
+    row0_1 = vec_perm(row0_1, above_right, sl1);
+    row1_0 = vec_perm(row1_0, row1_1, sl1);
+    row1_1 = vec_perm(row1_1, above_right, sl1);
+  }
+}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 8c65c17..affa9b1 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -154,7 +154,7 @@ add_proto qw/void vpx_d45_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, c
 specialize qw/vpx_d45_predictor_32x32 neon ssse3 vsx/;
 
 add_proto qw/void vpx_d63_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_d63_predictor_32x32 ssse3/;
+specialize qw/vpx_d63_predictor_32x32 ssse3 vsx/;
 
 add_proto qw/void vpx_h_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_h_predictor_32x32 neon msa sse2 vsx/;