From: Luca Barbato <lu_zero@gentoo.org>
Date: Sat, 8 Apr 2017 02:55:33 +0000 (+0000)
Subject: ppc: dc top and left predictor 16x16
X-Git-Tag: v1.7.0~545
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=ad9dea1f6d1cea9c704c4f936742288ac5c0c2c6;p=libvpx

ppc: dc top and left predictor 16x16

13x faster.

Change-Id: I1771ac39fda599153f933cb3f0506c9f97a6cbe6
---

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 4ba018227..533ca7198 100644
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -316,10 +316,10 @@ INTRA_PRED_TEST(MSA, TestIntraPred32, vpx_dc_predictor_32x32_msa,
 INTRA_PRED_TEST(VSX, TestIntraPred8, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
                 NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_8x8_vsx)
 
-INTRA_PRED_TEST(VSX, TestIntraPred16, NULL, NULL, NULL,
-                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)
+INTRA_PRED_TEST(VSX, TestIntraPred16, NULL, 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)
 
 INTRA_PRED_TEST(VSX, TestIntraPred32, NULL, NULL, NULL,
                 vpx_dc_128_predictor_32x32_vsx, vpx_v_predictor_32x32_vsx,
diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index 9350170f3..71c35e3e3 100644
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -398,3 +398,28 @@ void vpx_dc_128_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
 
   dc_fill_predictor_32x32(dst, stride, v128);
 }
+
+static uint8x16_t avg16(const uint8_t *values) {
+  const int32x4_t sum4s =
+      (int32x4_t)vec_sum4s(vec_vsx_ld(0, values), vec_splat_u32(0));
+  const uint32x4_t sum = (uint32x4_t)vec_sums(sum4s, vec_splat_s32(8));
+  const uint32x4_t avg = (uint32x4_t)vec_sr(sum, vec_splat_u32(4));
+
+  return vec_splat(vec_pack(vec_pack(avg, vec_splat_u32(0)), vec_splat_u16(0)),
+                   3);
+}
+
+void vpx_dc_left_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+
+  dc_fill_predictor_16x16(dst, stride, avg16(left));
+}
+
+void vpx_dc_top_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  dc_fill_predictor_16x16(dst, stride, avg16(above));
+}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index f347290c3..c6e8ae577 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -139,10 +139,10 @@ add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa sse2/;
 
 add_proto qw/void vpx_dc_top_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_top_predictor_16x16 neon msa sse2/;
+specialize qw/vpx_dc_top_predictor_16x16 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_left_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_left_predictor_16x16 neon msa sse2/;
+specialize qw/vpx_dc_left_predictor_16x16 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_128_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_dc_128_predictor_16x16 neon msa sse2 vsx/;