ppc: dc predictor 32x32

10x and 5x faster.

Change-Id: I7913c58c768334d818f541a5e219f1035791eeaf

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 53c3cfa7fd843209ddfb27ba7e71a2a5856eef55..fd27d71a9b358c265a6ff3f41656f4ca43737034 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -316,15 +316,17 @@ 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, 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, 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, NULL,
-                NULL, NULL, NULL, NULL, NULL, vpx_tm_predictor_32x32_vsx)
+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)
+
+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, NULL, NULL, NULL, NULL, NULL, NULL,
+                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 5347872bae1402ade29608dfd993ceeb0d1dbc21..23a5657565c4a7d69ec7eb1b609a4ba1dc9129cc 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -451,3 +451,41 @@ void vpx_dc_top_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
 
   dc_fill_predictor_32x32(dst, stride, avg32(above));
 }
+
+static uint8x16_t dc_avg16(const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t a0 = vec_vsx_ld(0, above);
+  const uint8x16_t l0 = vec_vsx_ld(0, left);
+  const int32x4_t v16 = vec_sl(vec_splat_s32(1), vec_splat_u32(4));
+  const int32x4_t sum4s =
+      (int32x4_t)vec_sum4s(l0, vec_sum4s(a0, vec_splat_u32(0)));
+  const uint32x4_t sum = (uint32x4_t)vec_sums(sum4s, v16);
+  const uint32x4_t avg = (uint32x4_t)vec_sr(sum, vec_splat_u32(5));
+
+  return vec_splat(vec_pack(vec_pack(avg, vec_splat_u32(0)), vec_splat_u16(0)),
+                   3);
+}
+
+void vpx_dc_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  dc_fill_predictor_16x16(dst, stride, dc_avg16(above, left));
+}
+
+static uint8x16_t dc_avg32(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 l0 = vec_vsx_ld(0, left);
+  const uint8x16_t l1 = vec_vsx_ld(16, left);
+  const int32x4_t v32 = vec_sl(vec_splat_s32(1), vec_splat_u32(5));
+  const uint32x4_t a_sum = vec_sum4s(a0, vec_sum4s(a1, vec_splat_u32(0)));
+  const int32x4_t sum4s = (int32x4_t)vec_sum4s(l0, vec_sum4s(l1, a_sum));
+  const uint32x4_t sum = (uint32x4_t)vec_sums(sum4s, v32);
+  const uint32x4_t avg = (uint32x4_t)vec_sr(sum, vec_splat_u32(6));
+
+  return vec_splat(vec_pack(vec_pack(avg, vec_splat_u32(0)), vec_splat_u16(0)),
+                   3);
+}
+
+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));
+}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index d84e461d358bc1fe4bff20240a8c8a2ce6cbdbd4..4b372c2a479727a5d5476a366377d02fade828c9 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -136,7 +136,7 @@ add_proto qw/void vpx_tm_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_tm_predictor_16x16 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa sse2/;
+specialize qw/vpx_dc_predictor_16x16 dspr2 neon msa sse2 vsx/;
 
 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 vsx/;
@@ -174,7 +174,7 @@ add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_tm_predictor_32x32 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_dc_predictor_32x32 msa neon sse2/;
+specialize qw/vpx_dc_predictor_32x32 msa neon sse2 vsx/;
 
 add_proto qw/void vpx_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_dc_top_predictor_32x32 msa neon sse2 vsx/;