ppc: tm predictor 4x4

Slightly faster.

Change-Id: I0ca43f309b3d9b50435d69bd5be64b53a99bd191

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 2881cf8..34b20cb 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -315,7 +315,7 @@ INTRA_PRED_TEST(MSA, TestIntraPred32, vpx_dc_predictor_32x32_msa,
 #if HAVE_VSX
 INTRA_PRED_TEST(VSX, TestIntraPred4, NULL, NULL, NULL, NULL, NULL,
                 vpx_h_predictor_4x4_vsx, NULL, NULL, NULL, NULL, NULL, NULL,
-                NULL)
+                vpx_tm_predictor_4x4_vsx)
 
 INTRA_PRED_TEST(VSX, TestIntraPred8, vpx_dc_predictor_8x8_vsx, NULL, NULL, NULL,
                 NULL, NULL, vpx_d45_predictor_8x8_vsx, NULL, NULL, NULL, NULL,

diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index 947c787..8216e3e 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -201,6 +201,38 @@ void vpx_h_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
   H_PREDICTOR_32(v15_1);
 }
 
+void vpx_tm_predictor_4x4_vsx(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int16x8_t tl = unpack_to_s16_h(vec_splat(vec_vsx_ld(-1, above), 0));
+  const int16x8_t l = unpack_to_s16_h(vec_vsx_ld(0, left));
+  const int16x8_t a = unpack_to_s16_h(vec_vsx_ld(0, above));
+  int16x8_t tmp, val;
+  uint8x16_t d;
+
+  d = vec_vsx_ld(0, dst);
+  tmp = unpack_to_s16_l(d);
+  val = vec_sub(vec_add(vec_splat(l, 0), a), tl);
+  vec_vsx_st(vec_sel(vec_packsu(val, tmp), d, (uint8x16_t)mask4), 0, dst);
+  dst += stride;
+
+  d = vec_vsx_ld(0, dst);
+  tmp = unpack_to_s16_l(d);
+  val = vec_sub(vec_add(vec_splat(l, 1), a), tl);
+  vec_vsx_st(vec_sel(vec_packsu(val, tmp), d, (uint8x16_t)mask4), 0, dst);
+  dst += stride;
+
+  d = vec_vsx_ld(0, dst);
+  tmp = unpack_to_s16_l(d);
+  val = vec_sub(vec_add(vec_splat(l, 2), a), tl);
+  vec_vsx_st(vec_sel(vec_packsu(val, tmp), d, (uint8x16_t)mask4), 0, dst);
+  dst += stride;
+
+  d = vec_vsx_ld(0, dst);
+  tmp = unpack_to_s16_l(d);
+  val = vec_sub(vec_add(vec_splat(l, 3), a), tl);
+  vec_vsx_st(vec_sel(vec_packsu(val, tmp), d, (uint8x16_t)mask4), 0, dst);
+}
+
 void vpx_tm_predictor_8x8_vsx(uint8_t *dst, ptrdiff_t stride,
                               const uint8_t *above, const uint8_t *left) {
   const int16x8_t tl = unpack_to_s16_h(vec_splat(vec_vsx_ld(-1, above), 0));

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index cbd1097..ba458c8 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -57,7 +57,7 @@ specialize qw/vpx_v_predictor_4x4 neon msa sse2/;
 add_proto qw/void vpx_ve_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 
 add_proto qw/void vpx_tm_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa sse2/;
+specialize qw/vpx_tm_predictor_4x4 neon dspr2 msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_predictor_4x4/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_dc_predictor_4x4 dspr2 msa neon sse2/;