ppc: tm predictor 32x32

About 8x faster.

Change-Id: I9bad827ccbdf47ec95406e961c74ac2ff45f80cf

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index d011beaf8186e4075f934117ebca468bc80a70e8..5411e2390005c58cf5dfae58a5f077bc9643c9a8 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -322,7 +322,7 @@ INTRA_PRED_TEST(VSX, TestIntraPred16, NULL, NULL, NULL, NULL,
 
 INTRA_PRED_TEST(VSX, TestIntraPred32, NULL, NULL, NULL, NULL,
                 vpx_v_predictor_32x32_vsx, vpx_h_predictor_32x32_vsx, NULL,
-                NULL, NULL, NULL, NULL, NULL, 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 cff6c7c2be4718116946d87740a31cf6fe854fa1..96491924e59f7f1297a7a2e5821827cf04ed1286 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -295,3 +295,69 @@ void vpx_tm_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
 
   tm_predictor_16x8(dst, stride, ll, ah, al, tl);
 }
+
+static INLINE void tm_predictor_32x1(uint8_t *dst, const int16x8_t ls,
+                                     const int16x8_t a0h, const int16x8_t a0l,
+                                     const int16x8_t a1h, const int16x8_t a1l,
+                                     const int16x8_t tl) {
+  int16x8_t vh, vl;
+
+  vh = vec_sub(vec_add(ls, a0h), tl);
+  vl = vec_sub(vec_add(ls, a0l), tl);
+  vec_vsx_st(vec_packsu(vh, vl), 0, dst);
+  vh = vec_sub(vec_add(ls, a1h), tl);
+  vl = vec_sub(vec_add(ls, a1l), tl);
+  vec_vsx_st(vec_packsu(vh, vl), 16, dst);
+}
+
+static void tm_predictor_32x8(uint8_t *dst, const ptrdiff_t stride,
+                              const int16x8_t l, const uint8x16_t a0,
+                              const uint8x16_t a1, const int16x8_t tl) {
+  const int16x8_t a0h = unpack_to_s16_h(a0);
+  const int16x8_t a0l = unpack_to_s16_l(a0);
+  const int16x8_t a1h = unpack_to_s16_h(a1);
+  const int16x8_t a1l = unpack_to_s16_l(a1);
+
+  tm_predictor_32x1(dst, vec_splat(l, 0), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 1), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 2), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 3), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 4), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 5), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 6), a0h, a0l, a1h, a1l, tl);
+  dst += stride;
+
+  tm_predictor_32x1(dst, vec_splat(l, 7), a0h, a0l, a1h, a1l, tl);
+}
+
+void vpx_tm_predictor_32x32_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 uint8x16_t l0 = vec_vsx_ld(0, left);
+  const uint8x16_t l1 = vec_vsx_ld(16, left);
+  const uint8x16_t a0 = vec_vsx_ld(0, above);
+  const uint8x16_t a1 = vec_vsx_ld(16, above);
+
+  tm_predictor_32x8(dst, stride, unpack_to_s16_h(l0), a0, a1, tl);
+  dst += stride * 8;
+
+  tm_predictor_32x8(dst, stride, unpack_to_s16_l(l0), a0, a1, tl);
+  dst += stride * 8;
+
+  tm_predictor_32x8(dst, stride, unpack_to_s16_h(l1), a0, a1, tl);
+  dst += stride * 8;
+
+  tm_predictor_32x8(dst, stride, unpack_to_s16_l(l1), a0, a1, tl);
+}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 5b009cffff7016c04c568063f9a0abf15d8b725a..680e278e557dbb92cfef9ba84ab293e6ced8e4e1 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -171,7 +171,7 @@ add_proto qw/void vpx_v_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, con
 specialize qw/vpx_v_predictor_32x32 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_tm_predictor_32x32 neon msa sse2/;
+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/;