ppc: tm predictor 8x8

About 5x faster.

Change-Id: I951230517f49c0dca9ac9eac2efa8916a303b85a

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 7dec13ecc2bb4b22694f6d8af197bb676de41d80..88b21cfacbeb2546ca8bc7d6148713bd875b55b4 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -313,6 +313,9 @@ INTRA_PRED_TEST(MSA, TestIntraPred32, vpx_dc_predictor_32x32_msa,
 #endif  // HAVE_MSA
 
 #if HAVE_VSX
+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, NULL,
                 vpx_v_predictor_16x16_vsx, vpx_h_predictor_16x16_vsx, NULL,
                 NULL, NULL, NULL, NULL, NULL, NULL)

diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index cce1ff96cf76bb3a4e827e257278b5b2379d36ec..f0e8d276b3306e9c65e3becc15fc923139895447 100644 (file)
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -179,3 +179,50 @@ void vpx_h_predictor_32x32_vsx(uint8_t *dst, ptrdiff_t stride,
   H_PREDICTOR_32(v14_1);
   H_PREDICTOR_32(v15_1);
 }
+
+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));
+  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;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 0), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 1), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 2), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 3), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 4), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 5), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 6), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+  dst += stride;
+
+  tmp = unpack_to_s16_l(vec_vsx_ld(0, dst));
+  val = vec_sub(vec_add(vec_splat(l, 7), a), tl);
+  vec_vsx_st(vec_packsu(val, tmp), 0, dst);
+}

diff --git a/vpx_dsp/ppc/types_vsx.h b/vpx_dsp/ppc/types_vsx.h
index 66a9a03fe283914584d3d1293a870b5dd90826dd..ff977c267803b5ef27e1100e0e12cdb5d013e7b0 100644 (file)
--- a/vpx_dsp/ppc/types_vsx.h
+++ b/vpx_dsp/ppc/types_vsx.h
@@ -20,4 +20,24 @@ typedef vector unsigned short uint16x8_t;
 typedef vector signed int int32x4_t;
 typedef vector unsigned int uint32x4_t;
 
+#ifdef WORDS_BIGENDIAN
+#define unpack_to_u16_h(v) \
+  (uint16x8_t) vec_mergeh(vec_splat_u8(0), (uint8x16_t)v)
+#define unpack_to_u16_l(v) \
+  (uint16x8_t) vec_mergel(vec_splat_u8(0), (uint8x16_t)v)
+#define unpack_to_s16_h(v) \
+  (int16x8_t) vec_mergeh(vec_splat_u8(0), (uint8x16_t)v)
+#define unpack_to_s16_l(v) \
+  (int16x8_t) vec_mergel(vec_splat_u8(0), (uint8x16_t)v)
+#else
+#define unpack_to_u16_h(v) \
+  (uint16x8_t) vec_mergeh((uint8x16_t)v, vec_splat_u8(0))
+#define unpack_to_u16_l(v) \
+  (uint16x8_t) vec_mergel((uint8x16_t)v, vec_splat_u8(0))
+#define unpack_to_s16_h(v) \
+  (int16x8_t) vec_mergeh((uint8x16_t)v, vec_splat_u8(0))
+#define unpack_to_s16_l(v) \
+  (int16x8_t) vec_mergel((uint8x16_t)v, vec_splat_u8(0))
+#endif
+
 #endif  // VPX_DSP_PPC_TYPES_VSX_H_

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index a7bc5bf78d90654862123ae130f7fa57c314091d..62377623d41b99923ee2568d95a6db3aa41983b5 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -95,7 +95,7 @@ add_proto qw/void vpx_v_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const
 specialize qw/vpx_v_predictor_8x8 neon msa sse2/;
 
 add_proto qw/void vpx_tm_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa sse2/;
+specialize qw/vpx_tm_predictor_8x8 neon dspr2 msa sse2 vsx/;
 
 add_proto qw/void vpx_dc_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_dc_predictor_8x8 dspr2 neon msa sse2/;