From: Luca Barbato <lu_zero@gentoo.org>
Date: Sun, 9 Apr 2017 11:07:22 +0000 (+0000)
Subject: ppc: d45 predictor 8x8
X-Git-Tag: v1.7.0~538
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=6b4a65e8b149e1eaa246ebf80b0887f388c9d378;p=libvpx

ppc: d45 predictor 8x8

11x faster.

Change-Id: I5b8f39213ee1f5260724fc254e3fb5c462435798
---

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 08dd3148c..c9031a197 100644
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -313,8 +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, TestIntraPred8, NULL, NULL, NULL, NULL, NULL, NULL,
+                vpx_d45_predictor_8x8_vsx, NULL, NULL, NULL, NULL, NULL,
+                vpx_tm_predictor_8x8_vsx)
 
 INTRA_PRED_TEST(VSX, TestIntraPred16, vpx_dc_predictor_16x16_vsx,
                 vpx_dc_left_predictor_16x16_vsx, vpx_dc_top_predictor_16x16_vsx,
diff --git a/vpx_dsp/ppc/intrapred_vsx.c b/vpx_dsp/ppc/intrapred_vsx.c
index 2006539f8..d4f555433 100644
--- a/vpx_dsp/ppc/intrapred_vsx.c
+++ b/vpx_dsp/ppc/intrapred_vsx.c
@@ -502,6 +502,25 @@ static uint8x16_t avg3(const uint8x16_t a, const uint8x16_t b,
 static const uint8x16_t sl1 = { 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8,
                                 0x9, 0xA, 0xB, 0xC, 0xD, 0xE, 0xF, 0x10 };
 
+void vpx_d45_predictor_8x8_vsx(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t af = vec_vsx_ld(0, above);
+  const uint8x16_t above_right = vec_splat(af, 7);
+  const uint8x16_t a = xxpermdi(af, above_right, 1);
+  const uint8x16_t b = vec_perm(a, above_right, sl1);
+  const uint8x16_t c = vec_perm(b, above_right, sl1);
+  uint8x16_t row = avg3(a, b, c);
+  int i;
+  (void)left;
+
+  for (i = 0; i < 8; i++) {
+    const uint8x16_t d = vec_vsx_ld(0, dst);
+    vec_vsx_st(xxpermdi(row, d, 1), 0, dst);
+    dst += stride;
+    row = vec_perm(row, above_right, sl1);
+  }
+}
+
 void vpx_d45_predictor_16x16_vsx(uint8_t *dst, ptrdiff_t stride,
                                  const uint8_t *above, const uint8_t *left) {
   const uint8x16_t a = vec_vsx_ld(0, above);
diff --git a/vpx_dsp/ppc/types_vsx.h b/vpx_dsp/ppc/types_vsx.h
index ff977c267..f611d02d2 100644
--- a/vpx_dsp/ppc/types_vsx.h
+++ b/vpx_dsp/ppc/types_vsx.h
@@ -20,6 +20,25 @@ typedef vector unsigned short uint16x8_t;
 typedef vector signed int int32x4_t;
 typedef vector unsigned int uint32x4_t;
 
+#ifdef __clang__
+static const uint8x16_t xxpermdi0_perm = { 0x00, 0x01, 0x02, 0x03, 0x04, 0x05,
+                                           0x06, 0x07, 0x10, 0x11, 0x12, 0x13,
+                                           0x14, 0x15, 0x16, 0x17 };
+static const uint8x16_t xxpermdi1_perm = { 0x00, 0x01, 0x02, 0x03, 0x04, 0x05,
+                                           0x06, 0x07, 0x18, 0x19, 0x1A, 0x1B,
+                                           0x1C, 0x1D, 0x1E, 0x1F };
+static const uint8x16_t xxpermdi2_perm = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,
+                                           0x0E, 0x0F, 0x10, 0x11, 0x12, 0x13,
+                                           0x14, 0x15, 0x16, 0x17 };
+static const uint8x16_t xxpermdi3_perm = { 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D,
+                                           0x0E, 0x0F, 0x18, 0x19, 0x1A, 0x1B,
+                                           0x1C, 0x1D, 0x1E, 0x1F };
+#define xxpermdi(a, b, c) vec_perm(a, b, xxpermdi##c##_perm)
+#elif defined(__GNUC__) && \
+    (__GNUC__ > 6 || (__GNUC__ == 6 && __GNUC_MINOR__ >= 3))
+#define xxpermdi(a, b, c) vec_xxpermdi(a, b, c)
+#endif
+
 #ifdef WORDS_BIGENDIAN
 #define unpack_to_u16_h(v) \
   (uint16x8_t) vec_mergeh(vec_splat_u8(0), (uint8x16_t)v)
@@ -29,6 +48,9 @@ typedef vector unsigned int uint32x4_t;
   (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)
+#ifndef xxpermdi
+#define xxpermdi(a, b, c) vec_xxpermdi(a, b, c)
+#endif
 #else
 #define unpack_to_u16_h(v) \
   (uint16x8_t) vec_mergeh((uint8x16_t)v, vec_splat_u8(0))
@@ -38,6 +60,9 @@ typedef vector unsigned int uint32x4_t;
   (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))
+#ifndef xxpermdi
+#define xxpermdi(a, b, c) vec_xxpermdi(b, a, ((c >> 1) | (c & 1) << 1) ^ 3)
+#endif
 #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 affa9b143..833787dfd 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -75,7 +75,7 @@ add_proto qw/void vpx_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vpx_d207_predictor_8x8 ssse3/;
 
 add_proto qw/void vpx_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vpx_d45_predictor_8x8 neon sse2/;
+specialize qw/vpx_d45_predictor_8x8 neon sse2 vsx/;
 
 add_proto qw/void vpx_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vpx_d63_predictor_8x8 ssse3/;