vpxdsp: [x86] add highbd_h_predictor functions

C vs SSE2 speed gains:
_4x4 : ~8.12x
_8x8 : ~9.71x
_16x16 : ~8.21x
_32x32 : ~5.0x

BUG=webm:1422

Change-Id: I5e8a1ed4db7b8dc539b3e2a728b0b34d8b4b1993

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 23fce335af9b6a10017d06f7af8484023a29717c..b35727f2c007d987b6531d85237aa6a8c298a161 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -482,25 +482,27 @@ HIGHBD_INTRA_PRED_TEST(
 #if HAVE_SSE2
 HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred4,
                        vpx_highbd_dc_predictor_4x4_sse2, NULL, NULL, NULL,
-                       vpx_highbd_v_predictor_4x4_sse2, NULL, NULL, NULL, NULL,
-                       NULL, NULL, NULL, vpx_highbd_tm_predictor_4x4_c)
+                       vpx_highbd_v_predictor_4x4_sse2,
+                       vpx_highbd_h_predictor_4x4_sse2, NULL, NULL, NULL, NULL,
+                       NULL, NULL, vpx_highbd_tm_predictor_4x4_c)
 
 HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred8,
                        vpx_highbd_dc_predictor_8x8_sse2, NULL, NULL, NULL,
-                       vpx_highbd_v_predictor_8x8_sse2, NULL, NULL, NULL, NULL,
-                       NULL, NULL, NULL, vpx_highbd_tm_predictor_8x8_sse2)
+                       vpx_highbd_v_predictor_8x8_sse2,
+                       vpx_highbd_h_predictor_8x8_sse2, NULL, NULL, NULL, NULL,
+                       NULL, NULL, vpx_highbd_tm_predictor_8x8_sse2)
 
 HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred16,
                        vpx_highbd_dc_predictor_16x16_sse2, NULL, NULL, NULL,
-                       vpx_highbd_v_predictor_16x16_sse2, NULL, NULL, NULL,
-                       NULL, NULL, NULL, NULL,
-                       vpx_highbd_tm_predictor_16x16_sse2)
+                       vpx_highbd_v_predictor_16x16_sse2,
+                       vpx_highbd_h_predictor_16x16_sse2, NULL, NULL, NULL,
+                       NULL, NULL, NULL, vpx_highbd_tm_predictor_16x16_sse2)
 
 HIGHBD_INTRA_PRED_TEST(SSE2, TestHighbdIntraPred32,
                        vpx_highbd_dc_predictor_32x32_sse2, NULL, NULL, NULL,
-                       vpx_highbd_v_predictor_32x32_sse2, NULL, NULL, NULL,
-                       NULL, NULL, NULL, NULL,
-                       vpx_highbd_tm_predictor_32x32_sse2)
+                       vpx_highbd_v_predictor_32x32_sse2,
+                       vpx_highbd_h_predictor_32x32_sse2, NULL, NULL, NULL,
+                       NULL, NULL, NULL, vpx_highbd_tm_predictor_32x32_sse2)
 #endif  // HAVE_SSE2
 
 #if HAVE_NEON

diff --git a/test/vp9_intrapred_test.cc b/test/vp9_intrapred_test.cc
index bee0213eabf1d31f7c47f9200704947951d7961d..79ba56199cffac3f5717303e03ccb7cffab2c81d 100644 (file)
--- a/test/vp9_intrapred_test.cc
+++ b/test/vp9_intrapred_test.cc
@@ -487,6 +487,14 @@ INSTANTIATE_TEST_CASE_P(
                              &vpx_highbd_tm_predictor_16x16_c, 16, 8),
         HighbdIntraPredParam(&vpx_highbd_tm_predictor_32x32_sse2,
                              &vpx_highbd_tm_predictor_32x32_c, 32, 8),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_4x4_sse2,
+                             &vpx_highbd_h_predictor_4x4_c, 4, 8),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_8x8_sse2,
+                             &vpx_highbd_h_predictor_8x8_c, 8, 8),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_16x16_sse2,
+                             &vpx_highbd_h_predictor_16x16_c, 16, 8),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_32x32_sse2,
+                             &vpx_highbd_h_predictor_32x32_c, 32, 8),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_4x4_sse2,
                              &vpx_highbd_v_predictor_4x4_c, 4, 8),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_8x8_sse2,
@@ -515,6 +523,14 @@ INSTANTIATE_TEST_CASE_P(
                              &vpx_highbd_tm_predictor_16x16_c, 16, 10),
         HighbdIntraPredParam(&vpx_highbd_tm_predictor_32x32_sse2,
                              &vpx_highbd_tm_predictor_32x32_c, 32, 10),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_4x4_sse2,
+                             &vpx_highbd_h_predictor_4x4_c, 4, 10),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_8x8_sse2,
+                             &vpx_highbd_h_predictor_8x8_c, 8, 10),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_16x16_sse2,
+                             &vpx_highbd_h_predictor_16x16_c, 16, 10),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_32x32_sse2,
+                             &vpx_highbd_h_predictor_32x32_c, 32, 10),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_4x4_sse2,
                              &vpx_highbd_v_predictor_4x4_c, 4, 10),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_8x8_sse2,
@@ -543,6 +559,14 @@ INSTANTIATE_TEST_CASE_P(
                              &vpx_highbd_tm_predictor_16x16_c, 16, 12),
         HighbdIntraPredParam(&vpx_highbd_tm_predictor_32x32_sse2,
                              &vpx_highbd_tm_predictor_32x32_c, 32, 12),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_4x4_sse2,
+                             &vpx_highbd_h_predictor_4x4_c, 4, 12),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_8x8_sse2,
+                             &vpx_highbd_h_predictor_8x8_c, 8, 12),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_16x16_sse2,
+                             &vpx_highbd_h_predictor_16x16_c, 16, 12),
+        HighbdIntraPredParam(&vpx_highbd_h_predictor_32x32_sse2,
+                             &vpx_highbd_h_predictor_32x32_c, 32, 12),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_4x4_sse2,
                              &vpx_highbd_v_predictor_4x4_c, 4, 12),
         HighbdIntraPredParam(&vpx_highbd_v_predictor_8x8_sse2,

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index aaf9e25ff9dd64dc10232efea4ca99771abc0a00..4bb942db5f0581c0c535e5a69a753d656e77fe17 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -56,6 +56,7 @@ DSP_SRCS-$(HAVE_VSX) += ppc/intrapred_vsx.c
 ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
 DSP_SRCS-$(HAVE_SSE)  += x86/highbd_intrapred_sse2.asm
 DSP_SRCS-$(HAVE_SSE2) += x86/highbd_intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_intrapred_intrin_sse2.c
 DSP_SRCS-$(HAVE_NEON) += arm/highbd_intrapred_neon.c
 endif  # CONFIG_VP9_HIGHBITDEPTH
 

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 51ec2e50ef901e3063f5846b134a33beb6cb6463..ede556f22479bf24ff3cb020f9290c980d36db7e 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -195,7 +195,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_highbd_d63_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
   add_proto qw/void vpx_highbd_h_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
-  specialize qw/vpx_highbd_h_predictor_4x4 neon/;
+  specialize qw/vpx_highbd_h_predictor_4x4 neon sse2/;
 
   add_proto qw/void vpx_highbd_d117_predictor_4x4/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
@@ -230,7 +230,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_highbd_d63_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
   add_proto qw/void vpx_highbd_h_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
-  specialize qw/vpx_highbd_h_predictor_8x8 neon/;
+  specialize qw/vpx_highbd_h_predictor_8x8 neon sse2/;
 
   add_proto qw/void vpx_highbd_d117_predictor_8x8/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
@@ -265,7 +265,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_highbd_d63_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
   add_proto qw/void vpx_highbd_h_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
-  specialize qw/vpx_highbd_h_predictor_16x16 neon/;
+  specialize qw/vpx_highbd_h_predictor_16x16 neon sse2/;
 
   add_proto qw/void vpx_highbd_d117_predictor_16x16/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
@@ -300,7 +300,7 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vpx_highbd_d63_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 
   add_proto qw/void vpx_highbd_h_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
-  specialize qw/vpx_highbd_h_predictor_32x32 neon/;
+  specialize qw/vpx_highbd_h_predictor_32x32 neon sse2/;
 
   add_proto qw/void vpx_highbd_d117_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
 

diff --git a/vpx_dsp/x86/highbd_intrapred_intrin_sse2.c b/vpx_dsp/x86/highbd_intrapred_intrin_sse2.c
new file mode 100644 (file)
index 0000000..059f900
--- /dev/null
+++ b/vpx_dsp/x86/highbd_intrapred_intrin_sse2.c
@@ -0,0 +1,159 @@
+/*
+ *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+// -----------------------------------------------------------------------------
+
+void vpx_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const __m128i left_u16 = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  (void)above;
+  (void)bd;
+  _mm_storel_epi64((__m128i *)dst, row0);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row1);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row2);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row3);
+}
+
+void vpx_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+  (void)above;
+  (void)bd;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7));
+}
+
+static INLINE void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpacklo_epi64(*row, *row);
+  _mm_store_si128((__m128i *)*dst, val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  *dst += stride;
+}
+
+static INLINE void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpackhi_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  *dst += stride;
+}
+
+void vpx_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 2; i++, left += 8) {
+    const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+    const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+    const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+    const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+    const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+    const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+    const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+    const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+    const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+    h_store_16_unpacklo(&dst, stride, &row0);
+    h_store_16_unpacklo(&dst, stride, &row1);
+    h_store_16_unpacklo(&dst, stride, &row2);
+    h_store_16_unpacklo(&dst, stride, &row3);
+    h_store_16_unpackhi(&dst, stride, &row4);
+    h_store_16_unpackhi(&dst, stride, &row5);
+    h_store_16_unpackhi(&dst, stride, &row6);
+    h_store_16_unpackhi(&dst, stride, &row7);
+  }
+}
+
+static INLINE void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpacklo_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  _mm_store_si128((__m128i *)(*dst + 16), val);
+  _mm_store_si128((__m128i *)(*dst + 24), val);
+  *dst += stride;
+}
+
+static INLINE void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpackhi_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  _mm_store_si128((__m128i *)(*dst + 16), val);
+  _mm_store_si128((__m128i *)(*dst + 24), val);
+  *dst += stride;
+}
+
+void vpx_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 4; i++, left += 8) {
+    const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+    const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+    const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+    const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+    const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+    const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+    const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+    const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+    const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+    h_store_32_unpacklo(&dst, stride, &row0);
+    h_store_32_unpacklo(&dst, stride, &row1);
+    h_store_32_unpacklo(&dst, stride, &row2);
+    h_store_32_unpacklo(&dst, stride, &row3);
+    h_store_32_unpackhi(&dst, stride, &row4);
+    h_store_32_unpackhi(&dst, stride, &row5);
+    h_store_32_unpackhi(&dst, stride, &row6);
+    h_store_32_unpackhi(&dst, stride, &row7);
+  }
+}