Added generic SIMD support for CLPF.

Change-Id: Ie03f9a5b0a4c708a586532198d755a1e7509f149

aom_dsp/aom_dsp_rtcd_defs.pl

@@ -624,6 +624,15 @@ specialize qw/aom_lpf_horizontal_4 mmx neon dspr2 msa/;
 add_proto qw/void aom_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 specialize qw/aom_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
 
+if (aom_config("CONFIG_CLPF") eq "yes") {
+  add_proto qw/void aom_clpf_block/, "const uint8_t *src, uint8_t *dst, int stride, int x0, int y0, int sizex, int sizey, int width, int height, unsigned int strength";
+  specialize qw/aom_clpf_block sse2 ssse3 sse4_1 neon/;
+  add_proto qw/void aom_clpf_detect/, "const uint8_t *rec, const uint8_t *org, int rstride, int ostride, int x0, int y0, int width, int height, int *sum0, int *sum1, unsigned int strength";
+  specialize qw/aom_clpf_detect sse2 ssse3 sse4_1 neon/;
+  add_proto qw/void aom_clpf_detect_multi/, "const uint8_t *rec, const uint8_t *org, int rstride, int ostride, int x0, int y0, int width, int height, int *sum";
+  specialize qw/aom_clpf_detect_multi sse2 ssse3 sse4_1 neon/;
+}
+
 if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void aom_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
   specialize qw/aom_highbd_lpf_vertical_16 sse2/;

av1/av1_common.mk

@@ -75,6 +75,11 @@ AV1_COMMON_SRCS-$(CONFIG_AV1_ENCODER) += common/av1_fwd_txfm.c
 ifeq ($(CONFIG_CLPF),yes)
 AV1_COMMON_SRCS-yes += common/clpf.c
 AV1_COMMON_SRCS-yes += common/clpf.h
+AV1_COMMON_SRCS-yes += common/clpf_simd.h
+AV1_COMMON_SRCS-$(HAVE_SSE2) += common/clpf_sse2.c
+AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/clpf_ssse3.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/clpf_sse4_1.c
+AV1_COMMON_SRCS-$(HAVE_NEON) += common/clpf_neon.c
 endif
 ifeq ($(CONFIG_DERING),yes)
 AV1_COMMON_SRCS-yes += common/od_dering.c

av1/av1_cx.mk

@@ -89,6 +89,11 @@ AV1_CX_SRCS-$(CONFIG_DERING) += encoder/pickdering.c
 ifeq ($(CONFIG_CLPF),yes)
 AV1_CX_SRCS-yes += encoder/clpf_rdo.c
 AV1_CX_SRCS-yes += encoder/clpf_rdo.h
+AV1_CX_SRCS-yes += encoder/clpf_rdo_simd.h
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/clpf_rdo_sse2.c
+AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/clpf_rdo_ssse3.c
+AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/clpf_rdo_sse4_1.c
+AV1_CX_SRCS-$(HAVE_NEON) += encoder/clpf_rdo_neon.c
 endif
 
 AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm

av1/common/clpf.c

@@ -9,6 +9,7 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 #include "av1/common/clpf.h"
+#include "./aom_dsp_rtcd.h"
 #include "aom_dsp/aom_dsp_common.h"
 
 int av1_clpf_maxbits(const AV1_COMMON *cm) {
@@ -27,9 +28,9 @@ int av1_clpf_sample(int X, int A, int B, int C, int D, int E, int F, int b) {
   return (8 + delta - (delta < 0)) >> 4;
 }
 
-static void clpf_block(const uint8_t *src, uint8_t *dst, int stride, int x0,
-                       int y0, int sizex, int sizey, int width, int height,
-                       unsigned int strength) {
+void aom_clpf_block_c(const uint8_t *src, uint8_t *dst, int stride, int x0,
+                      int y0, int sizex, int sizey, int width, int height,
+                      unsigned int strength) {
   int x, y;
   for (y = y0; y < y0 + sizey; y++) {
     for (x = x0; x < x0 + sizex; x++) {
@@ -102,8 +103,8 @@ int av1_clpf_frame(const YV12_BUFFER_CONFIG *dst, const YV12_BUFFER_CONFIG *rec,
             if (!cm->mi_grid_visible[ypos / bs * cm->mi_stride + xpos / bs]
                      ->mbmi.skip) {
               // Not skip block, apply the filter
-              clpf_block(rec->y_buffer, dst->y_buffer, stride_y, xpos, ypos, bs,
-                         bs, width, height, strength);
+              aom_clpf_block(rec->y_buffer, dst->y_buffer, stride_y, xpos, ypos,
+                             bs, bs, width, height, strength);
             } else {  // Skip block, copy instead
               for (c = 0; c < bs; c++)
                 *(uint64_t *)(dst->y_buffer + (ypos + c) * stride_y + xpos) =

av1/common/clpf_neon.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_neon
+#include "./clpf_simd.h"

av1/common/clpf_simd.h

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+
+static void clpf_block(const uint8_t *src, uint8_t *dst, int stride, int x0,
+                       int y0, int sizey, int width, int height,
+                       unsigned int strength) {
+  dst += x0 + y0 * stride;
+  src += x0 + y0 * stride;
+  {
+    int bottom = height - 2 - y0;
+    const v128 sp = v128_dup_8(strength);
+    const v128 sm = v128_dup_8(-(int)strength);
+    const v128 c8 = v128_dup_8(8);
+    const v128 c128 = v128_dup_8(128);
+
+    if (!x0) {  // Clip left
+      const v128 b_shuff = v128_from_v64(v64_from_64(0x0d0c0b0a09080808LL),
+                                         v64_from_64(0x0504030201000000LL));
+      const v128 c_shuff = v128_from_v64(v64_from_64(0x0e0d0c0b0a090808LL),
+                                         v64_from_64(0x0605040302010000LL));
+      int y;
+
+      for (y = 0; y < sizey; y += 2) {
+        const v64 l1 = v64_load_aligned(src);
+        const v64 l2 = v64_load_aligned(src + stride);
+        v128 o = v128_from_v64(l1, l2);
+        const v128 x = v128_add_8(c128, o);
+        const v128 a = v128_add_8(
+            c128,
+            v128_from_v64(v64_load_aligned(src - (y != -y0) * stride), l1));
+        const v128 b = v128_shuffle_8(x, b_shuff);
+        const v128 c = v128_shuffle_8(x, c_shuff);
+        const v128 d = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src + 1),
+                                v64_load_unaligned(src + 1 + stride)));
+        const v128 e = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src + 2),
+                                v64_load_unaligned(src + 2 + stride)));
+        const v128 f = v128_add_8(
+            c128, v128_from_v64(l2, v64_load_aligned(
+                                        src + ((y != bottom) + 1) * stride)));
+
+        const v128 tmp =
+            v128_add_8(v128_max_s8(v128_min_s8(v128_ssub_s8(c, x), sp), sm),
+                       v128_max_s8(v128_min_s8(v128_ssub_s8(d, x), sp), sm));
+        const v128 delta = v128_add_8(
+            v128_add_8(
+                v128_shl_8(
+                    v128_add_8(
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(a, x), sp), sm),
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(f, x), sp), sm)),
+                    2),
+                v128_add_8(
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(b, x), sp), sm),
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(e, x), sp), sm))),
+            v128_add_8(v128_add_8(tmp, tmp), tmp));
+        o = v128_add_8(
+            o, v128_shr_s8(
+                   v128_add_8(c8, v128_add_8(delta, v128_cmplt_s8(
+                                                        delta, v128_zero()))),
+                   4));
+        v64_store_aligned(dst, v128_high_v64(o));
+        v64_store_aligned(dst + stride, v128_low_v64(o));
+        src += stride * 2;
+        dst += stride * 2;
+      }
+    } else if (!(width - x0 - 8)) {  // Clip right
+      const v128 d_shuff = v128_from_v64(v64_from_64(0x0f0f0e0d0c0b0a09LL),
+                                         v64_from_64(0x0707060504030201LL));
+      const v128 e_shuff = v128_from_v64(v64_from_64(0x0f0f0f0e0d0c0b0aLL),
+                                         v64_from_64(0x0707070605040302LL));
+      int y;
+
+      for (y = 0; y < sizey; y += 2) {
+        const v64 l1 = v64_load_aligned(src);
+        const v64 l2 = v64_load_aligned(src + stride);
+        v128 o = v128_from_v64(l1, l2);
+        const v128 x = v128_add_8(c128, o);
+        const v128 a = v128_add_8(
+            c128,
+            v128_from_v64(v64_load_aligned(src - (y != -y0) * stride), l1));
+        const v128 b = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src - 2),
+                                v64_load_unaligned(src - 2 + stride)));
+        const v128 c = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src - 1),
+                                v64_load_unaligned(src - 1 + stride)));
+        const v128 d = v128_shuffle_8(x, d_shuff);
+        const v128 e = v128_shuffle_8(x, e_shuff);
+        const v128 f = v128_add_8(
+            c128, v128_from_v64(l2, v64_load_aligned(
+                                        src + ((y != bottom) + 1) * stride)));
+
+        const v128 tmp =
+            v128_add_8(v128_max_s8(v128_min_s8(v128_ssub_s8(c, x), sp), sm),
+                       v128_max_s8(v128_min_s8(v128_ssub_s8(d, x), sp), sm));
+        const v128 delta = v128_add_8(
+            v128_add_8(
+                v128_shl_8(
+                    v128_add_8(
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(a, x), sp), sm),
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(f, x), sp), sm)),
+                    2),
+                v128_add_8(
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(b, x), sp), sm),
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(e, x), sp), sm))),
+            v128_add_8(v128_add_8(tmp, tmp), tmp));
+        o = v128_add_8(
+            o, v128_shr_s8(
+                   v128_add_8(c8, v128_add_8(delta, v128_cmplt_s8(
+                                                        delta, v128_zero()))),
+                   4));
+        v64_store_aligned(dst, v128_high_v64(o));
+        v64_store_aligned(dst + stride, v128_low_v64(o));
+        src += stride * 2;
+        dst += stride * 2;
+      }
+    } else {  // No left/right clipping
+      int y;
+      for (y = 0; y < sizey; y += 2) {
+        const v64 l1 = v64_load_aligned(src);
+        const v64 l2 = v64_load_aligned(src + stride);
+        v128 o = v128_from_v64(l1, l2);
+        const v128 x = v128_add_8(c128, o);
+        const v128 a = v128_add_8(
+            c128,
+            v128_from_v64(v64_load_aligned(src - (y != -y0) * stride), l1));
+        const v128 b = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src - 2),
+                                v64_load_unaligned(src - 2 + stride)));
+        const v128 c = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src - 1),
+                                v64_load_unaligned(src - 1 + stride)));
+        const v128 d = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src + 1),
+                                v64_load_unaligned(src + 1 + stride)));
+        const v128 e = v128_add_8(
+            c128, v128_from_v64(v64_load_unaligned(src + 2),
+                                v64_load_unaligned(src + 2 + stride)));
+        const v128 f = v128_add_8(
+            c128, v128_from_v64(l2, v64_load_aligned(
+                                        src + ((y != bottom) + 1) * stride)));
+
+        const v128 tmp =
+            v128_add_8(v128_max_s8(v128_min_s8(v128_ssub_s8(c, x), sp), sm),
+                       v128_max_s8(v128_min_s8(v128_ssub_s8(d, x), sp), sm));
+        const v128 delta = v128_add_8(
+            v128_add_8(
+                v128_shl_8(
+                    v128_add_8(
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(a, x), sp), sm),
+                        v128_max_s8(v128_min_s8(v128_ssub_s8(f, x), sp), sm)),
+                    2),
+                v128_add_8(
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(b, x), sp), sm),
+                    v128_max_s8(v128_min_s8(v128_ssub_s8(e, x), sp), sm))),
+            v128_add_8(v128_add_8(tmp, tmp), tmp));
+        o = v128_add_8(
+            o, v128_shr_s8(
+                   v128_add_8(c8, v128_add_8(delta, v128_cmplt_s8(
+                                                        delta, v128_zero()))),
+                   4));
+        v64_store_aligned(dst, v128_high_v64(o));
+        v64_store_aligned(dst + stride, v128_low_v64(o));
+        src += stride * 2;
+        dst += stride * 2;
+      }
+    }
+  }
+}
+
+void SIMD_FUNC(aom_clpf_block)(const uint8_t *src, uint8_t *dst, int stride,
+                               int x0, int y0, int sizex, int sizey, int width,
+                               int height, unsigned int strength) {
+  // TODO(stemidts):
+  // A sizex different from 8 will only be needed if CLPF is extended to chroma.
+  // This will only be used if 4:2:0 and width not a multiple of 16 and along
+  // the right edge only, so we can fall back to the plain C implementation in
+  // this case.  If not extended to chroma, this test will be redundant.
+  if (sizex != 8 || width < 16) {  // Fallback to C if frame width < 16
+    aom_clpf_block_c(src, dst, stride, x0, y0, sizex, sizey, width, height,
+                     strength);
+  } else {
+    clpf_block(src, dst, stride, x0, y0, sizey, width, height, strength);
+  }
+}

av1/common/clpf_sse2.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse2
+#include "./clpf_simd.h"

av1/common/clpf_sse4_1.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse4_1
+#include "./clpf_simd.h"

av1/common/clpf_ssse3.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_ssse3
+#include "./clpf_simd.h"

av1/encoder/clpf_rdo.c

@@ -10,24 +10,25 @@
  */
 
 #include "av1/common/clpf.h"
+#include "./aom_dsp_rtcd.h"
 #include "aom/aom_integer.h"
 #include "av1/common/quant_common.h"
 
 // Calculate the error of a filtered and unfiltered block
-static void detect_clpf(const uint8_t *rec, const uint8_t *org, int x0, int y0,
-                        int width, int height, int so, int stride, int *sum0,
-                        int *sum1, unsigned int strength) {
+void aom_clpf_detect_c(const uint8_t *rec, const uint8_t *org, int rstride,
+                       int ostride, int x0, int y0, int width, int height,
+                       int *sum0, int *sum1, unsigned int strength) {
   int x, y;
   for (y = y0; y < y0 + 8; y++) {
     for (x = x0; x < x0 + 8; x++) {
-      int O = org[y * so + x];
-      int X = rec[y * stride + x];
-      int A = rec[AOMMAX(0, y - 1) * stride + x];
-      int B = rec[y * stride + AOMMAX(0, x - 2)];
-      int C = rec[y * stride + AOMMAX(0, x - 1)];
-      int D = rec[y * stride + AOMMIN(width - 1, x + 1)];
-      int E = rec[y * stride + AOMMIN(width - 1, x + 2)];
-      int F = rec[AOMMIN(height - 1, y + 1) * stride + x];
+      int O = org[y * ostride + x];
+      int X = rec[y * rstride + x];
+      int A = rec[AOMMAX(0, y - 1) * rstride + x];
+      int B = rec[y * rstride + AOMMAX(0, x - 2)];
+      int C = rec[y * rstride + AOMMAX(0, x - 1)];
+      int D = rec[y * rstride + AOMMIN(width - 1, x + 1)];
+      int E = rec[y * rstride + AOMMIN(width - 1, x + 2)];
+      int F = rec[AOMMIN(height - 1, y + 1) * rstride + x];
       int delta = av1_clpf_sample(X, A, B, C, D, E, F, strength);
       int Y = X + delta;
       *sum0 += (O - X) * (O - X);
@@ -36,21 +37,21 @@ static void detect_clpf(const uint8_t *rec, const uint8_t *org, int x0, int y0,
   }
 }
 
-static void detect_multi_clpf(const uint8_t *rec, const uint8_t *org, int x0,
-                              int y0, int width, int height, int so, int stride,
-                              int *sum) {
+void aom_clpf_detect_multi_c(const uint8_t *rec, const uint8_t *org,
+                             int rstride, int ostride, int x0, int y0,
+                             int width, int height, int *sum) {
   int x, y;
 
   for (y = y0; y < y0 + 8; y++) {
     for (x = x0; x < x0 + 8; x++) {
-      int O = org[y * so + x];
-      int X = rec[y * stride + x];
-      int A = rec[AOMMAX(0, y - 1) * stride + x];
-      int B = rec[y * stride + AOMMAX(0, x - 2)];
-      int C = rec[y * stride + AOMMAX(0, x - 1)];
-      int D = rec[y * stride + AOMMIN(width - 1, x + 1)];
-      int E = rec[y * stride + AOMMIN(width - 1, x + 2)];
-      int F = rec[AOMMIN(height - 1, y + 1) * stride + x];
+      int O = org[y * ostride + x];
+      int X = rec[y * rstride + x];
+      int A = rec[AOMMAX(0, y - 1) * rstride + x];
+      int B = rec[y * rstride + AOMMAX(0, x - 2)];
+      int C = rec[y * rstride + AOMMAX(0, x - 1)];
+      int D = rec[y * rstride + AOMMIN(width - 1, x + 1)];
+      int E = rec[y * rstride + AOMMIN(width - 1, x + 2)];
+      int F = rec[AOMMIN(height - 1, y + 1) * rstride + x];
       int delta1 = av1_clpf_sample(X, A, B, C, D, E, F, 1);
       int delta2 = av1_clpf_sample(X, A, B, C, D, E, F, 2);
       int delta3 = av1_clpf_sample(X, A, B, C, D, E, F, 4);
@@ -77,9 +78,9 @@ int av1_clpf_decision(int k, int l, const YV12_BUFFER_CONFIG *rec,
       const int bs = MAX_MIB_SIZE;
       if (!cm->mi_grid_visible[ypos / bs * cm->mi_stride + xpos / bs]
                ->mbmi.skip)
-        detect_clpf(rec->y_buffer, org->y_buffer, xpos, ypos, rec->y_crop_width,
-                    rec->y_crop_height, org->y_stride, rec->y_stride, &sum0,
-                    &sum1, strength);
+        aom_clpf_detect(rec->y_buffer, org->y_buffer, rec->y_stride,
+                        org->y_stride, xpos, ypos, rec->y_crop_width,
+                        rec->y_crop_height, &sum0, &sum1, strength);
     }
   }
   *res = sum1 < sum0;
@@ -144,9 +145,9 @@ static int clpf_rdo(int y, int x, const YV12_BUFFER_CONFIG *rec,
       if (!cm->mi_grid_visible[ypos / MAX_MIB_SIZE * cm->mi_stride +
                                xpos / MAX_MIB_SIZE]
                ->mbmi.skip) {
-        detect_multi_clpf(rec->y_buffer, org->y_buffer, xpos, ypos,
-                          rec->y_crop_width, rec->y_crop_height, org->y_stride,
-                          rec->y_stride, sum);
+        aom_clpf_detect_multi(rec->y_buffer, org->y_buffer, rec->y_stride,
+                              org->y_stride, xpos, ypos, rec->y_crop_width,
+                              rec->y_crop_height, sum);
         filtered = 1;
       }
     }

av1/encoder/clpf_rdo_neon.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_neon
+#include "./clpf_rdo_simd.h"

av1/encoder/clpf_rdo_sse2.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse2
+#include "./clpf_rdo_simd.h"

av1/encoder/clpf_rdo_sse4_1.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse4_1
+#include "./clpf_rdo_simd.h"

av1/encoder/clpf_rdo_ssse3.c

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_ssse3
+#include "./clpf_rdo_simd.h"

test/clpf_test.cc

+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/aom_timer.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+typedef void (*clpf_block_t)(const uint8_t *src, uint8_t *dst, int stride,
+                             int x0, int y0, int sizex, int sizey, int width,
+                             int height, unsigned int strength);
+
+typedef std::tr1::tuple<clpf_block_t, clpf_block_t, int, int>
+    clpf_block_param_t;
+
+class ClpfBlockTest : public ::testing::TestWithParam<clpf_block_param_t> {
+ public:
+  virtual ~ClpfBlockTest() {}
+  virtual void SetUp() {
+    clpf = GET_PARAM(0);
+    ref_clpf = GET_PARAM(1);
+    sizex = GET_PARAM(2);
+    sizey = GET_PARAM(3);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int sizex;
+  int sizey;
+  clpf_block_t clpf;
+  clpf_block_t ref_clpf;
+};
+
+typedef ClpfBlockTest ClpfSpeedTest;
+
+TEST_P(ClpfBlockTest, TestSIMDNoMismatch) {
+  int w = sizex;
+  int h = sizey;
+  const int size = 32;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, s[size * size]);
+  DECLARE_ALIGNED(16, uint8_t, d[size * size]);
+  DECLARE_ALIGNED(16, uint8_t, ref_d[size * size]);
+  memset(ref_d, 0, size * size);
+  memset(d, 0, size * size);
+
+  int error = 0;
+  int pos = 0;
+  int strength = 0;
+  int xpos = 0, ypos = 0;
+  int bits;
+  int level;
+
+  // Test every combination of:
+  // * Input with 1-8 bits of noise
+  // * Noise level around every value from 0 to 255
+  // * Blocks anywhere in the frame (along all egdes and also fully inside)
+  // * All strengths
+  for (level = 0; level < 256 && !error; level++) {
+    for (bits = 1; bits < 9 && !error; bits++) {
+      for (int i = 0; i < size * size; i++)
+        s[i] = clamp((rnd.Rand8() & ((1 << bits) - 1)) + level, 0, 255);
+
+      for (ypos = 0; ypos < size && !error; ypos += h * !error) {
+        for (xpos = 0; xpos < size && !error; xpos += w * !error) {
+          for (strength = 0; strength < 3 && !error; strength += !error) {
+            ref_clpf(s, ref_d, size, xpos, ypos, w, h, size, size,
+                     1 << strength);
+            ASM_REGISTER_STATE_CHECK(
+                clpf(s, d, size, xpos, ypos, w, h, size, size, 1 << strength));
+
+            for (pos = 0; pos < size * size && !error; pos++) {
+              error = ref_d[pos] != d[pos];
+            }
+          }
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, error)
+      << "Error: ClpfBlockTest, SIMD and C mismatch." << std::endl
+      << "First error at " << pos % size << "," << pos / size << " ("
+      << (int16_t)ref_d[pos] << " != " << (int16_t)d[pos] << ") " << std::endl
+      << "strength: " << (1 << strength) << std::endl
+      << "xpos: " << xpos << std::endl
+      << "ypos: " << ypos << std::endl
+      << "A=" << (pos > size ? (int16_t)s[pos - size] : -1) << std::endl
+      << "B=" << (pos % size - 2 >= 0 ? (int16_t)s[pos - 2] : -1) << std::endl
+      << "C=" << (pos % size - 1 >= 0 ? (int16_t)s[pos - 1] : -1) << std::endl
+      << "X=" << (int16_t)s[pos] << std::endl
+      << "D=" << (pos % size + 1 < size ? (int16_t)s[pos + 1] : -1) << std::endl
+      << "E=" << (pos % size + 2 < size ? (int16_t)s[pos + 2] : -1) << std::endl
+      << "F=" << (pos + size < size * size ? (int16_t)s[pos + size] : -1)
+      << std::endl;
+}
+
+TEST_P(ClpfSpeedTest, TestSpeed) {
+  int w = sizex;
+  int h = sizey;
+  const int size = 32;