Port SIMD optimization for obmc blending functions to av1

SIMD optimization for 1d blending functions in obmc mode, and some
code refactoring and cleanup.

(ped_1080p25.y4m, 150 frame, 2000 tb)
Encoding time overhead: +18.8% -> +18.1%
Decoding time overhead: +21.3% -> +8.7%
Change-Id: I9d856c32136e7e0e6e24ab5520ef901d7b1ee9c8

aom_dsp/aom_dsp.mk

@@ -15,6 +15,8 @@ DSP_SRCS-yes += aom_dsp_common.h
 
 DSP_SRCS-$(HAVE_MSA)    += mips/macros_msa.h
 
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/synonyms.h
+
 # bit reader
 DSP_SRCS-yes += prob.h
 DSP_SRCS-yes += prob.c
@@ -87,6 +89,16 @@ DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred16_dspr2.c
 DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.h
 DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.c
 
+# inter predictions
+DSP_SRCS-yes            += blend.h
+DSP_SRCS-yes            += blend_a64_mask.c
+DSP_SRCS-yes            += blend_a64_hmask.c
+DSP_SRCS-yes            += blend_a64_vmask.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_sse4.h
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_mask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_hmask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_vmask_sse4.c
+
 # interpolation filters
 DSP_SRCS-yes += aom_convolve.c
 DSP_SRCS-yes += aom_convolve.h

aom_dsp/aom_dsp_common.h

@@ -42,6 +42,22 @@ typedef int32_t tran_high_t;
 typedef int16_t tran_low_t;
 #endif  // CONFIG_AOM_HIGHBITDEPTH
 
+#define IMPLIES(a, b) (!(a) || (b))  //  Logical 'a implies b' (or 'a -> b')
+
+#define IS_POWER_OF_TWO(x) (((x) & ((x)-1)) == 0)
+
+// These can be used to give a hint about branch outcomes.
+// This can have an effect, even if your target processor has a
+// good branch predictor, as these hints can affect basic block
+// ordering by the compiler.
+#ifdef __GNUC__
+#define LIKELY(v) __builtin_expect(v, 1)
+#define UNLIKELY(v) __builtin_expect(v, 0)
+#else
+#define LIKELY(v) (v)
+#define UNLIKELY(v) (v)
+#endif
+
 static INLINE uint8_t clip_pixel(int val) {
   return (val > 255) ? 255 : (val < 0) ? 0 : val;
 }

aom_dsp/aom_dsp_rtcd_defs.pl

@@ -961,6 +961,25 @@ if (aom_config("CONFIG_AOM_QM") eq "yes") {
   }  # CONFIG_AV1_ENCODER
 } # CONFIG_AOM_QM
 
+#
+# Alpha blending with mask
+#
+add_proto qw/void aom_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx";
+add_proto qw/void aom_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
+add_proto qw/void aom_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
+specialize "aom_blend_a64_mask", qw/sse4_1/;
+specialize "aom_blend_a64_hmask", qw/sse4_1/;
+specialize "aom_blend_a64_vmask", qw/sse4_1/;
+
+if (aom_config("CONFIG_AOM_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void aom_highbd_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd";  
+  add_proto qw/void aom_highbd_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
+  add_proto qw/void aom_highbd_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
+  specialize "aom_highbd_blend_a64_mask", qw/sse4_1/;
+  specialize "aom_highbd_blend_a64_hmask", qw/sse4_1/;
+  specialize "aom_highbd_blend_a64_vmask", qw/sse4_1/;
+}
+
 if (aom_config("CONFIG_ENCODERS") eq "yes") {
 #
 # Block subtraction

aom_dsp/blend.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.
+ */
+
+#ifndef AOM_DSP_BLEND_H_
+#define AOM_DSP_BLEND_H_
+
+#include "aom_ports/mem.h"
+
+// Various blending functions and macros.
+// See also the aom_blend_* functions in aom_dsp_rtcd.h
+
+// Alpha blending with alpha values from the range [0, 64], where 64
+// means use the first input and 0 means use the second input.
+
+#define AOM_BLEND_A64_ROUND_BITS 6
+#define AOM_BLEND_A64_MAX_ALPHA (1 << AOM_BLEND_A64_ROUND_BITS)  // 64
+
+#define AOM_BLEND_A64(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A64_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A64_ROUND_BITS)
+
+// Alpha blending with alpha values from the range [0, 256], where 256
+// means use the first input and 0 means use the second input.
+#define AOM_BLEND_A256_ROUND_BITS 8
+#define AOM_BLEND_A256_MAX_ALPHA (1 << AOM_BLEND_A256_ROUND_BITS)  // 256
+
+#define AOM_BLEND_A256(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A256_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A256_ROUND_BITS)
+
+// Blending by averaging.
+#define AOM_BLEND_AVG(v0, v1) ROUND_POWER_OF_TWO((v0) + (v1), 1)
+
+#endif  // AOM_DSP_BLEND_H_

aom_dsp/blend_a64_hmask.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 <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_AOM_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_AOM_HIGHBITDEPTH

aom_dsp/blend_a64_mask.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 <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "./aom_dsp_rtcd.h"
+
+// Blending with alpha mask. Mask values come from the range [0, 64],
+// as described for AOM_BLEND_A64 in aom_dsp/blend.h. src0 or src1 can
+// be the same as dst, or dst can be different from both sources.
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride,
+                          const uint8_t *src0, uint32_t src0_stride,
+                          const uint8_t *src1, uint32_t src1_stride,
+                          const uint8_t *mask, uint32_t mask_stride, int h,
+                          int w, int subh, int subw) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+
+#if CONFIG_AOM_HIGHBITDEPTH
+void aom_highbd_blend_a64_mask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                 const uint8_t *src0_8, uint32_t src0_stride,
+                                 const uint8_t *src1_8, uint32_t src1_stride,
+                                 const uint8_t *mask, uint32_t mask_stride,
+                                 int h, int w, int subh, int subw, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+#endif  // CONFIG_AOM_HIGHBITDEPTH

aom_dsp/blend_a64_vmask.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 <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_AOM_HIGHBITDEPTH
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_AOM_HIGHBITDEPTH

aom_dsp/x86/blend_a64_hmask_sse4.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/aom_integer.h"
+
+#include "./aom_dsp_rtcd.h"
+
+// To start out, just dispatch to the function using the 2D mask and
+// pass mask stride as 0. This can be improved upon if necessary.
+
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  aom_blend_a64_mask_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                            src1_stride, mask, 0, h, w, 0, 0);
+}
+
+#if CONFIG_AOM_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_sse4_1(
+    uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8,
+    uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, int bd) {
+  aom_highbd_blend_a64_mask_sse4_1(dst_8, dst_stride, src0_8, src0_stride,
+                                   src1_8, src1_stride, mask, 0, h, w, 0, 0,
+                                   bd);
+}
+#endif  // CONFIG_AOM_HIGHBITDEPTH

aom_dsp/x86/blend_a64_vmask_sse4.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 <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/blend_sse4.h"
+
+#include "./aom_dsp_rtcd.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_vmask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w16n_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                        const uint8_t *src0,
+                                        uint32_t src0_stride,
+                                        const uint8_t *src1,
+                                        uint32_t src1_stride,
+                                        const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0_w, v_m1_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0_w, v_m1_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  typedef void (*blend_fn)(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w);
+
+  // Dimension: width_index
+  static const blend_fn blend[9] = {
+    blend_a64_vmask_w16n_sse4_1,  // w % 16 == 0
+    aom_blend_a64_vmask_c,        // w == 1
+    aom_blend_a64_vmask_c,        // w == 2
+    NULL,                         // INVALID
+    blend_a64_vmask_w4_sse4_1,    // w == 4
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    blend_a64_vmask_w8_sse4_1,    // w == 8
+  };
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  blend[w & 0xf](dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, h,
+                 w);
+}
+
+#if CONFIG_AOM_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_vmask_bn_w4_sse4_1(