Merge "Refactor x86/vpx_subpixel_8t_intrin_avx2.c"

3 files changed, 252 insertions, 405 deletions

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 2d071b96a..fa5feca16 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -89,6 +89,7 @@ DSP_SRCS-yes += vpx_filter.h
 DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/convolve.h
 DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/vpx_asm_stubs.c
 DSP_SRCS-$(HAVE_SSSE3) += x86/convolve_ssse3.h
+DSP_SRCS-$(HAVE_AVX2) += x86/convolve_avx2.h
 DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_subpixel_8t_sse2.asm
 DSP_SRCS-$(HAVE_SSE2)  += x86/vpx_subpixel_bilinear_sse2.asm
 DSP_SRCS-$(HAVE_SSSE3) += x86/vpx_subpixel_8t_ssse3.asm
diff --git a/vpx_dsp/x86/convolve_avx2.h b/vpx_dsp/x86/convolve_avx2.h
new file mode 100644
index 000000000..c2e83b53f
--- /dev/null
+++ b/vpx_dsp/x86/convolve_avx2.h
@@ -0,0 +1,99 @@
+/*
+ *  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.
+ */
+
+#ifndef VPX_DSP_X86_CONVOLVE_AVX2_H_
+#define VPX_DSP_X86_CONVOLVE_AVX2_H_
+
+#include <immintrin.h>  // AVX2
+
+#include "./vpx_config.h"
+
+#if defined(__clang__)
+#if (__clang_major__ > 0 && __clang_major__ < 3) ||            \
+    (__clang_major__ == 3 && __clang_minor__ <= 3) ||          \
+    (defined(__APPLE__) && defined(__apple_build_version__) && \
+     ((__clang_major__ == 4 && __clang_minor__ <= 2) ||        \
+      (__clang_major__ == 5 && __clang_minor__ == 0)))
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#else  // clang > 3.3, and not 5.0 on macosx.
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // clang <= 3.3
+#elif defined(__GNUC__)
+#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+#else  // gcc > 4.7
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // gcc <= 4.6
+#else   // !(gcc || clang)
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // __clang__
+
+static INLINE void shuffle_filter_avx2(const int16_t *const filter,
+                                       __m256i *const f) {
+  const __m256i f_values =
+      MM256_BROADCASTSI128_SI256(_mm_load_si128((const __m128i *)filter));
+  // pack and duplicate the filter values
+  f[0] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0200u));
+  f[1] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0604u));
+  f[2] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0a08u));
+  f[3] = _mm256_shuffle_epi8(f_values, _mm256_set1_epi16(0x0e0cu));
+}
+
+static INLINE __m256i convolve8_16_avx2(const __m256i *const s,
+                                        const __m256i *const f) {
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m256i k_64 = _mm256_set1_epi16(1 << 6);
+  const __m256i x0 = _mm256_maddubs_epi16(s[0], f[0]);
+  const __m256i x1 = _mm256_maddubs_epi16(s[1], f[1]);
+  const __m256i x2 = _mm256_maddubs_epi16(s[2], f[2]);
+  const __m256i x3 = _mm256_maddubs_epi16(s[3], f[3]);
+  // add and saturate the results together
+  const __m256i min_x2x1 = _mm256_min_epi16(x2, x1);
+  const __m256i max_x2x1 = _mm256_max_epi16(x2, x1);
+  __m256i temp = _mm256_adds_epi16(x0, x3);
+  temp = _mm256_adds_epi16(temp, min_x2x1);
+  temp = _mm256_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm256_adds_epi16(temp, k_64);
+  temp = _mm256_srai_epi16(temp, 7);
+  return temp;
+}
+
+static INLINE __m128i convolve8_8_avx2(const __m256i *const s,
+                                       const __m256i *const f) {
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i k_64 = _mm_set1_epi16(1 << 6);
+  const __m128i x0 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[0]),
+                                       _mm256_castsi256_si128(f[0]));
+  const __m128i x1 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[1]),
+                                       _mm256_castsi256_si128(f[1]));
+  const __m128i x2 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[2]),
+                                       _mm256_castsi256_si128(f[2]));
+  const __m128i x3 = _mm_maddubs_epi16(_mm256_castsi256_si128(s[3]),
+                                       _mm256_castsi256_si128(f[3]));
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_adds_epi16(temp, k_64);
+  temp = _mm_srai_epi16(temp, 7);
+  return temp;
+}
+
+#undef MM256_BROADCASTSI128_SI256
+
+#endif  // VPX_DSP_X86_CONVOLVE_AVX2_H_
diff --git a/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c b/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
index 4e851b58e..d0919695c 100644
--- a/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
+++ b/vpx_dsp/x86/vpx_subpixel_8t_intrin_avx2.c
@@ -12,9 +12,10 @@
 
 #include "./vpx_dsp_rtcd.h"
 #include "vpx_dsp/x86/convolve.h"
+#include "vpx_dsp/x86/convolve_avx2.h"
 #include "vpx_ports/mem.h"
 
-// filters for 16_h8 and 16_v8
+// filters for 16_h8
 DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
   0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
   0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
@@ -35,160 +36,68 @@ DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
   6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
 };
 
-#if defined(__clang__)
-#if (__clang_major__ > 0 && __clang_major__ < 3) ||            \
-    (__clang_major__ == 3 && __clang_minor__ <= 3) ||          \
-    (defined(__APPLE__) && defined(__apple_build_version__) && \
-     ((__clang_major__ == 4 && __clang_minor__ <= 2) ||        \
-      (__clang_major__ == 5 && __clang_minor__ == 0)))
-#define MM256_BROADCASTSI128_SI256(x) \
-  _mm_broadcastsi128_si256((__m128i const *)&(x))
-#else  // clang > 3.3, and not 5.0 on macosx.
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif  // clang <= 3.3
-#elif defined(__GNUC__)
-#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
-#define MM256_BROADCASTSI128_SI256(x) \
-  _mm_broadcastsi128_si256((__m128i const *)&(x))
-#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
-#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
-#else  // gcc > 4.7
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif  // gcc <= 4.6
-#else   // !(gcc || clang)
-#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
-#endif  // __clang__
-
-static INLINE void vpx_filter_block1d16_h8_X_avx2(
+static INLINE void vpx_filter_block1d16_h8_x_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
     ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter,
     const int avg) {
-  __m128i filtersReg, outReg1, outReg2;
-  __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
-  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
-  __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
-  __m256i srcReg32b1, srcReg32b2, filtersReg32;
+  __m128i outReg1, outReg2;
+  __m256i outReg32b1, outReg32b2;
   unsigned int i;
   ptrdiff_t src_stride, dst_stride;
+  __m256i f[4], filt[4], s[4];
 
-  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
-  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
-  filtersReg = _mm_loadu_si128((const __m128i *)filter);
-  // converting the 16 bit (short) to 8 bit (byte) and have the same data
-  // in both lanes of 128 bit register.
-  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
-  // have the same data in both lanes of a 256 bit register
-  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
-
-  // duplicate only the first 16 bits (first and second byte)
-  // across 256 bit register
-  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
-  // duplicate only the second 16 bits (third and forth byte)
-  // across 256 bit register
-  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
-  // duplicate only the third 16 bits (fifth and sixth byte)
-  // across 256 bit register
-  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
-  // duplicate only the forth 16 bits (seventh and eighth byte)
-  // across 256 bit register
-  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
-
-  filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
-  filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
-  filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
-  filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+  shuffle_filter_avx2(filter, f);
+  filt[0] = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+  filt[1] = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+  filt[2] = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+  filt[3] = _mm256_load_si256((__m256i const *)filt4_global_avx2);
 
   // multiple the size of the source and destination stride by two
   src_stride = src_pixels_per_line << 1;
   dst_stride = output_pitch << 1;
   for (i = output_height; i > 1; i -= 2) {
+    __m256i srcReg;
+
     // load the 2 strides of source
-    srcReg32b1 =
+    srcReg =
         _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3)));
-    srcReg32b1 = _mm256_inserti128_si256(
-        srcReg32b1,
+    srcReg = _mm256_inserti128_si256(
+        srcReg,
         _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)),
         1);
 
     // filter the source buffer
-    srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
-    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
-
-    // add and saturate the results together
-    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
-
-    // filter the source buffer
-    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
-    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
-
-    // add and saturate the results together
-    srcRegFilt32b1_1 = _mm256_adds_epi16(
-        srcRegFilt32b1_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+    s[0] = _mm256_shuffle_epi8(srcReg, filt[0]);
+    s[1] = _mm256_shuffle_epi8(srcReg, filt[1]);
+    s[2] = _mm256_shuffle_epi8(srcReg, filt[2]);
+    s[3] = _mm256_shuffle_epi8(srcReg, filt[3]);
+    outReg32b1 = convolve8_16_avx2(s, f);
 
     // reading 2 strides of the next 16 bytes
     // (part of it was being read by earlier read)
-    srcReg32b2 =
+    srcReg =
         _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5)));
-    srcReg32b2 = _mm256_inserti128_si256(
-        srcReg32b2,
+    srcReg = _mm256_inserti128_si256(
+        srcReg,
         _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)),
         1);
 
-    // add and saturate the results together
-    srcRegFilt32b1_1 = _mm256_adds_epi16(
-        srcRegFilt32b1_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
-
     // filter the source buffer
-    srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
-    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+    s[0] = _mm256_shuffle_epi8(srcReg, filt[0]);
+    s[1] = _mm256_shuffle_epi8(srcReg, filt[1]);
+    s[2] = _mm256_shuffle_epi8(srcReg, filt[2]);
+    s[3] = _mm256_shuffle_epi8(srcReg, filt[3]);
+    outReg32b2 = convolve8_16_avx2(s, f);
 
-    // add and saturate the results together
-    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
-
-    // filter the source buffer
-    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
-    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
-    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
-
-    // add and saturate the results together
-    srcRegFilt32b2_1 = _mm256_adds_epi16(
-        srcRegFilt32b2_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
-    srcRegFilt32b2_1 = _mm256_adds_epi16(
-        srcRegFilt32b2_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
-
-    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
-
-    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
-    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
-
-    // shrink to 8 bit each 16 bits, the first lane contain the first
-    // convolve result and the second lane contain the second convolve
-    // result
-    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+    // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+    // contain the first and second convolve result respectively
+    outReg32b1 = _mm256_packus_epi16(outReg32b1, outReg32b2);
 
     src_ptr += src_stride;
 
     // average if necessary
-    outReg1 = _mm256_castsi256_si128(srcRegFilt32b1_1);
-    outReg2 = _mm256_extractf128_si256(srcRegFilt32b1_1, 1);
+    outReg1 = _mm256_castsi256_si128(outReg32b1);
+    outReg2 = _mm256_extractf128_si256(outReg32b1, 1);
     if (avg) {
       outReg1 = _mm_avg_epu8(outReg1, _mm_load_si128((__m128i *)output_ptr));
       outReg2 = _mm_avg_epu8(
@@ -207,89 +116,40 @@ static INLINE void vpx_filter_block1d16_h8_X_avx2(
   // if the number of strides is odd.
   // process only 16 bytes
   if (i > 0) {
-    __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
-    __m128i srcRegFilt2, srcRegFilt3;
-
-    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
-
-    // filter the source buffer
-    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
-    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt1_1 =
-        _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters));
-    srcRegFilt2 =
-        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+    __m128i srcReg;
 
-    // add and saturate the results together
-    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+    // load the first 16 bytes of the last row
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
 
     // filter the source buffer
-    srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
-    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 =
-        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
-    srcRegFilt2 =
-        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
-
-    // add and saturate the results together
-    srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+    s[0] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[0])));
+    s[1] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[1])));
+    s[2] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[2])));
+    s[3] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[3])));
+    outReg1 = convolve8_8_avx2(s, f);
 
     // reading the next 16 bytes
     // (part of it was being read by earlier read)
-    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
-
-    // add and saturate the results together
-    srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
-
-    // filter the source buffer
-    srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg));
-    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt2_1 =
-        _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters));
-    srcRegFilt2 =
-        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
-
-    // add and saturate the results together
-    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
 
     // filter the source buffer
-    srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg));
-    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt3Reg));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 =
-        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
-    srcRegFilt2 =
-        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
-
-    // add and saturate the results together
-    srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
-    srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
-
-    srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg64));
-
-    srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg64));
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
-    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
-
-    // shrink to 8 bit each 16 bits, the first lane contain the first
-    // convolve result and the second lane contain the second convolve
-    // result
-    outReg1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+    s[0] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[0])));
+    s[1] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[1])));
+    s[2] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[2])));
+    s[3] = _mm256_castsi128_si256(
+        _mm_shuffle_epi8(srcReg, _mm256_castsi256_si128(filt[3])));
+    outReg2 = convolve8_8_avx2(s, f);
+
+    // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+    // contain the first and second convolve result respectively
+    outReg1 = _mm_packus_epi16(outReg1, outReg2);
 
     // average if necessary
     if (avg) {
@@ -304,169 +164,99 @@ static INLINE void vpx_filter_block1d16_h8_X_avx2(
 static void vpx_filter_block1d16_h8_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *output_ptr,
     ptrdiff_t dst_stride, uint32_t output_height, const int16_t *filter) {
-  vpx_filter_block1d16_h8_X_avx2(src_ptr, src_stride, output_ptr, dst_stride,
+  vpx_filter_block1d16_h8_x_avx2(src_ptr, src_stride, output_ptr, dst_stride,
                                  output_height, filter, 0);
 }
 
 static void vpx_filter_block1d16_h8_avg_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *output_ptr,
     ptrdiff_t dst_stride, uint32_t output_height, const int16_t *filter) {
-  vpx_filter_block1d16_h8_X_avx2(src_ptr, src_stride, output_ptr, dst_stride,
+  vpx_filter_block1d16_h8_x_avx2(src_ptr, src_stride, output_ptr, dst_stride,
                                  output_height, filter, 1);
 }
 
-static INLINE void vpx_filter_block1d16_v8_X_avx2(
+static INLINE void vpx_filter_block1d16_v8_x_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
     ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter,
     const int avg) {
-  __m128i filtersReg, outReg1, outReg2;
-  __m256i addFilterReg64;
-  __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
-  __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
-  __m256i srcReg32b11, srcReg32b12, filtersReg32;
-  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i outReg1, outReg2;
+  __m256i srcRegHead1;
   unsigned int i;
   ptrdiff_t src_stride, dst_stride;
+  __m256i f[4], s1[4], s2[4];
 
-  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
-  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
-  filtersReg = _mm_loadu_si128((const __m128i *)filter);
-  // converting the 16 bit (short) to  8 bit (byte) and have the
-  // same data in both lanes of 128 bit register.
-  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
-  // have the same data in both lanes of a 256 bit register
-  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
-
-  // duplicate only the first 16 bits (first and second byte)
-  // across 256 bit register
-  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
-  // duplicate only the second 16 bits (third and forth byte)
-  // across 256 bit register
-  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
-  // duplicate only the third 16 bits (fifth and sixth byte)
-  // across 256 bit register
-  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
-  // duplicate only the forth 16 bits (seventh and eighth byte)
-  // across 256 bit register
-  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+  shuffle_filter_avx2(filter, f);
 
   // multiple the size of the source and destination stride by two
   src_stride = src_pitch << 1;
   dst_stride = out_pitch << 1;
 
-  // load 16 bytes 7 times in stride of src_pitch
-  srcReg32b1 =
-      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr)));
-  srcReg32b2 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
-  srcReg32b3 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
-  srcReg32b4 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
-  srcReg32b5 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
-  srcReg32b6 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
-  srcReg32b7 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
-
-  // have each consecutive loads on the same 256 register
-  srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
-                                       _mm256_castsi256_si128(srcReg32b2), 1);
-  srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
-                                       _mm256_castsi256_si128(srcReg32b3), 1);
-  srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
-                                       _mm256_castsi256_si128(srcReg32b4), 1);
-  srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
-                                       _mm256_castsi256_si128(srcReg32b5), 1);
-  srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
-                                       _mm256_castsi256_si128(srcReg32b6), 1);
-  srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
-                                       _mm256_castsi256_si128(srcReg32b7), 1);
-
-  // merge every two consecutive registers except the last one
-  srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
-  srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
-
-  // save
-  srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
-
-  // save
-  srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
-
-  // save
-  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
-
-  // save
-  srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+  {
+    __m128i s[6];
+    __m256i s32b[6];
+
+    // load 16 bytes 7 times in stride of src_pitch
+    s[0] = _mm_loadu_si128((const __m128i *)(src_ptr + 0 * src_pitch));
+    s[1] = _mm_loadu_si128((const __m128i *)(src_ptr + 1 * src_pitch));
+    s[2] = _mm_loadu_si128((const __m128i *)(src_ptr + 2 * src_pitch));
+    s[3] = _mm_loadu_si128((const __m128i *)(src_ptr + 3 * src_pitch));
+    s[4] = _mm_loadu_si128((const __m128i *)(src_ptr + 4 * src_pitch));
+    s[5] = _mm_loadu_si128((const __m128i *)(src_ptr + 5 * src_pitch));
+    srcRegHead1 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + 6 * src_pitch)));
+
+    // have each consecutive loads on the same 256 register
+    s32b[0] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[0]), s[1], 1);
+    s32b[1] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[1]), s[2], 1);
+    s32b[2] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[2]), s[3], 1);
+    s32b[3] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[3]), s[4], 1);
+    s32b[4] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[4]), s[5], 1);
+    s32b[5] = _mm256_inserti128_si256(_mm256_castsi128_si256(s[5]),
+                                      _mm256_castsi256_si128(srcRegHead1), 1);
+
+    // merge every two consecutive registers except the last one
+    // the first lanes contain values for filtering odd rows (1,3,5...) and
+    // the second lanes contain values for filtering even rows (2,4,6...)
+    s1[0] = _mm256_unpacklo_epi8(s32b[0], s32b[1]);
+    s2[0] = _mm256_unpackhi_epi8(s32b[0], s32b[1]);
+    s1[1] = _mm256_unpacklo_epi8(s32b[2], s32b[3]);
+    s2[1] = _mm256_unpackhi_epi8(s32b[2], s32b[3]);
+    s1[2] = _mm256_unpacklo_epi8(s32b[4], s32b[5]);
+    s2[2] = _mm256_unpackhi_epi8(s32b[4], s32b[5]);
+  }
 
   for (i = output_height; i > 1; i -= 2) {
-    // load the last 2 loads of 16 bytes and have every two
+    __m256i srcRegHead2, srcRegHead3;
+
+    // load the next 2 loads of 16 bytes and have every two
     // consecutive loads in the same 256 bit register
-    srcReg32b8 = _mm256_castsi128_si256(
-        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
-    srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
-                                         _mm256_castsi256_si128(srcReg32b8), 1);
-    srcReg32b9 = _mm256_castsi128_si256(
-        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8)));
-    srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
-                                         _mm256_castsi256_si128(srcReg32b9), 1);
-
-    // merge every two consecutive registers
-    // save
-    srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
-    srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
-    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
-
-    // add and saturate the results together
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
-    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
-
-    // add and saturate the results together
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
-                                    _mm256_min_epi16(srcReg32b8, srcReg32b12));
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
-                                    _mm256_max_epi16(srcReg32b8, srcReg32b12));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
-    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
-
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
-    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
-
-    // add and saturate the results together
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
-                                   _mm256_min_epi16(srcReg32b8, srcReg32b12));
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
-                                   _mm256_max_epi16(srcReg32b8, srcReg32b12));
-
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
-
-    // shift by 7 bit each 16 bit
-    srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
-    srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
-
-    // shrink to 8 bit each 16 bits, the first lane contain the first
-    // convolve result and the second lane contain the second convolve
-    // result
-    srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+    srcRegHead2 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + 7 * src_pitch)));
+    srcRegHead1 = _mm256_inserti128_si256(
+        srcRegHead1, _mm256_castsi256_si128(srcRegHead2), 1);
+    srcRegHead3 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + 8 * src_pitch)));
+    srcRegHead2 = _mm256_inserti128_si256(
+        srcRegHead2, _mm256_castsi256_si128(srcRegHead3), 1);
+
+    // merge the two new consecutive registers
+    // the first lane contain values for filtering odd rows (1,3,5...) and
+    // the second lane contain values for filtering even rows (2,4,6...)
+    s1[3] = _mm256_unpacklo_epi8(srcRegHead1, srcRegHead2);
+    s2[3] = _mm256_unpackhi_epi8(srcRegHead1, srcRegHead2);
+
+    s1[0] = convolve8_16_avx2(s1, f);
+    s2[0] = convolve8_16_avx2(s2, f);
+
+    // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+    // contain the first and second convolve result respectively
+    s1[0] = _mm256_packus_epi16(s1[0], s2[0]);
 
     src_ptr += src_stride;
 
     // average if necessary
-    outReg1 = _mm256_castsi256_si128(srcReg32b1);
-    outReg2 = _mm256_extractf128_si256(srcReg32b1, 1);
+    outReg1 = _mm256_castsi256_si128(s1[0]);
+    outReg2 = _mm256_extractf128_si256(s1[0], 1);
     if (avg) {
       outReg1 = _mm_avg_epu8(outReg1, _mm_load_si128((__m128i *)output_ptr));
       outReg2 = _mm_avg_epu8(
@@ -481,78 +271,35 @@ static INLINE void vpx_filter_block1d16_v8_X_avx2(
 
     output_ptr += dst_stride;
 
-    // save part of the registers for next strides
-    srcReg32b10 = srcReg32b11;
-    srcReg32b1 = srcReg32b3;
-    srcReg32b11 = srcReg32b2;
-    srcReg32b3 = srcReg32b5;
-    srcReg32b2 = srcReg32b4;
-    srcReg32b5 = srcReg32b7;
-    srcReg32b7 = srcReg32b9;
+    // shift down by two rows
+    s1[0] = s1[1];
+    s2[0] = s2[1];
+    s1[1] = s1[2];
+    s2[1] = s2[2];
+    s1[2] = s1[3];
+    s2[2] = s2[3];
+    srcRegHead1 = srcRegHead3;
   }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
   if (i > 0) {
-    __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
-    __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
     // load the last 16 bytes
-    srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+    const __m128i srcRegHead2 =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
 
     // merge the last 2 results together
-    srcRegFilt4 =
-        _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
-    srcRegFilt7 =
-        _mm_unpackhi_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
-                                    _mm256_castsi256_si128(firstFilters));
-    srcRegFilt4 =
-        _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters));
-    srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
-                                    _mm256_castsi256_si128(firstFilters));
-    srcRegFilt7 =
-        _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters));
-
-    // add and saturate the results together
-    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
-    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
-                                    _mm256_castsi256_si128(secondFilters));
-    srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
-                                    _mm256_castsi256_si128(secondFilters));
-
-    // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
-                                    _mm256_castsi256_si128(thirdFilters));
-    srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
-                                    _mm256_castsi256_si128(thirdFilters));
-
-    // add and saturate the results together
-    srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm_min_epi16(srcRegFilt4, srcRegFilt6));
-    srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm_min_epi16(srcRegFilt5, srcRegFilt7));
-
-    // add and saturate the results together
-    srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm_max_epi16(srcRegFilt4, srcRegFilt6));
-    srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm_max_epi16(srcRegFilt5, srcRegFilt7));
-
-    srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg64));
-    srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg64));
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
-    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
-
-    // shrink to 8 bit each 16 bits, the first lane contain the first
-    // convolve result and the second lane contain the second convolve
-    // result
-    outReg1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+    s1[0] = _mm256_castsi128_si256(
+        _mm_unpacklo_epi8(_mm256_castsi256_si128(srcRegHead1), srcRegHead2));
+    s2[0] = _mm256_castsi128_si256(
+        _mm_unpackhi_epi8(_mm256_castsi256_si128(srcRegHead1), srcRegHead2));
+
+    outReg1 = convolve8_8_avx2(s1, f);
+    outReg2 = convolve8_8_avx2(s2, f);
+
+    // shrink to 8 bit each 16 bits, the low and high 64-bits of each lane
+    // contain the first and second convolve result respectively
+    outReg1 = _mm_packus_epi16(outReg1, outReg2);
 
     // average if necessary
     if (avg) {
@@ -568,14 +315,14 @@ static void vpx_filter_block1d16_v8_avx2(const uint8_t *src_ptr,
                                          ptrdiff_t src_stride, uint8_t *dst_ptr,
                                          ptrdiff_t dst_stride, uint32_t height,
                                          const int16_t *filter) {
-  vpx_filter_block1d16_v8_X_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
+  vpx_filter_block1d16_v8_x_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
                                  height, filter, 0);
 }
 
 static void vpx_filter_block1d16_v8_avg_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_stride, uint8_t *dst_ptr,
     ptrdiff_t dst_stride, uint32_t height, const int16_t *filter) {
-  vpx_filter_block1d16_v8_X_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
+  vpx_filter_block1d16_v8_x_avx2(src_ptr, src_stride, dst_ptr, dst_stride,
                                  height, filter, 1);
 }