libvpx: Pull from upstream

source/libvpx/vp9/common/x86/vp9_subpixel_8t_intrin_ssse3.c

 /*
  *  Copyright (c) 2010 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.
  */
 
@@ skipping 26 matching lines @@
 DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = {
   6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
 };
 
 void vp9_filter_block1d4_h8_intrin_ssse3(unsigned char *src_ptr,
                                          unsigned int src_pixels_per_line,
                                          unsigned char *output_ptr,
                                          unsigned int output_pitch,
                                          unsigned int output_height,
                                          int16_t *filter) {
-  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i firstFilters, secondFilters, shuffle1, shuffle2;
   __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
   __m128i addFilterReg64, filtersReg, srcReg, minReg;
   unsigned int i;
 
   // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
   addFilterReg64 =_mm_set1_epi32((int)0x0400040u);
   filtersReg = _mm_loadu_si128((__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);
 
   // duplicate only the first 16 bits in the filter into the first lane
   firstFilters = _mm_shufflelo_epi16(filtersReg, 0);
   // duplicate only the third 16 bit in the filter into the first lane
   secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu);
   // duplicate only the seconds 16 bits in the filter into the second lane
+  // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3
   firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u);
   // duplicate only the forth 16 bits in the filter into the second lane
+  // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7
   secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu);
 
   // loading the local filters
-  thirdFilters =_mm_load_si128((__m128i const *)filt1_4_h8);
-  forthFilters = _mm_load_si128((__m128i const *)filt2_4_h8);
+  shuffle1 =_mm_load_si128((__m128i const *)filt1_4_h8);
+  shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
 
   for (i = 0; i < output_height; i++) {
     srcReg = _mm_loadu_si128((__m128i *)(src_ptr-3));
 
     // filter the source buffer
-    srcRegFilt1= _mm_shuffle_epi8(srcReg, thirdFilters);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg, forthFilters);
+    srcRegFilt1= _mm_shuffle_epi8(srcReg, shuffle1);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg, shuffle2);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
 
     // extract the higher half of the lane
     srcRegFilt3 =  _mm_srli_si128(srcRegFilt1, 8);
     srcRegFilt4 =  _mm_srli_si128(srcRegFilt2, 8);
 
     minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2);
@@ skipping 69 matching lines @@
 
     // filter the source buffer
     srcRegFilt3= _mm_shuffle_epi8(srcReg, filt3Reg);
     srcRegFilt4= _mm_shuffle_epi8(srcReg, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters);
     srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters);
 
     // add and saturate all the results together
-    minReg = _mm_min_epi16(srcRegFilt4, srcRegFilt3);
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+
+    srcRegFilt2= _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
     srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
-
-    srcRegFilt4= _mm_max_epi16(srcRegFilt4, srcRegFilt3);
-    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
-    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
     srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
 
     // shift by 7 bit each 16 bits
     srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
 
     // shrink to 8 bit each 16 bits
     srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
 
     src_ptr+=src_pixels_per_line;
 
@@ skipping 39 matching lines @@
   filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
   filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
   filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
   filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
 
   for (i = 0; i < output_height; i++) {
     srcReg1 = _mm_loadu_si128((__m128i *)(src_ptr-3));
 
     // filter the source buffer
     srcRegFilt1_1= _mm_shuffle_epi8(srcReg1, filt1Reg);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt1_1 = _mm_maddubs_epi16(srcRegFilt1_1, firstFilters);
-    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
 
     // filter the source buffer
-    srcRegFilt3= _mm_shuffle_epi8(srcReg1, filt4Reg);
+    srcRegFilt3= _mm_shuffle_epi8(srcReg1, filt2Reg);
     srcRegFilt2= _mm_shuffle_epi8(srcReg1, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
 
     // add and saturate the results together
     srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1,
                                    _mm_min_epi16(srcRegFilt3, srcRegFilt2));
 
     // reading the next 16 bytes.
     // (part of it was being read by earlier read)
     srcReg2 = _mm_loadu_si128((__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, filt1Reg);
-    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt2Reg);
+    srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt4Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
     srcRegFilt2_1 = _mm_maddubs_epi16(srcRegFilt2_1, firstFilters);
-    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, forthFilters);
 
     // add and saturate the results together
     srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
 
     // filter the source buffer
-    srcRegFilt3= _mm_shuffle_epi8(srcReg2, filt4Reg);
+    srcRegFilt3= _mm_shuffle_epi8(srcReg2, filt2Reg);
     srcRegFilt2= _mm_shuffle_epi8(srcReg2, filt3Reg);
 
     // multiply 2 adjacent elements with the filter and add the result
-    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
     srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, 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, addFilterReg64);
     srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, addFilterReg64);
@@ skipping 193 matching lines @@
     srcRegFilt1 = _mm_packus_epi16(srcRegFilt5, srcRegFilt1);
 
     src_ptr+=src_pitch;
 
     // save 16 bytes convolve result
     _mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
 
     output_ptr+=out_pitch;
   }
 }