YUV scale filter columns improved filtering accuracy

source/scale_gcc.cc

 /*
  *  Copyright 2013 The LibYuv 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 803 matching lines @@
     "vzeroupper                                \n"
   : "+r"(src_ptr),     // %0
     "+r"(dst_ptr),     // %1
     "+r"(src_width)    // %2
   :
   : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
   );
 }
 #endif  // HAS_SCALEADDROW_AVX2
 
+// Constant for making pixels signed to avoid pmaddubsw
+// saturation.
+static uvec8 kFsub80 =
+  { 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80 };
+
+// Constant for making pixels unsigned and adding .5 for rounding.
+static uvec16 kFadd40 =
+  { 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040, 0x4040 };
+
 // Bilinear column filtering. SSSE3 version.
 void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
                            int dst_width, int x, int dx) {
   intptr_t x0, x1, temp_pixel;
   asm volatile (
     "movd      %6,%%xmm2                       \n"
     "movd      %7,%%xmm3                       \n"
     "movl      $0x04040000,%k2                 \n"
     "movd      %k2,%%xmm5                      \n"
     "pcmpeqb   %%xmm6,%%xmm6                   \n"
-    "psrlw     $0x9,%%xmm6                     \n"
+    "psrlw     $0x9,%%xmm6                     \n"  // 0x007f007f
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $15,%%xmm7                      \n"  // 0x00010001
+
     "pextrw    $0x1,%%xmm2,%k3                 \n"
     "subl      $0x2,%5                         \n"
     "jl        29f                             \n"
     "movdqa    %%xmm2,%%xmm0                   \n"
     "paddd     %%xmm3,%%xmm0                   \n"
     "punpckldq %%xmm0,%%xmm2                   \n"
     "punpckldq %%xmm3,%%xmm3                   \n"
     "paddd     %%xmm3,%%xmm3                   \n"
     "pextrw    $0x3,%%xmm2,%k4                 \n"
 
     LABELALIGN
   "2:                                          \n"
     "movdqa    %%xmm2,%%xmm1                   \n"
     "paddd     %%xmm3,%%xmm2                   \n"
     MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
     "movd      %k2,%%xmm0                      \n"
     "psrlw     $0x9,%%xmm1                     \n"
     MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
     "movd      %k2,%%xmm4                      \n"
     "pshufb    %%xmm5,%%xmm1                   \n"
     "punpcklwd %%xmm4,%%xmm0                   \n"
-    "pxor      %%xmm6,%%xmm1                   \n"
-    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
+    "pxor      %%xmm6,%%xmm1                   \n"  // 128 -f = (f ^ 127 ) + 1
+    "paddusb   %%xmm7,%%xmm1                   \n"
+    "pmaddubsw %%xmm0,%%xmm1                   \n"
     "pextrw    $0x1,%%xmm2,%k3                 \n"
     "pextrw    $0x3,%%xmm2,%k4                 \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movd      %%xmm0,%k2                      \n"
+    "paddw     %9,%%xmm1                       \n"  // make pixels unsigned.
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movd      %%xmm1,%k2                      \n"
     "mov       %w2," MEMACCESS(0) "            \n"
     "lea       " MEMLEA(0x2,0) ",%0            \n"
     "sub       $0x2,%5                         \n"
     "jge       2b                              \n"
 
     LABELALIGN
   "29:                                         \n"
     "addl      $0x1,%5                         \n"
     "jl        99f                             \n"
     MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
     "movd      %k2,%%xmm0                      \n"
     "psrlw     $0x9,%%xmm2                     \n"
     "pshufb    %%xmm5,%%xmm2                   \n"
+    "psubb     %8,%%xmm0                       \n"  // make pixels signed.
     "pxor      %%xmm6,%%xmm2                   \n"
-    "pmaddubsw %%xmm2,%%xmm0                   \n"
-    "psrlw     $0x7,%%xmm0                     \n"
-    "packuswb  %%xmm0,%%xmm0                   \n"
-    "movd      %%xmm0,%k2                      \n"
+    "paddusb   %%xmm7,%%xmm2                   \n"
+    "pmaddubsw %%xmm0,%%xmm2                   \n"
+    "paddw     %9,%%xmm2                       \n"  // make pixels unsigned.
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movd      %%xmm2,%k2                      \n"
     "mov       %b2," MEMACCESS(0) "            \n"
   "99:                                         \n"
   : "+r"(dst_ptr),      // %0
     "+r"(src_ptr),      // %1
     "=&a"(temp_pixel),  // %2
     "=&r"(x0),          // %3
     "=&r"(x1),          // %4
     "+rm"(dst_width)    // %5
   : "rm"(x),            // %6
-    "rm"(dx)            // %7
+    "rm"(dx),           // %7
+#if defined(__x86_64__)
+    "x"(kFsub80),       // %8
+    "x"(kFadd40)        // %9
+#else
+    "m"(kFsub80),       // %8
+    "m"(kFadd40)        // %9
+#endif
   : "memory", "cc", NACL_R14
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 
 // Reads 4 pixels, duplicates them and writes 8 pixels.
 // Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
 void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
                        int dst_width, int x, int dx) {
   asm volatile (
     LABELALIGN
   "1:                                          \n"
@@ skipping 380 matching lines @@
   );
   return num;
 }
 
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
 #endif