[x265] [PATCH] pixel8.inc: Replace weightUnidirPixel vector class function with intrinsic

[dnyaneshwar at multicorewareinc.com]

# HG changeset patch
# User Dnyaneshwar Gorade <dnyaneshwar at multicorewareinc.com>
# Date 1381308260 -19800
#      Wed Oct 09 14:14:20 2013 +0530
# Node ID 9250f0f7e2df00ef582d198b17a835583e2452e2
# Parent  bf91b2c911cb7fd258ee2a67f8dca1955d6ba562
pixel8.inc: Replace weightUnidirPixel vector class function with intrinsic.

diff -r bf91b2c911cb -r 9250f0f7e2df source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc	Wed Oct 09 12:31:10 2013 +0530
+++ b/source/common/vec/pixel8.inc	Wed Oct 09 14:14:20 2013 +0530
@@ -55,31 +55,48 @@
     }
 }
 
-void weightUnidirPixel(pixel *src, pixel *dst, intptr_t srcStride, intptr_t dstStride, int width, int height, int w0, int round, int shift, int offset)
-{
-    int x, y;
-    Vec16uc tmp;
+void weightUnidirPixel(pixel *source, pixel *dest, intptr_t sourceStride, intptr_t destStride, int width, int height, int w0, int arg_round, int shift, int offset)
+ {
+     int x, y;
+    __m128i temp;
+    __m128i vw0    = _mm_set1_epi32(w0);                // broadcast (32-bit integer) w0 to all elements of vw0
+    __m128i iofs   = _mm_set1_epi32(IF_INTERNAL_OFFS);
+    __m128i ofs    = _mm_set1_epi32(offset);
+    __m128i round  = _mm_set1_epi32(arg_round);
+    __m128i src, dst;
 
-    Vec4i vw0(w0), vsrc, iofs(IF_INTERNAL_OFFS), ofs(offset), vround(round), vdst;
-    for (y = height - 1; y >= 0; y--)
-    {
-        for (x = 0; x <= width - 4; x += 4)
-        {
-            tmp = load_partial(const_int(4), src + x);
-            // The intermediate results would outgrow 16 bits because internal offset is too high
-            vsrc = extend_low(extend_low(tmp));
-            vdst = ((vw0 * (vsrc + iofs) + vround) >> shift) + ofs;
-            store_partial(const_int(4), dst + x, compress_unsafe(compress_saturated(vdst, vdst), 0));
-        }
+     for (y = height - 1; y >= 0; y--)
+     {
+         for (x = 0; x <= width - 4; x += 4)
+         {
+             // The intermediate results would outgrow 16 bits because internal offset is too high
+            temp = _mm_cvtsi32_si128(*(uint32_t*)(source + x));
+            src = _mm_unpacklo_epi16(_mm_unpacklo_epi8(temp,_mm_setzero_si128()), _mm_setzero_si128());
+            dst = _mm_add_epi32((_mm_mullo_epi32(vw0, _mm_add_epi32(src, iofs))), round);
+            dst =  _mm_sra_epi32(dst, _mm_cvtsi32_si128(shift));
+            dst = _mm_add_epi32(dst, ofs);
+            *(uint32_t*)(dest + x) = _mm_cvtsi128_si32(_mm_packus_epi16(_mm_packs_epi32(dst, dst), _mm_setzero_si128()));
+         }
 
-        if (width > x)
-        {
-            tmp  = load_partial(const_int(4), src + x);
-            vsrc = extend_low(extend_low(tmp));
-            vdst = ((vw0 * (vsrc + iofs) + vround) >> shift) + ofs;
-            compress_unsafe(compress_saturated(vdst, vdst), 0).store_partial(2, dst + x);
-        }
-        src += srcStride;
-        dst += dstStride;
-    }
-}
+         if (width > x)
+         {
+            temp = _mm_cvtsi32_si128(*(uint32_t*)(source + x));
+            src = _mm_unpacklo_epi16(_mm_unpacklo_epi8(temp, _mm_setzero_si128()), _mm_setzero_si128());
+            dst = _mm_add_epi32((_mm_mullo_epi32(vw0, _mm_add_epi32(src, iofs))), round);
+            dst =  _mm_sra_epi32(dst, _mm_cvtsi32_si128(shift));
+            dst = _mm_add_epi32(dst, ofs);
+            temp = _mm_packus_epi16(_mm_packs_epi32(dst,dst), _mm_setzero_si128());
+
+            union 
+            {
+                int8_t  c[16];
+                int16_t s[8];
+            } u;
+
+            _mm_storeu_si128((__m128i*)u.c, temp);
+            ((int16_t*)(dest + x))[0] = u.s[0];
+         }
+        source += sourceStride;
+        dest += destStride;
+     }
+ }