[x265] [PATCH] sse.inc: Replace sse_ss48 vector class function with intrinsic

[dnyaneshwar at multicorewareinc.com]

# HG changeset patch
# User Dnyaneshwar Gorade <dnyaneshwar at multicorewareinc.com>
# Date 1381319957 -19800
#      Wed Oct 09 17:29:17 2013 +0530
# Node ID 933485037b324449660bb74192e1b889d861d08e
# Parent  cb9f29e64e64403a3d4c9c01fa55736b867f720e
sse.inc: Replace sse_ss48 vector class function with intrinsic.

diff -r cb9f29e64e64 -r 933485037b32 source/common/vec/sse.inc
--- a/source/common/vec/sse.inc	Wed Oct 09 17:25:23 2013 +0530
+++ b/source/common/vec/sse.inc	Wed Oct 09 17:29:17 2013 +0530
@@ -307,70 +307,75 @@
 {
     int rows = ly;
 
-    Vec4i diff(0);
-    Vec8s m1, n1;
-    Vec4i sum(0);
+    __m128i sum  = _mm_setzero_si128();
+    __m128i m1, n1, diff, sign1, sign2, tmp1, tmp2;
+
+#define PROCESS_SSE_SS48x1(BASE)\
+    m1 = _mm_loadu_si128((__m128i const*)(fenc + BASE)); \
+    n1 = _mm_loadu_si128((__m128i const*)(fref + BASE)); \
+    sign1 = _mm_srai_epi16(m1, 15); \
+    tmp1 = _mm_unpacklo_epi16(m1, sign1); \
+    sign2 = _mm_srai_epi16(n1, 15); \
+    tmp2 = _mm_unpacklo_epi16(n1, sign2); \
+    diff = _mm_sub_epi32(tmp1, tmp2); \
+    diff = _mm_mullo_epi32(diff, diff); \
+    sum = _mm_add_epi32(sum, diff)
+
     for (; rows != 0; rows--)
     {
-        m1.load(fenc);
-        n1.load(fref);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        PROCESS_SSE_SS48x1(0);
 
-        m1.load(fenc + 8);
-        n1.load(fref + 8);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
 
-        m1.load(fenc + 16);
-        n1.load(fref + 16);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        PROCESS_SSE_SS48x1(8);
 
-        m1.load(fenc + 24);
-        n1.load(fref + 24);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
 
-        m1.load(fenc + 32);
-        n1.load(fref + 32);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        PROCESS_SSE_SS48x1(16);
 
-        m1.load(fenc + 40);
-        n1.load(fref + 40);
-        diff = extend_low(m1) - extend_low(n1);
-        diff = diff * diff;
-        sum += diff;
-        diff = extend_high(m1) - extend_high(n1);
-        diff = diff * diff;
-        sum += diff;
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
+
+        PROCESS_SSE_SS48x1(24);
+
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
+
+        PROCESS_SSE_SS48x1(32);
+
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
+
+        PROCESS_SSE_SS48x1(40);
+
+        m1 = _mm_unpackhi_epi16(m1, sign1);
+        n1 = _mm_unpackhi_epi16(n1, sign2);
+        diff = _mm_sub_epi32(m1, n1);
+        diff = _mm_mullo_epi32(diff, diff);
+        sum = _mm_add_epi32(sum, diff);
 
         fenc += strideFenc;
         fref += strideFref;
     }
-
-    return horizontal_add(sum);
+    __m128i sum1  = _mm_hadd_epi32(sum, sum);             // horizontally add 4 elements
+    return _mm_cvtsi128_si32(_mm_hadd_epi32(sum1, sum1));
 }
 
 template<int ly>