[x265] [PATCH] intra-sse3.cpp: Replace PredIntraAng4_m_13 vector class function with intrinsic

[dnyaneshwar at multicorewareinc.com]

# HG changeset patch
# User Dnyaneshwar Gorade <dnyaneshwar at multicorewareinc.com>
# Date 1381486773 -19800
#      Fri Oct 11 15:49:33 2013 +0530
# Node ID 4738c4ccb89e0a121a2632c70425717516727ad2
# Parent  5194d5a5737b60192d06dfcdb8b0d632f30f1875
intra-sse3.cpp: Replace PredIntraAng4_m_13 vector class function with intrinsic.

diff -r 5194d5a5737b -r 4738c4ccb89e source/common/vec/intra-sse3.cpp
--- a/source/common/vec/intra-sse3.cpp	Fri Oct 11 15:46:32 2013 +0530
+++ b/source/common/vec/intra-sse3.cpp	Fri Oct 11 15:49:33 2013 +0530
@@ -1761,30 +1761,32 @@
 
 void PredIntraAng4_m_13(pixel* dst, int dstStride, pixel *refMain, int dirMode)
 {
-    PRED_INTRA_ANG4_START
-
-        tmp16_1 = (Vec16uc)load_partial(const_int(8), refMain - 1);
-
-    row41 = extend_low(tmp16_1);    //offsets(-2,-1,0,1)
-    tmp2uq = reinterpret_i(tmp16_1);
-    tmp2uq = tmp2uq >> 8;
-    tmp16_2 = reinterpret_i(tmp2uq);
-    row42 = extend_low(tmp16_2);    //offsets(-1,0,1,2)
+    PRED_INTRA_ANGLE_4_START
+
+    tmp16_1 = _mm_loadl_epi64((__m128i*)(refMain - 1));
+    row41 = _mm_unpacklo_epi8(tmp16_1, _mm_setzero_si128());
+
+    tmp2 = tmp16_1;
+    tmp2 = _mm_srl_epi64(tmp2,_mm_cvtsi32_si128(8));
+
+    tmp16_2 = tmp2;
+    row42 = _mm_unpacklo_epi8(tmp16_2, _mm_setzero_si128());
 
     row11 = row42;
-    tmp2uq = reinterpret_i(tmp16_1);
-    tmp2uq = tmp2uq >> 16;
-    tmp16_2 = reinterpret_i(tmp2uq);
-    row12 = extend_low(tmp16_2);    //offsets(-1,0,1,2)
-
-    row21 = row42;                  //offsets(0,1,2,3)
+    tmp2 = tmp16_1;
+    tmp2 = _mm_srl_epi64(tmp2,_mm_cvtsi32_si128(16));
+
+    tmp16_2 = tmp2;
+    row12 = _mm_unpacklo_epi8(tmp16_2, _mm_setzero_si128());
+
+    row21 = row42;
     row22 = row12;
     row31 = row41;
     row32 = row42;
 
-    v_deltaPos = v_ipAngle = -13;
-
-    PRED_INTRA_ANG4_END
+    deltaPos = ipAngle = _mm_set1_epi16(-13);
+
+    PRED_INTRA_ANGLE_4_END
 }
 
 void PredIntraAng4_m_17(pixel* dst, int dstStride, pixel *refMain, int dirMode)