[x265] [PATCH] replace sad_24 vector class function with intrinsic

dnyaneshwar at multicorewareinc.com dnyaneshwar at multicorewareinc.com
Thu Oct 3 15:39:05 CEST 2013


# HG changeset patch
# User Dnyaneshwar
# Date 1380807452 -19800
#      Thu Oct 03 19:07:32 2013 +0530
# Node ID 3c61b10dfb1370f0c5479b714258094e88583a01
# Parent  55bf7d1fc1716bc67aba7285deb27542796a56fd
replace sad_24 vector class function with intrinsic.

diff -r 55bf7d1fc171 -r 3c61b10dfb13 source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc	Thu Oct 03 18:55:00 2013 +0530
+++ b/source/common/vec/pixel8.inc	Thu Oct 03 19:07:32 2013 +0530
@@ -1406,60 +1406,504 @@
     return _mm_cvtsi128_si32(sum0);
 }
 
+template<int ly>
+int sad_24(pixel *fenc, intptr_t fencstride, pixel *fref, intptr_t frefstride)
+{
+    assert((ly % 4) == 0);
+    __m128i sum0 = _mm_setzero_si128();
+    __m128i sum1 = _mm_setzero_si128();
+
+    if (ly == 4)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T01 = _mm_load_si128((__m128i*)(fenc + fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref));
+        T11 = _mm_load_si128((__m128i*)(fref + frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (2) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (3) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        T20 = _mm_add_epi16(T20, T21);
+        T22 = _mm_add_epi16(T22, T23);
+        sum0 = _mm_add_epi16(T20, T22); 
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (fencstride + 16)));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((2) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((3) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((0) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((1) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((2) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((3) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+    }
+    else if (ly == 8)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T01 = _mm_load_si128((__m128i*)(fenc + fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref));
+        T11 = _mm_load_si128((__m128i*)(fref + frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (2) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (3) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        T20 = _mm_add_epi16(T20, T21);
+        T22 = _mm_add_epi16(T22, T23);
+        sum0 = _mm_add_epi16(T20, T22); 
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (fencstride + 16)));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((2) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((3) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((0) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((1) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((2) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((3) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (4) * fencstride));
+        T01 = _mm_load_si128((__m128i*)(fenc + (5) * fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (6) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (7) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref + (4) * frefstride));
+        T11 = _mm_load_si128((__m128i*)(fref + (5) * frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (6) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (7) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi16(sum0, T20);
+        sum0 = _mm_add_epi16(sum0, T21);
+        sum0 = _mm_add_epi16(sum0, T22);
+        sum0 = _mm_add_epi16(sum0, T23);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + ((4) * fencstride) + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + ((5) * fencstride) + 16));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((6) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((7) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((4) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((5) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((6) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((7) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+    }
+    else if (ly == 16)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T01 = _mm_load_si128((__m128i*)(fenc + fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref));
+        T11 = _mm_load_si128((__m128i*)(fref + frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (2) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (3) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        T20 = _mm_add_epi16(T20, T21);
+        T22 = _mm_add_epi16(T22, T23);
+        sum0 = _mm_add_epi16(T20, T22); 
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + fencstride + 16));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((2) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((3) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + (frefstride + 16)));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((2) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((3) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (4) * fencstride));
+        T01 = _mm_load_si128((__m128i*)(fenc + (5) * fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (6) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (7) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref + (4) * frefstride));
+        T11 = _mm_load_si128((__m128i*)(fref + (5) * frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (6) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (7) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi16(sum0, T20);
+        sum0 = _mm_add_epi16(sum0, T21);
+        sum0 = _mm_add_epi16(sum0, T22);
+        sum0 = _mm_add_epi16(sum0, T23);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + ((4) * fencstride) + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + ((5) * fencstride) + 16));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((6) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((7) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((4) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((5) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((6) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((7) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (8) * fencstride));
+        T01 = _mm_load_si128((__m128i*)(fenc + (9) * fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (10) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (11) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref + (8) * frefstride));
+        T11 = _mm_load_si128((__m128i*)(fref + (9) * frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (10) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (11) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi16(sum0, T20);
+        sum0 = _mm_add_epi16(sum0, T21);
+        sum0 = _mm_add_epi16(sum0, T22);
+        sum0 = _mm_add_epi16(sum0, T23);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + ((8) * fencstride) + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + ((9) * fencstride) + 16));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((10) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((11) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((8) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((9) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((10) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((11) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (12) * fencstride));
+        T01 = _mm_load_si128((__m128i*)(fenc + (13) * fencstride));
+        T02 = _mm_load_si128((__m128i*)(fenc + (14) * fencstride));
+        T03 = _mm_load_si128((__m128i*)(fenc + (15) * fencstride));
+
+        T10 = _mm_load_si128((__m128i*)(fref + (12) * frefstride));
+        T11 = _mm_load_si128((__m128i*)(fref + (13) * frefstride));
+        T12 = _mm_load_si128((__m128i*)(fref + (14) * frefstride));
+        T13 = _mm_load_si128((__m128i*)(fref + (15) * frefstride));
+
+        T20 = _mm_sad_epu8(T00, T10);
+        T21 = _mm_sad_epu8(T01, T11);
+        T22 = _mm_sad_epu8(T02, T12);
+        T23 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi16(sum0, T20);
+        sum0 = _mm_add_epi16(sum0, T21);
+        sum0 = _mm_add_epi16(sum0, T22);
+        sum0 = _mm_add_epi16(sum0, T23);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + ((12) * fencstride) + 16));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + ((13) * fencstride) + 16));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + ((14) * fencstride) + 16));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + ((15) * fencstride) + 16));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref + ((12) * frefstride) + 16));
+        T11 = _mm_loadl_epi64((__m128i*)(fref + ((13) * frefstride) + 16));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+
+        T12 = _mm_loadl_epi64((__m128i*)(fref + ((14) * frefstride) + 16));
+        T13 = _mm_loadl_epi64((__m128i*)(fref + ((15) * frefstride) + 16));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_setzero_si128();
+        T21 = _mm_setzero_si128();
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+
+        sum0 = _mm_add_epi32(sum0, T20);
+        sum0 = _mm_add_epi32(sum0, T21);
+    }
+    else if ((ly % 8) == 0)
+    {
+        for (int i = 0; i < ly; i += 8)
+        {
+            __m128i T00, T01, T02, T03;
+            __m128i T10, T11, T12, T13;
+            __m128i T20, T21, T22, T23;
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i) * fencstride));
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 1) * fencstride));
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 2) * fencstride));
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 3) * fencstride));
+
+            T10 = _mm_load_si128((__m128i*)(fref + (i) * frefstride));
+            T11 = _mm_load_si128((__m128i*)(fref + (i + 1) * frefstride));
+            T12 = _mm_load_si128((__m128i*)(fref + (i + 2) * frefstride));
+            T13 = _mm_load_si128((__m128i*)(fref + (i + 3) * frefstride));
+
+            T20 = _mm_sad_epu8(T00, T10);
+            T21 = _mm_sad_epu8(T01, T11);
+            T22 = _mm_sad_epu8(T02, T12);
+            T23 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi32(sum0, T20);
+            sum0 = _mm_add_epi32(sum0, T21);
+            sum0 = _mm_add_epi32(sum0, T22);
+            sum0 = _mm_add_epi32(sum0, T23);
+
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + ((i) * fencstride) + 16));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 1) * fencstride) + 16));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 2) * fencstride) + 16));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 3) * fencstride) + 16));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref + ((i) * frefstride) + 16));
+            T11 = _mm_loadl_epi64((__m128i*)(fref + ((i + 1) * frefstride) + 16));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+
+            T12 = _mm_loadl_epi64((__m128i*)(fref + ((i + 2) * frefstride) + 16));
+            T13 = _mm_loadl_epi64((__m128i*)(fref + ((i + 3) * frefstride) + 16));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_setzero_si128();
+            T21 = _mm_setzero_si128();
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi32(sum0, T20);
+            sum0 = _mm_add_epi32(sum0, T21);
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i + 4) * fencstride));
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 5) * fencstride));
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 6) * fencstride));
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 7) * fencstride));
+
+            T10 = _mm_load_si128((__m128i*)(fref + (i + 4) * frefstride));
+            T11 = _mm_load_si128((__m128i*)(fref + (i + 5) * frefstride));
+            T12 = _mm_load_si128((__m128i*)(fref + (i + 6) * frefstride));
+            T13 = _mm_load_si128((__m128i*)(fref + (i + 7) * frefstride));
+
+            T20 = _mm_sad_epu8(T00, T10);
+            T21 = _mm_sad_epu8(T01, T11);
+            T22 = _mm_sad_epu8(T02, T12);
+            T23 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi32(sum0, T20);
+            sum0 = _mm_add_epi32(sum0, T21);
+            sum0 = _mm_add_epi32(sum0, T22);
+            sum0 = _mm_add_epi32(sum0, T23);
+
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 4) * fencstride) + 16));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 5) * fencstride) + 16));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 6) * fencstride) + 16));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 7) * fencstride) + 16));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref + ((i + 4) * frefstride) + 16));
+            T11 = _mm_loadl_epi64((__m128i*)(fref + ((i + 5) * frefstride) + 16));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+
+            T12 = _mm_loadl_epi64((__m128i*)(fref + ((i + 6) * frefstride) + 16));
+            T13 = _mm_loadl_epi64((__m128i*)(fref + ((i + 7) * frefstride) + 16));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_setzero_si128();
+            T21 = _mm_setzero_si128();
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi32(sum0, T20);
+            sum0 = _mm_add_epi32(sum0, T21);
+        }
+    }
+    else
+    {
+        for (int i = 0; i < ly; i += 4)
+        {
+            __m128i T00, T01, T02, T03;
+            __m128i T10, T11, T12, T13;
+            __m128i T20, T21, T22, T23;
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i) * fencstride));
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 1) * fencstride));
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 2) * fencstride));
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 3) * fencstride));
+
+            T10 = _mm_load_si128((__m128i*)(fref + (i) * frefstride));
+            T11 = _mm_load_si128((__m128i*)(fref + (i + 1) * frefstride));
+            T12 = _mm_load_si128((__m128i*)(fref + (i + 2) * frefstride));
+            T13 = _mm_load_si128((__m128i*)(fref + (i + 3) * frefstride));
+
+            T20 = _mm_sad_epu8(T00, T10);
+            T21 = _mm_sad_epu8(T01, T11);
+            T22 = _mm_sad_epu8(T02, T12);
+            T23 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi16(sum0, T20);
+            sum0 = _mm_add_epi16(sum0, T21);
+            sum0 = _mm_add_epi16(sum0, T22);
+            sum0 = _mm_add_epi16(sum0, T23);
+
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + ((i) * fencstride) + 16));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 1) * fencstride) + 16));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 2) * fencstride) + 16));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + ((i + 3) * fencstride) + 16));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref + ((i) * frefstride) + 16));
+            T11 = _mm_loadl_epi64((__m128i*)(fref + ((i + 1) * frefstride) + 16));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+
+            T12 = _mm_loadl_epi64((__m128i*)(fref + ((i + 2) * frefstride) + 16));
+            T13 = _mm_loadl_epi64((__m128i*)(fref + ((i + 3) * frefstride) + 16));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_setzero_si128();
+            T21 = _mm_setzero_si128();
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+
+            sum0 = _mm_add_epi32(sum0, T20);
+            sum0 = _mm_add_epi32(sum0, T21);
+        }
+    }
+    sum1 = _mm_shuffle_epi32(sum0, 2);
+    sum0 = _mm_add_epi32(sum0, sum1);
+
+    return _mm_cvtsi128_si32(sum0);
+}
+
 #endif /* if INSTRSET >= X265_CPU_LEVEL_SSE41 */
 
-template<int ly>
-int sad_24(pixel *fenc, intptr_t fencstride, pixel *fref, intptr_t frefstride)
-{
-    Vec16uc m1, n1;
-
-    Vec4i sum(0);
-    Vec8us sad(0);
-    int max_iterators = (ly >> 4) << 4;
-    int row;
-
-    for (row = 0; row < max_iterators; row += 16)
-    {
-        for (int i = 0; i < 16; i++)
-        {
-            m1.load_a(fenc);
-            n1.load(fref);
-            sad.addSumAbsDiff(m1, n1);
-
-            m1.load_a(fenc + 16);
-            m1.cutoff(8);
-            n1.load(fref + 16);
-            n1.cutoff(8);
-            sad.addSumAbsDiff(m1, n1);
-
-            fenc += fencstride;
-            fref += frefstride;
-        }
-
-        sum += extend_low(sad) + extend_high(sad);
-        sad = 0;
-    }
-
-    while (row++ < ly)
-    {
-        m1.load_a(fenc);
-        n1.load(fref);
-        sad.addSumAbsDiff(m1, n1);
-
-        m1.load_a(fenc + 16);
-        m1.cutoff(8);
-        n1.load(fref + 16);
-        n1.cutoff(8);
-        sad.addSumAbsDiff(m1, n1);
-
-        fenc += fencstride;
-        fref += frefstride;
-    }
-
-    sum += extend_low(sad) + extend_high(sad);
-    return horizontal_add(sum);
-}
-
 template<int size>
 ALWAYSINLINE void unrollFunc_32(pixel *fenc, intptr_t fencstride, pixel *fref, intptr_t frefstride, Vec8us& sad)
 {


More information about the x265-devel mailing list