[x265] [PATCH] pixel8inc: sad_x3 8xN, replace vector with intrinsic

praveen at multicorewareinc.com praveen at multicorewareinc.com
Fri Aug 23 15:25:32 CEST 2013


# HG changeset patch
# User praveentiwari
# Date 1377264321 -19800
# Node ID 76b9c59e69aada33370b0d47a6d5c4ddbe3701d8
# Parent  7e5c568e480489d74c50673d538ddd323218cf97
pixel8inc: sad_x3 8xN, replace vector with intrinsic

diff -r 7e5c568e4804 -r 76b9c59e69aa source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc	Fri Aug 23 17:30:54 2013 +0530
+++ b/source/common/vec/pixel8.inc	Fri Aug 23 18:55:21 2013 +0530
@@ -1818,68 +1818,527 @@
     }
 }
 
-/* For performance - This function assumes that the *last load* can access 16 elements. */
 template<int ly>
 void sad_x3_8(pixel *fenc, pixel *fref1, pixel *fref2, pixel *fref3, intptr_t frefstride, int *res)
 {
-    Vec16uc m1, n1, n2, n3;
-
-    Vec4i sum1(0), sum2(0), sum3(0);
-    Vec8us sad1(0), sad2(0), sad3(0);
-    int max_iterators = (ly >> 4) << 4;
-    int row;
-
-    for (row = 0; row < max_iterators; row += 16)
+    assert((ly % 4) == 0);
+    __m128i sum0 = _mm_setzero_si128();
+
+    __m128i T00, T01, T02, T03;
+    __m128i T10, T11, T12, T13;
+    __m128i T20, T21;
+
+    if (ly == 4)
     {
-        for (int i = 0; i < 16; i++)
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (0) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (1) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = _mm_cvtsi128_si32(sum0);
+    }
+
+    if (ly == 8)
+    {
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (0) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (1) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (4) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (5) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (6) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (7) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+    }
+    if (ly == 16)
+    {
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (0) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (1) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (0) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (1) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (2) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (4) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (5) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (6) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (7) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (4) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (5) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (6) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (7) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (8) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (9) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (10) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (11) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (8) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (9) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (10) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (11) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (8) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (9) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (10) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (11) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (8) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (9) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (10) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (11) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_loadl_epi64((__m128i*)(fenc + (12) * FENC_STRIDE));
+        T01 = _mm_loadl_epi64((__m128i*)(fenc + (13) * FENC_STRIDE));
+        T01 = _mm_unpacklo_epi64(T00, T01);
+        T02 = _mm_loadl_epi64((__m128i*)(fenc + (14) * FENC_STRIDE));
+        T03 = _mm_loadl_epi64((__m128i*)(fenc + (15) * FENC_STRIDE));
+        T03 = _mm_unpacklo_epi64(T02, T03);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref1 + (12) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref1 + (13) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref1 + (14) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref1 + (15) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref2 + (12) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref2 + (13) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref2 + (14) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref2 + (15) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadl_epi64((__m128i*)(fref3 + (12) * frefstride));
+        T11 = _mm_loadl_epi64((__m128i*)(fref3 + (13) * frefstride));
+        T11 = _mm_unpacklo_epi64(T10, T11);
+        T12 = _mm_loadl_epi64((__m128i*)(fref3 + (14) * frefstride));
+        T13 = _mm_loadl_epi64((__m128i*)(fref3 + (15) * frefstride));
+        T13 = _mm_unpacklo_epi64(T12, T13);
+
+        T20 = _mm_sad_epu8(T01, T11);
+        T21 = _mm_sad_epu8(T03, T13);
+        T21 = _mm_add_epi32(T20, T21);
+        sum0 = _mm_shuffle_epi32(T21, 2);
+        sum0 = _mm_add_epi32(sum0, T21);
+        res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+    }
+
+    else if ((ly % 8) == 0)
+    {
+        res[0] = res[1] = res[2] = 0;
+        for (int i = 0; i < ly; i += 8)
         {
-            m1.load_a(fenc);
-            n1.load(fref1);
-            n2.load(fref2);
-            n3.load(fref3);
-
-            sad1.addSumAbsDiff(m1, n1);
-            sad2.addSumAbsDiff(m1, n2);
-            sad3.addSumAbsDiff(m1, n3);
-
-            fenc += FENC_STRIDE;
-            fref1 += frefstride;
-            fref2 += frefstride;
-            fref3 += frefstride;
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + (i + 0) * FENC_STRIDE));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + (i + 1) * FENC_STRIDE));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + (i + 2) * FENC_STRIDE));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + (i + 3) * FENC_STRIDE));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + (i + 4) * FENC_STRIDE));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + (i + 5) * FENC_STRIDE));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + (i + 6) * FENC_STRIDE));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + (i + 7) * FENC_STRIDE));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 4) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 5) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 6) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 7) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 4) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 5) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 6) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 7) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 4) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 5) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 6) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 7) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[2] = res[2] + _mm_cvtsi128_si32(sum0);
         }
-
-        sum1 += extend_low(sad1);
-        sum2 += extend_low(sad2);
-        sum3 += extend_low(sad3);
-        sad1 = 0;
-        sad2 = 0;
-        sad3 = 0;
     }
 
-    while (row++ < ly)
+    else
     {
-        m1.load_a(fenc);
-        n1.load(fref1);
-        n2.load(fref2);
-        n3.load(fref3);
-
-        sad1.addSumAbsDiff(m1, n1);
-        sad2.addSumAbsDiff(m1, n2);
-        sad3.addSumAbsDiff(m1, n3);
-
-        fenc += FENC_STRIDE;
-        fref1 += frefstride;
-        fref2 += frefstride;
-        fref3 += frefstride;
+        res[0] = res[1] = res[2] = 0;
+        for (int i = 0; i < ly; i += 4)
+        {
+            T00 = _mm_loadl_epi64((__m128i*)(fenc + (i + 0) * FENC_STRIDE));
+            T01 = _mm_loadl_epi64((__m128i*)(fenc + (i + 1) * FENC_STRIDE));
+            T01 = _mm_unpacklo_epi64(T00, T01);
+            T02 = _mm_loadl_epi64((__m128i*)(fenc + (i + 2) * FENC_STRIDE));
+            T03 = _mm_loadl_epi64((__m128i*)(fenc + (i + 3) * FENC_STRIDE));
+            T03 = _mm_unpacklo_epi64(T02, T03);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref1 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[0] = res[0] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref2 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[1] = res[1] + _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 0) * frefstride));
+            T11 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 1) * frefstride));
+            T11 = _mm_unpacklo_epi64(T10, T11);
+            T12 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 2) * frefstride));
+            T13 = _mm_loadl_epi64((__m128i*)(fref3 + (i + 3) * frefstride));
+            T13 = _mm_unpacklo_epi64(T12, T13);
+
+            T20 = _mm_sad_epu8(T01, T11);
+            T21 = _mm_sad_epu8(T03, T13);
+            T21 = _mm_add_epi32(T20, T21);
+            sum0 = _mm_shuffle_epi32(T21, 2);
+            sum0 = _mm_add_epi32(sum0, T21);
+            res[2] = res[2] + _mm_cvtsi128_si32(sum0);
+        }
     }
-
-    sum1 += extend_low(sad1);
-    sum2 += extend_low(sad2);
-    sum3 += extend_low(sad3);
-
-    res[0] = horizontal_add(sum1);
-    res[1] = horizontal_add(sum2);
-    res[2] = horizontal_add(sum3);
 }
 
 /* For performance - This function assumes that the *last load* can access 16 elements. */


More information about the x265-devel mailing list