[x265] [PATCH] replace sad_x3_12 vector class function with intrinsic

dnyaneshwar at multicorewareinc.com dnyaneshwar at multicorewareinc.com
Thu Oct 3 16:16:27 CEST 2013


# HG changeset patch
# User Dnyaneshwar
# Date 1380809737 -19800
#      Thu Oct 03 19:45:37 2013 +0530
# Node ID adeb4daba79fa52ae0171dc8d4b1dd4218e7d2f9
# Parent  76d82811cd42e019fb8e23e99804a240f6382c44
replace sad_x3_12 vector class function with intrinsic.

diff -r 76d82811cd42 -r adeb4daba79f source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc	Thu Oct 03 19:41:55 2013 +0530
+++ b/source/common/vec/pixel8.inc	Thu Oct 03 19:45:37 2013 +0530
@@ -3814,81 +3814,810 @@
 
 #endif /* if HAVE_MMX */
 
-#endif /* if INSTRSET >= X265_CPU_LEVEL_SSE41 */
-
 /* For performance - This function assumes that the *last load* can access 16 elements. */
-template<int ly>
-void sad_x3_12(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)
-    {
-        for (int i = 0; i < 16; i++)
-        {
-            m1.load_a(fenc);
-            m1.cutoff(12);
-            n1.load(fref1);
-            n1.cutoff(12);
-            n2.load(fref2);
-            n2.cutoff(12);
-            n3.load(fref3);
-            n3.cutoff(12);
-
-            sad1.addSumAbsDiff(m1, n1);
-            sad2.addSumAbsDiff(m1, n2);
-            sad3.addSumAbsDiff(m1, n3);
-
-            fenc += FENC_STRIDE;
-            fref1 += frefstride;
-            fref2 += frefstride;
-            fref3 += frefstride;
-        }
-
-        sum1 += extend_low(sad1) + extend_high(sad1);
-        sum2 += extend_low(sad2) + extend_high(sad2);
-        sum3 += extend_low(sad3) + extend_high(sad3);
-        sad1 = 0;
-        sad2 = 0;
-        sad3 = 0;
-    }
-
-    while (row++ < ly)
-    {
-        m1.load_a(fenc);
-        m1.cutoff(12);
-        n1.load(fref1);
-        n1.cutoff(12);
-        n2.load(fref2);
-        n2.cutoff(12);
-        n3.load(fref3);
-        n3.cutoff(12);
-
-        sad1.addSumAbsDiff(m1, n1);
-        sad2.addSumAbsDiff(m1, n2);
-        sad3.addSumAbsDiff(m1, n3);
-
-        fenc += FENC_STRIDE;
-        fref1 += frefstride;
-        fref2 += frefstride;
-        fref3 += frefstride;
-    }
-
-    sum1 += extend_low(sad1) + extend_high(sad1);
-    sum2 += extend_low(sad2) + extend_high(sad2);
-    sum3 += extend_low(sad3) + extend_high(sad3);
-
-    res[0] = horizontal_add(sum1);
-    res[1] = horizontal_add(sum2);
-    res[2] = horizontal_add(sum3);
+
+template<int ly>
+void sad_x3_12(pixel *fenc, pixel *fref1, pixel *fref2, pixel *fref3, intptr_t frefstride, int *res)
+{
+    assert((ly % 4) == 0);
+
+    __m128i mask = _mm_set_epi32(0x0, 0xffffffff, 0xffffffff, 0xffffffff);
+
+    if(ly == 4)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+        __m128i sum0, sum1;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] = _mm_cvtsi128_si32(sum0);
+    }
+    if (ly == 8)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+        __m128i sum0, sum1;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] = _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (4) * FENC_STRIDE));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + (5) * FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (6) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (7) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] += _mm_cvtsi128_si32(sum0);
+    }
+    if (ly == 16)
+    {
+        __m128i T00, T01, T02, T03;
+        __m128i T10, T11, T12, T13;
+        __m128i T20, T21, T22, T23;
+        __m128i sum0, sum1;
+
+        T00 = _mm_load_si128((__m128i*)(fenc));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (2) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (3) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] = _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (2) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (3) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] = _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (4) * FENC_STRIDE));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + (5) * FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (6) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (7) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3 + (4) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + (5) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (6) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (7) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] += _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (8) * FENC_STRIDE));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + (9) * FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (10) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (11) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1 + (8) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + (9) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (10) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (11) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2 + (8) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + (9) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (10) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (11) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3 + (8) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + (9) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (10) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (11) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] += _mm_cvtsi128_si32(sum0);
+
+        T00 = _mm_load_si128((__m128i*)(fenc + (12) * FENC_STRIDE));
+        T00 = _mm_and_si128(T00, mask);
+        T01 = _mm_load_si128((__m128i*)(fenc + (13) * FENC_STRIDE));
+        T01 = _mm_and_si128(T01, mask);
+        T02 = _mm_load_si128((__m128i*)(fenc + (14) * FENC_STRIDE));
+        T02 = _mm_and_si128(T02, mask);
+        T03 = _mm_load_si128((__m128i*)(fenc + (15) * FENC_STRIDE));
+        T03 = _mm_and_si128(T03, mask);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref1 + (12) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref1 + (13) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref1 + (14) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref1 + (15) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[0] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref2 + (12) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref2 + (13) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref2 + (14) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref2 + (15) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[1] += _mm_cvtsi128_si32(sum0);
+
+        T10 = _mm_loadu_si128((__m128i*)(fref3 + (12) * frefstride));
+        T10 = _mm_and_si128(T10, mask);
+        T11 = _mm_loadu_si128((__m128i*)(fref3 + (13) * frefstride));
+        T11 = _mm_and_si128(T11, mask);
+        T12 = _mm_loadu_si128((__m128i*)(fref3 + (14) * frefstride));
+        T12 = _mm_and_si128(T12, mask);
+        T13 = _mm_loadu_si128((__m128i*)(fref3 + (15) * frefstride));
+        T13 = _mm_and_si128(T13, mask);
+
+        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); 
+
+        sum1 = _mm_shuffle_epi32(sum0, 2);
+        sum0 = _mm_add_epi32(sum0, sum1);
+        res[2] += _mm_cvtsi128_si32(sum0);
+    }
+    if ((ly % 8) == 0)
+    {
+        res[0] = res[1] = res[2] = 0;
+        for(int i = 0; i < ly; i += 8)
+        {
+            __m128i T00, T01, T02, T03;
+            __m128i T10, T11, T12, T13;
+            __m128i T20, T21, T22, T23;
+            __m128i sum0, sum1;
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i) * FENC_STRIDE));
+            T00 = _mm_and_si128(T00, mask);
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 1) * FENC_STRIDE));
+            T01 = _mm_and_si128(T01, mask);
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 2) * FENC_STRIDE));
+            T02 = _mm_and_si128(T02, mask);
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 3) * FENC_STRIDE));
+            T03 = _mm_and_si128(T03, mask);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref1 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref1 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref1 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref1 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[0] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref2 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref2 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref2 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref2 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[1] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref3 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref3 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref3 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref3 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[2] += _mm_cvtsi128_si32(sum0);
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i + 4) * FENC_STRIDE));
+            T00 = _mm_and_si128(T00, mask);
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 5) * FENC_STRIDE));
+            T01 = _mm_and_si128(T01, mask);
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 6) * FENC_STRIDE));
+            T02 = _mm_and_si128(T02, mask);
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 7) * FENC_STRIDE));
+            T03 = _mm_and_si128(T03, mask);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref1 + (i + 4) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref1 + (i + 5) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref1 + (i + 6) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref1 + (i + 7) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[0] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref2 + (i + 4) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref2 + (i + 5) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref2 + (i + 6) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref2 + (i + 7) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[1] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref3 + (i + 4) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref3 + (i + 5) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref3 + (i + 6) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref3 + (i + 7) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[2] += _mm_cvtsi128_si32(sum0);
+        }
+    }
+    else
+    {
+        res[0] = res[1] = res[2] = 0;
+        for(int i = 0; i < ly; i += 4)
+        {
+            __m128i T00, T01, T02, T03;
+            __m128i T10, T11, T12, T13;
+            __m128i T20, T21, T22, T23;
+            __m128i sum0, sum1;
+
+            T00 = _mm_load_si128((__m128i*)(fenc + (i) * FENC_STRIDE));
+            T00 = _mm_and_si128(T00, mask);
+            T01 = _mm_load_si128((__m128i*)(fenc + (i + 1) * FENC_STRIDE));
+            T01 = _mm_and_si128(T01, mask);
+            T02 = _mm_load_si128((__m128i*)(fenc + (i + 2) * FENC_STRIDE));
+            T02 = _mm_and_si128(T02, mask);
+            T03 = _mm_load_si128((__m128i*)(fenc + (i + 3) * FENC_STRIDE));
+            T03 = _mm_and_si128(T03, mask);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref1 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref1 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref1 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref1 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[0] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref2 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref2 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref2 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref2 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[1] += _mm_cvtsi128_si32(sum0);
+
+            T10 = _mm_loadu_si128((__m128i*)(fref3 + (i) * frefstride));
+            T10 = _mm_and_si128(T10, mask);
+            T11 = _mm_loadu_si128((__m128i*)(fref3 + (i + 1) * frefstride));
+            T11 = _mm_and_si128(T11, mask);
+            T12 = _mm_loadu_si128((__m128i*)(fref3 + (i + 2) * frefstride));
+            T12 = _mm_and_si128(T12, mask);
+            T13 = _mm_loadu_si128((__m128i*)(fref3 + (i + 3) * frefstride));
+            T13 = _mm_and_si128(T13, mask);
+
+            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); 
+
+            sum1 = _mm_shuffle_epi32(sum0, 2);
+            sum0 = _mm_add_epi32(sum0, sum1);
+            res[2] += _mm_cvtsi128_si32(sum0);
+        }
+    }
 }
 
-#if INSTRSET >= X265_CPU_LEVEL_SSE41
 template<int ly>
 void sad_x3_16(pixel *fenc, pixel *fref1, pixel *fref2, pixel *fref3, intptr_t frefstride, int *res)
 {


More information about the x265-devel mailing list