[x265] [PATCH] sad[16xN]:Intrinsic
praveen at multicorewareinc.com
praveen at multicorewareinc.com
Fri Aug 23 14:01:06 CEST 2013
# HG changeset patch
# User praveentiwari
# Date 1377259254 -19800
# Node ID 7e5c568e480489d74c50673d538ddd323218cf97
# Parent 2892cde2c9c6ae9b2be004d9ae3420c210f4b6e8
sad[16xN]:Intrinsic
diff -r 2892cde2c9c6 -r 7e5c568e4804 source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc Fri Aug 23 17:19:33 2013 +0530
+++ b/source/common/vec/pixel8.inc Fri Aug 23 17:30:54 2013 +0530
@@ -876,53 +876,238 @@
return horizontal_add(sum);
}
-template<int size>
-ALWAYSINLINE void unrollFunc_16(pixel *fenc, intptr_t fencstride, pixel *fref, intptr_t frefstride, Vec8us& sad)
-{
- unrollFunc_16<1>(fenc, fencstride, fref, frefstride, sad);
- unrollFunc_16<size - 1>(fenc + fencstride, fencstride, fref + frefstride, frefstride, sad);
-}
-
-template<>
-ALWAYSINLINE void unrollFunc_16<1>(pixel *fenc, intptr_t, pixel *fref, intptr_t, Vec8us& sad)
-{
- Vec16uc m1, n1;
-
- m1.load_a(fenc);
- n1.load(fref);
- sad.addSumAbsDiff(m1, n1);
-}
-
template<int ly>
int sad_16(pixel * fenc, intptr_t fencstride, pixel * fref, intptr_t frefstride)
{
- Vec4i sum(0);
- Vec8us sad(0);
- int max_iterators = (ly >> 4) << 4;
- int row = 0;
-
- if (ly < 16)
+ assert((ly % 4) == 0);
+
+ __m128i sum0 = _mm_setzero_si128();
+ __m128i sum1 = _mm_setzero_si128();
+ __m128i T00, T01, T02, T03;
+ __m128i T10, T11, T12, T13;
+ __m128i T20, T21, T22, T23;
+
+ if (ly == 4)
{
- unrollFunc_16<ly>(fenc, fencstride, fref, frefstride, sad);
- sum += extend_low(sad) + extend_high(sad);
- return horizontal_add(sum);
+ T00 = _mm_load_si128((__m128i*)(fenc + (0) * fencstride));
+ T01 = _mm_load_si128((__m128i*)(fenc + (1) * fencstride));
+ T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+ T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+ T10 = _mm_load_si128((__m128i*)(fref + (0) * frefstride));
+ T11 = _mm_load_si128((__m128i*)(fref + (1) * 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);
+
+ sum0 = _mm_add_epi16(sum0, T20);
+ sum0 = _mm_add_epi16(sum0, T21);
+ sum0 = _mm_add_epi16(sum0, T22);
+ sum0 = _mm_add_epi16(sum0, T23);
}
- for (row = 0; row < max_iterators; row += 16)
+ else if (ly == 8)
{
- unrollFunc_16<16>(fenc, fencstride, fref, frefstride, sad);
- sum += extend_low(sad) + extend_high(sad);
- sad = 0;
- fenc += fencstride * 16;
- fref += frefstride * 16;
+ T00 = _mm_load_si128((__m128i*)(fenc + (0) * fencstride));
+ T01 = _mm_load_si128((__m128i*)(fenc + (1) * fencstride));
+ T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+ T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+ T10 = _mm_load_si128((__m128i*)(fref + (0) * frefstride));
+ T11 = _mm_load_si128((__m128i*)(fref + (1) * 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);
+
+ 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_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);
}
-
- if (ly & 8)
+ else if (ly == 16)
{
- unrollFunc_16<8>(fenc, fencstride, fref, frefstride, sad);
- sum += extend_low(sad) + extend_high(sad);
- return horizontal_add(sum);
+ T00 = _mm_load_si128((__m128i*)(fenc + (0) * fencstride));
+ T01 = _mm_load_si128((__m128i*)(fenc + (1) * fencstride));
+ T02 = _mm_load_si128((__m128i*)(fenc + (2) * fencstride));
+ T03 = _mm_load_si128((__m128i*)(fenc + (3) * fencstride));
+
+ T10 = _mm_load_si128((__m128i*)(fref + (0) * frefstride));
+ T11 = _mm_load_si128((__m128i*)(fref + (1) * 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);
+
+ 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_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_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_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);
}
- return horizontal_add(sum);
+ else if ((ly % 8) == 0)
+ {
+ for (int i = 0; i < ly; i += 8)
+ {
+ T00 = _mm_load_si128((__m128i*)(fenc + (i + 0) * 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 + 0) * 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_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_epi16(sum0, T20);
+ sum0 = _mm_add_epi16(sum0, T21);
+ sum0 = _mm_add_epi16(sum0, T22);
+ sum0 = _mm_add_epi16(sum0, T23);
+ }
+ }
+ else
+ {
+ for (int i = 0; i < ly; i += 4)
+ {
+ T00 = _mm_load_si128((__m128i*)(fenc + (i + 0) * 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 + 0) * 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);
+ }
+ }
+
+ sum1 = _mm_shuffle_epi32(sum0, 2);
+ sum0 = _mm_add_epi32(sum0, sum1);
+
+ return _mm_cvtsi128_si32(sum0);
}
template<int ly>
More information about the x265-devel
mailing list