[x265] [PATCH] pixel8.inc: sad_x3_8 integrated fast MMX 32-bit build code
praveen at multicorewareinc.com
praveen at multicorewareinc.com
Sun Aug 25 07:27:25 CEST 2013
# HG changeset patch
# User praveentiwari
# Date 1377408435 -19800
# Node ID d00c0ff31813bbfae5a928b7850533b7d7902378
# Parent 2d9f2ffb4b5da6e1a1f5a156ee26c2747161a749
pixel8.inc: sad_x3_8 integrated fast MMX 32-bit build code
diff -r 2d9f2ffb4b5d -r d00c0ff31813 source/common/vec/pixel8.inc
--- a/source/common/vec/pixel8.inc Sun Aug 25 10:46:41 2013 +0530
+++ b/source/common/vec/pixel8.inc Sun Aug 25 10:57:15 2013 +0530
@@ -2184,6 +2184,7 @@
#endif /* ifdef X86_64 */
+#ifdef x86_64
template<int ly>
void sad_x3_8(pixel *fenc, pixel *fref1, pixel *fref2, pixel *fref3, intptr_t frefstride, int *res)
{
@@ -2705,6 +2706,361 @@
}
}
+#else /* ifdef x86_64 */
+template<int ly>
+void sad_x3_8(pixel *fenc, pixel *fref1, pixel *fref2, pixel *fref3, intptr_t frefstride, int *res)
+{
+ assert((ly % 4) == 0);
+
+ __m64 sum0 = _mm_setzero_si64();
+ __m64 sum1 = _mm_setzero_si64();
+ __m64 sum2 = _mm_setzero_si64();
+
+ __m64 T00, T01, T02, T03, T04, T05, T06, T07;
+ __m64 T0, T1, T2, T3, T4, T5, T6, T7;
+ __m64 T10, T11, T12, T13, T14, T15, T16, T17;
+ __m64 T20, T21, T22, T23, T24, T25, T26, T27;
+
+ if (16 == ly)
+ {
+ T00 = (*(__m64*)(fenc + 0 * FENC_STRIDE));
+ T01 = (*(__m64*)(fenc + 1 * FENC_STRIDE));
+ T02 = (*(__m64*)(fenc + 2 * FENC_STRIDE));
+ T03 = (*(__m64*)(fenc + 3 * FENC_STRIDE));
+ T04 = (*(__m64*)(fenc + 4 * FENC_STRIDE));
+ T05 = (*(__m64*)(fenc + 5 * FENC_STRIDE));
+ T06 = (*(__m64*)(fenc + 6 * FENC_STRIDE));
+ T07 = (*(__m64*)(fenc + 7 * FENC_STRIDE));
+ T0 = (*(__m64*)(fenc + 8 * FENC_STRIDE));
+ T1 = (*(__m64*)(fenc + 9 * FENC_STRIDE));
+ T2 = (*(__m64*)(fenc + 10 * FENC_STRIDE));
+ T3 = (*(__m64*)(fenc + 11 * FENC_STRIDE));
+ T4 = (*(__m64*)(fenc + 12 * FENC_STRIDE));
+ T5 = (*(__m64*)(fenc + 13 * FENC_STRIDE));
+ T6 = (*(__m64*)(fenc + 14 * FENC_STRIDE));
+ T7 = (*(__m64*)(fenc + 15 * FENC_STRIDE));
+
+ T10 = (*(__m64*)(fref1 + 0 * frefstride));
+ T11 = (*(__m64*)(fref1 + 1 * frefstride));
+ T12 = (*(__m64*)(fref1 + 2 * frefstride));
+ T13 = (*(__m64*)(fref1 + 3 * frefstride));
+ T14 = (*(__m64*)(fref1 + 4 * frefstride));
+ T15 = (*(__m64*)(fref1 + 5 * frefstride));
+ T16 = (*(__m64*)(fref1 + 6 * frefstride));
+ T17 = (*(__m64*)(fref1 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum0 = _mm_add_pi16(sum0, T20);
+ sum0 = _mm_add_pi16(sum0, T21);
+ sum0 = _mm_add_pi16(sum0, T22);
+ sum0 = _mm_add_pi16(sum0, T23);
+ sum0 = _mm_add_pi16(sum0, T24);
+ sum0 = _mm_add_pi16(sum0, T25);
+ sum0 = _mm_add_pi16(sum0, T26);
+ sum0 = _mm_add_pi16(sum0, T27);
+
+ T10 = (*(__m64*)(fref1 + 8 * frefstride));
+ T11 = (*(__m64*)(fref1 + 9 * frefstride));
+ T12 = (*(__m64*)(fref1 + 10 * frefstride));
+ T13 = (*(__m64*)(fref1 + 11 * frefstride));
+ T14 = (*(__m64*)(fref1 + 12 * frefstride));
+ T15 = (*(__m64*)(fref1 + 13 * frefstride));
+ T16 = (*(__m64*)(fref1 + 14 * frefstride));
+ T17 = (*(__m64*)(fref1 + 15 * frefstride));
+
+ T20 = _mm_sad_pu8(T0, T10);
+ T21 = _mm_sad_pu8(T1, T11);
+ T22 = _mm_sad_pu8(T2, T12);
+ T23 = _mm_sad_pu8(T3, T13);
+ T24 = _mm_sad_pu8(T4, T14);
+ T25 = _mm_sad_pu8(T5, T15);
+ T26 = _mm_sad_pu8(T6, T16);
+ T27 = _mm_sad_pu8(T7, T17);
+
+ sum0 = _mm_add_pi16(sum0, T20);
+ sum0 = _mm_add_pi16(sum0, T21);
+ sum0 = _mm_add_pi16(sum0, T22);
+ sum0 = _mm_add_pi16(sum0, T23);
+ sum0 = _mm_add_pi16(sum0, T24);
+ sum0 = _mm_add_pi16(sum0, T25);
+ sum0 = _mm_add_pi16(sum0, T26);
+ sum0 = _mm_add_pi16(sum0, T27);
+ res[0] = _m_to_int(sum0);
+
+ T10 = (*(__m64*)(fref2 + 0 * frefstride));
+ T11 = (*(__m64*)(fref2 + 1 * frefstride));
+ T12 = (*(__m64*)(fref2 + 2 * frefstride));
+ T13 = (*(__m64*)(fref2 + 3 * frefstride));
+ T14 = (*(__m64*)(fref2 + 4 * frefstride));
+ T15 = (*(__m64*)(fref2 + 5 * frefstride));
+ T16 = (*(__m64*)(fref2 + 6 * frefstride));
+ T17 = (*(__m64*)(fref2 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum1 = _mm_add_pi16(sum1, T20);
+ sum1 = _mm_add_pi16(sum1, T21);
+ sum1 = _mm_add_pi16(sum1, T22);
+ sum1 = _mm_add_pi16(sum1, T23);
+ sum1 = _mm_add_pi16(sum1, T24);
+ sum1 = _mm_add_pi16(sum1, T25);
+ sum1 = _mm_add_pi16(sum1, T26);
+ sum1 = _mm_add_pi16(sum1, T27);
+
+ T10 = (*(__m64*)(fref2 + 8 * frefstride));
+ T11 = (*(__m64*)(fref2 + 9 * frefstride));
+ T12 = (*(__m64*)(fref2 + 10 * frefstride));
+ T13 = (*(__m64*)(fref2 + 11 * frefstride));
+ T14 = (*(__m64*)(fref2 + 12 * frefstride));
+ T15 = (*(__m64*)(fref2 + 13 * frefstride));
+ T16 = (*(__m64*)(fref2 + 14 * frefstride));
+ T17 = (*(__m64*)(fref2 + 15 * frefstride));
+
+ T20 = _mm_sad_pu8(T0, T10);
+ T21 = _mm_sad_pu8(T1, T11);
+ T22 = _mm_sad_pu8(T2, T12);
+ T23 = _mm_sad_pu8(T3, T13);
+ T24 = _mm_sad_pu8(T4, T14);
+ T25 = _mm_sad_pu8(T5, T15);
+ T26 = _mm_sad_pu8(T6, T16);
+ T27 = _mm_sad_pu8(T7, T17);
+
+ sum1 = _mm_add_pi16(sum1, T20);
+ sum1 = _mm_add_pi16(sum1, T21);
+ sum1 = _mm_add_pi16(sum1, T22);
+ sum1 = _mm_add_pi16(sum1, T23);
+ sum1 = _mm_add_pi16(sum1, T24);
+ sum1 = _mm_add_pi16(sum1, T25);
+ sum1 = _mm_add_pi16(sum1, T26);
+ sum1 = _mm_add_pi16(sum1, T27);
+ res[1] = _m_to_int(sum1);
+
+ T10 = (*(__m64*)(fref3 + 0 * frefstride));
+ T11 = (*(__m64*)(fref3 + 1 * frefstride));
+ T12 = (*(__m64*)(fref3 + 2 * frefstride));
+ T13 = (*(__m64*)(fref3 + 3 * frefstride));
+ T14 = (*(__m64*)(fref3 + 4 * frefstride));
+ T15 = (*(__m64*)(fref3 + 5 * frefstride));
+ T16 = (*(__m64*)(fref3 + 6 * frefstride));
+ T17 = (*(__m64*)(fref3 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum2 = _mm_add_pi16(sum2, T20);
+ sum2 = _mm_add_pi16(sum2, T21);
+ sum2 = _mm_add_pi16(sum2, T22);
+ sum2 = _mm_add_pi16(sum2, T23);
+ sum2 = _mm_add_pi16(sum2, T24);
+ sum2 = _mm_add_pi16(sum2, T25);
+ sum2 = _mm_add_pi16(sum2, T26);
+ sum2 = _mm_add_pi16(sum2, T27);
+
+ T10 = (*(__m64*)(fref3 + 8 * frefstride));
+ T11 = (*(__m64*)(fref3 + 9 * frefstride));
+ T12 = (*(__m64*)(fref3 + 10 * frefstride));
+ T13 = (*(__m64*)(fref3 + 11 * frefstride));
+ T14 = (*(__m64*)(fref3 + 12 * frefstride));
+ T15 = (*(__m64*)(fref3 + 13 * frefstride));
+ T16 = (*(__m64*)(fref3 + 14 * frefstride));
+ T17 = (*(__m64*)(fref3 + 15 * frefstride));
+
+ T20 = _mm_sad_pu8(T0, T10);
+ T21 = _mm_sad_pu8(T1, T11);
+ T22 = _mm_sad_pu8(T2, T12);
+ T23 = _mm_sad_pu8(T3, T13);
+ T24 = _mm_sad_pu8(T4, T14);
+ T25 = _mm_sad_pu8(T5, T15);
+ T26 = _mm_sad_pu8(T6, T16);
+ T27 = _mm_sad_pu8(T7, T17);
+
+ sum2 = _mm_add_pi16(sum2, T20);
+ sum2 = _mm_add_pi16(sum2, T21);
+ sum2 = _mm_add_pi16(sum2, T22);
+ sum2 = _mm_add_pi16(sum2, T23);
+ sum2 = _mm_add_pi16(sum2, T24);
+ sum2 = _mm_add_pi16(sum2, T25);
+ sum2 = _mm_add_pi16(sum2, T26);
+ sum2 = _mm_add_pi16(sum2, T27);
+ res[2] = _m_to_int(sum2);
+ }
+ else if (8 == ly)
+ {
+ T00 = (*(__m64*)(fenc + 0 * FENC_STRIDE));
+ T01 = (*(__m64*)(fenc + 1 * FENC_STRIDE));
+ T02 = (*(__m64*)(fenc + 2 * FENC_STRIDE));
+ T03 = (*(__m64*)(fenc + 3 * FENC_STRIDE));
+ T04 = (*(__m64*)(fenc + 4 * FENC_STRIDE));
+ T05 = (*(__m64*)(fenc + 5 * FENC_STRIDE));
+ T06 = (*(__m64*)(fenc + 6 * FENC_STRIDE));
+ T07 = (*(__m64*)(fenc + 7 * FENC_STRIDE));
+
+ T10 = (*(__m64*)(fref1 + 0 * frefstride));
+ T11 = (*(__m64*)(fref1 + 1 * frefstride));
+ T12 = (*(__m64*)(fref1 + 2 * frefstride));
+ T13 = (*(__m64*)(fref1 + 3 * frefstride));
+ T14 = (*(__m64*)(fref1 + 4 * frefstride));
+ T15 = (*(__m64*)(fref1 + 5 * frefstride));
+ T16 = (*(__m64*)(fref1 + 6 * frefstride));
+ T17 = (*(__m64*)(fref1 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum0 = _mm_add_pi16(sum0, T20);
+ sum0 = _mm_add_pi16(sum0, T21);
+ sum0 = _mm_add_pi16(sum0, T22);
+ sum0 = _mm_add_pi16(sum0, T23);
+ sum0 = _mm_add_pi16(sum0, T24);
+ sum0 = _mm_add_pi16(sum0, T25);
+ sum0 = _mm_add_pi16(sum0, T26);
+ sum0 = _mm_add_pi16(sum0, T27);
+ res[0] = _m_to_int(sum0);
+
+ T10 = (*(__m64*)(fref2 + 0 * frefstride));
+ T11 = (*(__m64*)(fref2 + 1 * frefstride));
+ T12 = (*(__m64*)(fref2 + 2 * frefstride));
+ T13 = (*(__m64*)(fref2 + 3 * frefstride));
+ T14 = (*(__m64*)(fref2 + 4 * frefstride));
+ T15 = (*(__m64*)(fref2 + 5 * frefstride));
+ T16 = (*(__m64*)(fref2 + 6 * frefstride));
+ T17 = (*(__m64*)(fref2 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum1 = _mm_add_pi16(sum1, T20);
+ sum1 = _mm_add_pi16(sum1, T21);
+ sum1 = _mm_add_pi16(sum1, T22);
+ sum1 = _mm_add_pi16(sum1, T23);
+ sum1 = _mm_add_pi16(sum1, T24);
+ sum1 = _mm_add_pi16(sum1, T25);
+ sum1 = _mm_add_pi16(sum1, T26);
+ sum1 = _mm_add_pi16(sum1, T27);
+ res[1] = _m_to_int(sum1);
+
+ T10 = (*(__m64*)(fref3 + 0 * frefstride));
+ T11 = (*(__m64*)(fref3 + 1 * frefstride));
+ T12 = (*(__m64*)(fref3 + 2 * frefstride));
+ T13 = (*(__m64*)(fref3 + 3 * frefstride));
+ T14 = (*(__m64*)(fref3 + 4 * frefstride));
+ T15 = (*(__m64*)(fref3 + 5 * frefstride));
+ T16 = (*(__m64*)(fref3 + 6 * frefstride));
+ T17 = (*(__m64*)(fref3 + 7 * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+ T24 = _mm_sad_pu8(T04, T14);
+ T25 = _mm_sad_pu8(T05, T15);
+ T26 = _mm_sad_pu8(T06, T16);
+ T27 = _mm_sad_pu8(T07, T17);
+
+ sum2 = _mm_add_pi16(sum2, T20);
+ sum2 = _mm_add_pi16(sum2, T21);
+ sum2 = _mm_add_pi16(sum2, T22);
+ sum2 = _mm_add_pi16(sum2, T23);
+ sum2 = _mm_add_pi16(sum2, T24);
+ sum2 = _mm_add_pi16(sum2, T25);
+ sum2 = _mm_add_pi16(sum2, T26);
+ sum2 = _mm_add_pi16(sum2, T27);
+ res[2] = _m_to_int(sum2);
+ }
+ else
+ {
+ for (int i = 0; i < ly; i += 4)
+ {
+ T00 = (*(__m64*)(fenc + (i + 0) * FENC_STRIDE));
+ T01 = (*(__m64*)(fenc + (i + 1) * FENC_STRIDE));
+ T02 = (*(__m64*)(fenc + (i + 2) * FENC_STRIDE));
+ T03 = (*(__m64*)(fenc + (i + 3) * FENC_STRIDE));
+
+ T10 = (*(__m64*)(fref1 + (i + 0) * frefstride));
+ T11 = (*(__m64*)(fref1 + (i + 1) * frefstride));
+ T12 = (*(__m64*)(fref1 + (i + 2) * frefstride));
+ T13 = (*(__m64*)(fref1 + (i + 3) * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+
+ sum0 = _mm_add_pi16(sum0, T20);
+ sum0 = _mm_add_pi16(sum0, T21);
+ sum0 = _mm_add_pi16(sum0, T22);
+ sum0 = _mm_add_pi16(sum0, T23);
+ res[0] = _m_to_int(sum0);
+
+ T10 = (*(__m64*)(fref2 + (i + 0) * frefstride));
+ T11 = (*(__m64*)(fref2 + (i + 1) * frefstride));
+ T12 = (*(__m64*)(fref2 + (i + 2) * frefstride));
+ T13 = (*(__m64*)(fref2 + (i + 3) * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+
+ sum1 = _mm_add_pi16(sum1, T20);
+ sum1 = _mm_add_pi16(sum1, T21);
+ sum1 = _mm_add_pi16(sum1, T22);
+ sum1 = _mm_add_pi16(sum1, T23);
+ res[1] = _m_to_int(sum1);
+
+ T10 = (*(__m64*)(fref3 + (i + 0) * frefstride));
+ T11 = (*(__m64*)(fref3 + (i + 1) * frefstride));
+ T12 = (*(__m64*)(fref3 + (i + 2) * frefstride));
+ T13 = (*(__m64*)(fref3 + (i + 3) * frefstride));
+
+ T20 = _mm_sad_pu8(T00, T10);
+ T21 = _mm_sad_pu8(T01, T11);
+ T22 = _mm_sad_pu8(T02, T12);
+ T23 = _mm_sad_pu8(T03, T13);
+
+ sum2 = _mm_add_pi16(sum2, T20);
+ sum2 = _mm_add_pi16(sum2, T21);
+ sum2 = _mm_add_pi16(sum2, T22);
+ sum2 = _mm_add_pi16(sum2, T23);
+ res[2] = _m_to_int(sum2);
+ }
+ }
+}
+
+#endif /* ifdef x86_64 */
#endif /* if INSTRSET >= X265_CPU_LEVEL_SSE41 */
/* For performance - This function assumes that the *last load* can access 16 elements. */
More information about the x265-devel
mailing list