sl@0: /*
sl@0:  * Copyright (c) 2005
sl@0:  *	Eric Anholt.  All rights reserved.
sl@0:  *
sl@0:  * Redistribution and use in source and binary forms, with or without
sl@0:  * modification, are permitted provided that the following conditions
sl@0:  * are met:
sl@0:  * 1. Redistributions of source code must retain the above copyright
sl@0:  *    notice, this list of conditions and the following disclaimer.
sl@0:  * 2. Redistributions in binary form must reproduce the above copyright
sl@0:  *    notice, this list of conditions and the following disclaimer in the
sl@0:  *    documentation and/or other materials provided with the distribution.
sl@0:  *
sl@0:  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND
sl@0:  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
sl@0:  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
sl@0:  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE
sl@0:  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
sl@0:  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
sl@0:  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
sl@0:  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
sl@0:  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
sl@0:  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
sl@0:  * SUCH DAMAGE.
sl@0:  */
sl@0: 
sl@0: #ifdef HAVE_CONFIG_H
sl@0: #include "config.h"
sl@0: #endif
sl@0: #include <liboil/liboilclasses.h>
sl@0: #include <liboil/liboilfunction.h>
sl@0: #include <emmintrin.h>
sl@0: 
sl@0: #define SSE_FUNCTION __attribute__((force_align_arg_pointer))
sl@0: 
sl@0: #ifdef ENABLE_BROKEN_IMPLS
sl@0: union m128_int {
sl@0:   __m128i m128;
sl@0:   uint32_t i[4];
sl@0:   uint16_t s[8];
sl@0: };
sl@0: 
sl@0: SSE_FUNCTION static void
sl@0: sad8x8_u8_sse (uint32_t *dest, uint8_t *src1, int sstr1, uint8_t *src2,
sl@0:     int sstr2)
sl@0: {
sl@0:   int i;
sl@0:   __m128i sum = _mm_setzero_si128();
sl@0:   union m128_int sumi;
sl@0: 
sl@0:   for (i = 0; i < 4; i++) {
sl@0:     __m128i xmm0, xmm1, xmm2, xmm3;
sl@0:     xmm0 = _mm_loadl_epi64((__m128i *)src1);
sl@0:     xmm1 = _mm_loadl_epi64((__m128i *)(src1 + sstr1));
sl@0:     xmm2 = _mm_loadl_epi64((__m128i *)src2);
sl@0:     xmm3 = _mm_loadl_epi64((__m128i *)(src2 + sstr2));
sl@0:     xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
sl@0:     xmm2 = _mm_unpacklo_epi8(xmm2, xmm3);
sl@0:     sum = _mm_add_epi64(sum, _mm_sad_epu8(xmm0, xmm2));
sl@0:     src1 += 2 * sstr1;
sl@0:     src2 += 2 * sstr2;
sl@0:   }
sl@0:   sumi.m128 = sum;
sl@0:   *dest = sumi.i[0] + sumi.i[2];
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL_WRAPPER(sad8x8_u8_sse, sad8x8_u8, OIL_IMPL_FLAG_SSE2);
sl@0: #endif
sl@0: