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: //Portions Copyright (c)  2008-2009 Nokia Corporation and/or its subsidiary(-ies). All rights reserved. 
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: 
sl@0: #ifdef HAVE_GCC_ASM      
sl@0: #include <mmintrin.h>
sl@0: #endif
sl@0: 
sl@0: #ifdef ENABLE_BROKEN_IMPLS
sl@0: union m64_int {
sl@0:   __m64 m64;
sl@0:   uint64_t ull;
sl@0: };
sl@0: 
sl@0: static const struct _MMXData {
sl@0:   union m64_int mmx_4x0080;
sl@0: } c = {
sl@0:     .mmx_4x0080.ull =	0x0080008000800080ULL,
sl@0: };
sl@0: 
sl@0: #define MC(x) (c.mmx_##x.m64)
sl@0: 
sl@0: static void
sl@0: recon8x8_intra_mmx (uint8_t *dest, int ds, int16_t *change)
sl@0: {
sl@0:   int i;
sl@0:   __m64 offset = MC(4x0080);
sl@0: 
sl@0:   for (i = 8; i; i--) {
sl@0:     __m64 mm0, mm1, c0, c1;
sl@0:     c0 = ((__m64 *)change)[0];
sl@0:     c1 = ((__m64 *)change)[1];
sl@0:     mm0 = _mm_adds_pi16(c0, offset);
sl@0:     mm1 = _mm_adds_pi16(c1, offset);
sl@0:     *(__m64 *)dest = _mm_packs_pu16(mm0, mm1);
sl@0: 
sl@0:     dest += ds;
sl@0:     change += 8;
sl@0:   }
sl@0:   _mm_empty();
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (recon8x8_intra_mmx, recon8x8_intra, OIL_IMPL_FLAG_MMX);
sl@0: #endif
sl@0: 
sl@0: static void
sl@0: recon8x8_inter_mmx (uint8_t *dest, int ds, uint8_t *src, int ss,
sl@0:     int16_t *change, int dss)
sl@0: {
sl@0: #ifdef HAVE_GCC_ASM          
sl@0:   int i;
sl@0: 
sl@0:   for (i = 8; i; i--) {
sl@0:     __m64 mm0, mm1, c0, c1;
sl@0:     c0 = ((__m64 *)change)[0];
sl@0:     c1 = ((__m64 *)change)[1];
sl@0:     mm0 = _mm_unpacklo_pi8(*(__m64 *)src, _mm_setzero_si64());
sl@0:     mm1 = _mm_unpackhi_pi8(*(__m64 *)src, _mm_setzero_si64());
sl@0:     mm0 = _mm_adds_pi16(mm0, c0);
sl@0:     mm1 = _mm_adds_pi16(mm1, c1);
sl@0:     *(__m64 *)dest = _mm_packs_pu16(mm0, mm1);
sl@0: 
sl@0:     change += 8;
sl@0:     dest += ds;
sl@0:     src += ss;
sl@0:   }
sl@0:   _mm_empty();
sl@0: #endif  
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (recon8x8_inter_mmx, recon8x8_inter, OIL_IMPL_FLAG_MMX);
sl@0: 
sl@0: static void
sl@0: recon8x8_inter2_mmx (uint8_t *dest, int ds, uint8_t *s1, int ss1, uint8_t *s2,
sl@0:     int ss2, int16_t *change)
sl@0: {
sl@0:   int i;
sl@0: #ifdef HAVE_GCC_ASM   
sl@0:   for (i = 8; i; i--) {
sl@0:     __m64 mm0, mm1, c0, c1;
sl@0:     mm0 = _mm_adds_pu16(
sl@0: 	_mm_unpacklo_pi8(*(__m64 *)s1, _mm_setzero_si64()),
sl@0: 	_mm_unpacklo_pi8(*(__m64 *)s2, _mm_setzero_si64()));
sl@0:     mm1 = _mm_adds_pu16(
sl@0: 	_mm_unpackhi_pi8(*(__m64 *)s1, _mm_setzero_si64()),
sl@0: 	_mm_unpackhi_pi8(*(__m64 *)s2, _mm_setzero_si64()));
sl@0:     c0 = ((__m64 *)change)[0];
sl@0:     c1 = ((__m64 *)change)[1];
sl@0:     mm0 = _mm_srli_pi16(mm0, 1);
sl@0:     mm1 = _mm_srli_pi16(mm1, 1);
sl@0:     mm0 = _mm_adds_pi16(mm0, c0);
sl@0:     mm1 = _mm_adds_pi16(mm1, c1);
sl@0:     *(__m64 *)dest = _mm_packs_pu16(mm0, mm1);
sl@0:     change += 8;
sl@0:     dest += ds;
sl@0:     s1 += ss1;
sl@0:     s2 += ss2;
sl@0:   }
sl@0:   _mm_empty();
sl@0: #endif
sl@0: }
sl@0: OIL_DEFINE_IMPL_FULL (recon8x8_inter2_mmx, recon8x8_inter2, OIL_IMPL_FLAG_MMX);
sl@0: 
sl@0: #if 0
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_recon8x8_intra_mmx() {
sl@0: 		return &_oil_function_impl_recon8x8_intra_mmx;
sl@0: }
sl@0: #endif
sl@0: #endif
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_recon8x8_inter_mmx() {
sl@0: 		return &_oil_function_impl_recon8x8_inter_mmx;
sl@0: }
sl@0: #endif
sl@0: 
sl@0: #ifdef	__SYMBIAN32__
sl@0:  
sl@0: OilFunctionImpl* __oil_function_impl_recon8x8_inter2_mmx() {
sl@0: 		return &_oil_function_impl_recon8x8_inter2_mmx;
sl@0: }
sl@0: #endif
sl@0: