sad_inline.h revision 29a84457aed4c45bc900998b5e11c03023264208
129a84457aed4c45bc900998b5e11c03023264208James Dong/* ------------------------------------------------------------------ 229a84457aed4c45bc900998b5e11c03023264208James Dong * Copyright (C) 1998-2009 PacketVideo 329a84457aed4c45bc900998b5e11c03023264208James Dong * 429a84457aed4c45bc900998b5e11c03023264208James Dong * Licensed under the Apache License, Version 2.0 (the "License"); 529a84457aed4c45bc900998b5e11c03023264208James Dong * you may not use this file except in compliance with the License. 629a84457aed4c45bc900998b5e11c03023264208James Dong * You may obtain a copy of the License at 729a84457aed4c45bc900998b5e11c03023264208James Dong * 829a84457aed4c45bc900998b5e11c03023264208James Dong * http://www.apache.org/licenses/LICENSE-2.0 929a84457aed4c45bc900998b5e11c03023264208James Dong * 1029a84457aed4c45bc900998b5e11c03023264208James Dong * Unless required by applicable law or agreed to in writing, software 1129a84457aed4c45bc900998b5e11c03023264208James Dong * distributed under the License is distributed on an "AS IS" BASIS, 1229a84457aed4c45bc900998b5e11c03023264208James Dong * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either 1329a84457aed4c45bc900998b5e11c03023264208James Dong * express or implied. 1429a84457aed4c45bc900998b5e11c03023264208James Dong * See the License for the specific language governing permissions 1529a84457aed4c45bc900998b5e11c03023264208James Dong * and limitations under the License. 1629a84457aed4c45bc900998b5e11c03023264208James Dong * ------------------------------------------------------------------- 1729a84457aed4c45bc900998b5e11c03023264208James Dong */ 1829a84457aed4c45bc900998b5e11c03023264208James Dong#ifndef _SAD_INLINE_H_ 1929a84457aed4c45bc900998b5e11c03023264208James Dong#define _SAD_INLINE_H_ 2029a84457aed4c45bc900998b5e11c03023264208James Dong 2129a84457aed4c45bc900998b5e11c03023264208James Dong#ifdef __cplusplus 2229a84457aed4c45bc900998b5e11c03023264208James Dongextern "C" 2329a84457aed4c45bc900998b5e11c03023264208James Dong{ 2429a84457aed4c45bc900998b5e11c03023264208James Dong#endif 2529a84457aed4c45bc900998b5e11c03023264208James Dong 2629a84457aed4c45bc900998b5e11c03023264208James Dong#if defined(__GNUC__) && defined(__arm__) /* ARM GNU COMPILER */ 2729a84457aed4c45bc900998b5e11c03023264208James Dong 2829a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 SUB_SAD(int32 sad, int32 tmp, int32 tmp2) 2929a84457aed4c45bc900998b5e11c03023264208James Dong { 3029a84457aed4c45bc900998b5e11c03023264208James Dong tmp = tmp - tmp2; 3129a84457aed4c45bc900998b5e11c03023264208James Dong if (tmp > 0) sad += tmp; 3229a84457aed4c45bc900998b5e11c03023264208James Dong else sad -= tmp; 3329a84457aed4c45bc900998b5e11c03023264208James Dong 3429a84457aed4c45bc900998b5e11c03023264208James Dong return sad; 3529a84457aed4c45bc900998b5e11c03023264208James Dong } 3629a84457aed4c45bc900998b5e11c03023264208James Dong 3729a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 sad_4pixel(int32 src1, int32 src2, int32 mask) 3829a84457aed4c45bc900998b5e11c03023264208James Dong { 3929a84457aed4c45bc900998b5e11c03023264208James Dong int32 x7; 4029a84457aed4c45bc900998b5e11c03023264208James Dong 4129a84457aed4c45bc900998b5e11c03023264208James Dong x7 = src2 ^ src1; /* check odd/even combination */ 4229a84457aed4c45bc900998b5e11c03023264208James Dong if ((uint32)src2 >= (uint32)src1) 4329a84457aed4c45bc900998b5e11c03023264208James Dong { 4429a84457aed4c45bc900998b5e11c03023264208James Dong src1 = src2 - src1; /* subs */ 4529a84457aed4c45bc900998b5e11c03023264208James Dong } 4629a84457aed4c45bc900998b5e11c03023264208James Dong else 4729a84457aed4c45bc900998b5e11c03023264208James Dong { 4829a84457aed4c45bc900998b5e11c03023264208James Dong src1 = src1 - src2; 4929a84457aed4c45bc900998b5e11c03023264208James Dong } 5029a84457aed4c45bc900998b5e11c03023264208James Dong x7 = x7 ^ src1; /* only odd bytes need to add carry */ 5129a84457aed4c45bc900998b5e11c03023264208James Dong x7 = mask & ((uint32)x7 >> 1); 5229a84457aed4c45bc900998b5e11c03023264208James Dong x7 = (x7 << 8) - x7; 5329a84457aed4c45bc900998b5e11c03023264208James Dong src1 = src1 + (x7 >> 7); /* add 0xFF to the negative byte, add back carry */ 5429a84457aed4c45bc900998b5e11c03023264208James Dong src1 = src1 ^(x7 >> 7); /* take absolute value of negative byte */ 5529a84457aed4c45bc900998b5e11c03023264208James Dong 5629a84457aed4c45bc900998b5e11c03023264208James Dong return src1; 5729a84457aed4c45bc900998b5e11c03023264208James Dong } 5829a84457aed4c45bc900998b5e11c03023264208James Dong 5929a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 3 6029a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 24 6129a84457aed4c45bc900998b5e11c03023264208James Dong 6229a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 6329a84457aed4c45bc900998b5e11c03023264208James Dong 6429a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 6529a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 2 6629a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 6729a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 16 6829a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 6929a84457aed4c45bc900998b5e11c03023264208James Dong 7029a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 7129a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 1 7229a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 7329a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 8 7429a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 7529a84457aed4c45bc900998b5e11c03023264208James Dong 7629a84457aed4c45bc900998b5e11c03023264208James Dong 7729a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 simd_sad_mb(uint8 *ref, uint8 *blk, int dmin, int lx) 7829a84457aed4c45bc900998b5e11c03023264208James Dong { 7929a84457aed4c45bc900998b5e11c03023264208James Dong int32 x4, x5, x6, x8, x9, x10, x11, x12, x14; 8029a84457aed4c45bc900998b5e11c03023264208James Dong 8129a84457aed4c45bc900998b5e11c03023264208James Dong x9 = 0x80808080; /* const. */ 8229a84457aed4c45bc900998b5e11c03023264208James Dong 8329a84457aed4c45bc900998b5e11c03023264208James Dong x8 = (uint32)ref & 0x3; 8429a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 3) 8529a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset3; 8629a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 2) 8729a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset2; 8829a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 1) 8929a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset1; 9029a84457aed4c45bc900998b5e11c03023264208James Dong 9129a84457aed4c45bc900998b5e11c03023264208James Dong// x5 = (x4<<8)-x4; /* x5 = x4*255; */ 9229a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x5 = 0; 9329a84457aed4c45bc900998b5e11c03023264208James Dong 9429a84457aed4c45bc900998b5e11c03023264208James Dong x6 = 0xFFFF00FF; 9529a84457aed4c45bc900998b5e11c03023264208James Dong 9629a84457aed4c45bc900998b5e11c03023264208James Dong ref -= lx; 9729a84457aed4c45bc900998b5e11c03023264208James Dong blk -= 16; 9829a84457aed4c45bc900998b5e11c03023264208James Dong 9929a84457aed4c45bc900998b5e11c03023264208James Dong x8 = 16; 10029a84457aed4c45bc900998b5e11c03023264208James Dong 10129a84457aed4c45bc900998b5e11c03023264208James DongLOOP_SAD0: 10229a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 10329a84457aed4c45bc900998b5e11c03023264208James Dong x10 = *((uint32*)(ref += lx)); 10429a84457aed4c45bc900998b5e11c03023264208James Dong x11 = *((uint32*)(ref + 4)); 10529a84457aed4c45bc900998b5e11c03023264208James Dong x12 = *((uint32*)(blk += 16)); 10629a84457aed4c45bc900998b5e11c03023264208James Dong x14 = *((uint32*)(blk + 4)); 10729a84457aed4c45bc900998b5e11c03023264208James Dong 10829a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 10929a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 11029a84457aed4c45bc900998b5e11c03023264208James Dong 11129a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 11229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 11329a84457aed4c45bc900998b5e11c03023264208James Dong 11429a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 11529a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 11629a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 11729a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 11829a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 11929a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 12029a84457aed4c45bc900998b5e11c03023264208James Dong 12129a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 12229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = *((uint32*)(ref + 8)); 12329a84457aed4c45bc900998b5e11c03023264208James Dong x11 = *((uint32*)(ref + 12)); 12429a84457aed4c45bc900998b5e11c03023264208James Dong x12 = *((uint32*)(blk + 8)); 12529a84457aed4c45bc900998b5e11c03023264208James Dong x14 = *((uint32*)(blk + 12)); 12629a84457aed4c45bc900998b5e11c03023264208James Dong 12729a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 12829a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 12929a84457aed4c45bc900998b5e11c03023264208James Dong 13029a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 13129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 13229a84457aed4c45bc900998b5e11c03023264208James Dong 13329a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 13429a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 13529a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 13629a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 13729a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 13829a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 13929a84457aed4c45bc900998b5e11c03023264208James Dong 14029a84457aed4c45bc900998b5e11c03023264208James Dong /****************/ 14129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x5 - (x4 << 8); /* extract low bytes */ 14229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + x4; /* add with high bytes */ 14329a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + (x10 << 16); /* add with lower half word */ 14429a84457aed4c45bc900998b5e11c03023264208James Dong 14529a84457aed4c45bc900998b5e11c03023264208James Dong if ((int)((uint32)x10 >> 16) <= dmin) /* compare with dmin */ 14629a84457aed4c45bc900998b5e11c03023264208James Dong { 14729a84457aed4c45bc900998b5e11c03023264208James Dong if (--x8) 14829a84457aed4c45bc900998b5e11c03023264208James Dong { 14929a84457aed4c45bc900998b5e11c03023264208James Dong goto LOOP_SAD0; 15029a84457aed4c45bc900998b5e11c03023264208James Dong } 15129a84457aed4c45bc900998b5e11c03023264208James Dong 15229a84457aed4c45bc900998b5e11c03023264208James Dong } 15329a84457aed4c45bc900998b5e11c03023264208James Dong 15429a84457aed4c45bc900998b5e11c03023264208James Dong return ((uint32)x10 >> 16); 15529a84457aed4c45bc900998b5e11c03023264208James Dong 15629a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset3: 15729a84457aed4c45bc900998b5e11c03023264208James Dong 15829a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset3(ref, blk, lx, dmin); 15929a84457aed4c45bc900998b5e11c03023264208James Dong 16029a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset2: 16129a84457aed4c45bc900998b5e11c03023264208James Dong 16229a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset2(ref, blk, lx, dmin); 16329a84457aed4c45bc900998b5e11c03023264208James Dong 16429a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset1: 16529a84457aed4c45bc900998b5e11c03023264208James Dong 16629a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset1(ref, blk, lx, dmin); 16729a84457aed4c45bc900998b5e11c03023264208James Dong 16829a84457aed4c45bc900998b5e11c03023264208James Dong } 16929a84457aed4c45bc900998b5e11c03023264208James Dong 17029a84457aed4c45bc900998b5e11c03023264208James Dong#elif defined(__CC_ARM) /* only work with arm v5 */ 17129a84457aed4c45bc900998b5e11c03023264208James Dong 17229a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 SUB_SAD(int32 sad, int32 tmp, int32 tmp2) 17329a84457aed4c45bc900998b5e11c03023264208James Dong { 17429a84457aed4c45bc900998b5e11c03023264208James Dong __asm 17529a84457aed4c45bc900998b5e11c03023264208James Dong { 17629a84457aed4c45bc900998b5e11c03023264208James Dong rsbs tmp, tmp, tmp2 ; 17729a84457aed4c45bc900998b5e11c03023264208James Dong rsbmi tmp, tmp, #0 ; 17829a84457aed4c45bc900998b5e11c03023264208James Dong add sad, sad, tmp ; 17929a84457aed4c45bc900998b5e11c03023264208James Dong } 18029a84457aed4c45bc900998b5e11c03023264208James Dong 18129a84457aed4c45bc900998b5e11c03023264208James Dong return sad; 18229a84457aed4c45bc900998b5e11c03023264208James Dong } 18329a84457aed4c45bc900998b5e11c03023264208James Dong 18429a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 sad_4pixel(int32 src1, int32 src2, int32 mask) 18529a84457aed4c45bc900998b5e11c03023264208James Dong { 18629a84457aed4c45bc900998b5e11c03023264208James Dong int32 x7; 18729a84457aed4c45bc900998b5e11c03023264208James Dong 18829a84457aed4c45bc900998b5e11c03023264208James Dong __asm 18929a84457aed4c45bc900998b5e11c03023264208James Dong { 19029a84457aed4c45bc900998b5e11c03023264208James Dong EOR x7, src2, src1; /* check odd/even combination */ 19129a84457aed4c45bc900998b5e11c03023264208James Dong SUBS src1, src2, src1; 19229a84457aed4c45bc900998b5e11c03023264208James Dong EOR x7, x7, src1; 19329a84457aed4c45bc900998b5e11c03023264208James Dong AND x7, mask, x7, lsr #1; 19429a84457aed4c45bc900998b5e11c03023264208James Dong ORRCC x7, x7, #0x80000000; 19529a84457aed4c45bc900998b5e11c03023264208James Dong RSB x7, x7, x7, lsl #8; 19629a84457aed4c45bc900998b5e11c03023264208James Dong ADD src1, src1, x7, asr #7; /* add 0xFF to the negative byte, add back carry */ 19729a84457aed4c45bc900998b5e11c03023264208James Dong EOR src1, src1, x7, asr #7; /* take absolute value of negative byte */ 19829a84457aed4c45bc900998b5e11c03023264208James Dong } 19929a84457aed4c45bc900998b5e11c03023264208James Dong 20029a84457aed4c45bc900998b5e11c03023264208James Dong return src1; 20129a84457aed4c45bc900998b5e11c03023264208James Dong } 20229a84457aed4c45bc900998b5e11c03023264208James Dong 20329a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 sad_4pixelN(int32 src1, int32 src2, int32 mask) 20429a84457aed4c45bc900998b5e11c03023264208James Dong { 20529a84457aed4c45bc900998b5e11c03023264208James Dong int32 x7; 20629a84457aed4c45bc900998b5e11c03023264208James Dong 20729a84457aed4c45bc900998b5e11c03023264208James Dong __asm 20829a84457aed4c45bc900998b5e11c03023264208James Dong { 20929a84457aed4c45bc900998b5e11c03023264208James Dong EOR x7, src2, src1; /* check odd/even combination */ 21029a84457aed4c45bc900998b5e11c03023264208James Dong ADDS src1, src2, src1; 21129a84457aed4c45bc900998b5e11c03023264208James Dong EOR x7, x7, src1; /* only odd bytes need to add carry */ 21229a84457aed4c45bc900998b5e11c03023264208James Dong ANDS x7, mask, x7, rrx; 21329a84457aed4c45bc900998b5e11c03023264208James Dong RSB x7, x7, x7, lsl #8; 21429a84457aed4c45bc900998b5e11c03023264208James Dong SUB src1, src1, x7, asr #7; /* add 0xFF to the negative byte, add back carry */ 21529a84457aed4c45bc900998b5e11c03023264208James Dong EOR src1, src1, x7, asr #7; /* take absolute value of negative byte */ 21629a84457aed4c45bc900998b5e11c03023264208James Dong } 21729a84457aed4c45bc900998b5e11c03023264208James Dong 21829a84457aed4c45bc900998b5e11c03023264208James Dong return src1; 21929a84457aed4c45bc900998b5e11c03023264208James Dong } 22029a84457aed4c45bc900998b5e11c03023264208James Dong 22129a84457aed4c45bc900998b5e11c03023264208James Dong#define sum_accumulate __asm{ SBC x5, x5, x10; /* accumulate low bytes */ \ 22229a84457aed4c45bc900998b5e11c03023264208James Dong BIC x10, x6, x10; /* x10 & 0xFF00FF00 */ \ 22329a84457aed4c45bc900998b5e11c03023264208James Dong ADD x4, x4, x10,lsr #8; /* accumulate high bytes */ \ 22429a84457aed4c45bc900998b5e11c03023264208James Dong SBC x5, x5, x11; /* accumulate low bytes */ \ 22529a84457aed4c45bc900998b5e11c03023264208James Dong BIC x11, x6, x11; /* x11 & 0xFF00FF00 */ \ 22629a84457aed4c45bc900998b5e11c03023264208James Dong ADD x4, x4, x11,lsr #8; } /* accumulate high bytes */ 22729a84457aed4c45bc900998b5e11c03023264208James Dong 22829a84457aed4c45bc900998b5e11c03023264208James Dong 22929a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 3 23029a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 24 23129a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x08000001 23229a84457aed4c45bc900998b5e11c03023264208James Dong 23329a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 23429a84457aed4c45bc900998b5e11c03023264208James Dong 23529a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 23629a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 2 23729a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 23829a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 16 23929a84457aed4c45bc900998b5e11c03023264208James Dong#undef INC_X8 24029a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x10000001 24129a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 24229a84457aed4c45bc900998b5e11c03023264208James Dong 24329a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 24429a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 1 24529a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 24629a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 8 24729a84457aed4c45bc900998b5e11c03023264208James Dong#undef INC_X8 24829a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x08000001 24929a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 25029a84457aed4c45bc900998b5e11c03023264208James Dong 25129a84457aed4c45bc900998b5e11c03023264208James Dong 25229a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 simd_sad_mb(uint8 *ref, uint8 *blk, int dmin, int lx) 25329a84457aed4c45bc900998b5e11c03023264208James Dong { 25429a84457aed4c45bc900998b5e11c03023264208James Dong int32 x4, x5, x6, x8, x9, x10, x11, x12, x14; 25529a84457aed4c45bc900998b5e11c03023264208James Dong 25629a84457aed4c45bc900998b5e11c03023264208James Dong x9 = 0x80808080; /* const. */ 25729a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x5 = 0; 25829a84457aed4c45bc900998b5e11c03023264208James Dong 25929a84457aed4c45bc900998b5e11c03023264208James Dong __asm 26029a84457aed4c45bc900998b5e11c03023264208James Dong { 26129a84457aed4c45bc900998b5e11c03023264208James Dong MOVS x8, ref, lsl #31 ; 26229a84457aed4c45bc900998b5e11c03023264208James Dong BHI SadMBOffset3; 26329a84457aed4c45bc900998b5e11c03023264208James Dong BCS SadMBOffset2; 26429a84457aed4c45bc900998b5e11c03023264208James Dong BMI SadMBOffset1; 26529a84457aed4c45bc900998b5e11c03023264208James Dong 26629a84457aed4c45bc900998b5e11c03023264208James Dong MVN x6, #0xFF00; 26729a84457aed4c45bc900998b5e11c03023264208James Dong } 26829a84457aed4c45bc900998b5e11c03023264208James DongLOOP_SAD0: 26929a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 27029a84457aed4c45bc900998b5e11c03023264208James Dong x11 = *((int32*)(ref + 12)); 27129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = *((int32*)(ref + 8)); 27229a84457aed4c45bc900998b5e11c03023264208James Dong x14 = *((int32*)(blk + 12)); 27329a84457aed4c45bc900998b5e11c03023264208James Dong x12 = *((int32*)(blk + 8)); 27429a84457aed4c45bc900998b5e11c03023264208James Dong 27529a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 27629a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 27729a84457aed4c45bc900998b5e11c03023264208James Dong 27829a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 27929a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 28029a84457aed4c45bc900998b5e11c03023264208James Dong 28129a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 28229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 28329a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 28429a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 28529a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 28629a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 28729a84457aed4c45bc900998b5e11c03023264208James Dong 28829a84457aed4c45bc900998b5e11c03023264208James Dong __asm 28929a84457aed4c45bc900998b5e11c03023264208James Dong { 29029a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 29129a84457aed4c45bc900998b5e11c03023264208James Dong LDR x11, [ref, #4]; 29229a84457aed4c45bc900998b5e11c03023264208James Dong LDR x10, [ref], lx ; 29329a84457aed4c45bc900998b5e11c03023264208James Dong LDR x14, [blk, #4]; 29429a84457aed4c45bc900998b5e11c03023264208James Dong LDR x12, [blk], #16 ; 29529a84457aed4c45bc900998b5e11c03023264208James Dong } 29629a84457aed4c45bc900998b5e11c03023264208James Dong 29729a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 29829a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 29929a84457aed4c45bc900998b5e11c03023264208James Dong 30029a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 30129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 30229a84457aed4c45bc900998b5e11c03023264208James Dong 30329a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 30429a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 30529a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 30629a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 30729a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 30829a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 30929a84457aed4c45bc900998b5e11c03023264208James Dong 31029a84457aed4c45bc900998b5e11c03023264208James Dong /****************/ 31129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x5 - (x4 << 8); /* extract low bytes */ 31229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + x4; /* add with high bytes */ 31329a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + (x10 << 16); /* add with lower half word */ 31429a84457aed4c45bc900998b5e11c03023264208James Dong 31529a84457aed4c45bc900998b5e11c03023264208James Dong __asm 31629a84457aed4c45bc900998b5e11c03023264208James Dong { 31729a84457aed4c45bc900998b5e11c03023264208James Dong /****************/ 31829a84457aed4c45bc900998b5e11c03023264208James Dong RSBS x11, dmin, x10, lsr #16; 31929a84457aed4c45bc900998b5e11c03023264208James Dong ADDLSS x8, x8, #0x10000001; 32029a84457aed4c45bc900998b5e11c03023264208James Dong BLS LOOP_SAD0; 32129a84457aed4c45bc900998b5e11c03023264208James Dong } 32229a84457aed4c45bc900998b5e11c03023264208James Dong 32329a84457aed4c45bc900998b5e11c03023264208James Dong return ((uint32)x10 >> 16); 32429a84457aed4c45bc900998b5e11c03023264208James Dong 32529a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset3: 32629a84457aed4c45bc900998b5e11c03023264208James Dong 32729a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset3(ref, blk, lx, dmin, x8); 32829a84457aed4c45bc900998b5e11c03023264208James Dong 32929a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset2: 33029a84457aed4c45bc900998b5e11c03023264208James Dong 33129a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset2(ref, blk, lx, dmin, x8); 33229a84457aed4c45bc900998b5e11c03023264208James Dong 33329a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset1: 33429a84457aed4c45bc900998b5e11c03023264208James Dong 33529a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset1(ref, blk, lx, dmin, x8); 33629a84457aed4c45bc900998b5e11c03023264208James Dong } 33729a84457aed4c45bc900998b5e11c03023264208James Dong 33829a84457aed4c45bc900998b5e11c03023264208James Dong 33929a84457aed4c45bc900998b5e11c03023264208James Dong#elif defined(__GNUC__) && defined(__arm__) /* ARM GNU COMPILER */ 34029a84457aed4c45bc900998b5e11c03023264208James Dong 34129a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 SUB_SAD(int32 sad, int32 tmp, int32 tmp2) 34229a84457aed4c45bc900998b5e11c03023264208James Dong { 34329a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("rsbs %1, %1, %2\n\trsbmi %1, %1, #0\n\tadd %0, %0, %1": "=r"(sad): "r"(tmp), "r"(tmp2)); 34429a84457aed4c45bc900998b5e11c03023264208James Dong return sad; 34529a84457aed4c45bc900998b5e11c03023264208James Dong } 34629a84457aed4c45bc900998b5e11c03023264208James Dong 34729a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 sad_4pixel(int32 src1, int32 src2, int32 mask) 34829a84457aed4c45bc900998b5e11c03023264208James Dong { 34929a84457aed4c45bc900998b5e11c03023264208James Dong int32 x7; 35029a84457aed4c45bc900998b5e11c03023264208James Dong 35129a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("EOR %1, %2, %0\n\tSUBS %0, %2, %0\n\tEOR %1, %1, %0\n\tAND %1, %3, %1, lsr #1\n\tORRCC %1, %1, #0x80000000\n\tRSB %1, %1, %1, lsl #8\n\tADD %0, %0, %1, asr #7\n\tEOR %0, %0, %1, asr #7": "=r"(src1), "=&r"(x7): "r"(src2), "r"(mask)); 35229a84457aed4c45bc900998b5e11c03023264208James Dong 35329a84457aed4c45bc900998b5e11c03023264208James Dong return src1; 35429a84457aed4c45bc900998b5e11c03023264208James Dong } 35529a84457aed4c45bc900998b5e11c03023264208James Dong 35629a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 sad_4pixelN(int32 src1, int32 src2, int32 mask) 35729a84457aed4c45bc900998b5e11c03023264208James Dong { 35829a84457aed4c45bc900998b5e11c03023264208James Dong int32 x7; 35929a84457aed4c45bc900998b5e11c03023264208James Dong 36029a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("EOR %1, %2, %0\n\tADDS %0, %2, %0\n\tEOR %1, %1, %0\n\tANDS %1, %3, %1, rrx\n\tRSB %1, %1, %1, lsl #8\n\tSUB %0, %0, %1, asr #7\n\tEOR %0, %0, %1, asr #7": "=r"(src1), "=&r"(x7): "r"(src2), "r"(mask)); 36129a84457aed4c45bc900998b5e11c03023264208James Dong 36229a84457aed4c45bc900998b5e11c03023264208James Dong return src1; 36329a84457aed4c45bc900998b5e11c03023264208James Dong } 36429a84457aed4c45bc900998b5e11c03023264208James Dong 36529a84457aed4c45bc900998b5e11c03023264208James Dong#define sum_accumulate __asm__ volatile("SBC %0, %0, %1\n\tBIC %1, %4, %1\n\tADD %2, %2, %1, lsr #8\n\tSBC %0, %0, %3\n\tBIC %3, %4, %3\n\tADD %2, %2, %3, lsr #8": "=&r" (x5), "=&r" (x10), "=&r" (x4), "=&r" (x11): "r" (x6)); 36629a84457aed4c45bc900998b5e11c03023264208James Dong 36729a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 3 36829a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 24 36929a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x08000001 37029a84457aed4c45bc900998b5e11c03023264208James Dong 37129a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 37229a84457aed4c45bc900998b5e11c03023264208James Dong 37329a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 37429a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 2 37529a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 37629a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 16 37729a84457aed4c45bc900998b5e11c03023264208James Dong#undef INC_X8 37829a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x10000001 37929a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 38029a84457aed4c45bc900998b5e11c03023264208James Dong 38129a84457aed4c45bc900998b5e11c03023264208James Dong#undef NUMBER 38229a84457aed4c45bc900998b5e11c03023264208James Dong#define NUMBER 1 38329a84457aed4c45bc900998b5e11c03023264208James Dong#undef SHIFT 38429a84457aed4c45bc900998b5e11c03023264208James Dong#define SHIFT 8 38529a84457aed4c45bc900998b5e11c03023264208James Dong#undef INC_X8 38629a84457aed4c45bc900998b5e11c03023264208James Dong#define INC_X8 0x08000001 38729a84457aed4c45bc900998b5e11c03023264208James Dong#include "sad_mb_offset.h" 38829a84457aed4c45bc900998b5e11c03023264208James Dong 38929a84457aed4c45bc900998b5e11c03023264208James Dong 39029a84457aed4c45bc900998b5e11c03023264208James Dong __inline int32 simd_sad_mb(uint8 *ref, uint8 *blk, int dmin, int lx) 39129a84457aed4c45bc900998b5e11c03023264208James Dong { 39229a84457aed4c45bc900998b5e11c03023264208James Dong int32 x4, x5, x6, x8, x9, x10, x11, x12, x14; 39329a84457aed4c45bc900998b5e11c03023264208James Dong 39429a84457aed4c45bc900998b5e11c03023264208James Dong x9 = 0x80808080; /* const. */ 39529a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x5 = 0; 39629a84457aed4c45bc900998b5e11c03023264208James Dong 39729a84457aed4c45bc900998b5e11c03023264208James Dong x8 = (uint32)ref & 0x3; 39829a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 3) 39929a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset3; 40029a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 2) 40129a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset2; 40229a84457aed4c45bc900998b5e11c03023264208James Dong if (x8 == 1) 40329a84457aed4c45bc900998b5e11c03023264208James Dong goto SadMBOffset1; 40429a84457aed4c45bc900998b5e11c03023264208James Dong 40529a84457aed4c45bc900998b5e11c03023264208James Dong x8 = 16; 40629a84457aed4c45bc900998b5e11c03023264208James Dong/// 40729a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("MVN %0, #0xFF00": "=r"(x6)); 40829a84457aed4c45bc900998b5e11c03023264208James Dong 40929a84457aed4c45bc900998b5e11c03023264208James DongLOOP_SAD0: 41029a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 41129a84457aed4c45bc900998b5e11c03023264208James Dong x11 = *((int32*)(ref + 12)); 41229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = *((int32*)(ref + 8)); 41329a84457aed4c45bc900998b5e11c03023264208James Dong x14 = *((int32*)(blk + 12)); 41429a84457aed4c45bc900998b5e11c03023264208James Dong x12 = *((int32*)(blk + 8)); 41529a84457aed4c45bc900998b5e11c03023264208James Dong 41629a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 41729a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 41829a84457aed4c45bc900998b5e11c03023264208James Dong 41929a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 42029a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 42129a84457aed4c45bc900998b5e11c03023264208James Dong 42229a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 42329a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 42429a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 42529a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 42629a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 42729a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 42829a84457aed4c45bc900998b5e11c03023264208James Dong 42929a84457aed4c45bc900998b5e11c03023264208James Dong /****** process 8 pixels ******/ 43029a84457aed4c45bc900998b5e11c03023264208James Dong x11 = *((int32*)(ref + 4)); 43129a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("LDR %0, [%1], %2": "=&r"(x10), "=r"(ref): "r"(lx)); 43229a84457aed4c45bc900998b5e11c03023264208James Dong //x10 = *((int32*)ref); ref+=lx; 43329a84457aed4c45bc900998b5e11c03023264208James Dong x14 = *((int32*)(blk + 4)); 43429a84457aed4c45bc900998b5e11c03023264208James Dong__asm__ volatile("LDR %0, [%1], #16": "=&r"(x12), "=r"(blk)); 43529a84457aed4c45bc900998b5e11c03023264208James Dong 43629a84457aed4c45bc900998b5e11c03023264208James Dong /* process x11 & x14 */ 43729a84457aed4c45bc900998b5e11c03023264208James Dong x11 = sad_4pixel(x11, x14, x9); 43829a84457aed4c45bc900998b5e11c03023264208James Dong 43929a84457aed4c45bc900998b5e11c03023264208James Dong /* process x12 & x10 */ 44029a84457aed4c45bc900998b5e11c03023264208James Dong x10 = sad_4pixel(x10, x12, x9); 44129a84457aed4c45bc900998b5e11c03023264208James Dong 44229a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x10; /* accumulate low bytes */ 44329a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 & (x6 << 8); /* x10 & 0xFF00FF00 */ 44429a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x10 >> 8); /* accumulate high bytes */ 44529a84457aed4c45bc900998b5e11c03023264208James Dong x5 = x5 + x11; /* accumulate low bytes */ 44629a84457aed4c45bc900998b5e11c03023264208James Dong x11 = x11 & (x6 << 8); /* x11 & 0xFF00FF00 */ 44729a84457aed4c45bc900998b5e11c03023264208James Dong x4 = x4 + ((uint32)x11 >> 8); /* accumulate high bytes */ 44829a84457aed4c45bc900998b5e11c03023264208James Dong 44929a84457aed4c45bc900998b5e11c03023264208James Dong /****************/ 45029a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x5 - (x4 << 8); /* extract low bytes */ 45129a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + x4; /* add with high bytes */ 45229a84457aed4c45bc900998b5e11c03023264208James Dong x10 = x10 + (x10 << 16); /* add with lower half word */ 45329a84457aed4c45bc900998b5e11c03023264208James Dong 45429a84457aed4c45bc900998b5e11c03023264208James Dong /****************/ 45529a84457aed4c45bc900998b5e11c03023264208James Dong 45629a84457aed4c45bc900998b5e11c03023264208James Dong if (((uint32)x10 >> 16) <= dmin) /* compare with dmin */ 45729a84457aed4c45bc900998b5e11c03023264208James Dong { 45829a84457aed4c45bc900998b5e11c03023264208James Dong if (--x8) 45929a84457aed4c45bc900998b5e11c03023264208James Dong { 46029a84457aed4c45bc900998b5e11c03023264208James Dong goto LOOP_SAD0; 46129a84457aed4c45bc900998b5e11c03023264208James Dong } 46229a84457aed4c45bc900998b5e11c03023264208James Dong 46329a84457aed4c45bc900998b5e11c03023264208James Dong } 46429a84457aed4c45bc900998b5e11c03023264208James Dong 46529a84457aed4c45bc900998b5e11c03023264208James Dong return ((uint32)x10 >> 16); 46629a84457aed4c45bc900998b5e11c03023264208James Dong 46729a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset3: 46829a84457aed4c45bc900998b5e11c03023264208James Dong 46929a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset3(ref, blk, lx, dmin); 47029a84457aed4c45bc900998b5e11c03023264208James Dong 47129a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset2: 47229a84457aed4c45bc900998b5e11c03023264208James Dong 47329a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset2(ref, blk, lx, dmin); 47429a84457aed4c45bc900998b5e11c03023264208James Dong 47529a84457aed4c45bc900998b5e11c03023264208James DongSadMBOffset1: 47629a84457aed4c45bc900998b5e11c03023264208James Dong 47729a84457aed4c45bc900998b5e11c03023264208James Dong return sad_mb_offset1(ref, blk, lx, dmin); 47829a84457aed4c45bc900998b5e11c03023264208James Dong } 47929a84457aed4c45bc900998b5e11c03023264208James Dong 48029a84457aed4c45bc900998b5e11c03023264208James Dong 48129a84457aed4c45bc900998b5e11c03023264208James Dong#endif 48229a84457aed4c45bc900998b5e11c03023264208James Dong 48329a84457aed4c45bc900998b5e11c03023264208James Dong#ifdef __cplusplus 48429a84457aed4c45bc900998b5e11c03023264208James Dong} 48529a84457aed4c45bc900998b5e11c03023264208James Dong#endif 48629a84457aed4c45bc900998b5e11c03023264208James Dong 48729a84457aed4c45bc900998b5e11c03023264208James Dong#endif // _SAD_INLINE_H_ 48829a84457aed4c45bc900998b5e11c03023264208James Dong 489