yading@11: /* yading@11: * Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at) yading@11: * yading@11: * This file is part of FFmpeg. yading@11: * yading@11: * FFmpeg is free software; you can redistribute it and/or modify yading@11: * it under the terms of the GNU General Public License as published by yading@11: * the Free Software Foundation; either version 2 of the License, or yading@11: * (at your option) any later version. yading@11: * yading@11: * FFmpeg is distributed in the hope that it will be useful, yading@11: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@11: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the yading@11: * GNU General Public License for more details. yading@11: * yading@11: * You should have received a copy of the GNU General Public License yading@11: * along with FFmpeg; if not, write to the Free Software yading@11: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@11: */ yading@11: yading@11: /** yading@11: * @file yading@11: * mmx/mmx2/3dnow postprocess code. yading@11: */ yading@11: yading@11: #include "libavutil/x86/asm.h" yading@11: yading@11: /* A single TEMPLATE_PP_* should be defined (to 1) when this template is yading@11: * included. The following macros will define its dependencies to 1 as well yading@11: * (like MMX2 depending on MMX), and will define to 0 all the others. Every yading@11: * TEMPLATE_PP_* need to be undef at the end. */ yading@11: yading@11: #ifdef TEMPLATE_PP_C yading@11: # define RENAME(a) a ## _C yading@11: #else yading@11: # define TEMPLATE_PP_C 0 yading@11: #endif yading@11: yading@11: #ifdef TEMPLATE_PP_ALTIVEC yading@11: # define RENAME(a) a ## _altivec yading@11: #else yading@11: # define TEMPLATE_PP_ALTIVEC 0 yading@11: #endif yading@11: yading@11: #ifdef TEMPLATE_PP_MMX yading@11: # define RENAME(a) a ## _MMX yading@11: #else yading@11: # define TEMPLATE_PP_MMX 0 yading@11: #endif yading@11: yading@11: #ifdef TEMPLATE_PP_MMXEXT yading@11: # undef TEMPLATE_PP_MMX yading@11: # define TEMPLATE_PP_MMX 1 yading@11: # define RENAME(a) a ## _MMX2 yading@11: #else yading@11: # define TEMPLATE_PP_MMXEXT 0 yading@11: #endif yading@11: yading@11: #ifdef TEMPLATE_PP_3DNOW yading@11: # undef TEMPLATE_PP_MMX yading@11: # define TEMPLATE_PP_MMX 1 yading@11: # define RENAME(a) a ## _3DNow yading@11: #else yading@11: # define TEMPLATE_PP_3DNOW 0 yading@11: #endif yading@11: yading@11: #ifdef TEMPLATE_PP_SSE2 yading@11: # undef TEMPLATE_PP_MMX yading@11: # define TEMPLATE_PP_MMX 1 yading@11: # undef TEMPLATE_PP_MMXEXT yading@11: # define TEMPLATE_PP_MMXEXT 1 yading@11: # define RENAME(a) a ## _SSE2 yading@11: #else yading@11: # define TEMPLATE_PP_SSE2 0 yading@11: #endif yading@11: yading@11: #undef REAL_PAVGB yading@11: #undef PAVGB yading@11: #undef PMINUB yading@11: #undef PMAXUB yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: #define REAL_PAVGB(a,b) "pavgb " #a ", " #b " \n\t" yading@11: #elif TEMPLATE_PP_3DNOW yading@11: #define REAL_PAVGB(a,b) "pavgusb " #a ", " #b " \n\t" yading@11: #endif yading@11: #define PAVGB(a,b) REAL_PAVGB(a,b) yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t" yading@11: #elif TEMPLATE_PP_MMX yading@11: #define PMINUB(b,a,t) \ yading@11: "movq " #a ", " #t " \n\t"\ yading@11: "psubusb " #b ", " #t " \n\t"\ yading@11: "psubb " #t ", " #a " \n\t" yading@11: #endif yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t" yading@11: #elif TEMPLATE_PP_MMX yading@11: #define PMAXUB(a,b) \ yading@11: "psubusb " #a ", " #b " \n\t"\ yading@11: "paddb " #a ", " #b " \n\t" yading@11: #endif yading@11: yading@11: //FIXME? |255-0| = 1 (should not be a problem ...) yading@11: #if TEMPLATE_PP_MMX yading@11: /** yading@11: * Check if the middle 8x8 Block in the given 8x16 block is flat yading@11: */ yading@11: static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){ yading@11: int numEq= 0, dcOk; yading@11: src+= stride*4; // src points to begin of the 8x8 Block yading@11: __asm__ volatile( yading@11: "movq %0, %%mm7 \n\t" yading@11: "movq %1, %%mm6 \n\t" yading@11: : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) yading@11: ); yading@11: yading@11: __asm__ volatile( yading@11: "lea (%2, %3), %%"REG_a" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 yading@11: yading@11: "movq (%2), %%mm0 \n\t" yading@11: "movq (%%"REG_a"), %%mm1 \n\t" yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "movq %%mm0, %%mm4 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm0 \n\t" // mm0 = difference yading@11: "paddb %%mm7, %%mm0 \n\t" yading@11: "pcmpgtb %%mm6, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a",%3), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3, 2), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: yading@11: "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" yading@11: yading@11: "movq (%2, %3, 4), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a"), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3, 2), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: "psubusb %%mm3, %%mm4 \n\t" yading@11: yading@11: " \n\t" yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: "psadbw %%mm7, %%mm0 \n\t" yading@11: #else yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "psrlw $8, %%mm0 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "psrlq $16, %%mm0 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: #endif yading@11: "movq %4, %%mm7 \n\t" // QP,..., QP yading@11: "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP yading@11: "psubusb %%mm7, %%mm4 \n\t" // Diff <= 2QP -> 0 yading@11: "packssdw %%mm4, %%mm4 \n\t" yading@11: "movd %%mm0, %0 \n\t" yading@11: "movd %%mm4, %1 \n\t" yading@11: yading@11: : "=r" (numEq), "=r" (dcOk) yading@11: : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb) yading@11: : "%"REG_a yading@11: ); yading@11: yading@11: numEq= (-numEq) &0xFF; yading@11: if(numEq > c->ppMode.flatnessThreshold){ yading@11: if(dcOk) return 0; yading@11: else return 1; yading@11: }else{ yading@11: return 2; yading@11: } yading@11: } yading@11: #endif //TEMPLATE_PP_MMX yading@11: yading@11: /** yading@11: * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle) yading@11: * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 yading@11: */ yading@11: #if !TEMPLATE_PP_ALTIVEC yading@11: static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= stride*3; yading@11: __asm__ volatile( //"movv %0 %1 %2\n\t" yading@11: "movq %2, %%mm0 \n\t" // QP,..., QP yading@11: "pxor %%mm4, %%mm4 \n\t" yading@11: yading@11: "movq (%0), %%mm6 \n\t" yading@11: "movq (%0, %1), %%mm5 \n\t" yading@11: "movq %%mm5, %%mm1 \n\t" yading@11: "movq %%mm6, %%mm2 \n\t" yading@11: "psubusb %%mm6, %%mm5 \n\t" yading@11: "psubusb %%mm1, %%mm2 \n\t" yading@11: "por %%mm5, %%mm2 \n\t" // ABS Diff of lines yading@11: "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 yading@11: "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF yading@11: yading@11: "pand %%mm2, %%mm6 \n\t" yading@11: "pandn %%mm1, %%mm2 \n\t" yading@11: "por %%mm2, %%mm6 \n\t"// First Line to Filter yading@11: yading@11: "movq (%0, %1, 8), %%mm5 \n\t" yading@11: "lea (%0, %1, 4), %%"REG_a" \n\t" yading@11: "lea (%0, %1, 8), %%"REG_c" \n\t" yading@11: "sub %1, %%"REG_c" \n\t" yading@11: "add %1, %0 \n\t" // %0 points to line 1 not 0 yading@11: "movq (%0, %1, 8), %%mm7 \n\t" yading@11: "movq %%mm5, %%mm1 \n\t" yading@11: "movq %%mm7, %%mm2 \n\t" yading@11: "psubusb %%mm7, %%mm5 \n\t" yading@11: "psubusb %%mm1, %%mm2 \n\t" yading@11: "por %%mm5, %%mm2 \n\t" // ABS Diff of lines yading@11: "psubusb %%mm0, %%mm2 \n\t" // diff <= QP -> 0 yading@11: "pcmpeqb %%mm4, %%mm2 \n\t" // diff <= QP -> FF yading@11: yading@11: "pand %%mm2, %%mm7 \n\t" yading@11: "pandn %%mm1, %%mm2 \n\t" yading@11: "por %%mm2, %%mm7 \n\t" // First Line to Filter yading@11: yading@11: yading@11: // 1 2 3 4 5 6 7 8 yading@11: // %0 %0+%1 %0+2%1 eax %0+4%1 eax+2%1 ecx eax+4%1 yading@11: // 6 4 2 2 1 1 yading@11: // 6 4 4 2 yading@11: // 6 8 2 yading@11: yading@11: "movq (%0, %1), %%mm0 \n\t" // 1 yading@11: "movq %%mm0, %%mm1 \n\t" // 1 yading@11: PAVGB(%%mm6, %%mm0) //1 1 /2 yading@11: PAVGB(%%mm6, %%mm0) //3 1 /4 yading@11: yading@11: "movq (%0, %1, 4), %%mm2 \n\t" // 1 yading@11: "movq %%mm2, %%mm5 \n\t" // 1 yading@11: PAVGB((%%REGa), %%mm2) // 11 /2 yading@11: PAVGB((%0, %1, 2), %%mm2) // 211 /4 yading@11: "movq %%mm2, %%mm3 \n\t" // 211 /4 yading@11: "movq (%0), %%mm4 \n\t" // 1 yading@11: PAVGB(%%mm4, %%mm3) // 4 211 /8 yading@11: PAVGB(%%mm0, %%mm3) //642211 /16 yading@11: "movq %%mm3, (%0) \n\t" // X yading@11: // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9 yading@11: "movq %%mm1, %%mm0 \n\t" // 1 yading@11: PAVGB(%%mm6, %%mm0) //1 1 /2 yading@11: "movq %%mm4, %%mm3 \n\t" // 1 yading@11: PAVGB((%0,%1,2), %%mm3) // 1 1 /2 yading@11: PAVGB((%%REGa,%1,2), %%mm5) // 11 /2 yading@11: PAVGB((%%REGa), %%mm5) // 211 /4 yading@11: PAVGB(%%mm5, %%mm3) // 2 2211 /8 yading@11: PAVGB(%%mm0, %%mm3) //4242211 /16 yading@11: "movq %%mm3, (%0,%1) \n\t" // X yading@11: // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9 yading@11: PAVGB(%%mm4, %%mm6) //11 /2 yading@11: "movq (%%"REG_c"), %%mm0 \n\t" // 1 yading@11: PAVGB((%%REGa, %1, 2), %%mm0) // 11/2 yading@11: "movq %%mm0, %%mm3 \n\t" // 11/2 yading@11: PAVGB(%%mm1, %%mm0) // 2 11/4 yading@11: PAVGB(%%mm6, %%mm0) //222 11/8 yading@11: PAVGB(%%mm2, %%mm0) //22242211/16 yading@11: "movq (%0, %1, 2), %%mm2 \n\t" // 1 yading@11: "movq %%mm0, (%0, %1, 2) \n\t" // X yading@11: // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9 yading@11: "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 yading@11: PAVGB((%%REGc), %%mm0) // 11 /2 yading@11: PAVGB(%%mm0, %%mm6) //11 11 /4 yading@11: PAVGB(%%mm1, %%mm4) // 11 /2 yading@11: PAVGB(%%mm2, %%mm1) // 11 /2 yading@11: PAVGB(%%mm1, %%mm6) //1122 11 /8 yading@11: PAVGB(%%mm5, %%mm6) //112242211 /16 yading@11: "movq (%%"REG_a"), %%mm5 \n\t" // 1 yading@11: "movq %%mm6, (%%"REG_a") \n\t" // X yading@11: // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9 yading@11: "movq (%%"REG_a", %1, 4), %%mm6 \n\t" // 1 yading@11: PAVGB(%%mm7, %%mm6) // 11 /2 yading@11: PAVGB(%%mm4, %%mm6) // 11 11 /4 yading@11: PAVGB(%%mm3, %%mm6) // 11 2211 /8 yading@11: PAVGB(%%mm5, %%mm2) // 11 /2 yading@11: "movq (%0, %1, 4), %%mm4 \n\t" // 1 yading@11: PAVGB(%%mm4, %%mm2) // 112 /4 yading@11: PAVGB(%%mm2, %%mm6) // 112242211 /16 yading@11: "movq %%mm6, (%0, %1, 4) \n\t" // X yading@11: // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9 yading@11: PAVGB(%%mm7, %%mm1) // 11 2 /4 yading@11: PAVGB(%%mm4, %%mm5) // 11 /2 yading@11: PAVGB(%%mm5, %%mm0) // 11 11 /4 yading@11: "movq (%%"REG_a", %1, 2), %%mm6 \n\t" // 1 yading@11: PAVGB(%%mm6, %%mm1) // 11 4 2 /8 yading@11: PAVGB(%%mm0, %%mm1) // 11224222 /16 yading@11: "movq %%mm1, (%%"REG_a", %1, 2) \n\t" // X yading@11: // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9 yading@11: PAVGB((%%REGc), %%mm2) // 112 4 /8 yading@11: "movq (%%"REG_a", %1, 4), %%mm0 \n\t" // 1 yading@11: PAVGB(%%mm0, %%mm6) // 1 1 /2 yading@11: PAVGB(%%mm7, %%mm6) // 1 12 /4 yading@11: PAVGB(%%mm2, %%mm6) // 1122424 /4 yading@11: "movq %%mm6, (%%"REG_c") \n\t" // X yading@11: // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9 yading@11: PAVGB(%%mm7, %%mm5) // 11 2 /4 yading@11: PAVGB(%%mm7, %%mm5) // 11 6 /8 yading@11: yading@11: PAVGB(%%mm3, %%mm0) // 112 /4 yading@11: PAVGB(%%mm0, %%mm5) // 112246 /16 yading@11: "movq %%mm5, (%%"REG_a", %1, 4) \n\t" // X yading@11: "sub %1, %0 \n\t" yading@11: yading@11: : yading@11: : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb) yading@11: : "%"REG_a, "%"REG_c yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: const int l1= stride; yading@11: const int l2= stride + l1; yading@11: const int l3= stride + l2; yading@11: const int l4= stride + l3; yading@11: const int l5= stride + l4; yading@11: const int l6= stride + l5; yading@11: const int l7= stride + l6; yading@11: const int l8= stride + l7; yading@11: const int l9= stride + l8; yading@11: int x; yading@11: src+= stride*3; yading@11: for(x=0; xQP ? src[0] : src[l1]; yading@11: const int last= FFABS(src[l8] - src[l9]) < c->QP ? src[l9] : src[l8]; yading@11: yading@11: int sums[10]; yading@11: sums[0] = 4*first + src[l1] + src[l2] + src[l3] + 4; yading@11: sums[1] = sums[0] - first + src[l4]; yading@11: sums[2] = sums[1] - first + src[l5]; yading@11: sums[3] = sums[2] - first + src[l6]; yading@11: sums[4] = sums[3] - first + src[l7]; yading@11: sums[5] = sums[4] - src[l1] + src[l8]; yading@11: sums[6] = sums[5] - src[l2] + last; yading@11: sums[7] = sums[6] - src[l3] + last; yading@11: sums[8] = sums[7] - src[l4] + last; yading@11: sums[9] = sums[8] - src[l5] + last; yading@11: yading@11: src[l1]= (sums[0] + sums[2] + 2*src[l1])>>4; yading@11: src[l2]= (sums[1] + sums[3] + 2*src[l2])>>4; yading@11: src[l3]= (sums[2] + sums[4] + 2*src[l3])>>4; yading@11: src[l4]= (sums[3] + sums[5] + 2*src[l4])>>4; yading@11: src[l5]= (sums[4] + sums[6] + 2*src[l5])>>4; yading@11: src[l6]= (sums[5] + sums[7] + 2*src[l6])>>4; yading@11: src[l7]= (sums[6] + sums[8] + 2*src[l7])>>4; yading@11: src[l8]= (sums[7] + sums[9] + 2*src[l8])>>4; yading@11: yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: #endif //TEMPLATE_PP_ALTIVEC yading@11: yading@11: /** yading@11: * Experimental Filter 1 yading@11: * will not damage linear gradients yading@11: * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter yading@11: * can only smooth blocks at the expected locations (it cannot smooth them if they did move) yading@11: * MMX2 version does correct clipping C version does not yading@11: */ yading@11: static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= stride*3; yading@11: yading@11: __asm__ volatile( yading@11: "pxor %%mm7, %%mm7 \n\t" // 0 yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 yading@11: "movq (%0, %1, 4), %%mm1 \n\t" // line 4 yading@11: "movq %%mm1, %%mm2 \n\t" // line 4 yading@11: "psubusb %%mm0, %%mm1 \n\t" yading@11: "psubusb %%mm2, %%mm0 \n\t" yading@11: "por %%mm1, %%mm0 \n\t" // |l2 - l3| yading@11: "movq (%%"REG_c"), %%mm3 \n\t" // line 5 yading@11: "movq (%%"REG_c", %1), %%mm4 \n\t" // line 6 yading@11: "movq %%mm3, %%mm5 \n\t" // line 5 yading@11: "psubusb %%mm4, %%mm3 \n\t" yading@11: "psubusb %%mm5, %%mm4 \n\t" yading@11: "por %%mm4, %%mm3 \n\t" // |l5 - l6| yading@11: PAVGB(%%mm3, %%mm0) // (|l2 - l3| + |l5 - l6|)/2 yading@11: "movq %%mm2, %%mm1 \n\t" // line 4 yading@11: "psubusb %%mm5, %%mm2 \n\t" yading@11: "movq %%mm2, %%mm4 \n\t" yading@11: "pcmpeqb %%mm7, %%mm2 \n\t" // (l4 - l5) <= 0 ? -1 : 0 yading@11: "psubusb %%mm1, %%mm5 \n\t" yading@11: "por %%mm5, %%mm4 \n\t" // |l4 - l5| yading@11: "psubusb %%mm0, %%mm4 \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2) yading@11: "movq %%mm4, %%mm3 \n\t" // d yading@11: "movq %2, %%mm0 \n\t" yading@11: "paddusb %%mm0, %%mm0 \n\t" yading@11: "psubusb %%mm0, %%mm4 \n\t" yading@11: "pcmpeqb %%mm7, %%mm4 \n\t" // d <= QP ? -1 : 0 yading@11: "psubusb "MANGLE(b01)", %%mm3 \n\t" yading@11: "pand %%mm4, %%mm3 \n\t" // d <= QP ? d : 0 yading@11: yading@11: PAVGB(%%mm7, %%mm3) // d/2 yading@11: "movq %%mm3, %%mm1 \n\t" // d/2 yading@11: PAVGB(%%mm7, %%mm3) // d/4 yading@11: PAVGB(%%mm1, %%mm3) // 3*d/8 yading@11: yading@11: "movq (%0, %1, 4), %%mm0 \n\t" // line 4 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 yading@11: "psubusb %%mm3, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%0, %1, 4) \n\t" // line 4 yading@11: yading@11: "movq (%%"REG_c"), %%mm0 \n\t" // line 5 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 yading@11: "paddusb %%mm3, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_c") \n\t" // line 5 yading@11: yading@11: PAVGB(%%mm7, %%mm1) // d/4 yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" // line 3 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4 yading@11: "psubusb %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_a", %1, 2) \n\t" // line 3 yading@11: yading@11: "movq (%%"REG_c", %1), %%mm0 \n\t" // line 6 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5 yading@11: "paddusb %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_c", %1) \n\t" // line 6 yading@11: yading@11: PAVGB(%%mm7, %%mm1) // d/8 yading@11: yading@11: "movq (%%"REG_a", %1), %%mm0 \n\t" // line 2 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2 yading@11: "psubusb %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_a", %1) \n\t" // line 2 yading@11: yading@11: "movq (%%"REG_c", %1, 2), %%mm0 \n\t" // line 7 yading@11: "pxor %%mm2, %%mm0 \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7 yading@11: "paddusb %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_c", %1, 2) \n\t" // line 7 yading@11: yading@11: : yading@11: : "r" (src), "r" ((x86_reg)stride), "m" (co->pQPb) yading@11: : "%"REG_a, "%"REG_c yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: yading@11: const int l1= stride; yading@11: const int l2= stride + l1; yading@11: const int l3= stride + l2; yading@11: const int l4= stride + l3; yading@11: const int l5= stride + l4; yading@11: const int l6= stride + l5; yading@11: const int l7= stride + l6; yading@11: // const int l8= stride + l7; yading@11: // const int l9= stride + l8; yading@11: int x; yading@11: yading@11: src+= stride*3; yading@11: for(x=0; x>1); yading@11: d= FFMAX(d, 0); yading@11: yading@11: if(d < co->QP*2){ yading@11: int v = d * FFSIGN(-b); yading@11: yading@11: src[l2] +=v>>3; yading@11: src[l3] +=v>>2; yading@11: src[l4] +=(3*v)>>3; yading@11: src[l5] -=(3*v)>>3; yading@11: src[l6] -=v>>2; yading@11: src[l7] -=v>>3; yading@11: } yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: yading@11: #if !TEMPLATE_PP_ALTIVEC yading@11: static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: /* yading@11: uint8_t tmp[16]; yading@11: const int l1= stride; yading@11: const int l2= stride + l1; yading@11: const int l3= stride + l2; yading@11: const int l4= (int)tmp - (int)src - stride*3; yading@11: const int l5= (int)tmp - (int)src - stride*3 + 8; yading@11: const int l6= stride*3 + l3; yading@11: const int l7= stride + l6; yading@11: const int l8= stride + l7; yading@11: yading@11: memcpy(tmp, src+stride*7, 8); yading@11: memcpy(tmp+8, src+stride*8, 8); yading@11: */ yading@11: src+= stride*4; yading@11: __asm__ volatile( yading@11: yading@11: #if 0 //slightly more accurate and slightly slower yading@11: "pxor %%mm7, %%mm7 \n\t" // 0 yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" yading@11: // 0 1 2 3 4 5 6 7 yading@11: // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 yading@11: yading@11: yading@11: "movq (%0, %1, 2), %%mm0 \n\t" // l2 yading@11: "movq (%0), %%mm1 \n\t" // l0 yading@11: "movq %%mm0, %%mm2 \n\t" // l2 yading@11: PAVGB(%%mm7, %%mm0) // ~l2/2 yading@11: PAVGB(%%mm1, %%mm0) // ~(l2 + 2l0)/4 yading@11: PAVGB(%%mm2, %%mm0) // ~(5l2 + 2l0)/8 yading@11: yading@11: "movq (%%"REG_a"), %%mm1 \n\t" // l1 yading@11: "movq (%%"REG_a", %1, 2), %%mm3 \n\t" // l3 yading@11: "movq %%mm1, %%mm4 \n\t" // l1 yading@11: PAVGB(%%mm7, %%mm1) // ~l1/2 yading@11: PAVGB(%%mm3, %%mm1) // ~(l1 + 2l3)/4 yading@11: PAVGB(%%mm4, %%mm1) // ~(5l1 + 2l3)/8 yading@11: yading@11: "movq %%mm0, %%mm4 \n\t" // ~(5l2 + 2l0)/8 yading@11: "psubusb %%mm1, %%mm0 \n\t" yading@11: "psubusb %%mm4, %%mm1 \n\t" yading@11: "por %%mm0, %%mm1 \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8 yading@11: // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0 yading@11: yading@11: "movq (%0, %1, 4), %%mm0 \n\t" // l4 yading@11: "movq %%mm0, %%mm4 \n\t" // l4 yading@11: PAVGB(%%mm7, %%mm0) // ~l4/2 yading@11: PAVGB(%%mm2, %%mm0) // ~(l4 + 2l2)/4 yading@11: PAVGB(%%mm4, %%mm0) // ~(5l4 + 2l2)/8 yading@11: yading@11: "movq (%%"REG_c"), %%mm2 \n\t" // l5 yading@11: "movq %%mm3, %%mm5 \n\t" // l3 yading@11: PAVGB(%%mm7, %%mm3) // ~l3/2 yading@11: PAVGB(%%mm2, %%mm3) // ~(l3 + 2l5)/4 yading@11: PAVGB(%%mm5, %%mm3) // ~(5l3 + 2l5)/8 yading@11: yading@11: "movq %%mm0, %%mm6 \n\t" // ~(5l4 + 2l2)/8 yading@11: "psubusb %%mm3, %%mm0 \n\t" yading@11: "psubusb %%mm6, %%mm3 \n\t" yading@11: "por %%mm0, %%mm3 \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8 yading@11: "pcmpeqb %%mm7, %%mm0 \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5) yading@11: // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0 yading@11: yading@11: "movq (%%"REG_c", %1), %%mm6 \n\t" // l6 yading@11: "movq %%mm6, %%mm5 \n\t" // l6 yading@11: PAVGB(%%mm7, %%mm6) // ~l6/2 yading@11: PAVGB(%%mm4, %%mm6) // ~(l6 + 2l4)/4 yading@11: PAVGB(%%mm5, %%mm6) // ~(5l6 + 2l4)/8 yading@11: yading@11: "movq (%%"REG_c", %1, 2), %%mm5 \n\t" // l7 yading@11: "movq %%mm2, %%mm4 \n\t" // l5 yading@11: PAVGB(%%mm7, %%mm2) // ~l5/2 yading@11: PAVGB(%%mm5, %%mm2) // ~(l5 + 2l7)/4 yading@11: PAVGB(%%mm4, %%mm2) // ~(5l5 + 2l7)/8 yading@11: yading@11: "movq %%mm6, %%mm4 \n\t" // ~(5l6 + 2l4)/8 yading@11: "psubusb %%mm2, %%mm6 \n\t" yading@11: "psubusb %%mm4, %%mm2 \n\t" yading@11: "por %%mm6, %%mm2 \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8 yading@11: // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0 yading@11: yading@11: yading@11: PMINUB(%%mm2, %%mm1, %%mm4) // MIN(|lenergy|,|renergy|)/8 yading@11: "movq %2, %%mm4 \n\t" // QP //FIXME QP+1 ? yading@11: "paddusb "MANGLE(b01)", %%mm4 \n\t" yading@11: "pcmpgtb %%mm3, %%mm4 \n\t" // |menergy|/8 < QP yading@11: "psubusb %%mm1, %%mm3 \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8 yading@11: "pand %%mm4, %%mm3 \n\t" yading@11: yading@11: "movq %%mm3, %%mm1 \n\t" yading@11: // "psubusb "MANGLE(b01)", %%mm3 \n\t" yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm7, %%mm3) yading@11: "paddusb %%mm1, %%mm3 \n\t" yading@11: // "paddusb "MANGLE(b01)", %%mm3 \n\t" yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm6 \n\t" //l3 yading@11: "movq (%0, %1, 4), %%mm5 \n\t" //l4 yading@11: "movq (%0, %1, 4), %%mm4 \n\t" //l4 yading@11: "psubusb %%mm6, %%mm5 \n\t" yading@11: "psubusb %%mm4, %%mm6 \n\t" yading@11: "por %%mm6, %%mm5 \n\t" // |l3-l4| yading@11: "pcmpeqb %%mm7, %%mm6 \n\t" // SIGN(l3-l4) yading@11: "pxor %%mm6, %%mm0 \n\t" yading@11: "pand %%mm0, %%mm3 \n\t" yading@11: PMINUB(%%mm5, %%mm3, %%mm0) yading@11: yading@11: "psubusb "MANGLE(b01)", %%mm3 \n\t" yading@11: PAVGB(%%mm7, %%mm3) yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" yading@11: "movq (%0, %1, 4), %%mm2 \n\t" yading@11: "pxor %%mm6, %%mm0 \n\t" yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "psubb %%mm3, %%mm0 \n\t" yading@11: "paddb %%mm3, %%mm2 \n\t" yading@11: "pxor %%mm6, %%mm0 \n\t" yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "movq %%mm0, (%%"REG_a", %1, 2) \n\t" yading@11: "movq %%mm2, (%0, %1, 4) \n\t" yading@11: #endif //0 yading@11: yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "pcmpeqb %%mm6, %%mm6 \n\t" // -1 yading@11: // 0 1 2 3 4 5 6 7 yading@11: // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 ecx+%1 ecx+2%1 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 yading@11: yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // l3 yading@11: "movq (%0, %1, 4), %%mm0 \n\t" // l4 yading@11: "pxor %%mm6, %%mm1 \n\t" // -l3-1 yading@11: PAVGB(%%mm1, %%mm0) // -q+128 = (l4-l3+256)/2 yading@11: // mm1=-l3-1, mm0=128-q yading@11: yading@11: "movq (%%"REG_a", %1, 4), %%mm2 \n\t" // l5 yading@11: "movq (%%"REG_a", %1), %%mm3 \n\t" // l2 yading@11: "pxor %%mm6, %%mm2 \n\t" // -l5-1 yading@11: "movq %%mm2, %%mm5 \n\t" // -l5-1 yading@11: "movq "MANGLE(b80)", %%mm4 \n\t" // 128 yading@11: "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" yading@11: PAVGB(%%mm3, %%mm2) // (l2-l5+256)/2 yading@11: PAVGB(%%mm0, %%mm4) // ~(l4-l3)/4 + 128 yading@11: PAVGB(%%mm2, %%mm4) // ~(l2-l5)/4 +(l4-l3)/8 + 128 yading@11: PAVGB(%%mm0, %%mm4) // ~(l2-l5)/8 +5(l4-l3)/16 + 128 yading@11: // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1 yading@11: yading@11: "movq (%%"REG_a"), %%mm2 \n\t" // l1 yading@11: "pxor %%mm6, %%mm2 \n\t" // -l1-1 yading@11: PAVGB(%%mm3, %%mm2) // (l2-l1+256)/2 yading@11: PAVGB((%0), %%mm1) // (l0-l3+256)/2 yading@11: "movq "MANGLE(b80)", %%mm3 \n\t" // 128 yading@11: PAVGB(%%mm2, %%mm3) // ~(l2-l1)/4 + 128 yading@11: PAVGB(%%mm1, %%mm3) // ~(l0-l3)/4 +(l2-l1)/8 + 128 yading@11: PAVGB(%%mm2, %%mm3) // ~(l0-l3)/8 +5(l2-l1)/16 + 128 yading@11: // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1 yading@11: yading@11: PAVGB((%%REGc, %1), %%mm5) // (l6-l5+256)/2 yading@11: "movq (%%"REG_c", %1, 2), %%mm1 \n\t" // l7 yading@11: "pxor %%mm6, %%mm1 \n\t" // -l7-1 yading@11: PAVGB((%0, %1, 4), %%mm1) // (l4-l7+256)/2 yading@11: "movq "MANGLE(b80)", %%mm2 \n\t" // 128 yading@11: PAVGB(%%mm5, %%mm2) // ~(l6-l5)/4 + 128 yading@11: PAVGB(%%mm1, %%mm2) // ~(l4-l7)/4 +(l6-l5)/8 + 128 yading@11: PAVGB(%%mm5, %%mm2) // ~(l4-l7)/8 +5(l6-l5)/16 + 128 yading@11: // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128 yading@11: yading@11: "movq "MANGLE(b00)", %%mm1 \n\t" // 0 yading@11: "movq "MANGLE(b00)", %%mm5 \n\t" // 0 yading@11: "psubb %%mm2, %%mm1 \n\t" // 128 - renergy/16 yading@11: "psubb %%mm3, %%mm5 \n\t" // 128 - lenergy/16 yading@11: PMAXUB(%%mm1, %%mm2) // 128 + |renergy/16| yading@11: PMAXUB(%%mm5, %%mm3) // 128 + |lenergy/16| yading@11: PMINUB(%%mm2, %%mm3, %%mm1) // 128 + MIN(|lenergy|,|renergy|)/16 yading@11: yading@11: // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128 yading@11: yading@11: "movq "MANGLE(b00)", %%mm7 \n\t" // 0 yading@11: "movq %2, %%mm2 \n\t" // QP yading@11: PAVGB(%%mm6, %%mm2) // 128 + QP/2 yading@11: "psubb %%mm6, %%mm2 \n\t" yading@11: yading@11: "movq %%mm4, %%mm1 \n\t" yading@11: "pcmpgtb %%mm7, %%mm1 \n\t" // SIGN(menergy) yading@11: "pxor %%mm1, %%mm4 \n\t" yading@11: "psubb %%mm1, %%mm4 \n\t" // 128 + |menergy|/16 yading@11: "pcmpgtb %%mm4, %%mm2 \n\t" // |menergy|/16 < QP/2 yading@11: "psubusb %%mm3, %%mm4 \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16 yading@11: // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16 yading@11: yading@11: "movq %%mm4, %%mm3 \n\t" // d yading@11: "psubusb "MANGLE(b01)", %%mm4 \n\t" yading@11: PAVGB(%%mm7, %%mm4) // d/32 yading@11: PAVGB(%%mm7, %%mm4) // (d + 32)/64 yading@11: "paddb %%mm3, %%mm4 \n\t" // 5d/64 yading@11: "pand %%mm2, %%mm4 \n\t" yading@11: yading@11: "movq "MANGLE(b80)", %%mm5 \n\t" // 128 yading@11: "psubb %%mm0, %%mm5 \n\t" // q yading@11: "paddsb %%mm6, %%mm5 \n\t" // fix bad rounding yading@11: "pcmpgtb %%mm5, %%mm7 \n\t" // SIGN(q) yading@11: "pxor %%mm7, %%mm5 \n\t" yading@11: yading@11: PMINUB(%%mm5, %%mm4, %%mm3) // MIN(|q|, 5d/64) yading@11: "pxor %%mm1, %%mm7 \n\t" // SIGN(d*q) yading@11: yading@11: "pand %%mm7, %%mm4 \n\t" yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" yading@11: "movq (%0, %1, 4), %%mm2 \n\t" yading@11: "pxor %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm4, %%mm0 \n\t" yading@11: "psubb %%mm4, %%mm2 \n\t" yading@11: "pxor %%mm1, %%mm0 \n\t" yading@11: "pxor %%mm1, %%mm2 \n\t" yading@11: "movq %%mm0, (%%"REG_a", %1, 2) \n\t" yading@11: "movq %%mm2, (%0, %1, 4) \n\t" yading@11: yading@11: : yading@11: : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb) yading@11: : "%"REG_a, "%"REG_c yading@11: ); yading@11: yading@11: /* yading@11: { yading@11: int x; yading@11: src-= stride; yading@11: for(x=0; x> 6; yading@11: d*= FFSIGN(-middleEnergy); yading@11: yading@11: if(q>0){ yading@11: d= d<0 ? 0 : d; yading@11: d= d>q ? q : d; yading@11: }else{ yading@11: d= d>0 ? 0 : d; yading@11: d= d3 ? 1 : 0; yading@11: if(ad>3){ yading@11: src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255; yading@11: } yading@11: if(y==4) bias+=d; yading@11: num++; yading@11: if(num%1000000 == 0){ yading@11: av_log(c, AV_LOG_INFO, " %d %d %d %d\n", num, sum, max, bias); yading@11: } yading@11: } yading@11: } yading@11: } yading@11: */ yading@11: #elif TEMPLATE_PP_MMX yading@11: DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars yading@11: src+= stride*4; yading@11: __asm__ volatile( yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: // 0 1 2 3 4 5 6 7 yading@11: // %0 %0+%1 %0+2%1 eax+2%1 %0+4%1 eax+4%1 edx+%1 edx+2%1 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 yading@11: yading@11: "movq (%0), %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 yading@11: "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 yading@11: yading@11: "movq (%0, %1), %%mm2 \n\t" yading@11: "lea (%0, %1, 2), %%"REG_a" \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 yading@11: yading@11: "movq (%%"REG_a"), %%mm4 \n\t" yading@11: "movq %%mm4, %%mm5 \n\t" yading@11: "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 yading@11: "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 yading@11: yading@11: "paddw %%mm0, %%mm0 \n\t" // 2L0 yading@11: "paddw %%mm1, %%mm1 \n\t" // 2H0 yading@11: "psubw %%mm4, %%mm2 \n\t" // L1 - L2 yading@11: "psubw %%mm5, %%mm3 \n\t" // H1 - H2 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 yading@11: yading@11: "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 yading@11: "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 yading@11: yading@11: "movq (%%"REG_a", %1), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L3 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H3 yading@11: yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: "movq %%mm0, (%3) \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "movq %%mm1, 8(%3) \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "punpcklbw %%mm7, %%mm0 \n\t" // L4 yading@11: "punpckhbw %%mm7, %%mm1 \n\t" // H4 yading@11: yading@11: "psubw %%mm0, %%mm2 \n\t" // L3 - L4 yading@11: "psubw %%mm1, %%mm3 \n\t" // H3 - H4 yading@11: "movq %%mm2, 16(%3) \n\t" // L3 - L4 yading@11: "movq %%mm3, 24(%3) \n\t" // H3 - H4 yading@11: "paddw %%mm4, %%mm4 \n\t" // 2L2 yading@11: "paddw %%mm5, %%mm5 \n\t" // 2H2 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 yading@11: yading@11: "lea (%%"REG_a", %1), %0 \n\t" yading@11: "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 yading@11: "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 yading@11: //50 opcodes so far yading@11: "movq (%0, %1, 2), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L5 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H5 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 yading@11: yading@11: "movq (%%"REG_a", %1, 4), %%mm6 \n\t" yading@11: "punpcklbw %%mm7, %%mm6 \n\t" // L6 yading@11: "psubw %%mm6, %%mm2 \n\t" // L5 - L6 yading@11: "movq (%%"REG_a", %1, 4), %%mm6 \n\t" yading@11: "punpckhbw %%mm7, %%mm6 \n\t" // H6 yading@11: "psubw %%mm6, %%mm3 \n\t" // H5 - H6 yading@11: yading@11: "paddw %%mm0, %%mm0 \n\t" // 2L4 yading@11: "paddw %%mm1, %%mm1 \n\t" // 2H4 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 yading@11: yading@11: "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 yading@11: "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 yading@11: yading@11: "movq (%0, %1, 4), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L7 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H7 yading@11: yading@11: "paddw %%mm2, %%mm2 \n\t" // 2L7 yading@11: "paddw %%mm3, %%mm3 \n\t" // 2H7 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 yading@11: yading@11: "movq (%3), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "movq 8(%3), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm0, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm1, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm2, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm3, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| yading@11: #else yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm0, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm1, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm1 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm2, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm3, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm3 \n\t" yading@11: "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| yading@11: #endif yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "pminsw %%mm2, %%mm0 \n\t" yading@11: "pminsw %%mm3, %%mm1 \n\t" yading@11: #else yading@11: "movq %%mm0, %%mm6 \n\t" yading@11: "psubusw %%mm2, %%mm6 \n\t" yading@11: "psubw %%mm6, %%mm0 \n\t" yading@11: "movq %%mm1, %%mm6 \n\t" yading@11: "psubusw %%mm3, %%mm6 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: #endif yading@11: yading@11: "movd %2, %%mm2 \n\t" // QP yading@11: "punpcklbw %%mm7, %%mm2 \n\t" yading@11: yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) yading@11: "pxor %%mm6, %%mm4 \n\t" yading@11: "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| yading@11: "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) yading@11: "pxor %%mm7, %%mm5 \n\t" yading@11: "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| yading@11: // 100 opcodes yading@11: "psllw $3, %%mm2 \n\t" // 8QP yading@11: "movq %%mm2, %%mm3 \n\t" // 8QP yading@11: "pcmpgtw %%mm4, %%mm2 \n\t" yading@11: "pcmpgtw %%mm5, %%mm3 \n\t" yading@11: "pand %%mm2, %%mm4 \n\t" yading@11: "pand %%mm3, %%mm5 \n\t" yading@11: yading@11: yading@11: "psubusw %%mm0, %%mm4 \n\t" // hd yading@11: "psubusw %%mm1, %%mm5 \n\t" // ld yading@11: yading@11: yading@11: "movq "MANGLE(w05)", %%mm2 \n\t" // 5 yading@11: "pmullw %%mm2, %%mm4 \n\t" yading@11: "pmullw %%mm2, %%mm5 \n\t" yading@11: "movq "MANGLE(w20)", %%mm2 \n\t" // 32 yading@11: "paddw %%mm2, %%mm4 \n\t" yading@11: "paddw %%mm2, %%mm5 \n\t" yading@11: "psrlw $6, %%mm4 \n\t" yading@11: "psrlw $6, %%mm5 \n\t" yading@11: yading@11: "movq 16(%3), %%mm0 \n\t" // L3 - L4 yading@11: "movq 24(%3), %%mm1 \n\t" // H3 - H4 yading@11: yading@11: "pxor %%mm2, %%mm2 \n\t" yading@11: "pxor %%mm3, %%mm3 \n\t" yading@11: yading@11: "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) yading@11: "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "pxor %%mm3, %%mm1 \n\t" yading@11: "psubw %%mm2, %%mm0 \n\t" // |L3-L4| yading@11: "psubw %%mm3, %%mm1 \n\t" // |H3-H4| yading@11: "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 yading@11: "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 yading@11: yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "pxor %%mm7, %%mm3 \n\t" yading@11: "pand %%mm2, %%mm4 \n\t" yading@11: "pand %%mm3, %%mm5 \n\t" yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "pminsw %%mm0, %%mm4 \n\t" yading@11: "pminsw %%mm1, %%mm5 \n\t" yading@11: #else yading@11: "movq %%mm4, %%mm2 \n\t" yading@11: "psubusw %%mm0, %%mm2 \n\t" yading@11: "psubw %%mm2, %%mm4 \n\t" yading@11: "movq %%mm5, %%mm2 \n\t" yading@11: "psubusw %%mm1, %%mm2 \n\t" yading@11: "psubw %%mm2, %%mm5 \n\t" yading@11: #endif yading@11: "pxor %%mm6, %%mm4 \n\t" yading@11: "pxor %%mm7, %%mm5 \n\t" yading@11: "psubw %%mm6, %%mm4 \n\t" yading@11: "psubw %%mm7, %%mm5 \n\t" yading@11: "packsswb %%mm5, %%mm4 \n\t" yading@11: "movq (%0), %%mm0 \n\t" yading@11: "paddb %%mm4, %%mm0 \n\t" yading@11: "movq %%mm0, (%0) \n\t" yading@11: "movq (%0, %1), %%mm0 \n\t" yading@11: "psubb %%mm4, %%mm0 \n\t" yading@11: "movq %%mm0, (%0, %1) \n\t" yading@11: yading@11: : "+r" (src) yading@11: : "r" ((x86_reg)stride), "m" (c->pQPb), "r"(tmp) yading@11: : "%"REG_a yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: const int l1= stride; yading@11: const int l2= stride + l1; yading@11: const int l3= stride + l2; yading@11: const int l4= stride + l3; yading@11: const int l5= stride + l4; yading@11: const int l6= stride + l5; yading@11: const int l7= stride + l6; yading@11: const int l8= stride + l7; yading@11: // const int l9= stride + l8; yading@11: int x; yading@11: src+= stride*3; yading@11: for(x=0; xQP){ yading@11: const int q=(src[l4] - src[l5])/2; yading@11: const int leftEnergy= 5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]); yading@11: const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]); yading@11: yading@11: int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) ); yading@11: d= FFMAX(d, 0); yading@11: yading@11: d= (5*d + 32) >> 6; yading@11: d*= FFSIGN(-middleEnergy); yading@11: yading@11: if(q>0){ yading@11: d= d<0 ? 0 : d; yading@11: d= d>q ? q : d; yading@11: }else{ yading@11: d= d>0 ? 0 : d; yading@11: d= d a ? 0 : -1 yading@11: "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L20 > a ? 0 : -1 yading@11: "pcmpeqb "MANGLE(b00)", %%mm3 \n\t" // L00 > a ? 0 : -1 yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: "paddb %%mm3, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a"), %%mm2 \n\t" // L11 yading@11: "movq %%mm2, %%mm3 \n\t" // L11 yading@11: "movq %%mm2, %%mm4 \n\t" // L11 yading@11: "psllq $8, %%mm3 \n\t" yading@11: "psrlq $8, %%mm4 \n\t" yading@11: "movd -4(%%"REG_a"), %%mm5 \n\t" yading@11: "movd 8(%%"REG_a"), %%mm6 \n\t" yading@11: "psrlq $24, %%mm5 \n\t" yading@11: "psllq $56, %%mm6 \n\t" yading@11: "por %%mm5, %%mm3 \n\t" // L01 yading@11: "por %%mm6, %%mm4 \n\t" // L21 yading@11: "movq %%mm3, %%mm5 \n\t" // L01 yading@11: PAVGB(%%mm4, %%mm3) // (L21 + L01)/2 yading@11: PAVGB(%%mm2, %%mm3) // (L21 + L01 + 2L11)/4 yading@11: "psubusb %%mm7, %%mm2 \n\t" yading@11: "psubusb %%mm7, %%mm4 \n\t" yading@11: "psubusb %%mm7, %%mm5 \n\t" yading@11: "pcmpeqb "MANGLE(b00)", %%mm2 \n\t" // L11 > a ? 0 : -1 yading@11: "pcmpeqb "MANGLE(b00)", %%mm4 \n\t" // L21 > a ? 0 : -1 yading@11: "pcmpeqb "MANGLE(b00)", %%mm5 \n\t" // L01 > a ? 0 : -1 yading@11: "paddb %%mm4, %%mm2 \n\t" yading@11: "paddb %%mm5, %%mm2 \n\t" yading@11: // 0, 2, 3, 1 yading@11: #define REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ yading@11: "movq " #src ", " #sx " \n\t" /* src[0] */\ yading@11: "movq " #sx ", " #lx " \n\t" /* src[0] */\ yading@11: "movq " #sx ", " #t0 " \n\t" /* src[0] */\ yading@11: "psllq $8, " #lx " \n\t"\ yading@11: "psrlq $8, " #t0 " \n\t"\ yading@11: "movd -4" #src ", " #t1 " \n\t"\ yading@11: "psrlq $24, " #t1 " \n\t"\ yading@11: "por " #t1 ", " #lx " \n\t" /* src[-1] */\ yading@11: "movd 8" #src ", " #t1 " \n\t"\ yading@11: "psllq $56, " #t1 " \n\t"\ yading@11: "por " #t1 ", " #t0 " \n\t" /* src[+1] */\ yading@11: "movq " #lx ", " #t1 " \n\t" /* src[-1] */\ yading@11: PAVGB(t0, lx) /* (src[-1] + src[+1])/2 */\ yading@11: PAVGB(sx, lx) /* (src[-1] + 2src[0] + src[+1])/4 */\ yading@11: PAVGB(lx, pplx) \ yading@11: "movq " #lx ", 8(%4) \n\t"\ yading@11: "movq (%4), " #lx " \n\t"\ yading@11: "psubusb " #lx ", " #t1 " \n\t"\ yading@11: "psubusb " #lx ", " #t0 " \n\t"\ yading@11: "psubusb " #lx ", " #sx " \n\t"\ yading@11: "movq "MANGLE(b00)", " #lx " \n\t"\ yading@11: "pcmpeqb " #lx ", " #t1 " \n\t" /* src[-1] > a ? 0 : -1*/\ yading@11: "pcmpeqb " #lx ", " #t0 " \n\t" /* src[+1] > a ? 0 : -1*/\ yading@11: "pcmpeqb " #lx ", " #sx " \n\t" /* src[0] > a ? 0 : -1*/\ yading@11: "paddb " #t1 ", " #t0 " \n\t"\ yading@11: "paddb " #t0 ", " #sx " \n\t"\ yading@11: \ yading@11: PAVGB(plx, pplx) /* filtered */\ yading@11: "movq " #dst ", " #t0 " \n\t" /* dst */\ yading@11: "movq " #t0 ", " #t1 " \n\t" /* dst */\ yading@11: "psubusb %3, " #t0 " \n\t"\ yading@11: "paddusb %3, " #t1 " \n\t"\ yading@11: PMAXUB(t0, pplx)\ yading@11: PMINUB(t1, pplx, t0)\ yading@11: "paddb " #sx ", " #ppsx " \n\t"\ yading@11: "paddb " #psx ", " #ppsx " \n\t"\ yading@11: "#paddb "MANGLE(b02)", " #ppsx " \n\t"\ yading@11: "pand "MANGLE(b08)", " #ppsx " \n\t"\ yading@11: "pcmpeqb " #lx ", " #ppsx " \n\t"\ yading@11: "pand " #ppsx ", " #pplx " \n\t"\ yading@11: "pandn " #dst ", " #ppsx " \n\t"\ yading@11: "por " #pplx ", " #ppsx " \n\t"\ yading@11: "movq " #ppsx ", " #dst " \n\t"\ yading@11: "movq 8(%4), " #lx " \n\t" yading@11: yading@11: #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \ yading@11: REAL_DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) yading@11: /* yading@11: 0000000 yading@11: 1111111 yading@11: yading@11: 1111110 yading@11: 1111101 yading@11: 1111100 yading@11: 1111011 yading@11: 1111010 yading@11: 1111001 yading@11: yading@11: 1111000 yading@11: 1110111 yading@11: yading@11: */ yading@11: //DERING_CORE(dst ,src ,ppsx ,psx ,sx ,pplx ,plx ,lx ,t0 ,t1) yading@11: DERING_CORE((%%REGa) ,(%%REGa, %1) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) yading@11: DERING_CORE((%%REGa, %1) ,(%%REGa, %1, 2),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) yading@11: DERING_CORE((%%REGa, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) yading@11: DERING_CORE((%0, %1, 4) ,(%%REGd) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) yading@11: DERING_CORE((%%REGd) ,(%%REGd, %1) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) yading@11: DERING_CORE((%%REGd, %1) ,(%%REGd, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7) yading@11: DERING_CORE((%%REGd, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7) yading@11: DERING_CORE((%0, %1, 8) ,(%%REGd, %1, 4),%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7) yading@11: yading@11: "1: \n\t" yading@11: : : "r" (src), "r" ((x86_reg)stride), "m" (c->pQPb), "m"(c->pQPb2), "q"(tmp) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #else // HAVE_7REGS && (TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW) yading@11: int y; yading@11: int min=255; yading@11: int max=0; yading@11: int avg; yading@11: uint8_t *p; yading@11: int s[10]; yading@11: const int QP2= c->QP/2 + 1; yading@11: yading@11: src --; yading@11: for(y=1; y<9; y++){ yading@11: int x; yading@11: p= src + stride*y; yading@11: for(x=1; x<9; x++){ yading@11: p++; yading@11: if(*p > max) max= *p; yading@11: if(*p < min) min= *p; yading@11: } yading@11: } yading@11: avg= (min + max + 1)>>1; yading@11: yading@11: if(max - min avg) t+= 1; yading@11: if(src[stride*y + 1] > avg) t+= 2; yading@11: if(src[stride*y + 2] > avg) t+= 4; yading@11: if(src[stride*y + 3] > avg) t+= 8; yading@11: if(src[stride*y + 4] > avg) t+= 16; yading@11: if(src[stride*y + 5] > avg) t+= 32; yading@11: if(src[stride*y + 6] > avg) t+= 64; yading@11: if(src[stride*y + 7] > avg) t+= 128; yading@11: if(src[stride*y + 8] > avg) t+= 256; yading@11: if(src[stride*y + 9] > avg) t+= 512; yading@11: yading@11: t |= (~t)<<16; yading@11: t &= (t<<1) & (t>>1); yading@11: s[y] = t; yading@11: } yading@11: yading@11: for(y=1; y<9; y++){ yading@11: int t = s[y-1] & s[y] & s[y+1]; yading@11: t|= t>>16; yading@11: s[y-1]= t; yading@11: } yading@11: yading@11: for(y=1; y<9; y++){ yading@11: int x; yading@11: int t = s[y-1]; yading@11: yading@11: p= src + stride*y; yading@11: for(x=1; x<9; x++){ yading@11: p++; yading@11: if(t & (1<>4; yading@11: yading@11: #ifdef DEBUG_DERING_THRESHOLD yading@11: __asm__ volatile("emms\n\t":); yading@11: { yading@11: static long long numPixels=0; yading@11: if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++; yading@11: // if((max-min)<20 || (max-min)*QP<200) yading@11: // if((max-min)*QP < 500) yading@11: // if(max-min worstDiff){ yading@11: worstDiff= absDiff; yading@11: worstQP= QP; yading@11: worstRange= max-min; yading@11: } yading@11: errorSum+= error; yading@11: yading@11: if(1024LL*1024LL*1024LL % numSkipped == 0){ yading@11: av_log(c, AV_LOG_INFO, "sum:%1.3f, skip:%d, wQP:%d, " yading@11: "wRange:%d, wDiff:%d, relSkip:%1.3f\n", yading@11: (float)errorSum/numSkipped, numSkipped, worstQP, worstRange, yading@11: worstDiff, (float)numSkipped/numPixels); yading@11: } yading@11: } yading@11: } yading@11: #endif yading@11: if (*p + QP2 < f) *p= *p + QP2; yading@11: else if(*p - QP2 > f) *p= *p - QP2; yading@11: else *p=f; yading@11: } yading@11: } yading@11: } yading@11: #ifdef DEBUG_DERING_THRESHOLD yading@11: if(max-min < 20){ yading@11: for(y=1; y<9; y++){ yading@11: int x; yading@11: int t = 0; yading@11: p= src + stride*y; yading@11: for(x=1; x<9; x++){ yading@11: p++; yading@11: *p = FFMIN(*p + 20, 255); yading@11: } yading@11: } yading@11: // src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255; yading@11: } yading@11: #endif yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: #endif //TEMPLATE_PP_ALTIVEC yading@11: yading@11: /** yading@11: * Deinterlace the given block by linearly interpolating every second line. yading@11: * will be called for every 8x8 block and can read & write from line 4-15 yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: */ yading@11: static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= 4*stride; yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_c" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %0+8%1 ecx+4%1 yading@11: yading@11: "movq (%0), %%mm0 \n\t" yading@11: "movq (%%"REG_a", %1), %%mm1 \n\t" yading@11: PAVGB(%%mm1, %%mm0) yading@11: "movq %%mm0, (%%"REG_a") \n\t" yading@11: "movq (%0, %1, 4), %%mm0 \n\t" yading@11: PAVGB(%%mm0, %%mm1) yading@11: "movq %%mm1, (%%"REG_a", %1, 2) \n\t" yading@11: "movq (%%"REG_c", %1), %%mm1 \n\t" yading@11: PAVGB(%%mm1, %%mm0) yading@11: "movq %%mm0, (%%"REG_c") \n\t" yading@11: "movq (%0, %1, 8), %%mm0 \n\t" yading@11: PAVGB(%%mm0, %%mm1) yading@11: "movq %%mm1, (%%"REG_c", %1, 2) \n\t" yading@11: yading@11: : : "r" (src), "r" ((x86_reg)stride) yading@11: : "%"REG_a, "%"REG_c yading@11: ); yading@11: #else yading@11: int a, b, x; yading@11: src+= 4*stride; yading@11: yading@11: for(x=0; x<2; x++){ yading@11: a= *(uint32_t*)&src[stride*0]; yading@11: b= *(uint32_t*)&src[stride*2]; yading@11: *(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: a= *(uint32_t*)&src[stride*4]; yading@11: *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: b= *(uint32_t*)&src[stride*6]; yading@11: *(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: a= *(uint32_t*)&src[stride*8]; yading@11: *(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: src += 4; yading@11: } yading@11: #endif yading@11: } yading@11: yading@11: /** yading@11: * Deinterlace the given block by cubic interpolating every second line. yading@11: * will be called for every 8x8 block and can read & write from line 4-15 yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: * this filter will read lines 3-15 and write 7-13 yading@11: */ yading@11: static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride) yading@11: { yading@11: #if TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= stride*3; yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: "lea (%%"REG_d", %1, 4), %%"REG_c" \n\t" yading@11: "add %1, %%"REG_c" \n\t" yading@11: #if TEMPLATE_PP_SSE2 yading@11: "pxor %%xmm7, %%xmm7 \n\t" yading@11: #define REAL_DEINT_CUBIC(a,b,c,d,e)\ yading@11: "movq " #a ", %%xmm0 \n\t"\ yading@11: "movq " #b ", %%xmm1 \n\t"\ yading@11: "movq " #d ", %%xmm2 \n\t"\ yading@11: "movq " #e ", %%xmm3 \n\t"\ yading@11: "pavgb %%xmm2, %%xmm1 \n\t"\ yading@11: "pavgb %%xmm3, %%xmm0 \n\t"\ yading@11: "punpcklbw %%xmm7, %%xmm0 \n\t"\ yading@11: "punpcklbw %%xmm7, %%xmm1 \n\t"\ yading@11: "psubw %%xmm1, %%xmm0 \n\t"\ yading@11: "psraw $3, %%xmm0 \n\t"\ yading@11: "psubw %%xmm0, %%xmm1 \n\t"\ yading@11: "packuswb %%xmm1, %%xmm1 \n\t"\ yading@11: "movlps %%xmm1, " #c " \n\t" yading@11: #else //TEMPLATE_PP_SSE2 yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 10 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx yading@11: yading@11: #define REAL_DEINT_CUBIC(a,b,c,d,e)\ yading@11: "movq " #a ", %%mm0 \n\t"\ yading@11: "movq " #b ", %%mm1 \n\t"\ yading@11: "movq " #d ", %%mm2 \n\t"\ yading@11: "movq " #e ", %%mm3 \n\t"\ yading@11: PAVGB(%%mm2, %%mm1) /* (b+d) /2 */\ yading@11: PAVGB(%%mm3, %%mm0) /* (a+e) /2 */\ yading@11: "movq %%mm0, %%mm2 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm0 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm2 \n\t"\ yading@11: "movq %%mm1, %%mm3 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm1 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm3 \n\t"\ yading@11: "psubw %%mm1, %%mm0 \n\t" /* L(a+e - (b+d))/2 */\ yading@11: "psubw %%mm3, %%mm2 \n\t" /* H(a+e - (b+d))/2 */\ yading@11: "psraw $3, %%mm0 \n\t" /* L(a+e - (b+d))/16 */\ yading@11: "psraw $3, %%mm2 \n\t" /* H(a+e - (b+d))/16 */\ yading@11: "psubw %%mm0, %%mm1 \n\t" /* L(9b + 9d - a - e)/16 */\ yading@11: "psubw %%mm2, %%mm3 \n\t" /* H(9b + 9d - a - e)/16 */\ yading@11: "packuswb %%mm3, %%mm1 \n\t"\ yading@11: "movq %%mm1, " #c " \n\t" yading@11: #endif //TEMPLATE_PP_SSE2 yading@11: #define DEINT_CUBIC(a,b,c,d,e) REAL_DEINT_CUBIC(a,b,c,d,e) yading@11: yading@11: DEINT_CUBIC((%0) , (%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd, %1)) yading@11: DEINT_CUBIC((%%REGa, %1), (%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%0, %1, 8)) yading@11: DEINT_CUBIC((%0, %1, 4) , (%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGc)) yading@11: DEINT_CUBIC((%%REGd, %1), (%0, %1, 8) , (%%REGd, %1, 4), (%%REGc) , (%%REGc, %1, 2)) yading@11: yading@11: : : "r" (src), "r" ((x86_reg)stride) yading@11: : yading@11: #if TEMPLATE_PP_SSE2 yading@11: XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm7",) yading@11: #endif yading@11: "%"REG_a, "%"REG_d, "%"REG_c yading@11: ); yading@11: #undef REAL_DEINT_CUBIC yading@11: #else //TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: int x; yading@11: src+= stride*3; yading@11: for(x=0; x<8; x++){ yading@11: src[stride*3] = CLIP((-src[0] + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4); yading@11: src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4); yading@11: src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4); yading@11: src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4); yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_SSE2 || TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: yading@11: /** yading@11: * Deinterlace the given block by filtering every second line with a (-1 4 2 4 -1) filter. yading@11: * will be called for every 8x8 block and can read & write from line 4-15 yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: * this filter will read lines 4-13 and write 5-11 yading@11: */ yading@11: static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= stride*4; yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: "movq (%2), %%mm0 \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 10 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx yading@11: yading@11: #define REAL_DEINT_FF(a,b,c,d)\ yading@11: "movq " #a ", %%mm1 \n\t"\ yading@11: "movq " #b ", %%mm2 \n\t"\ yading@11: "movq " #c ", %%mm3 \n\t"\ yading@11: "movq " #d ", %%mm4 \n\t"\ yading@11: PAVGB(%%mm3, %%mm1) \ yading@11: PAVGB(%%mm4, %%mm0) \ yading@11: "movq %%mm0, %%mm3 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm0 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm3 \n\t"\ yading@11: "movq %%mm1, %%mm4 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm1 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm4 \n\t"\ yading@11: "psllw $2, %%mm1 \n\t"\ yading@11: "psllw $2, %%mm4 \n\t"\ yading@11: "psubw %%mm0, %%mm1 \n\t"\ yading@11: "psubw %%mm3, %%mm4 \n\t"\ yading@11: "movq %%mm2, %%mm5 \n\t"\ yading@11: "movq %%mm2, %%mm0 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm2 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm5 \n\t"\ yading@11: "paddw %%mm2, %%mm1 \n\t"\ yading@11: "paddw %%mm5, %%mm4 \n\t"\ yading@11: "psraw $2, %%mm1 \n\t"\ yading@11: "psraw $2, %%mm4 \n\t"\ yading@11: "packuswb %%mm4, %%mm1 \n\t"\ yading@11: "movq %%mm1, " #b " \n\t"\ yading@11: yading@11: #define DEINT_FF(a,b,c,d) REAL_DEINT_FF(a,b,c,d) yading@11: yading@11: DEINT_FF((%0) , (%%REGa) , (%%REGa, %1), (%%REGa, %1, 2)) yading@11: DEINT_FF((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) ) yading@11: DEINT_FF((%0, %1, 4) , (%%REGd) , (%%REGd, %1), (%%REGd, %1, 2)) yading@11: DEINT_FF((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) yading@11: yading@11: "movq %%mm0, (%2) \n\t" yading@11: : : "r" (src), "r" ((x86_reg)stride), "r"(tmp) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: int x; yading@11: src+= stride*4; yading@11: for(x=0; x<8; x++){ yading@11: int t1= tmp[x]; yading@11: int t2= src[stride*1]; yading@11: yading@11: src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3); yading@11: t1= src[stride*4]; yading@11: src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3); yading@11: t2= src[stride*6]; yading@11: src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3); yading@11: t1= src[stride*8]; yading@11: src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3); yading@11: tmp[x]= t1; yading@11: yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: yading@11: /** yading@11: * Deinterlace the given block by filtering every line with a (-1 2 6 2 -1) filter. yading@11: * will be called for every 8x8 block and can read & write from line 4-15 yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: * this filter will read lines 4-13 and write 4-11 yading@11: */ yading@11: static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= stride*4; yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: "movq (%2), %%mm0 \n\t" yading@11: "movq (%3), %%mm1 \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 10 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 ecx yading@11: yading@11: #define REAL_DEINT_L5(t1,t2,a,b,c)\ yading@11: "movq " #a ", %%mm2 \n\t"\ yading@11: "movq " #b ", %%mm3 \n\t"\ yading@11: "movq " #c ", %%mm4 \n\t"\ yading@11: PAVGB(t2, %%mm3) \ yading@11: PAVGB(t1, %%mm4) \ yading@11: "movq %%mm2, %%mm5 \n\t"\ yading@11: "movq %%mm2, " #t1 " \n\t"\ yading@11: "punpcklbw %%mm7, %%mm2 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm5 \n\t"\ yading@11: "movq %%mm2, %%mm6 \n\t"\ yading@11: "paddw %%mm2, %%mm2 \n\t"\ yading@11: "paddw %%mm6, %%mm2 \n\t"\ yading@11: "movq %%mm5, %%mm6 \n\t"\ yading@11: "paddw %%mm5, %%mm5 \n\t"\ yading@11: "paddw %%mm6, %%mm5 \n\t"\ yading@11: "movq %%mm3, %%mm6 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm3 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm6 \n\t"\ yading@11: "paddw %%mm3, %%mm3 \n\t"\ yading@11: "paddw %%mm6, %%mm6 \n\t"\ yading@11: "paddw %%mm3, %%mm2 \n\t"\ yading@11: "paddw %%mm6, %%mm5 \n\t"\ yading@11: "movq %%mm4, %%mm6 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm4 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm6 \n\t"\ yading@11: "psubw %%mm4, %%mm2 \n\t"\ yading@11: "psubw %%mm6, %%mm5 \n\t"\ yading@11: "psraw $2, %%mm2 \n\t"\ yading@11: "psraw $2, %%mm5 \n\t"\ yading@11: "packuswb %%mm5, %%mm2 \n\t"\ yading@11: "movq %%mm2, " #a " \n\t"\ yading@11: yading@11: #define DEINT_L5(t1,t2,a,b,c) REAL_DEINT_L5(t1,t2,a,b,c) yading@11: yading@11: DEINT_L5(%%mm0, %%mm1, (%0) , (%%REGa) , (%%REGa, %1) ) yading@11: DEINT_L5(%%mm1, %%mm0, (%%REGa) , (%%REGa, %1) , (%%REGa, %1, 2)) yading@11: DEINT_L5(%%mm0, %%mm1, (%%REGa, %1) , (%%REGa, %1, 2), (%0, %1, 4) ) yading@11: DEINT_L5(%%mm1, %%mm0, (%%REGa, %1, 2), (%0, %1, 4) , (%%REGd) ) yading@11: DEINT_L5(%%mm0, %%mm1, (%0, %1, 4) , (%%REGd) , (%%REGd, %1) ) yading@11: DEINT_L5(%%mm1, %%mm0, (%%REGd) , (%%REGd, %1) , (%%REGd, %1, 2)) yading@11: DEINT_L5(%%mm0, %%mm1, (%%REGd, %1) , (%%REGd, %1, 2), (%0, %1, 8) ) yading@11: DEINT_L5(%%mm1, %%mm0, (%%REGd, %1, 2), (%0, %1, 8) , (%%REGd, %1, 4)) yading@11: yading@11: "movq %%mm0, (%2) \n\t" yading@11: "movq %%mm1, (%3) \n\t" yading@11: : : "r" (src), "r" ((x86_reg)stride), "r"(tmp), "r"(tmp2) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: int x; yading@11: src+= stride*4; yading@11: for(x=0; x<8; x++){ yading@11: int t1= tmp[x]; yading@11: int t2= tmp2[x]; yading@11: int t3= src[0]; yading@11: yading@11: src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3); yading@11: t1= src[stride*1]; yading@11: src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3); yading@11: t2= src[stride*2]; yading@11: src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3); yading@11: t3= src[stride*3]; yading@11: src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3); yading@11: t1= src[stride*4]; yading@11: src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3); yading@11: t2= src[stride*5]; yading@11: src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3); yading@11: t3= src[stride*6]; yading@11: src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3); yading@11: t1= src[stride*7]; yading@11: src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3); yading@11: yading@11: tmp[x]= t3; yading@11: tmp2[x]= t1; yading@11: yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: yading@11: /** yading@11: * Deinterlace the given block by filtering all lines with a (1 2 1) filter. yading@11: * will be called for every 8x8 block and can read & write from line 4-15 yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: * this filter will read lines 4-13 and write 4-11 yading@11: */ yading@11: static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride, uint8_t *tmp) yading@11: { yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: src+= 4*stride; yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 yading@11: yading@11: "movq (%2), %%mm0 \n\t" // L0 yading@11: "movq (%%"REG_a"), %%mm1 \n\t" // L2 yading@11: PAVGB(%%mm1, %%mm0) // L0+L2 yading@11: "movq (%0), %%mm2 \n\t" // L1 yading@11: PAVGB(%%mm2, %%mm0) yading@11: "movq %%mm0, (%0) \n\t" yading@11: "movq (%%"REG_a", %1), %%mm0 \n\t" // L3 yading@11: PAVGB(%%mm0, %%mm2) // L1+L3 yading@11: PAVGB(%%mm1, %%mm2) // 2L2 + L1 + L3 yading@11: "movq %%mm2, (%%"REG_a") \n\t" yading@11: "movq (%%"REG_a", %1, 2), %%mm2 \n\t" // L4 yading@11: PAVGB(%%mm2, %%mm1) // L2+L4 yading@11: PAVGB(%%mm0, %%mm1) // 2L3 + L2 + L4 yading@11: "movq %%mm1, (%%"REG_a", %1) \n\t" yading@11: "movq (%0, %1, 4), %%mm1 \n\t" // L5 yading@11: PAVGB(%%mm1, %%mm0) // L3+L5 yading@11: PAVGB(%%mm2, %%mm0) // 2L4 + L3 + L5 yading@11: "movq %%mm0, (%%"REG_a", %1, 2) \n\t" yading@11: "movq (%%"REG_d"), %%mm0 \n\t" // L6 yading@11: PAVGB(%%mm0, %%mm2) // L4+L6 yading@11: PAVGB(%%mm1, %%mm2) // 2L5 + L4 + L6 yading@11: "movq %%mm2, (%0, %1, 4) \n\t" yading@11: "movq (%%"REG_d", %1), %%mm2 \n\t" // L7 yading@11: PAVGB(%%mm2, %%mm1) // L5+L7 yading@11: PAVGB(%%mm0, %%mm1) // 2L6 + L5 + L7 yading@11: "movq %%mm1, (%%"REG_d") \n\t" yading@11: "movq (%%"REG_d", %1, 2), %%mm1 \n\t" // L8 yading@11: PAVGB(%%mm1, %%mm0) // L6+L8 yading@11: PAVGB(%%mm2, %%mm0) // 2L7 + L6 + L8 yading@11: "movq %%mm0, (%%"REG_d", %1) \n\t" yading@11: "movq (%0, %1, 8), %%mm0 \n\t" // L9 yading@11: PAVGB(%%mm0, %%mm2) // L7+L9 yading@11: PAVGB(%%mm1, %%mm2) // 2L8 + L7 + L9 yading@11: "movq %%mm2, (%%"REG_d", %1, 2) \n\t" yading@11: "movq %%mm1, (%2) \n\t" yading@11: yading@11: : : "r" (src), "r" ((x86_reg)stride), "r" (tmp) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: int a, b, c, x; yading@11: src+= 4*stride; yading@11: yading@11: for(x=0; x<2; x++){ yading@11: a= *(uint32_t*)&tmp[stride*0]; yading@11: b= *(uint32_t*)&src[stride*0]; yading@11: c= *(uint32_t*)&src[stride*1]; yading@11: a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: a= *(uint32_t*)&src[stride*2]; yading@11: b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: b= *(uint32_t*)&src[stride*3]; yading@11: c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); yading@11: yading@11: c= *(uint32_t*)&src[stride*4]; yading@11: a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: a= *(uint32_t*)&src[stride*5]; yading@11: b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: b= *(uint32_t*)&src[stride*6]; yading@11: c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1); yading@11: yading@11: c= *(uint32_t*)&src[stride*7]; yading@11: a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: a= *(uint32_t*)&src[stride*8]; yading@11: b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1); yading@11: *(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1); yading@11: yading@11: *(uint32_t*)&tmp[stride*0]= c; yading@11: src += 4; yading@11: tmp += 4; yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: yading@11: /** yading@11: * Deinterlace the given block by applying a median filter to every second line. yading@11: * will be called for every 8x8 block and can read & write from line 4-15, yading@11: * lines 0-3 have been passed through the deblock / dering filters already, but can be read, too. yading@11: * lines 4-12 will be read into the deblocking filter and should be deinterlaced yading@11: */ yading@11: static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride) yading@11: { yading@11: #if TEMPLATE_PP_MMX yading@11: src+= 4*stride; yading@11: #if TEMPLATE_PP_MMXEXT yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 yading@11: yading@11: "movq (%0), %%mm0 \n\t" // yading@11: "movq (%%"REG_a", %1), %%mm2 \n\t" // yading@11: "movq (%%"REG_a"), %%mm1 \n\t" // yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "pmaxub %%mm1, %%mm0 \n\t" // yading@11: "pminub %%mm3, %%mm1 \n\t" // yading@11: "pmaxub %%mm2, %%mm1 \n\t" // yading@11: "pminub %%mm1, %%mm0 \n\t" yading@11: "movq %%mm0, (%%"REG_a") \n\t" yading@11: yading@11: "movq (%0, %1, 4), %%mm0 \n\t" // yading@11: "movq (%%"REG_a", %1, 2), %%mm1 \n\t" // yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "pmaxub %%mm1, %%mm2 \n\t" // yading@11: "pminub %%mm3, %%mm1 \n\t" // yading@11: "pmaxub %%mm0, %%mm1 \n\t" // yading@11: "pminub %%mm1, %%mm2 \n\t" yading@11: "movq %%mm2, (%%"REG_a", %1, 2) \n\t" yading@11: yading@11: "movq (%%"REG_d"), %%mm2 \n\t" // yading@11: "movq (%%"REG_d", %1), %%mm1 \n\t" // yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "pmaxub %%mm0, %%mm2 \n\t" // yading@11: "pminub %%mm3, %%mm0 \n\t" // yading@11: "pmaxub %%mm1, %%mm0 \n\t" // yading@11: "pminub %%mm0, %%mm2 \n\t" yading@11: "movq %%mm2, (%%"REG_d") \n\t" yading@11: yading@11: "movq (%%"REG_d", %1, 2), %%mm2 \n\t" // yading@11: "movq (%0, %1, 8), %%mm0 \n\t" // yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "pmaxub %%mm0, %%mm2 \n\t" // yading@11: "pminub %%mm3, %%mm0 \n\t" // yading@11: "pmaxub %%mm1, %%mm0 \n\t" // yading@11: "pminub %%mm0, %%mm2 \n\t" yading@11: "movq %%mm2, (%%"REG_d", %1, 2) \n\t" yading@11: yading@11: yading@11: : : "r" (src), "r" ((x86_reg)stride) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: yading@11: #else // MMX without MMX2 yading@11: __asm__ volatile( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a", %1, 4), %%"REG_d" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: yading@11: #define REAL_MEDIAN(a,b,c)\ yading@11: "movq " #a ", %%mm0 \n\t"\ yading@11: "movq " #b ", %%mm2 \n\t"\ yading@11: "movq " #c ", %%mm1 \n\t"\ yading@11: "movq %%mm0, %%mm3 \n\t"\ yading@11: "movq %%mm1, %%mm4 \n\t"\ yading@11: "movq %%mm2, %%mm5 \n\t"\ yading@11: "psubusb %%mm1, %%mm3 \n\t"\ yading@11: "psubusb %%mm2, %%mm4 \n\t"\ yading@11: "psubusb %%mm0, %%mm5 \n\t"\ yading@11: "pcmpeqb %%mm7, %%mm3 \n\t"\ yading@11: "pcmpeqb %%mm7, %%mm4 \n\t"\ yading@11: "pcmpeqb %%mm7, %%mm5 \n\t"\ yading@11: "movq %%mm3, %%mm6 \n\t"\ yading@11: "pxor %%mm4, %%mm3 \n\t"\ yading@11: "pxor %%mm5, %%mm4 \n\t"\ yading@11: "pxor %%mm6, %%mm5 \n\t"\ yading@11: "por %%mm3, %%mm1 \n\t"\ yading@11: "por %%mm4, %%mm2 \n\t"\ yading@11: "por %%mm5, %%mm0 \n\t"\ yading@11: "pand %%mm2, %%mm0 \n\t"\ yading@11: "pand %%mm1, %%mm0 \n\t"\ yading@11: "movq %%mm0, " #b " \n\t" yading@11: #define MEDIAN(a,b,c) REAL_MEDIAN(a,b,c) yading@11: yading@11: MEDIAN((%0) , (%%REGa) , (%%REGa, %1)) yading@11: MEDIAN((%%REGa, %1), (%%REGa, %1, 2), (%0, %1, 4)) yading@11: MEDIAN((%0, %1, 4) , (%%REGd) , (%%REGd, %1)) yading@11: MEDIAN((%%REGd, %1), (%%REGd, %1, 2), (%0, %1, 8)) yading@11: yading@11: : : "r" (src), "r" ((x86_reg)stride) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #endif //TEMPLATE_PP_MMXEXT yading@11: #else //TEMPLATE_PP_MMX yading@11: int x, y; yading@11: src+= 4*stride; yading@11: // FIXME - there should be a way to do a few columns in parallel like w/mmx yading@11: for(x=0; x<8; x++){ yading@11: uint8_t *colsrc = src; yading@11: for (y=0; y<4; y++){ yading@11: int a, b, c, d, e, f; yading@11: a = colsrc[0 ]; yading@11: b = colsrc[stride ]; yading@11: c = colsrc[stride*2]; yading@11: d = (a-b)>>31; yading@11: e = (b-c)>>31; yading@11: f = (c-a)>>31; yading@11: colsrc[stride ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f)); yading@11: colsrc += stride*2; yading@11: } yading@11: src++; yading@11: } yading@11: #endif //TEMPLATE_PP_MMX yading@11: } yading@11: yading@11: #if TEMPLATE_PP_MMX yading@11: /** yading@11: * Transpose and shift the given 8x8 Block into dst1 and dst2. yading@11: */ yading@11: static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride) yading@11: { yading@11: __asm__( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 yading@11: "movq (%0), %%mm0 \n\t" // 12345678 yading@11: "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh yading@11: "movq %%mm0, %%mm2 \n\t" // 12345678 yading@11: "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d yading@11: "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h yading@11: yading@11: "movq (%%"REG_a", %1), %%mm1 \n\t" yading@11: "movq (%%"REG_a", %1, 2), %%mm3 \n\t" yading@11: "movq %%mm1, %%mm4 \n\t" yading@11: "punpcklbw %%mm3, %%mm1 \n\t" yading@11: "punpckhbw %%mm3, %%mm4 \n\t" yading@11: yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "punpcklwd %%mm1, %%mm0 \n\t" yading@11: "punpckhwd %%mm1, %%mm3 \n\t" yading@11: "movq %%mm2, %%mm1 \n\t" yading@11: "punpcklwd %%mm4, %%mm2 \n\t" yading@11: "punpckhwd %%mm4, %%mm1 \n\t" yading@11: yading@11: "movd %%mm0, 128(%2) \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "movd %%mm0, 144(%2) \n\t" yading@11: "movd %%mm3, 160(%2) \n\t" yading@11: "psrlq $32, %%mm3 \n\t" yading@11: "movd %%mm3, 176(%2) \n\t" yading@11: "movd %%mm3, 48(%3) \n\t" yading@11: "movd %%mm2, 192(%2) \n\t" yading@11: "movd %%mm2, 64(%3) \n\t" yading@11: "psrlq $32, %%mm2 \n\t" yading@11: "movd %%mm2, 80(%3) \n\t" yading@11: "movd %%mm1, 96(%3) \n\t" yading@11: "psrlq $32, %%mm1 \n\t" yading@11: "movd %%mm1, 112(%3) \n\t" yading@11: yading@11: "lea (%%"REG_a", %1, 4), %%"REG_a" \n\t" yading@11: yading@11: "movq (%0, %1, 4), %%mm0 \n\t" // 12345678 yading@11: "movq (%%"REG_a"), %%mm1 \n\t" // abcdefgh yading@11: "movq %%mm0, %%mm2 \n\t" // 12345678 yading@11: "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d yading@11: "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h yading@11: yading@11: "movq (%%"REG_a", %1), %%mm1 \n\t" yading@11: "movq (%%"REG_a", %1, 2), %%mm3 \n\t" yading@11: "movq %%mm1, %%mm4 \n\t" yading@11: "punpcklbw %%mm3, %%mm1 \n\t" yading@11: "punpckhbw %%mm3, %%mm4 \n\t" yading@11: yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "punpcklwd %%mm1, %%mm0 \n\t" yading@11: "punpckhwd %%mm1, %%mm3 \n\t" yading@11: "movq %%mm2, %%mm1 \n\t" yading@11: "punpcklwd %%mm4, %%mm2 \n\t" yading@11: "punpckhwd %%mm4, %%mm1 \n\t" yading@11: yading@11: "movd %%mm0, 132(%2) \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "movd %%mm0, 148(%2) \n\t" yading@11: "movd %%mm3, 164(%2) \n\t" yading@11: "psrlq $32, %%mm3 \n\t" yading@11: "movd %%mm3, 180(%2) \n\t" yading@11: "movd %%mm3, 52(%3) \n\t" yading@11: "movd %%mm2, 196(%2) \n\t" yading@11: "movd %%mm2, 68(%3) \n\t" yading@11: "psrlq $32, %%mm2 \n\t" yading@11: "movd %%mm2, 84(%3) \n\t" yading@11: "movd %%mm1, 100(%3) \n\t" yading@11: "psrlq $32, %%mm1 \n\t" yading@11: "movd %%mm1, 116(%3) \n\t" yading@11: yading@11: yading@11: :: "r" (src), "r" ((x86_reg)srcStride), "r" (dst1), "r" (dst2) yading@11: : "%"REG_a yading@11: ); yading@11: } yading@11: yading@11: /** yading@11: * Transpose the given 8x8 block. yading@11: */ yading@11: static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src) yading@11: { yading@11: __asm__( yading@11: "lea (%0, %1), %%"REG_a" \n\t" yading@11: "lea (%%"REG_a",%1,4), %%"REG_d" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 edx edx+%1 edx+2%1 %0+8%1 edx+4%1 yading@11: "movq (%2), %%mm0 \n\t" // 12345678 yading@11: "movq 16(%2), %%mm1 \n\t" // abcdefgh yading@11: "movq %%mm0, %%mm2 \n\t" // 12345678 yading@11: "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d yading@11: "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h yading@11: yading@11: "movq 32(%2), %%mm1 \n\t" yading@11: "movq 48(%2), %%mm3 \n\t" yading@11: "movq %%mm1, %%mm4 \n\t" yading@11: "punpcklbw %%mm3, %%mm1 \n\t" yading@11: "punpckhbw %%mm3, %%mm4 \n\t" yading@11: yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "punpcklwd %%mm1, %%mm0 \n\t" yading@11: "punpckhwd %%mm1, %%mm3 \n\t" yading@11: "movq %%mm2, %%mm1 \n\t" yading@11: "punpcklwd %%mm4, %%mm2 \n\t" yading@11: "punpckhwd %%mm4, %%mm1 \n\t" yading@11: yading@11: "movd %%mm0, (%0) \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "movd %%mm0, (%%"REG_a") \n\t" yading@11: "movd %%mm3, (%%"REG_a", %1) \n\t" yading@11: "psrlq $32, %%mm3 \n\t" yading@11: "movd %%mm3, (%%"REG_a", %1, 2) \n\t" yading@11: "movd %%mm2, (%0, %1, 4) \n\t" yading@11: "psrlq $32, %%mm2 \n\t" yading@11: "movd %%mm2, (%%"REG_d") \n\t" yading@11: "movd %%mm1, (%%"REG_d", %1) \n\t" yading@11: "psrlq $32, %%mm1 \n\t" yading@11: "movd %%mm1, (%%"REG_d", %1, 2) \n\t" yading@11: yading@11: yading@11: "movq 64(%2), %%mm0 \n\t" // 12345678 yading@11: "movq 80(%2), %%mm1 \n\t" // abcdefgh yading@11: "movq %%mm0, %%mm2 \n\t" // 12345678 yading@11: "punpcklbw %%mm1, %%mm0 \n\t" // 1a2b3c4d yading@11: "punpckhbw %%mm1, %%mm2 \n\t" // 5e6f7g8h yading@11: yading@11: "movq 96(%2), %%mm1 \n\t" yading@11: "movq 112(%2), %%mm3 \n\t" yading@11: "movq %%mm1, %%mm4 \n\t" yading@11: "punpcklbw %%mm3, %%mm1 \n\t" yading@11: "punpckhbw %%mm3, %%mm4 \n\t" yading@11: yading@11: "movq %%mm0, %%mm3 \n\t" yading@11: "punpcklwd %%mm1, %%mm0 \n\t" yading@11: "punpckhwd %%mm1, %%mm3 \n\t" yading@11: "movq %%mm2, %%mm1 \n\t" yading@11: "punpcklwd %%mm4, %%mm2 \n\t" yading@11: "punpckhwd %%mm4, %%mm1 \n\t" yading@11: yading@11: "movd %%mm0, 4(%0) \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "movd %%mm0, 4(%%"REG_a") \n\t" yading@11: "movd %%mm3, 4(%%"REG_a", %1) \n\t" yading@11: "psrlq $32, %%mm3 \n\t" yading@11: "movd %%mm3, 4(%%"REG_a", %1, 2) \n\t" yading@11: "movd %%mm2, 4(%0, %1, 4) \n\t" yading@11: "psrlq $32, %%mm2 \n\t" yading@11: "movd %%mm2, 4(%%"REG_d") \n\t" yading@11: "movd %%mm1, 4(%%"REG_d", %1) \n\t" yading@11: "psrlq $32, %%mm1 \n\t" yading@11: "movd %%mm1, 4(%%"REG_d", %1, 2) \n\t" yading@11: yading@11: :: "r" (dst), "r" ((x86_reg)dstStride), "r" (src) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: } yading@11: #endif //TEMPLATE_PP_MMX yading@11: //static long test=0; yading@11: yading@11: #if !TEMPLATE_PP_ALTIVEC yading@11: static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride, yading@11: uint8_t *tempBlurred, uint32_t *tempBlurredPast, int *maxNoise) yading@11: { yading@11: // to save a register (FIXME do this outside of the loops) yading@11: tempBlurredPast[127]= maxNoise[0]; yading@11: tempBlurredPast[128]= maxNoise[1]; yading@11: tempBlurredPast[129]= maxNoise[2]; yading@11: yading@11: #define FAST_L2_DIFF yading@11: //#define L1_DIFF //u should change the thresholds too if u try that one yading@11: #if TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: __asm__ volatile( yading@11: "lea (%2, %2, 2), %%"REG_a" \n\t" // 3*stride yading@11: "lea (%2, %2, 4), %%"REG_d" \n\t" // 5*stride yading@11: "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %x %x+%2 %x+2%2 %x+eax %x+4%2 %x+edx %x+2eax %x+ecx %x+8%2 yading@11: //FIXME reorder? yading@11: #ifdef L1_DIFF //needs mmx2 yading@11: "movq (%0), %%mm0 \n\t" // L0 yading@11: "psadbw (%1), %%mm0 \n\t" // |L0-R0| yading@11: "movq (%0, %2), %%mm1 \n\t" // L1 yading@11: "psadbw (%1, %2), %%mm1 \n\t" // |L1-R1| yading@11: "movq (%0, %2, 2), %%mm2 \n\t" // L2 yading@11: "psadbw (%1, %2, 2), %%mm2 \n\t" // |L2-R2| yading@11: "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 yading@11: "psadbw (%1, %%"REG_a"), %%mm3 \n\t" // |L3-R3| yading@11: yading@11: "movq (%0, %2, 4), %%mm4 \n\t" // L4 yading@11: "paddw %%mm1, %%mm0 \n\t" yading@11: "psadbw (%1, %2, 4), %%mm4 \n\t" // |L4-R4| yading@11: "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 yading@11: "paddw %%mm2, %%mm0 \n\t" yading@11: "psadbw (%1, %%"REG_d"), %%mm5 \n\t" // |L5-R5| yading@11: "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 yading@11: "paddw %%mm3, %%mm0 \n\t" yading@11: "psadbw (%1, %%"REG_a", 2), %%mm6 \n\t" // |L6-R6| yading@11: "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 yading@11: "paddw %%mm4, %%mm0 \n\t" yading@11: "psadbw (%1, %%"REG_c"), %%mm7 \n\t" // |L7-R7| yading@11: "paddw %%mm5, %%mm6 \n\t" yading@11: "paddw %%mm7, %%mm6 \n\t" yading@11: "paddw %%mm6, %%mm0 \n\t" yading@11: #else //L1_DIFF yading@11: #if defined (FAST_L2_DIFF) yading@11: "pcmpeqb %%mm7, %%mm7 \n\t" yading@11: "movq "MANGLE(b80)", %%mm6 \n\t" yading@11: "pxor %%mm0, %%mm0 \n\t" yading@11: #define REAL_L2_DIFF_CORE(a, b)\ yading@11: "movq " #a ", %%mm5 \n\t"\ yading@11: "movq " #b ", %%mm2 \n\t"\ yading@11: "pxor %%mm7, %%mm2 \n\t"\ yading@11: PAVGB(%%mm2, %%mm5)\ yading@11: "paddb %%mm6, %%mm5 \n\t"\ yading@11: "movq %%mm5, %%mm2 \n\t"\ yading@11: "psllw $8, %%mm5 \n\t"\ yading@11: "pmaddwd %%mm5, %%mm5 \n\t"\ yading@11: "pmaddwd %%mm2, %%mm2 \n\t"\ yading@11: "paddd %%mm2, %%mm5 \n\t"\ yading@11: "psrld $14, %%mm5 \n\t"\ yading@11: "paddd %%mm5, %%mm0 \n\t" yading@11: yading@11: #else //defined (FAST_L2_DIFF) yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: "pxor %%mm0, %%mm0 \n\t" yading@11: #define REAL_L2_DIFF_CORE(a, b)\ yading@11: "movq " #a ", %%mm5 \n\t"\ yading@11: "movq " #b ", %%mm2 \n\t"\ yading@11: "movq %%mm5, %%mm1 \n\t"\ yading@11: "movq %%mm2, %%mm3 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm5 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm1 \n\t"\ yading@11: "punpcklbw %%mm7, %%mm2 \n\t"\ yading@11: "punpckhbw %%mm7, %%mm3 \n\t"\ yading@11: "psubw %%mm2, %%mm5 \n\t"\ yading@11: "psubw %%mm3, %%mm1 \n\t"\ yading@11: "pmaddwd %%mm5, %%mm5 \n\t"\ yading@11: "pmaddwd %%mm1, %%mm1 \n\t"\ yading@11: "paddd %%mm1, %%mm5 \n\t"\ yading@11: "paddd %%mm5, %%mm0 \n\t" yading@11: yading@11: #endif //defined (FAST_L2_DIFF) yading@11: yading@11: #define L2_DIFF_CORE(a, b) REAL_L2_DIFF_CORE(a, b) yading@11: yading@11: L2_DIFF_CORE((%0) , (%1)) yading@11: L2_DIFF_CORE((%0, %2) , (%1, %2)) yading@11: L2_DIFF_CORE((%0, %2, 2) , (%1, %2, 2)) yading@11: L2_DIFF_CORE((%0, %%REGa) , (%1, %%REGa)) yading@11: L2_DIFF_CORE((%0, %2, 4) , (%1, %2, 4)) yading@11: L2_DIFF_CORE((%0, %%REGd) , (%1, %%REGd)) yading@11: L2_DIFF_CORE((%0, %%REGa,2), (%1, %%REGa,2)) yading@11: L2_DIFF_CORE((%0, %%REGc) , (%1, %%REGc)) yading@11: yading@11: #endif //L1_DIFF yading@11: yading@11: "movq %%mm0, %%mm4 \n\t" yading@11: "psrlq $32, %%mm0 \n\t" yading@11: "paddd %%mm0, %%mm4 \n\t" yading@11: "movd %%mm4, %%ecx \n\t" yading@11: "shll $2, %%ecx \n\t" yading@11: "mov %3, %%"REG_d" \n\t" yading@11: "addl -4(%%"REG_d"), %%ecx \n\t" yading@11: "addl 4(%%"REG_d"), %%ecx \n\t" yading@11: "addl -1024(%%"REG_d"), %%ecx \n\t" yading@11: "addl $4, %%ecx \n\t" yading@11: "addl 1024(%%"REG_d"), %%ecx \n\t" yading@11: "shrl $3, %%ecx \n\t" yading@11: "movl %%ecx, (%%"REG_d") \n\t" yading@11: yading@11: // "mov %3, %%"REG_c" \n\t" yading@11: // "mov %%"REG_c", test \n\t" yading@11: // "jmp 4f \n\t" yading@11: "cmpl 512(%%"REG_d"), %%ecx \n\t" yading@11: " jb 2f \n\t" yading@11: "cmpl 516(%%"REG_d"), %%ecx \n\t" yading@11: " jb 1f \n\t" yading@11: yading@11: "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride yading@11: "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride yading@11: "movq (%0), %%mm0 \n\t" // L0 yading@11: "movq (%0, %2), %%mm1 \n\t" // L1 yading@11: "movq (%0, %2, 2), %%mm2 \n\t" // L2 yading@11: "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 yading@11: "movq (%0, %2, 4), %%mm4 \n\t" // L4 yading@11: "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 yading@11: "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 yading@11: "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 yading@11: "movq %%mm0, (%1) \n\t" // L0 yading@11: "movq %%mm1, (%1, %2) \n\t" // L1 yading@11: "movq %%mm2, (%1, %2, 2) \n\t" // L2 yading@11: "movq %%mm3, (%1, %%"REG_a") \n\t" // L3 yading@11: "movq %%mm4, (%1, %2, 4) \n\t" // L4 yading@11: "movq %%mm5, (%1, %%"REG_d") \n\t" // L5 yading@11: "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // L6 yading@11: "movq %%mm7, (%1, %%"REG_c") \n\t" // L7 yading@11: "jmp 4f \n\t" yading@11: yading@11: "1: \n\t" yading@11: "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride yading@11: "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride yading@11: "movq (%0), %%mm0 \n\t" // L0 yading@11: PAVGB((%1), %%mm0) // L0 yading@11: "movq (%0, %2), %%mm1 \n\t" // L1 yading@11: PAVGB((%1, %2), %%mm1) // L1 yading@11: "movq (%0, %2, 2), %%mm2 \n\t" // L2 yading@11: PAVGB((%1, %2, 2), %%mm2) // L2 yading@11: "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 yading@11: PAVGB((%1, %%REGa), %%mm3) // L3 yading@11: "movq (%0, %2, 4), %%mm4 \n\t" // L4 yading@11: PAVGB((%1, %2, 4), %%mm4) // L4 yading@11: "movq (%0, %%"REG_d"), %%mm5 \n\t" // L5 yading@11: PAVGB((%1, %%REGd), %%mm5) // L5 yading@11: "movq (%0, %%"REG_a", 2), %%mm6 \n\t" // L6 yading@11: PAVGB((%1, %%REGa, 2), %%mm6) // L6 yading@11: "movq (%0, %%"REG_c"), %%mm7 \n\t" // L7 yading@11: PAVGB((%1, %%REGc), %%mm7) // L7 yading@11: "movq %%mm0, (%1) \n\t" // R0 yading@11: "movq %%mm1, (%1, %2) \n\t" // R1 yading@11: "movq %%mm2, (%1, %2, 2) \n\t" // R2 yading@11: "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 yading@11: "movq %%mm4, (%1, %2, 4) \n\t" // R4 yading@11: "movq %%mm5, (%1, %%"REG_d") \n\t" // R5 yading@11: "movq %%mm6, (%1, %%"REG_a", 2) \n\t" // R6 yading@11: "movq %%mm7, (%1, %%"REG_c") \n\t" // R7 yading@11: "movq %%mm0, (%0) \n\t" // L0 yading@11: "movq %%mm1, (%0, %2) \n\t" // L1 yading@11: "movq %%mm2, (%0, %2, 2) \n\t" // L2 yading@11: "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 yading@11: "movq %%mm4, (%0, %2, 4) \n\t" // L4 yading@11: "movq %%mm5, (%0, %%"REG_d") \n\t" // L5 yading@11: "movq %%mm6, (%0, %%"REG_a", 2) \n\t" // L6 yading@11: "movq %%mm7, (%0, %%"REG_c") \n\t" // L7 yading@11: "jmp 4f \n\t" yading@11: yading@11: "2: \n\t" yading@11: "cmpl 508(%%"REG_d"), %%ecx \n\t" yading@11: " jb 3f \n\t" yading@11: yading@11: "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride yading@11: "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride yading@11: "movq (%0), %%mm0 \n\t" // L0 yading@11: "movq (%0, %2), %%mm1 \n\t" // L1 yading@11: "movq (%0, %2, 2), %%mm2 \n\t" // L2 yading@11: "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 yading@11: "movq (%1), %%mm4 \n\t" // R0 yading@11: "movq (%1, %2), %%mm5 \n\t" // R1 yading@11: "movq (%1, %2, 2), %%mm6 \n\t" // R2 yading@11: "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: "movq %%mm0, (%1) \n\t" // R0 yading@11: "movq %%mm1, (%1, %2) \n\t" // R1 yading@11: "movq %%mm2, (%1, %2, 2) \n\t" // R2 yading@11: "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 yading@11: "movq %%mm0, (%0) \n\t" // L0 yading@11: "movq %%mm1, (%0, %2) \n\t" // L1 yading@11: "movq %%mm2, (%0, %2, 2) \n\t" // L2 yading@11: "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 yading@11: yading@11: "movq (%0, %2, 4), %%mm0 \n\t" // L4 yading@11: "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 yading@11: "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 yading@11: "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 yading@11: "movq (%1, %2, 4), %%mm4 \n\t" // R4 yading@11: "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 yading@11: "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 yading@11: "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: "movq %%mm0, (%1, %2, 4) \n\t" // R4 yading@11: "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 yading@11: "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 yading@11: "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 yading@11: "movq %%mm0, (%0, %2, 4) \n\t" // L4 yading@11: "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 yading@11: "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 yading@11: "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 yading@11: "jmp 4f \n\t" yading@11: yading@11: "3: \n\t" yading@11: "lea (%%"REG_a", %2, 2), %%"REG_d" \n\t" // 5*stride yading@11: "lea (%%"REG_d", %2, 2), %%"REG_c" \n\t" // 7*stride yading@11: "movq (%0), %%mm0 \n\t" // L0 yading@11: "movq (%0, %2), %%mm1 \n\t" // L1 yading@11: "movq (%0, %2, 2), %%mm2 \n\t" // L2 yading@11: "movq (%0, %%"REG_a"), %%mm3 \n\t" // L3 yading@11: "movq (%1), %%mm4 \n\t" // R0 yading@11: "movq (%1, %2), %%mm5 \n\t" // R1 yading@11: "movq (%1, %2, 2), %%mm6 \n\t" // R2 yading@11: "movq (%1, %%"REG_a"), %%mm7 \n\t" // R3 yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: "movq %%mm0, (%1) \n\t" // R0 yading@11: "movq %%mm1, (%1, %2) \n\t" // R1 yading@11: "movq %%mm2, (%1, %2, 2) \n\t" // R2 yading@11: "movq %%mm3, (%1, %%"REG_a") \n\t" // R3 yading@11: "movq %%mm0, (%0) \n\t" // L0 yading@11: "movq %%mm1, (%0, %2) \n\t" // L1 yading@11: "movq %%mm2, (%0, %2, 2) \n\t" // L2 yading@11: "movq %%mm3, (%0, %%"REG_a") \n\t" // L3 yading@11: yading@11: "movq (%0, %2, 4), %%mm0 \n\t" // L4 yading@11: "movq (%0, %%"REG_d"), %%mm1 \n\t" // L5 yading@11: "movq (%0, %%"REG_a", 2), %%mm2 \n\t" // L6 yading@11: "movq (%0, %%"REG_c"), %%mm3 \n\t" // L7 yading@11: "movq (%1, %2, 4), %%mm4 \n\t" // R4 yading@11: "movq (%1, %%"REG_d"), %%mm5 \n\t" // R5 yading@11: "movq (%1, %%"REG_a", 2), %%mm6 \n\t" // R6 yading@11: "movq (%1, %%"REG_c"), %%mm7 \n\t" // R7 yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: PAVGB(%%mm4, %%mm0) yading@11: PAVGB(%%mm5, %%mm1) yading@11: PAVGB(%%mm6, %%mm2) yading@11: PAVGB(%%mm7, %%mm3) yading@11: "movq %%mm0, (%1, %2, 4) \n\t" // R4 yading@11: "movq %%mm1, (%1, %%"REG_d") \n\t" // R5 yading@11: "movq %%mm2, (%1, %%"REG_a", 2) \n\t" // R6 yading@11: "movq %%mm3, (%1, %%"REG_c") \n\t" // R7 yading@11: "movq %%mm0, (%0, %2, 4) \n\t" // L4 yading@11: "movq %%mm1, (%0, %%"REG_d") \n\t" // L5 yading@11: "movq %%mm2, (%0, %%"REG_a", 2) \n\t" // L6 yading@11: "movq %%mm3, (%0, %%"REG_c") \n\t" // L7 yading@11: yading@11: "4: \n\t" yading@11: yading@11: :: "r" (src), "r" (tempBlurred), "r"((x86_reg)stride), "m" (tempBlurredPast) yading@11: : "%"REG_a, "%"REG_d, "%"REG_c, "memory" yading@11: ); yading@11: #else //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: { yading@11: int y; yading@11: int d=0; yading@11: // int sysd=0; yading@11: int i; yading@11: yading@11: for(y=0; y<8; y++){ yading@11: int x; yading@11: for(x=0; x<8; x++){ yading@11: int ref= tempBlurred[ x + y*stride ]; yading@11: int cur= src[ x + y*stride ]; yading@11: int d1=ref - cur; yading@11: // if(x==0 || x==7) d1+= d1>>1; yading@11: // if(y==0 || y==7) d1+= d1>>1; yading@11: // d+= FFABS(d1); yading@11: d+= d1*d1; yading@11: // sysd+= d1; yading@11: } yading@11: } yading@11: i=d; yading@11: d= ( yading@11: 4*d yading@11: +(*(tempBlurredPast-256)) yading@11: +(*(tempBlurredPast-1))+ (*(tempBlurredPast+1)) yading@11: +(*(tempBlurredPast+256)) yading@11: +4)>>3; yading@11: *tempBlurredPast=i; yading@11: // ((*tempBlurredPast)*3 + d + 2)>>2; yading@11: yading@11: /* yading@11: Switch between yading@11: 1 0 0 0 0 0 0 (0) yading@11: 64 32 16 8 4 2 1 (1) yading@11: 64 48 36 27 20 15 11 (33) (approx) yading@11: 64 56 49 43 37 33 29 (200) (approx) yading@11: */ yading@11: if(d > maxNoise[1]){ yading@11: if(d < maxNoise[2]){ yading@11: for(y=0; y<8; y++){ yading@11: int x; yading@11: for(x=0; x<8; x++){ yading@11: int ref= tempBlurred[ x + y*stride ]; yading@11: int cur= src[ x + y*stride ]; yading@11: tempBlurred[ x + y*stride ]= yading@11: src[ x + y*stride ]= yading@11: (ref + cur + 1)>>1; yading@11: } yading@11: } yading@11: }else{ yading@11: for(y=0; y<8; y++){ yading@11: int x; yading@11: for(x=0; x<8; x++){ yading@11: tempBlurred[ x + y*stride ]= src[ x + y*stride ]; yading@11: } yading@11: } yading@11: } yading@11: }else{ yading@11: if(d < maxNoise[0]){ yading@11: for(y=0; y<8; y++){ yading@11: int x; yading@11: for(x=0; x<8; x++){ yading@11: int ref= tempBlurred[ x + y*stride ]; yading@11: int cur= src[ x + y*stride ]; yading@11: tempBlurred[ x + y*stride ]= yading@11: src[ x + y*stride ]= yading@11: (ref*7 + cur + 4)>>3; yading@11: } yading@11: } yading@11: }else{ yading@11: for(y=0; y<8; y++){ yading@11: int x; yading@11: for(x=0; x<8; x++){ yading@11: int ref= tempBlurred[ x + y*stride ]; yading@11: int cur= src[ x + y*stride ]; yading@11: tempBlurred[ x + y*stride ]= yading@11: src[ x + y*stride ]= yading@11: (ref*3 + cur + 2)>>2; yading@11: } yading@11: } yading@11: } yading@11: } yading@11: } yading@11: #endif //TEMPLATE_PP_MMXEXT || TEMPLATE_PP_3DNOW yading@11: } yading@11: #endif //TEMPLATE_PP_ALTIVEC yading@11: yading@11: #if TEMPLATE_PP_MMX yading@11: /** yading@11: * accurate deblock filter yading@11: */ yading@11: static av_always_inline void RENAME(do_a_deblock)(uint8_t *src, int step, int stride, PPContext *c){ yading@11: int64_t dc_mask, eq_mask, both_masks; yading@11: int64_t sums[10*8*2]; yading@11: src+= step*3; // src points to begin of the 8x8 Block yading@11: //{ START_TIMER yading@11: __asm__ volatile( yading@11: "movq %0, %%mm7 \n\t" yading@11: "movq %1, %%mm6 \n\t" yading@11: : : "m" (c->mmxDcOffset[c->nonBQP]), "m" (c->mmxDcThreshold[c->nonBQP]) yading@11: ); yading@11: yading@11: __asm__ volatile( yading@11: "lea (%2, %3), %%"REG_a" \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %1 eax eax+%2 eax+2%2 %1+4%2 ecx ecx+%2 ecx+2%2 %1+8%2 ecx+4%2 yading@11: yading@11: "movq (%2), %%mm0 \n\t" yading@11: "movq (%%"REG_a"), %%mm1 \n\t" yading@11: "movq %%mm1, %%mm3 \n\t" yading@11: "movq %%mm1, %%mm4 \n\t" yading@11: "psubb %%mm1, %%mm0 \n\t" // mm0 = difference yading@11: "paddb %%mm7, %%mm0 \n\t" yading@11: "pcmpgtb %%mm6, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a",%3), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3, 2), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: yading@11: "lea (%%"REG_a", %3, 4), %%"REG_a" \n\t" yading@11: yading@11: "movq (%2, %3, 4), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a"), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3, 2), %%mm1 \n\t" yading@11: PMAXUB(%%mm1, %%mm4) yading@11: PMINUB(%%mm1, %%mm3, %%mm5) yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: yading@11: "movq (%2, %3, 8), %%mm2 \n\t" yading@11: PMAXUB(%%mm2, %%mm4) yading@11: PMINUB(%%mm2, %%mm3, %%mm5) yading@11: "psubb %%mm2, %%mm1 \n\t" yading@11: "paddb %%mm7, %%mm1 \n\t" yading@11: "pcmpgtb %%mm6, %%mm1 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: yading@11: "movq (%%"REG_a", %3, 4), %%mm1 \n\t" yading@11: "psubb %%mm1, %%mm2 \n\t" yading@11: "paddb %%mm7, %%mm2 \n\t" yading@11: "pcmpgtb %%mm6, %%mm2 \n\t" yading@11: "paddb %%mm2, %%mm0 \n\t" yading@11: "psubusb %%mm3, %%mm4 \n\t" yading@11: yading@11: "pxor %%mm6, %%mm6 \n\t" yading@11: "movq %4, %%mm7 \n\t" // QP,..., QP yading@11: "paddusb %%mm7, %%mm7 \n\t" // 2QP ... 2QP yading@11: "psubusb %%mm4, %%mm7 \n\t" // Diff >=2QP -> 0 yading@11: "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 yading@11: "pcmpeqb %%mm6, %%mm7 \n\t" // Diff < 2QP -> 0 yading@11: "movq %%mm7, %1 \n\t" yading@11: yading@11: "movq %5, %%mm7 \n\t" yading@11: "punpcklbw %%mm7, %%mm7 \n\t" yading@11: "punpcklbw %%mm7, %%mm7 \n\t" yading@11: "punpcklbw %%mm7, %%mm7 \n\t" yading@11: "psubb %%mm0, %%mm6 \n\t" yading@11: "pcmpgtb %%mm7, %%mm6 \n\t" yading@11: "movq %%mm6, %0 \n\t" yading@11: yading@11: : "=m" (eq_mask), "=m" (dc_mask) yading@11: : "r" (src), "r" ((x86_reg)step), "m" (c->pQPb), "m"(c->ppMode.flatnessThreshold) yading@11: : "%"REG_a yading@11: ); yading@11: yading@11: both_masks = dc_mask & eq_mask; yading@11: yading@11: if(both_masks){ yading@11: x86_reg offset= -8*step; yading@11: int64_t *temp_sums= sums; yading@11: yading@11: __asm__ volatile( yading@11: "movq %2, %%mm0 \n\t" // QP,..., QP yading@11: "pxor %%mm4, %%mm4 \n\t" yading@11: yading@11: "movq (%0), %%mm6 \n\t" yading@11: "movq (%0, %1), %%mm5 \n\t" yading@11: "movq %%mm5, %%mm1 \n\t" yading@11: "movq %%mm6, %%mm2 \n\t" yading@11: "psubusb %%mm6, %%mm5 \n\t" yading@11: "psubusb %%mm1, %%mm2 \n\t" yading@11: "por %%mm5, %%mm2 \n\t" // ABS Diff of lines yading@11: "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 yading@11: "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF yading@11: yading@11: "pxor %%mm6, %%mm1 \n\t" yading@11: "pand %%mm0, %%mm1 \n\t" yading@11: "pxor %%mm1, %%mm6 \n\t" yading@11: // 0:QP 6:First yading@11: yading@11: "movq (%0, %1, 8), %%mm5 \n\t" yading@11: "add %1, %0 \n\t" // %0 points to line 1 not 0 yading@11: "movq (%0, %1, 8), %%mm7 \n\t" yading@11: "movq %%mm5, %%mm1 \n\t" yading@11: "movq %%mm7, %%mm2 \n\t" yading@11: "psubusb %%mm7, %%mm5 \n\t" yading@11: "psubusb %%mm1, %%mm2 \n\t" yading@11: "por %%mm5, %%mm2 \n\t" // ABS Diff of lines yading@11: "movq %2, %%mm0 \n\t" // QP,..., QP yading@11: "psubusb %%mm2, %%mm0 \n\t" // diff >= QP -> 0 yading@11: "pcmpeqb %%mm4, %%mm0 \n\t" // diff >= QP -> FF yading@11: yading@11: "pxor %%mm7, %%mm1 \n\t" yading@11: "pand %%mm0, %%mm1 \n\t" yading@11: "pxor %%mm1, %%mm7 \n\t" yading@11: yading@11: "movq %%mm6, %%mm5 \n\t" yading@11: "punpckhbw %%mm4, %%mm6 \n\t" yading@11: "punpcklbw %%mm4, %%mm5 \n\t" yading@11: // 4:0 5/6:First 7:Last yading@11: yading@11: "movq %%mm5, %%mm0 \n\t" yading@11: "movq %%mm6, %%mm1 \n\t" yading@11: "psllw $2, %%mm0 \n\t" yading@11: "psllw $2, %%mm1 \n\t" yading@11: "paddw "MANGLE(w04)", %%mm0 \n\t" yading@11: "paddw "MANGLE(w04)", %%mm1 \n\t" yading@11: yading@11: #define NEXT\ yading@11: "movq (%0), %%mm2 \n\t"\ yading@11: "movq (%0), %%mm3 \n\t"\ yading@11: "add %1, %0 \n\t"\ yading@11: "punpcklbw %%mm4, %%mm2 \n\t"\ yading@11: "punpckhbw %%mm4, %%mm3 \n\t"\ yading@11: "paddw %%mm2, %%mm0 \n\t"\ yading@11: "paddw %%mm3, %%mm1 \n\t" yading@11: yading@11: #define PREV\ yading@11: "movq (%0), %%mm2 \n\t"\ yading@11: "movq (%0), %%mm3 \n\t"\ yading@11: "add %1, %0 \n\t"\ yading@11: "punpcklbw %%mm4, %%mm2 \n\t"\ yading@11: "punpckhbw %%mm4, %%mm3 \n\t"\ yading@11: "psubw %%mm2, %%mm0 \n\t"\ yading@11: "psubw %%mm3, %%mm1 \n\t" yading@11: yading@11: yading@11: NEXT //0 yading@11: NEXT //1 yading@11: NEXT //2 yading@11: "movq %%mm0, (%3) \n\t" yading@11: "movq %%mm1, 8(%3) \n\t" yading@11: yading@11: NEXT //3 yading@11: "psubw %%mm5, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: "movq %%mm0, 16(%3) \n\t" yading@11: "movq %%mm1, 24(%3) \n\t" yading@11: yading@11: NEXT //4 yading@11: "psubw %%mm5, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: "movq %%mm0, 32(%3) \n\t" yading@11: "movq %%mm1, 40(%3) \n\t" yading@11: yading@11: NEXT //5 yading@11: "psubw %%mm5, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: "movq %%mm0, 48(%3) \n\t" yading@11: "movq %%mm1, 56(%3) \n\t" yading@11: yading@11: NEXT //6 yading@11: "psubw %%mm5, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: "movq %%mm0, 64(%3) \n\t" yading@11: "movq %%mm1, 72(%3) \n\t" yading@11: yading@11: "movq %%mm7, %%mm6 \n\t" yading@11: "punpckhbw %%mm4, %%mm7 \n\t" yading@11: "punpcklbw %%mm4, %%mm6 \n\t" yading@11: yading@11: NEXT //7 yading@11: "mov %4, %0 \n\t" yading@11: "add %1, %0 \n\t" yading@11: PREV //0 yading@11: "movq %%mm0, 80(%3) \n\t" yading@11: "movq %%mm1, 88(%3) \n\t" yading@11: yading@11: PREV //1 yading@11: "paddw %%mm6, %%mm0 \n\t" yading@11: "paddw %%mm7, %%mm1 \n\t" yading@11: "movq %%mm0, 96(%3) \n\t" yading@11: "movq %%mm1, 104(%3) \n\t" yading@11: yading@11: PREV //2 yading@11: "paddw %%mm6, %%mm0 \n\t" yading@11: "paddw %%mm7, %%mm1 \n\t" yading@11: "movq %%mm0, 112(%3) \n\t" yading@11: "movq %%mm1, 120(%3) \n\t" yading@11: yading@11: PREV //3 yading@11: "paddw %%mm6, %%mm0 \n\t" yading@11: "paddw %%mm7, %%mm1 \n\t" yading@11: "movq %%mm0, 128(%3) \n\t" yading@11: "movq %%mm1, 136(%3) \n\t" yading@11: yading@11: PREV //4 yading@11: "paddw %%mm6, %%mm0 \n\t" yading@11: "paddw %%mm7, %%mm1 \n\t" yading@11: "movq %%mm0, 144(%3) \n\t" yading@11: "movq %%mm1, 152(%3) \n\t" yading@11: yading@11: "mov %4, %0 \n\t" //FIXME yading@11: yading@11: : "+&r"(src) yading@11: : "r" ((x86_reg)step), "m" (c->pQPb), "r"(sums), "g"(src) yading@11: ); yading@11: yading@11: src+= step; // src points to begin of the 8x8 Block yading@11: yading@11: __asm__ volatile( yading@11: "movq %4, %%mm6 \n\t" yading@11: "pcmpeqb %%mm5, %%mm5 \n\t" yading@11: "pxor %%mm6, %%mm5 \n\t" yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: yading@11: "1: \n\t" yading@11: "movq (%1), %%mm0 \n\t" yading@11: "movq 8(%1), %%mm1 \n\t" yading@11: "paddw 32(%1), %%mm0 \n\t" yading@11: "paddw 40(%1), %%mm1 \n\t" yading@11: "movq (%0, %3), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "movq %%mm2, %%mm4 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" yading@11: "punpckhbw %%mm7, %%mm3 \n\t" yading@11: "paddw %%mm2, %%mm0 \n\t" yading@11: "paddw %%mm3, %%mm1 \n\t" yading@11: "paddw %%mm2, %%mm0 \n\t" yading@11: "paddw %%mm3, %%mm1 \n\t" yading@11: "psrlw $4, %%mm0 \n\t" yading@11: "psrlw $4, %%mm1 \n\t" yading@11: "packuswb %%mm1, %%mm0 \n\t" yading@11: "pand %%mm6, %%mm0 \n\t" yading@11: "pand %%mm5, %%mm4 \n\t" yading@11: "por %%mm4, %%mm0 \n\t" yading@11: "movq %%mm0, (%0, %3) \n\t" yading@11: "add $16, %1 \n\t" yading@11: "add %2, %0 \n\t" yading@11: " js 1b \n\t" yading@11: yading@11: : "+r"(offset), "+r"(temp_sums) yading@11: : "r" ((x86_reg)step), "r"(src - offset), "m"(both_masks) yading@11: ); yading@11: }else yading@11: src+= step; // src points to begin of the 8x8 Block yading@11: yading@11: if(eq_mask != -1LL){ yading@11: uint8_t *temp_src= src; yading@11: DECLARE_ALIGNED(8, uint64_t, tmp)[4]; // make space for 4 8-byte vars yading@11: __asm__ volatile( yading@11: "pxor %%mm7, %%mm7 \n\t" yading@11: // 0 1 2 3 4 5 6 7 8 9 yading@11: // %0 eax eax+%1 eax+2%1 %0+4%1 ecx ecx+%1 ecx+2%1 %1+8%1 ecx+4%1 yading@11: yading@11: "movq (%0), %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "punpcklbw %%mm7, %%mm0 \n\t" // low part of line 0 yading@11: "punpckhbw %%mm7, %%mm1 \n\t" // high part of line 0 yading@11: yading@11: "movq (%0, %1), %%mm2 \n\t" yading@11: "lea (%0, %1, 2), %%"REG_a" \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // low part of line 1 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // high part of line 1 yading@11: yading@11: "movq (%%"REG_a"), %%mm4 \n\t" yading@11: "movq %%mm4, %%mm5 \n\t" yading@11: "punpcklbw %%mm7, %%mm4 \n\t" // low part of line 2 yading@11: "punpckhbw %%mm7, %%mm5 \n\t" // high part of line 2 yading@11: yading@11: "paddw %%mm0, %%mm0 \n\t" // 2L0 yading@11: "paddw %%mm1, %%mm1 \n\t" // 2H0 yading@11: "psubw %%mm4, %%mm2 \n\t" // L1 - L2 yading@11: "psubw %%mm5, %%mm3 \n\t" // H1 - H2 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - L1 + L2 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - H1 + H2 yading@11: yading@11: "psllw $2, %%mm2 \n\t" // 4L1 - 4L2 yading@11: "psllw $2, %%mm3 \n\t" // 4H1 - 4H2 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 yading@11: yading@11: "movq (%%"REG_a", %1), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L3 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H3 yading@11: yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - L3 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - H3 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: "movq %%mm0, (%4) \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "movq %%mm1, 8(%4) \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: yading@11: "movq (%%"REG_a", %1, 2), %%mm0 \n\t" yading@11: "movq %%mm0, %%mm1 \n\t" yading@11: "punpcklbw %%mm7, %%mm0 \n\t" // L4 yading@11: "punpckhbw %%mm7, %%mm1 \n\t" // H4 yading@11: yading@11: "psubw %%mm0, %%mm2 \n\t" // L3 - L4 yading@11: "psubw %%mm1, %%mm3 \n\t" // H3 - H4 yading@11: "movq %%mm2, 16(%4) \n\t" // L3 - L4 yading@11: "movq %%mm3, 24(%4) \n\t" // H3 - H4 yading@11: "paddw %%mm4, %%mm4 \n\t" // 2L2 yading@11: "paddw %%mm5, %%mm5 \n\t" // 2H2 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - L3 + L4 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - H3 + H4 yading@11: yading@11: "lea (%%"REG_a", %1), %0 \n\t" yading@11: "psllw $2, %%mm2 \n\t" // 4L3 - 4L4 yading@11: "psllw $2, %%mm3 \n\t" // 4H3 - 4H4 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 yading@11: //50 opcodes so far yading@11: "movq (%0, %1, 2), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L5 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H5 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - L5 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - H5 yading@11: "psubw %%mm2, %%mm4 \n\t" // 2L2 - 5L3 + 5L4 - 2L5 yading@11: "psubw %%mm3, %%mm5 \n\t" // 2H2 - 5H3 + 5H4 - 2H5 yading@11: yading@11: "movq (%%"REG_a", %1, 4), %%mm6 \n\t" yading@11: "punpcklbw %%mm7, %%mm6 \n\t" // L6 yading@11: "psubw %%mm6, %%mm2 \n\t" // L5 - L6 yading@11: "movq (%%"REG_a", %1, 4), %%mm6 \n\t" yading@11: "punpckhbw %%mm7, %%mm6 \n\t" // H6 yading@11: "psubw %%mm6, %%mm3 \n\t" // H5 - H6 yading@11: yading@11: "paddw %%mm0, %%mm0 \n\t" // 2L4 yading@11: "paddw %%mm1, %%mm1 \n\t" // 2H4 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - L5 + L6 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - H5 + H6 yading@11: yading@11: "psllw $2, %%mm2 \n\t" // 4L5 - 4L6 yading@11: "psllw $2, %%mm3 \n\t" // 4H5 - 4H6 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 yading@11: yading@11: "movq (%0, %1, 4), %%mm2 \n\t" yading@11: "movq %%mm2, %%mm3 \n\t" yading@11: "punpcklbw %%mm7, %%mm2 \n\t" // L7 yading@11: "punpckhbw %%mm7, %%mm3 \n\t" // H7 yading@11: yading@11: "paddw %%mm2, %%mm2 \n\t" // 2L7 yading@11: "paddw %%mm3, %%mm3 \n\t" // 2H7 yading@11: "psubw %%mm2, %%mm0 \n\t" // 2L4 - 5L5 + 5L6 - 2L7 yading@11: "psubw %%mm3, %%mm1 \n\t" // 2H4 - 5H5 + 5H6 - 2H7 yading@11: yading@11: "movq (%4), %%mm2 \n\t" // 2L0 - 5L1 + 5L2 - 2L3 yading@11: "movq 8(%4), %%mm3 \n\t" // 2H0 - 5H1 + 5H2 - 2H3 yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm0, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm1, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm2, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "psubw %%mm3, %%mm6 \n\t" yading@11: "pmaxsw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| yading@11: #else yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm0, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm0 \n\t" yading@11: "psubw %%mm6, %%mm0 \n\t" // |2L4 - 5L5 + 5L6 - 2L7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm1, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm1 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" // |2H4 - 5H5 + 5H6 - 2H7| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm2, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "psubw %%mm6, %%mm2 \n\t" // |2L0 - 5L1 + 5L2 - 2L3| yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm3, %%mm6 \n\t" yading@11: "pxor %%mm6, %%mm3 \n\t" yading@11: "psubw %%mm6, %%mm3 \n\t" // |2H0 - 5H1 + 5H2 - 2H3| yading@11: #endif yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "pminsw %%mm2, %%mm0 \n\t" yading@11: "pminsw %%mm3, %%mm1 \n\t" yading@11: #else yading@11: "movq %%mm0, %%mm6 \n\t" yading@11: "psubusw %%mm2, %%mm6 \n\t" yading@11: "psubw %%mm6, %%mm0 \n\t" yading@11: "movq %%mm1, %%mm6 \n\t" yading@11: "psubusw %%mm3, %%mm6 \n\t" yading@11: "psubw %%mm6, %%mm1 \n\t" yading@11: #endif yading@11: yading@11: "movd %2, %%mm2 \n\t" // QP yading@11: "punpcklbw %%mm7, %%mm2 \n\t" yading@11: yading@11: "movq %%mm7, %%mm6 \n\t" // 0 yading@11: "pcmpgtw %%mm4, %%mm6 \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5) yading@11: "pxor %%mm6, %%mm4 \n\t" yading@11: "psubw %%mm6, %%mm4 \n\t" // |2L2 - 5L3 + 5L4 - 2L5| yading@11: "pcmpgtw %%mm5, %%mm7 \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5) yading@11: "pxor %%mm7, %%mm5 \n\t" yading@11: "psubw %%mm7, %%mm5 \n\t" // |2H2 - 5H3 + 5H4 - 2H5| yading@11: // 100 opcodes yading@11: "psllw $3, %%mm2 \n\t" // 8QP yading@11: "movq %%mm2, %%mm3 \n\t" // 8QP yading@11: "pcmpgtw %%mm4, %%mm2 \n\t" yading@11: "pcmpgtw %%mm5, %%mm3 \n\t" yading@11: "pand %%mm2, %%mm4 \n\t" yading@11: "pand %%mm3, %%mm5 \n\t" yading@11: yading@11: yading@11: "psubusw %%mm0, %%mm4 \n\t" // hd yading@11: "psubusw %%mm1, %%mm5 \n\t" // ld yading@11: yading@11: yading@11: "movq "MANGLE(w05)", %%mm2 \n\t" // 5 yading@11: "pmullw %%mm2, %%mm4 \n\t" yading@11: "pmullw %%mm2, %%mm5 \n\t" yading@11: "movq "MANGLE(w20)", %%mm2 \n\t" // 32 yading@11: "paddw %%mm2, %%mm4 \n\t" yading@11: "paddw %%mm2, %%mm5 \n\t" yading@11: "psrlw $6, %%mm4 \n\t" yading@11: "psrlw $6, %%mm5 \n\t" yading@11: yading@11: "movq 16(%4), %%mm0 \n\t" // L3 - L4 yading@11: "movq 24(%4), %%mm1 \n\t" // H3 - H4 yading@11: yading@11: "pxor %%mm2, %%mm2 \n\t" yading@11: "pxor %%mm3, %%mm3 \n\t" yading@11: yading@11: "pcmpgtw %%mm0, %%mm2 \n\t" // sign (L3-L4) yading@11: "pcmpgtw %%mm1, %%mm3 \n\t" // sign (H3-H4) yading@11: "pxor %%mm2, %%mm0 \n\t" yading@11: "pxor %%mm3, %%mm1 \n\t" yading@11: "psubw %%mm2, %%mm0 \n\t" // |L3-L4| yading@11: "psubw %%mm3, %%mm1 \n\t" // |H3-H4| yading@11: "psrlw $1, %%mm0 \n\t" // |L3 - L4|/2 yading@11: "psrlw $1, %%mm1 \n\t" // |H3 - H4|/2 yading@11: yading@11: "pxor %%mm6, %%mm2 \n\t" yading@11: "pxor %%mm7, %%mm3 \n\t" yading@11: "pand %%mm2, %%mm4 \n\t" yading@11: "pand %%mm3, %%mm5 \n\t" yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: "pminsw %%mm0, %%mm4 \n\t" yading@11: "pminsw %%mm1, %%mm5 \n\t" yading@11: #else yading@11: "movq %%mm4, %%mm2 \n\t" yading@11: "psubusw %%mm0, %%mm2 \n\t" yading@11: "psubw %%mm2, %%mm4 \n\t" yading@11: "movq %%mm5, %%mm2 \n\t" yading@11: "psubusw %%mm1, %%mm2 \n\t" yading@11: "psubw %%mm2, %%mm5 \n\t" yading@11: #endif yading@11: "pxor %%mm6, %%mm4 \n\t" yading@11: "pxor %%mm7, %%mm5 \n\t" yading@11: "psubw %%mm6, %%mm4 \n\t" yading@11: "psubw %%mm7, %%mm5 \n\t" yading@11: "packsswb %%mm5, %%mm4 \n\t" yading@11: "movq %3, %%mm1 \n\t" yading@11: "pandn %%mm4, %%mm1 \n\t" yading@11: "movq (%0), %%mm0 \n\t" yading@11: "paddb %%mm1, %%mm0 \n\t" yading@11: "movq %%mm0, (%0) \n\t" yading@11: "movq (%0, %1), %%mm0 \n\t" yading@11: "psubb %%mm1, %%mm0 \n\t" yading@11: "movq %%mm0, (%0, %1) \n\t" yading@11: yading@11: : "+r" (temp_src) yading@11: : "r" ((x86_reg)step), "m" (c->pQPb), "m"(eq_mask), "r"(tmp) yading@11: : "%"REG_a yading@11: ); yading@11: } yading@11: /*if(step==16){ yading@11: STOP_TIMER("step16") yading@11: }else{ yading@11: STOP_TIMER("stepX") yading@11: } yading@11: } */ yading@11: } yading@11: #endif //TEMPLATE_PP_MMX yading@11: yading@11: static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, yading@11: const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c); yading@11: yading@11: /** yading@11: * Copy a block from src to dst and fixes the blacklevel. yading@11: * levelFix == 0 -> do not touch the brightness & contrast yading@11: */ yading@11: #undef REAL_SCALED_CPY yading@11: #undef SCALED_CPY yading@11: yading@11: static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, const uint8_t src[], int srcStride, yading@11: int levelFix, int64_t *packedOffsetAndScale) yading@11: { yading@11: #if !TEMPLATE_PP_MMX yading@11: int i; yading@11: #endif yading@11: if(levelFix){ yading@11: #if TEMPLATE_PP_MMX yading@11: __asm__ volatile( yading@11: "movq (%%"REG_a"), %%mm2 \n\t" // packedYOffset yading@11: "movq 8(%%"REG_a"), %%mm3 \n\t" // packedYScale yading@11: "lea (%2,%4), %%"REG_a" \n\t" yading@11: "lea (%3,%5), %%"REG_d" \n\t" yading@11: "pxor %%mm4, %%mm4 \n\t" yading@11: #if TEMPLATE_PP_MMXEXT yading@11: #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ yading@11: "movq " #src1 ", %%mm0 \n\t"\ yading@11: "movq " #src1 ", %%mm5 \n\t"\ yading@11: "movq " #src2 ", %%mm1 \n\t"\ yading@11: "movq " #src2 ", %%mm6 \n\t"\ yading@11: "punpcklbw %%mm0, %%mm0 \n\t"\ yading@11: "punpckhbw %%mm5, %%mm5 \n\t"\ yading@11: "punpcklbw %%mm1, %%mm1 \n\t"\ yading@11: "punpckhbw %%mm6, %%mm6 \n\t"\ yading@11: "pmulhuw %%mm3, %%mm0 \n\t"\ yading@11: "pmulhuw %%mm3, %%mm5 \n\t"\ yading@11: "pmulhuw %%mm3, %%mm1 \n\t"\ yading@11: "pmulhuw %%mm3, %%mm6 \n\t"\ yading@11: "psubw %%mm2, %%mm0 \n\t"\ yading@11: "psubw %%mm2, %%mm5 \n\t"\ yading@11: "psubw %%mm2, %%mm1 \n\t"\ yading@11: "psubw %%mm2, %%mm6 \n\t"\ yading@11: "packuswb %%mm5, %%mm0 \n\t"\ yading@11: "packuswb %%mm6, %%mm1 \n\t"\ yading@11: "movq %%mm0, " #dst1 " \n\t"\ yading@11: "movq %%mm1, " #dst2 " \n\t"\ yading@11: yading@11: #else //TEMPLATE_PP_MMXEXT yading@11: #define REAL_SCALED_CPY(src1, src2, dst1, dst2) \ yading@11: "movq " #src1 ", %%mm0 \n\t"\ yading@11: "movq " #src1 ", %%mm5 \n\t"\ yading@11: "punpcklbw %%mm4, %%mm0 \n\t"\ yading@11: "punpckhbw %%mm4, %%mm5 \n\t"\ yading@11: "psubw %%mm2, %%mm0 \n\t"\ yading@11: "psubw %%mm2, %%mm5 \n\t"\ yading@11: "movq " #src2 ", %%mm1 \n\t"\ yading@11: "psllw $6, %%mm0 \n\t"\ yading@11: "psllw $6, %%mm5 \n\t"\ yading@11: "pmulhw %%mm3, %%mm0 \n\t"\ yading@11: "movq " #src2 ", %%mm6 \n\t"\ yading@11: "pmulhw %%mm3, %%mm5 \n\t"\ yading@11: "punpcklbw %%mm4, %%mm1 \n\t"\ yading@11: "punpckhbw %%mm4, %%mm6 \n\t"\ yading@11: "psubw %%mm2, %%mm1 \n\t"\ yading@11: "psubw %%mm2, %%mm6 \n\t"\ yading@11: "psllw $6, %%mm1 \n\t"\ yading@11: "psllw $6, %%mm6 \n\t"\ yading@11: "pmulhw %%mm3, %%mm1 \n\t"\ yading@11: "pmulhw %%mm3, %%mm6 \n\t"\ yading@11: "packuswb %%mm5, %%mm0 \n\t"\ yading@11: "packuswb %%mm6, %%mm1 \n\t"\ yading@11: "movq %%mm0, " #dst1 " \n\t"\ yading@11: "movq %%mm1, " #dst2 " \n\t"\ yading@11: yading@11: #endif //TEMPLATE_PP_MMXEXT yading@11: #define SCALED_CPY(src1, src2, dst1, dst2)\ yading@11: REAL_SCALED_CPY(src1, src2, dst1, dst2) yading@11: yading@11: SCALED_CPY((%2) , (%2, %4) , (%3) , (%3, %5)) yading@11: SCALED_CPY((%2, %4, 2), (%%REGa, %4, 2), (%3, %5, 2), (%%REGd, %5, 2)) yading@11: SCALED_CPY((%2, %4, 4), (%%REGa, %4, 4), (%3, %5, 4), (%%REGd, %5, 4)) yading@11: "lea (%%"REG_a",%4,4), %%"REG_a" \n\t" yading@11: "lea (%%"REG_d",%5,4), %%"REG_d" \n\t" yading@11: SCALED_CPY((%%REGa, %4), (%%REGa, %4, 2), (%%REGd, %5), (%%REGd, %5, 2)) yading@11: yading@11: yading@11: : "=&a" (packedOffsetAndScale) yading@11: : "0" (packedOffsetAndScale), yading@11: "r"(src), yading@11: "r"(dst), yading@11: "r" ((x86_reg)srcStride), yading@11: "r" ((x86_reg)dstStride) yading@11: : "%"REG_d yading@11: ); yading@11: #else //TEMPLATE_PP_MMX yading@11: for(i=0; i<8; i++) yading@11: memcpy( &(dst[dstStride*i]), yading@11: &(src[srcStride*i]), BLOCK_SIZE); yading@11: #endif //TEMPLATE_PP_MMX yading@11: }else{ yading@11: #if TEMPLATE_PP_MMX yading@11: __asm__ volatile( yading@11: "lea (%0,%2), %%"REG_a" \n\t" yading@11: "lea (%1,%3), %%"REG_d" \n\t" yading@11: yading@11: #define REAL_SIMPLE_CPY(src1, src2, dst1, dst2) \ yading@11: "movq " #src1 ", %%mm0 \n\t"\ yading@11: "movq " #src2 ", %%mm1 \n\t"\ yading@11: "movq %%mm0, " #dst1 " \n\t"\ yading@11: "movq %%mm1, " #dst2 " \n\t"\ yading@11: yading@11: #define SIMPLE_CPY(src1, src2, dst1, dst2)\ yading@11: REAL_SIMPLE_CPY(src1, src2, dst1, dst2) yading@11: yading@11: SIMPLE_CPY((%0) , (%0, %2) , (%1) , (%1, %3)) yading@11: SIMPLE_CPY((%0, %2, 2), (%%REGa, %2, 2), (%1, %3, 2), (%%REGd, %3, 2)) yading@11: SIMPLE_CPY((%0, %2, 4), (%%REGa, %2, 4), (%1, %3, 4), (%%REGd, %3, 4)) yading@11: "lea (%%"REG_a",%2,4), %%"REG_a" \n\t" yading@11: "lea (%%"REG_d",%3,4), %%"REG_d" \n\t" yading@11: SIMPLE_CPY((%%REGa, %2), (%%REGa, %2, 2), (%%REGd, %3), (%%REGd, %3, 2)) yading@11: yading@11: : : "r" (src), yading@11: "r" (dst), yading@11: "r" ((x86_reg)srcStride), yading@11: "r" ((x86_reg)dstStride) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: #else //TEMPLATE_PP_MMX yading@11: for(i=0; i<8; i++) yading@11: memcpy( &(dst[dstStride*i]), yading@11: &(src[srcStride*i]), BLOCK_SIZE); yading@11: #endif //TEMPLATE_PP_MMX yading@11: } yading@11: } yading@11: yading@11: /** yading@11: * Duplicate the given 8 src pixels ? times upward yading@11: */ yading@11: static inline void RENAME(duplicate)(uint8_t src[], int stride) yading@11: { yading@11: #if TEMPLATE_PP_MMX yading@11: __asm__ volatile( yading@11: "movq (%0), %%mm0 \n\t" yading@11: "movq %%mm0, (%0, %1, 4) \n\t" yading@11: "add %1, %0 \n\t" yading@11: "movq %%mm0, (%0) \n\t" yading@11: "movq %%mm0, (%0, %1) \n\t" yading@11: "movq %%mm0, (%0, %1, 2) \n\t" yading@11: "movq %%mm0, (%0, %1, 4) \n\t" yading@11: : "+r" (src) yading@11: : "r" ((x86_reg)-stride) yading@11: ); yading@11: #else yading@11: int i; yading@11: uint8_t *p=src; yading@11: for(i=0; i<5; i++){ yading@11: p-= stride; yading@11: memcpy(p, src, 8); yading@11: } yading@11: #endif yading@11: } yading@11: yading@11: /** yading@11: * Filter array of bytes (Y or U or V values) yading@11: */ yading@11: static void RENAME(postProcess)(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height, yading@11: const QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2) yading@11: { yading@11: DECLARE_ALIGNED(8, PPContext, c)= *c2; //copy to stack for faster access yading@11: int x,y; yading@11: #ifdef TEMPLATE_PP_TIME_MODE yading@11: const int mode= TEMPLATE_PP_TIME_MODE; yading@11: #else yading@11: const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode; yading@11: #endif yading@11: int black=0, white=255; // blackest black and whitest white in the picture yading@11: int QPCorrecture= 256*256; yading@11: yading@11: int copyAhead; yading@11: #if TEMPLATE_PP_MMX yading@11: int i; yading@11: #endif yading@11: yading@11: const int qpHShift= isColor ? 4-c.hChromaSubSample : 4; yading@11: const int qpVShift= isColor ? 4-c.vChromaSubSample : 4; yading@11: yading@11: //FIXME remove yading@11: uint64_t * const yHistogram= c.yHistogram; yading@11: uint8_t * const tempSrc= srcStride > 0 ? c.tempSrc : c.tempSrc - 23*srcStride; yading@11: uint8_t * const tempDst= (dstStride > 0 ? c.tempDst : c.tempDst - 23*dstStride) + 32; yading@11: //const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4; yading@11: yading@11: #if TEMPLATE_PP_MMX yading@11: for(i=0; i<57; i++){ yading@11: int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1; yading@11: int threshold= offset*2 + 1; yading@11: c.mmxDcOffset[i]= 0x7F - offset; yading@11: c.mmxDcThreshold[i]= 0x7F - threshold; yading@11: c.mmxDcOffset[i]*= 0x0101010101010101LL; yading@11: c.mmxDcThreshold[i]*= 0x0101010101010101LL; yading@11: } yading@11: #endif yading@11: yading@11: if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16; yading@11: else if( (mode & LINEAR_BLEND_DEINT_FILTER) yading@11: || (mode & FFMPEG_DEINT_FILTER) yading@11: || (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14; yading@11: else if( (mode & V_DEBLOCK) yading@11: || (mode & LINEAR_IPOL_DEINT_FILTER) yading@11: || (mode & MEDIAN_DEINT_FILTER) yading@11: || (mode & V_A_DEBLOCK)) copyAhead=13; yading@11: else if(mode & V_X1_FILTER) copyAhead=11; yading@11: // else if(mode & V_RK1_FILTER) copyAhead=10; yading@11: else if(mode & DERING) copyAhead=9; yading@11: else copyAhead=8; yading@11: yading@11: copyAhead-= 8; yading@11: yading@11: if(!isColor){ yading@11: uint64_t sum= 0; yading@11: int i; yading@11: uint64_t maxClipped; yading@11: uint64_t clipped; yading@11: double scale; yading@11: yading@11: c.frameNum++; yading@11: // first frame is fscked so we ignore it yading@11: if(c.frameNum == 1) yHistogram[0]= width*(uint64_t)height/64*15/256; yading@11: yading@11: for(i=0; i<256; i++){ yading@11: sum+= yHistogram[i]; yading@11: } yading@11: yading@11: /* We always get a completely black picture first. */ yading@11: maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold); yading@11: yading@11: clipped= sum; yading@11: for(black=255; black>0; black--){ yading@11: if(clipped < maxClipped) break; yading@11: clipped-= yHistogram[black]; yading@11: } yading@11: yading@11: clipped= sum; yading@11: for(white=0; white<256; white++){ yading@11: if(clipped < maxClipped) break; yading@11: clipped-= yHistogram[white]; yading@11: } yading@11: yading@11: scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black); yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: c.packedYScale= (uint16_t)(scale*256.0 + 0.5); yading@11: c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF; yading@11: #else yading@11: c.packedYScale= (uint16_t)(scale*1024.0 + 0.5); yading@11: c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF; yading@11: #endif yading@11: yading@11: c.packedYOffset|= c.packedYOffset<<32; yading@11: c.packedYOffset|= c.packedYOffset<<16; yading@11: yading@11: c.packedYScale|= c.packedYScale<<32; yading@11: c.packedYScale|= c.packedYScale<<16; yading@11: yading@11: if(mode & LEVEL_FIX) QPCorrecture= (int)(scale*256*256 + 0.5); yading@11: else QPCorrecture= 256*256; yading@11: }else{ yading@11: c.packedYScale= 0x0100010001000100LL; yading@11: c.packedYOffset= 0; yading@11: QPCorrecture= 256*256; yading@11: } yading@11: yading@11: /* copy & deinterlace first row of blocks */ yading@11: y=-BLOCK_SIZE; yading@11: { yading@11: const uint8_t *srcBlock= &(src[y*srcStride]); yading@11: uint8_t *dstBlock= tempDst + dstStride; yading@11: yading@11: // From this point on it is guaranteed that we can read and write 16 lines downward yading@11: // finish 1 block before the next otherwise we might have a problem yading@11: // with the L1 Cache of the P4 ... or only a few blocks at a time or something yading@11: for(x=0; x>2)&6) + 5)*srcStride + 32); yading@11: prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); yading@11: prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); yading@11: prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); yading@11: */ yading@11: yading@11: __asm__( yading@11: "mov %4, %%"REG_a" \n\t" yading@11: "shr $2, %%"REG_a" \n\t" yading@11: "and $6, %%"REG_a" \n\t" yading@11: "add %5, %%"REG_a" \n\t" yading@11: "mov %%"REG_a", %%"REG_d" \n\t" yading@11: "imul %1, %%"REG_a" \n\t" yading@11: "imul %3, %%"REG_d" \n\t" yading@11: "prefetchnta 32(%%"REG_a", %0) \n\t" yading@11: "prefetcht0 32(%%"REG_d", %2) \n\t" yading@11: "add %1, %%"REG_a" \n\t" yading@11: "add %3, %%"REG_d" \n\t" yading@11: "prefetchnta 32(%%"REG_a", %0) \n\t" yading@11: "prefetcht0 32(%%"REG_d", %2) \n\t" yading@11: :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride), yading@11: "g" ((x86_reg)x), "g" ((x86_reg)copyAhead) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: yading@11: #elif TEMPLATE_PP_3DNOW yading@11: //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... yading@11: /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); yading@11: prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); yading@11: prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); yading@11: prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); yading@11: */ yading@11: #endif yading@11: yading@11: RENAME(blockCopy)(dstBlock + dstStride*8, dstStride, yading@11: srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset); yading@11: yading@11: RENAME(duplicate)(dstBlock + dstStride*8, dstStride); yading@11: yading@11: if(mode & LINEAR_IPOL_DEINT_FILTER) yading@11: RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); yading@11: else if(mode & LINEAR_BLEND_DEINT_FILTER) yading@11: RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); yading@11: else if(mode & MEDIAN_DEINT_FILTER) yading@11: RENAME(deInterlaceMedian)(dstBlock, dstStride); yading@11: else if(mode & CUBIC_IPOL_DEINT_FILTER) yading@11: RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); yading@11: else if(mode & FFMPEG_DEINT_FILTER) yading@11: RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); yading@11: else if(mode & LOWPASS5_DEINT_FILTER) yading@11: RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); yading@11: /* else if(mode & CUBIC_BLEND_DEINT_FILTER) yading@11: RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); yading@11: */ yading@11: dstBlock+=8; yading@11: srcBlock+=8; yading@11: } yading@11: if(width==FFABS(dstStride)) yading@11: linecpy(dst, tempDst + 9*dstStride, copyAhead, dstStride); yading@11: else{ yading@11: int i; yading@11: for(i=0; i>qpVShift)*QPStride]; yading@11: int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*FFABS(QPStride)]; yading@11: int QP=0; yading@11: /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards yading@11: if not than use a temporary buffer */ yading@11: if(y+15 >= height){ yading@11: int i; yading@11: /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with yading@11: blockcopy to dst later */ yading@11: linecpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead, yading@11: FFMAX(height-y-copyAhead, 0), srcStride); yading@11: yading@11: /* duplicate last line of src to fill the void up to line (copyAhead+7) */ yading@11: for(i=FFMAX(height-y, 8); i>qpHShift]; yading@11: c.nonBQP= nonBQPptr[x>>qpHShift]; yading@11: }else{ yading@11: QP= QPptr[x>>4]; yading@11: QP= (QP* QPCorrecture + 256*128)>>16; yading@11: c.nonBQP= nonBQPptr[x>>4]; yading@11: c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16; yading@11: yHistogram[ srcBlock[srcStride*12 + 4] ]++; yading@11: } yading@11: c.QP= QP; yading@11: #if TEMPLATE_PP_MMX yading@11: __asm__ volatile( yading@11: "movd %1, %%mm7 \n\t" yading@11: "packuswb %%mm7, %%mm7 \n\t" // 0, 0, 0, QP, 0, 0, 0, QP yading@11: "packuswb %%mm7, %%mm7 \n\t" // 0,QP, 0, QP, 0,QP, 0, QP yading@11: "packuswb %%mm7, %%mm7 \n\t" // QP,..., QP yading@11: "movq %%mm7, %0 \n\t" yading@11: : "=m" (c.pQPb) yading@11: : "r" (QP) yading@11: ); yading@11: #endif yading@11: yading@11: yading@11: #if TEMPLATE_PP_MMXEXT yading@11: /* yading@11: prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32); yading@11: prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32); yading@11: prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32); yading@11: prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32); yading@11: */ yading@11: yading@11: __asm__( yading@11: "mov %4, %%"REG_a" \n\t" yading@11: "shr $2, %%"REG_a" \n\t" yading@11: "and $6, %%"REG_a" \n\t" yading@11: "add %5, %%"REG_a" \n\t" yading@11: "mov %%"REG_a", %%"REG_d" \n\t" yading@11: "imul %1, %%"REG_a" \n\t" yading@11: "imul %3, %%"REG_d" \n\t" yading@11: "prefetchnta 32(%%"REG_a", %0) \n\t" yading@11: "prefetcht0 32(%%"REG_d", %2) \n\t" yading@11: "add %1, %%"REG_a" \n\t" yading@11: "add %3, %%"REG_d" \n\t" yading@11: "prefetchnta 32(%%"REG_a", %0) \n\t" yading@11: "prefetcht0 32(%%"REG_d", %2) \n\t" yading@11: :: "r" (srcBlock), "r" ((x86_reg)srcStride), "r" (dstBlock), "r" ((x86_reg)dstStride), yading@11: "g" ((x86_reg)x), "g" ((x86_reg)copyAhead) yading@11: : "%"REG_a, "%"REG_d yading@11: ); yading@11: yading@11: #elif TEMPLATE_PP_3DNOW yading@11: //FIXME check if this is faster on an 3dnow chip or if it is faster without the prefetch or ... yading@11: /* prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32); yading@11: prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32); yading@11: prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32); yading@11: prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32); yading@11: */ yading@11: #endif yading@11: yading@11: RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride, yading@11: srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset); yading@11: yading@11: if(mode & LINEAR_IPOL_DEINT_FILTER) yading@11: RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride); yading@11: else if(mode & LINEAR_BLEND_DEINT_FILTER) yading@11: RENAME(deInterlaceBlendLinear)(dstBlock, dstStride, c.deintTemp + x); yading@11: else if(mode & MEDIAN_DEINT_FILTER) yading@11: RENAME(deInterlaceMedian)(dstBlock, dstStride); yading@11: else if(mode & CUBIC_IPOL_DEINT_FILTER) yading@11: RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride); yading@11: else if(mode & FFMPEG_DEINT_FILTER) yading@11: RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x); yading@11: else if(mode & LOWPASS5_DEINT_FILTER) yading@11: RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x); yading@11: /* else if(mode & CUBIC_BLEND_DEINT_FILTER) yading@11: RENAME(deInterlaceBlendCubic)(dstBlock, dstStride); yading@11: */ yading@11: yading@11: /* only deblock if we have 2 blocks */ yading@11: if(y + 8 < height){ yading@11: if(mode & V_X1_FILTER) yading@11: RENAME(vertX1Filter)(dstBlock, stride, &c); yading@11: else if(mode & V_DEBLOCK){ yading@11: const int t= RENAME(vertClassify)(dstBlock, stride, &c); yading@11: yading@11: if(t==1) yading@11: RENAME(doVertLowPass)(dstBlock, stride, &c); yading@11: else if(t==2) yading@11: RENAME(doVertDefFilter)(dstBlock, stride, &c); yading@11: }else if(mode & V_A_DEBLOCK){ yading@11: RENAME(do_a_deblock)(dstBlock, stride, 1, &c); yading@11: } yading@11: } yading@11: yading@11: #if TEMPLATE_PP_MMX yading@11: RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride); yading@11: #endif yading@11: /* check if we have a previous block to deblock it with dstBlock */ yading@11: if(x - 8 >= 0){ yading@11: #if TEMPLATE_PP_MMX yading@11: if(mode & H_X1_FILTER) yading@11: RENAME(vertX1Filter)(tempBlock1, 16, &c); yading@11: else if(mode & H_DEBLOCK){ yading@11: //START_TIMER yading@11: const int t= RENAME(vertClassify)(tempBlock1, 16, &c); yading@11: //STOP_TIMER("dc & minmax") yading@11: if(t==1) yading@11: RENAME(doVertLowPass)(tempBlock1, 16, &c); yading@11: else if(t==2) yading@11: RENAME(doVertDefFilter)(tempBlock1, 16, &c); yading@11: }else if(mode & H_A_DEBLOCK){ yading@11: RENAME(do_a_deblock)(tempBlock1, 16, 1, &c); yading@11: } yading@11: yading@11: RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16); yading@11: yading@11: #else yading@11: if(mode & H_X1_FILTER) yading@11: horizX1Filter(dstBlock-4, stride, QP); yading@11: else if(mode & H_DEBLOCK){ yading@11: #if TEMPLATE_PP_ALTIVEC yading@11: DECLARE_ALIGNED(16, unsigned char, tempBlock)[272]; yading@11: int t; yading@11: transpose_16x8_char_toPackedAlign_altivec(tempBlock, dstBlock - (4 + 1), stride); yading@11: yading@11: t = vertClassify_altivec(tempBlock-48, 16, &c); yading@11: if(t==1) { yading@11: doVertLowPass_altivec(tempBlock-48, 16, &c); yading@11: transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); yading@11: } yading@11: else if(t==2) { yading@11: doVertDefFilter_altivec(tempBlock-48, 16, &c); yading@11: transpose_8x16_char_fromPackedAlign_altivec(dstBlock - (4 + 1), tempBlock, stride); yading@11: } yading@11: #else yading@11: const int t= RENAME(horizClassify)(dstBlock-4, stride, &c); yading@11: yading@11: if(t==1) yading@11: RENAME(doHorizLowPass)(dstBlock-4, stride, &c); yading@11: else if(t==2) yading@11: RENAME(doHorizDefFilter)(dstBlock-4, stride, &c); yading@11: #endif yading@11: }else if(mode & H_A_DEBLOCK){ yading@11: RENAME(do_a_deblock)(dstBlock-8, 1, stride, &c); yading@11: } yading@11: #endif //TEMPLATE_PP_MMX yading@11: if(mode & DERING){ yading@11: //FIXME filter first line yading@11: if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c); yading@11: } yading@11: yading@11: if(mode & TEMP_NOISE_FILTER) yading@11: { yading@11: RENAME(tempNoiseReducer)(dstBlock-8, stride, yading@11: c.tempBlurred[isColor] + y*dstStride + x, yading@11: c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3) + 256, yading@11: c.ppMode.maxTmpNoise); yading@11: } yading@11: } yading@11: yading@11: dstBlock+=8; yading@11: srcBlock+=8; yading@11: yading@11: #if TEMPLATE_PP_MMX yading@11: tmpXchg= tempBlock1; yading@11: tempBlock1= tempBlock2; yading@11: tempBlock2 = tmpXchg; yading@11: #endif yading@11: } yading@11: yading@11: if(mode & DERING){ yading@11: if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c); yading@11: } yading@11: yading@11: if((mode & TEMP_NOISE_FILTER)){ yading@11: RENAME(tempNoiseReducer)(dstBlock-8, dstStride, yading@11: c.tempBlurred[isColor] + y*dstStride + x, yading@11: c.tempBlurredPast[isColor] + (y>>3)*256 + (x>>3) + 256, yading@11: c.ppMode.maxTmpNoise); yading@11: } yading@11: yading@11: /* did we use a tmp buffer for the last lines*/ yading@11: if(y+15 >= height){ yading@11: uint8_t *dstBlock= &(dst[y*dstStride]); yading@11: if(width==FFABS(dstStride)) yading@11: linecpy(dstBlock, tempDst + dstStride, height-y, dstStride); yading@11: else{ yading@11: int i; yading@11: for(i=0; i max) max=yHistogram[i]; yading@11: yading@11: for(i=1; i<256; i++){ yading@11: int x; yading@11: int start=yHistogram[i-1]/(max/256+1); yading@11: int end=yHistogram[i]/(max/256+1); yading@11: int inc= end > start ? 1 : -1; yading@11: for(x=start; x!=end+inc; x+=inc) yading@11: dst[ i*dstStride + x]+=128; yading@11: } yading@11: yading@11: for(i=0; i<100; i+=2){ yading@11: dst[ (white)*dstStride + i]+=128; yading@11: dst[ (black)*dstStride + i]+=128; yading@11: } yading@11: } yading@11: #endif yading@11: yading@11: *c2= c; //copy local context back yading@11: yading@11: } yading@11: yading@11: #undef RENAME yading@11: #undef TEMPLATE_PP_C yading@11: #undef TEMPLATE_PP_ALTIVEC yading@11: #undef TEMPLATE_PP_MMX yading@11: #undef TEMPLATE_PP_MMXEXT yading@11: #undef TEMPLATE_PP_3DNOW yading@11: #undef TEMPLATE_PP_SSE2