FFmpeg: libavcodec/x86/mpegvideo_mmx

00001 /*
00002  * MPEG video MMX templates
00003  *
00004  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
00005  *
00006  * This file is part of FFmpeg.
00007  *
00008  * FFmpeg is free software; you can redistribute it and/or
00009  * modify it under the terms of the GNU Lesser General Public
00010  * License as published by the Free Software Foundation; either
00011  * version 2.1 of the License, or (at your option) any later version.
00012  *
00013  * FFmpeg is distributed in the hope that it will be useful,
00014  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00015  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00016  * Lesser General Public License for more details.
00017  *
00018  * You should have received a copy of the GNU Lesser General Public
00019  * License along with FFmpeg; if not, write to the Free Software
00020  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00021  */
00022 
00023 #undef MMREG_WIDTH
00024 #undef MM
00025 #undef MOVQ
00026 #undef SPREADW
00027 #undef PMAXW
00028 #undef PMAX
00029 #undef SAVE_SIGN
00030 #undef RESTORE_SIGN
00031 
00032 #if HAVE_SSE2
00033 #define MMREG_WIDTH "16"
00034 #define MM "%%xmm"
00035 #define MOVQ "movdqa"
00036 #define SPREADW(a) \
00037             "pshuflw $0, "a", "a"       \n\t"\
00038             "punpcklwd "a", "a"         \n\t"
00039 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
00040 #define PMAX(a,b) \
00041             "movhlps "a", "b"           \n\t"\
00042             PMAXW(b, a)\
00043             "pshuflw $0x0E, "a", "b"    \n\t"\
00044             PMAXW(b, a)\
00045             "pshuflw $0x01, "a", "b"    \n\t"\
00046             PMAXW(b, a)
00047 #else
00048 #define MMREG_WIDTH "8"
00049 #define MM "%%mm"
00050 #define MOVQ "movq"
00051 #if HAVE_MMX2
00052 #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
00053 #define PMAXW(a,b) "pmaxsw "a", "b"     \n\t"
00054 #define PMAX(a,b) \
00055             "pshufw $0x0E, "a", "b"     \n\t"\
00056             PMAXW(b, a)\
00057             "pshufw $0x01, "a", "b"     \n\t"\
00058             PMAXW(b, a)
00059 #else
00060 #define SPREADW(a) \
00061             "punpcklwd "a", "a"         \n\t"\
00062             "punpcklwd "a", "a"         \n\t"
00063 #define PMAXW(a,b) \
00064             "psubusw "a", "b"           \n\t"\
00065             "paddw "a", "b"             \n\t"
00066 #define PMAX(a,b)  \
00067             "movq "a", "b"              \n\t"\
00068             "psrlq $32, "a"             \n\t"\
00069             PMAXW(b, a)\
00070             "movq "a", "b"              \n\t"\
00071             "psrlq $16, "a"             \n\t"\
00072             PMAXW(b, a)
00073 
00074 #endif
00075 #endif
00076 
00077 #if HAVE_SSSE3
00078 #define SAVE_SIGN(a,b) \
00079             "movdqa "b", "a"            \n\t"\
00080             "pabsw  "b", "b"            \n\t"
00081 #define RESTORE_SIGN(a,b) \
00082             "psignw "a", "b"            \n\t"
00083 #else
00084 #define SAVE_SIGN(a,b) \
00085             "pxor "a", "a"              \n\t"\
00086             "pcmpgtw "b", "a"           \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
00087             "pxor "a", "b"              \n\t"\
00088             "psubw "a", "b"             \n\t" /* ABS(block[i]) */
00089 #define RESTORE_SIGN(a,b) \
00090             "pxor "a", "b"              \n\t"\
00091             "psubw "a", "b"             \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
00092 #endif
00093 
00094 static int RENAME(dct_quantize)(MpegEncContext *s,
00095                             DCTELEM *block, int n,
00096                             int qscale, int *overflow)
00097 {
00098     x86_reg last_non_zero_p1;
00099     int level=0, q; //=0 is because gcc says uninitialized ...
00100     const uint16_t *qmat, *bias;
00101     LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
00102 
00103     assert((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
00104 
00105     //s->fdct (block);
00106     RENAMEl(ff_fdct) (block); //cannot be anything else ...
00107 
00108     if(s->dct_error_sum)
00109         s->denoise_dct(s, block);
00110 
00111     if (s->mb_intra) {
00112         int dummy;
00113         if (n < 4){
00114             q = s->y_dc_scale;
00115             bias = s->q_intra_matrix16[qscale][1];
00116             qmat = s->q_intra_matrix16[qscale][0];
00117         }else{
00118             q = s->c_dc_scale;
00119             bias = s->q_chroma_intra_matrix16[qscale][1];
00120             qmat = s->q_chroma_intra_matrix16[qscale][0];
00121         }
00122         /* note: block[0] is assumed to be positive */
00123         if (!s->h263_aic) {
00124         __asm__ volatile (
00125                 "mul %%ecx                \n\t"
00126                 : "=d" (level), "=a"(dummy)
00127                 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
00128         );
00129         } else
00130             /* For AIC we skip quant/dequant of INTRADC */
00131             level = (block[0] + 4)>>3;
00132 
00133         block[0]=0; //avoid fake overflow
00134 //        temp_block[0] = (block[0] + (q >> 1)) / q;
00135         last_non_zero_p1 = 1;
00136     } else {
00137         last_non_zero_p1 = 0;
00138         bias = s->q_inter_matrix16[qscale][1];
00139         qmat = s->q_inter_matrix16[qscale][0];
00140     }
00141 
00142     if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
00143 
00144         __asm__ volatile(
00145             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
00146             SPREADW(MM"3")
00147             "pxor "MM"7, "MM"7                  \n\t" // 0
00148             "pxor "MM"4, "MM"4                  \n\t" // 0
00149             MOVQ" (%2), "MM"5                   \n\t" // qmat[0]
00150             "pxor "MM"6, "MM"6                  \n\t"
00151             "psubw (%3), "MM"6                  \n\t" // -bias[0]
00152             "mov $-128, %%"REG_a"               \n\t"
00153             ".p2align 4                         \n\t"
00154             "1:                                 \n\t"
00155             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
00156             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
00157             "psubusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
00158             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
00159             "por "MM"0, "MM"4                   \n\t"
00160             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
00161             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
00162             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
00163             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
00164             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
00165             "pandn "MM"1, "MM"0                 \n\t"
00166             PMAXW(MM"0", MM"3")
00167             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
00168             " js 1b                             \n\t"
00169             PMAX(MM"3", MM"0")
00170             "movd "MM"3, %%"REG_a"              \n\t"
00171             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
00172             : "+a" (last_non_zero_p1)
00173             : "r" (block+64), "r" (qmat), "r" (bias),
00174               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
00175               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
00176                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
00177         );
00178     }else{ // FMT_H263
00179         __asm__ volatile(
00180             "movd %%"REG_a", "MM"3              \n\t" // last_non_zero_p1
00181             SPREADW(MM"3")
00182             "pxor "MM"7, "MM"7                  \n\t" // 0
00183             "pxor "MM"4, "MM"4                  \n\t" // 0
00184             "mov $-128, %%"REG_a"               \n\t"
00185             ".p2align 4                         \n\t"
00186             "1:                                 \n\t"
00187             MOVQ" (%1, %%"REG_a"), "MM"0        \n\t" // block[i]
00188             SAVE_SIGN(MM"1", MM"0")                   // ABS(block[i])
00189             MOVQ" (%3, %%"REG_a"), "MM"6        \n\t" // bias[0]
00190             "paddusw "MM"6, "MM"0               \n\t" // ABS(block[i]) + bias[0]
00191             MOVQ" (%2, %%"REG_a"), "MM"5        \n\t" // qmat[i]
00192             "pmulhw "MM"5, "MM"0                \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
00193             "por "MM"0, "MM"4                   \n\t"
00194             RESTORE_SIGN(MM"1", MM"0")                // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
00195             MOVQ" "MM"0, (%5, %%"REG_a")        \n\t"
00196             "pcmpeqw "MM"7, "MM"0               \n\t" // out==0 ? 0xFF : 0x00
00197             MOVQ" (%4, %%"REG_a"), "MM"1        \n\t"
00198             MOVQ" "MM"7, (%1, %%"REG_a")        \n\t" // 0
00199             "pandn "MM"1, "MM"0                 \n\t"
00200             PMAXW(MM"0", MM"3")
00201             "add $"MMREG_WIDTH", %%"REG_a"      \n\t"
00202             " js 1b                             \n\t"
00203             PMAX(MM"3", MM"0")
00204             "movd "MM"3, %%"REG_a"              \n\t"
00205             "movzb %%al, %%"REG_a"              \n\t" // last_non_zero_p1
00206             : "+a" (last_non_zero_p1)
00207             : "r" (block+64), "r" (qmat+64), "r" (bias+64),
00208               "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
00209               XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
00210                                 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
00211         );
00212     }
00213     __asm__ volatile(
00214         "movd %1, "MM"1                     \n\t" // max_qcoeff
00215         SPREADW(MM"1")
00216         "psubusw "MM"1, "MM"4               \n\t"
00217         "packuswb "MM"4, "MM"4              \n\t"
00218 #if HAVE_SSE2
00219         "packuswb "MM"4, "MM"4              \n\t"
00220 #endif
00221         "movd "MM"4, %0                     \n\t" // *overflow
00222         : "=g" (*overflow)
00223         : "g" (s->max_qcoeff)
00224     );
00225 
00226     if(s->mb_intra) block[0]= level;
00227     else            block[0]= temp_block[0];
00228 
00229     if(s->dsp.idct_permutation_type == FF_SIMPLE_IDCT_PERM){
00230         if(last_non_zero_p1 <= 1) goto end;
00231         block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
00232         block[0x20] = temp_block[0x10];
00233         if(last_non_zero_p1 <= 4) goto end;
00234         block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
00235         block[0x09] = temp_block[0x03];
00236         if(last_non_zero_p1 <= 7) goto end;
00237         block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
00238         block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
00239         if(last_non_zero_p1 <= 11) goto end;
00240         block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
00241         block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
00242         block[0x0C] = temp_block[0x05];
00243         if(last_non_zero_p1 <= 16) goto end;
00244         block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
00245         block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
00246         block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
00247         block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
00248         if(last_non_zero_p1 <= 24) goto end;
00249         block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
00250         block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
00251         block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
00252         block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
00253         if(last_non_zero_p1 <= 32) goto end;
00254         block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
00255         block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
00256         block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
00257         block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
00258         if(last_non_zero_p1 <= 40) goto end;
00259         block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
00260         block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
00261         block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
00262         block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
00263         if(last_non_zero_p1 <= 48) goto end;
00264         block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
00265         block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
00266         block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
00267         block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
00268         if(last_non_zero_p1 <= 56) goto end;
00269         block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
00270         block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
00271         block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
00272         block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
00273     }else if(s->dsp.idct_permutation_type == FF_LIBMPEG2_IDCT_PERM){
00274         if(last_non_zero_p1 <= 1) goto end;
00275         block[0x04] = temp_block[0x01];
00276         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
00277         if(last_non_zero_p1 <= 4) goto end;
00278         block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
00279         block[0x05] = temp_block[0x03];
00280         if(last_non_zero_p1 <= 7) goto end;
00281         block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
00282         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
00283         if(last_non_zero_p1 <= 11) goto end;
00284         block[0x1C] = temp_block[0x19];
00285         block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
00286         block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
00287         if(last_non_zero_p1 <= 16) goto end;
00288         block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
00289         block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
00290         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
00291         block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
00292         if(last_non_zero_p1 <= 24) goto end;
00293         block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
00294         block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
00295         block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
00296         block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
00297         if(last_non_zero_p1 <= 32) goto end;
00298         block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
00299         block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
00300         block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
00301         block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
00302         if(last_non_zero_p1 <= 40) goto end;
00303         block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
00304         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
00305         block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
00306         block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
00307         if(last_non_zero_p1 <= 48) goto end;
00308         block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
00309         block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
00310             block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
00311         block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
00312         if(last_non_zero_p1 <= 56) goto end;
00313         block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
00314         block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
00315         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
00316         block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
00317     }else{
00318         if(last_non_zero_p1 <= 1) goto end;
00319         block[0x01] = temp_block[0x01];
00320         block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
00321         if(last_non_zero_p1 <= 4) goto end;
00322         block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
00323         block[0x03] = temp_block[0x03];
00324         if(last_non_zero_p1 <= 7) goto end;
00325         block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
00326         block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
00327         if(last_non_zero_p1 <= 11) goto end;
00328         block[0x19] = temp_block[0x19];
00329         block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
00330         block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
00331         if(last_non_zero_p1 <= 16) goto end;
00332         block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
00333         block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
00334         block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
00335         block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
00336         if(last_non_zero_p1 <= 24) goto end;
00337         block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
00338         block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
00339         block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
00340         block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
00341         if(last_non_zero_p1 <= 32) goto end;
00342         block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
00343         block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
00344         block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
00345         block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
00346         if(last_non_zero_p1 <= 40) goto end;
00347         block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
00348         block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
00349         block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
00350         block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
00351         if(last_non_zero_p1 <= 48) goto end;
00352         block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
00353         block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
00354         block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
00355         block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
00356         if(last_non_zero_p1 <= 56) goto end;
00357         block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
00358         block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
00359         block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
00360         block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
00361     }
00362     end:
00363 /*
00364     for(i=0; i<last_non_zero_p1; i++)
00365     {
00366        int j= zigzag_direct_noperm[i];
00367        block[block_permute_op(j)]= temp_block[j];
00368     }
00369 */
00370 
00371     return last_non_zero_p1 - 1;
00372 }
libavcodec/x86/mpegvideo_mmx_template.c
