00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
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" \
00087 "pxor "a", "b" \n\t"\
00088 "psubw "a", "b" \n\t"
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;
00100 const uint16_t *qmat, *bias;
00101 LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
00102
00103 assert((7&(int)(&temp_block[0])) == 0);
00104
00105
00106 RENAMEl(ff_fdct) (block);
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 else
00116 q = s->c_dc_scale;
00117
00118 if (!s->h263_aic) {
00119 __asm__ volatile (
00120 "mul %%ecx \n\t"
00121 : "=d" (level), "=a"(dummy)
00122 : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
00123 );
00124 } else
00125
00126 level = (block[0] + 4)>>3;
00127
00128 block[0]=0;
00129
00130 last_non_zero_p1 = 1;
00131 bias = s->q_intra_matrix16[qscale][1];
00132 qmat = s->q_intra_matrix16[qscale][0];
00133 } else {
00134 last_non_zero_p1 = 0;
00135 bias = s->q_inter_matrix16[qscale][1];
00136 qmat = s->q_inter_matrix16[qscale][0];
00137 }
00138
00139 if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
00140
00141 __asm__ volatile(
00142 "movd %%"REG_a", "MM"3 \n\t"
00143 SPREADW(MM"3")
00144 "pxor "MM"7, "MM"7 \n\t"
00145 "pxor "MM"4, "MM"4 \n\t"
00146 MOVQ" (%2), "MM"5 \n\t"
00147 "pxor "MM"6, "MM"6 \n\t"
00148 "psubw (%3), "MM"6 \n\t"
00149 "mov $-128, %%"REG_a" \n\t"
00150 ".p2align 4 \n\t"
00151 "1: \n\t"
00152 MOVQ" (%1, %%"REG_a"), "MM"0 \n\t"
00153 SAVE_SIGN(MM"1", MM"0")
00154 "psubusw "MM"6, "MM"0 \n\t"
00155 "pmulhw "MM"5, "MM"0 \n\t"
00156 "por "MM"0, "MM"4 \n\t"
00157 RESTORE_SIGN(MM"1", MM"0")
00158 MOVQ" "MM"0, (%5, %%"REG_a") \n\t"
00159 "pcmpeqw "MM"7, "MM"0 \n\t"
00160 MOVQ" (%4, %%"REG_a"), "MM"1 \n\t"
00161 MOVQ" "MM"7, (%1, %%"REG_a") \n\t"
00162 "pandn "MM"1, "MM"0 \n\t"
00163 PMAXW(MM"0", MM"3")
00164 "add $"MMREG_WIDTH", %%"REG_a" \n\t"
00165 " js 1b \n\t"
00166 PMAX(MM"3", MM"0")
00167 "movd "MM"3, %%"REG_a" \n\t"
00168 "movzb %%al, %%"REG_a" \n\t"
00169 : "+a" (last_non_zero_p1)
00170 : "r" (block+64), "r" (qmat), "r" (bias),
00171 "r" (inv_zigzag_direct16+64), "r" (temp_block+64)
00172 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
00173 "%xmm4", "%xmm5", "%xmm6", "%xmm7")
00174 );
00175 }else{
00176 __asm__ volatile(
00177 "movd %%"REG_a", "MM"3 \n\t"
00178 SPREADW(MM"3")
00179 "pxor "MM"7, "MM"7 \n\t"
00180 "pxor "MM"4, "MM"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>"4 \n\t"
00>