1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libtheora/lib/x86_vc/mmxloop.h Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,219 @@
1.4 +#if !defined(_x86_vc_mmxloop_H)
1.5 +# define _x86_vc_mmxloop_H (1)
1.6 +# include <stddef.h>
1.7 +# include "x86int.h"
1.8 +
1.9 +#if defined(OC_X86_ASM)
1.10 +
1.11 +/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
1.12 + On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
1.13 + mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}; mm0 and mm3 are clobbered.*/
1.14 +#define OC_LOOP_FILTER8_MMX __asm{ \
1.15 + /*mm7=0*/ \
1.16 + __asm pxor mm7,mm7 \
1.17 + /*mm6:mm0={a0,...,a7}*/ \
1.18 + __asm movq mm6,mm0 \
1.19 + __asm punpcklbw mm0,mm7 \
1.20 + __asm punpckhbw mm6,mm7 \
1.21 + /*mm3:mm5={d0,...,d7}*/ \
1.22 + __asm movq mm5,mm3 \
1.23 + __asm punpcklbw mm3,mm7 \
1.24 + __asm punpckhbw mm5,mm7 \
1.25 + /*mm6:mm0={a0-d0,...,a7-d7}*/ \
1.26 + __asm psubw mm0,mm3 \
1.27 + __asm psubw mm6,mm5 \
1.28 + /*mm3:mm1={b0,...,b7}*/ \
1.29 + __asm movq mm3,mm1 \
1.30 + __asm punpcklbw mm1,mm7 \
1.31 + __asm movq mm4,mm2 \
1.32 + __asm punpckhbw mm3,mm7 \
1.33 + /*mm5:mm4={c0,...,c7}*/ \
1.34 + __asm movq mm5,mm2 \
1.35 + __asm punpcklbw mm4,mm7 \
1.36 + __asm punpckhbw mm5,mm7 \
1.37 + /*mm7={3}x4 \
1.38 + mm5:mm4={c0-b0,...,c7-b7}*/ \
1.39 + __asm pcmpeqw mm7,mm7 \
1.40 + __asm psubw mm4,mm1 \
1.41 + __asm psrlw mm7,14 \
1.42 + __asm psubw mm5,mm3 \
1.43 + /*Scale by 3.*/ \
1.44 + __asm pmullw mm4,mm7 \
1.45 + __asm pmullw mm5,mm7 \
1.46 + /*mm7={4}x4 \
1.47 + mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
1.48 + __asm psrlw mm7,1 \
1.49 + __asm paddw mm4,mm0 \
1.50 + __asm psllw mm7,2 \
1.51 + __asm movq mm0,[LL] \
1.52 + __asm paddw mm5,mm6 \
1.53 + /*R_i has the range [-127,128], so we compute -R_i instead. \
1.54 + mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
1.55 + __asm psubw mm4,mm7 \
1.56 + __asm psubw mm5,mm7 \
1.57 + __asm psraw mm4,3 \
1.58 + __asm psraw mm5,3 \
1.59 + __asm pcmpeqb mm7,mm7 \
1.60 + __asm packsswb mm4,mm5 \
1.61 + __asm pxor mm6,mm6 \
1.62 + __asm pxor mm4,mm7 \
1.63 + __asm packuswb mm1,mm3 \
1.64 + /*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
1.65 + /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
1.66 + we have to split things by sign (the other option is to work in 16 bits, \
1.67 + but working in 8 bits gives much better parallelism). \
1.68 + We compute abs(R_i), but save a mask of which terms were negative in mm6. \
1.69 + Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
1.70 + Finally, we split mm4 into positive and negative pieces using the mask in \
1.71 + mm6, and add and subtract them as appropriate.*/ \
1.72 + /*mm4=abs(-R_i)*/ \
1.73 + /*mm7=255-2*L*/ \
1.74 + __asm pcmpgtb mm6,mm4 \
1.75 + __asm psubb mm7,mm0 \
1.76 + __asm pxor mm4,mm6 \
1.77 + __asm psubb mm7,mm0 \
1.78 + __asm psubb mm4,mm6 \
1.79 + /*mm7=255-max(2*L-abs(R_i),0)*/ \
1.80 + __asm paddusb mm7,mm4 \
1.81 + /*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
1.82 + __asm paddusb mm4,mm7 \
1.83 + __asm psubusb mm4,mm7 \
1.84 + /*Now split mm4 by the original sign of -R_i.*/ \
1.85 + __asm movq mm5,mm4 \
1.86 + __asm pand mm4,mm6 \
1.87 + __asm pandn mm6,mm5 \
1.88 + /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
1.89 + /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
1.90 + __asm paddusb mm1,mm4 \
1.91 + __asm psubusb mm2,mm4 \
1.92 + __asm psubusb mm1,mm6 \
1.93 + __asm paddusb mm2,mm6 \
1.94 +}
1.95 +
1.96 +#define OC_LOOP_FILTER_V_MMX(_pix,_ystride,_ll) \
1.97 + do{ \
1.98 + /*Used local variable pix__ in order to fix compilation errors like: \
1.99 + "error C2425: 'SHL' : non-constant expression in 'second operand'".*/ \
1.100 + unsigned char *pix__; \
1.101 + unsigned char *ll__; \
1.102 + ll__=(_ll); \
1.103 + pix__=(_pix); \
1.104 + __asm mov YSTRIDE,_ystride \
1.105 + __asm mov LL,ll__ \
1.106 + __asm mov PIX,pix__ \
1.107 + __asm sub PIX,YSTRIDE \
1.108 + __asm sub PIX,YSTRIDE \
1.109 + /*mm0={a0,...,a7}*/ \
1.110 + __asm movq mm0,[PIX] \
1.111 + /*ystride3=_ystride*3*/ \
1.112 + __asm lea YSTRIDE3,[YSTRIDE+YSTRIDE*2] \
1.113 + /*mm3={d0,...,d7}*/ \
1.114 + __asm movq mm3,[PIX+YSTRIDE3] \
1.115 + /*mm1={b0,...,b7}*/ \
1.116 + __asm movq mm1,[PIX+YSTRIDE] \
1.117 + /*mm2={c0,...,c7}*/ \
1.118 + __asm movq mm2,[PIX+YSTRIDE*2] \
1.119 + OC_LOOP_FILTER8_MMX \
1.120 + /*Write it back out.*/ \
1.121 + __asm movq [PIX+YSTRIDE],mm1 \
1.122 + __asm movq [PIX+YSTRIDE*2],mm2 \
1.123 + } \
1.124 + while(0)
1.125 +
1.126 +#define OC_LOOP_FILTER_H_MMX(_pix,_ystride,_ll) \
1.127 + do{ \
1.128 + /*Used local variable ll__ in order to fix compilation errors like: \
1.129 + "error C2443: operand size conflict".*/ \
1.130 + unsigned char *ll__; \
1.131 + unsigned char *pix__; \
1.132 + ll__=(_ll); \
1.133 + pix__=(_pix)-2; \
1.134 + __asm mov PIX,pix__ \
1.135 + __asm mov YSTRIDE,_ystride \
1.136 + __asm mov LL,ll__ \
1.137 + /*x x x x d0 c0 b0 a0*/ \
1.138 + __asm movd mm0,[PIX] \
1.139 + /*x x x x d1 c1 b1 a1*/ \
1.140 + __asm movd mm1,[PIX+YSTRIDE] \
1.141 + /*ystride3=_ystride*3*/ \
1.142 + __asm lea YSTRIDE3,[YSTRIDE+YSTRIDE*2] \
1.143 + /*x x x x d2 c2 b2 a2*/ \
1.144 + __asm movd mm2,[PIX+YSTRIDE*2] \
1.145 + /*x x x x d3 c3 b3 a3*/ \
1.146 + __asm lea D,[PIX+YSTRIDE*4] \
1.147 + __asm movd mm3,[PIX+YSTRIDE3] \
1.148 + /*x x x x d4 c4 b4 a4*/ \
1.149 + __asm movd mm4,[D] \
1.150 + /*x x x x d5 c5 b5 a5*/ \
1.151 + __asm movd mm5,[D+YSTRIDE] \
1.152 + /*x x x x d6 c6 b6 a6*/ \
1.153 + __asm movd mm6,[D+YSTRIDE*2] \
1.154 + /*x x x x d7 c7 b7 a7*/ \
1.155 + __asm movd mm7,[D+YSTRIDE3] \
1.156 + /*mm0=d1 d0 c1 c0 b1 b0 a1 a0*/ \
1.157 + __asm punpcklbw mm0,mm1 \
1.158 + /*mm2=d3 d2 c3 c2 b3 b2 a3 a2*/ \
1.159 + __asm punpcklbw mm2,mm3 \
1.160 + /*mm3=d1 d0 c1 c0 b1 b0 a1 a0*/ \
1.161 + __asm movq mm3,mm0 \
1.162 + /*mm0=b3 b2 b1 b0 a3 a2 a1 a0*/ \
1.163 + __asm punpcklwd mm0,mm2 \
1.164 + /*mm3=d3 d2 d1 d0 c3 c2 c1 c0*/ \
1.165 + __asm punpckhwd mm3,mm2 \
1.166 + /*mm1=b3 b2 b1 b0 a3 a2 a1 a0*/ \
1.167 + __asm movq mm1,mm0 \
1.168 + /*mm4=d5 d4 c5 c4 b5 b4 a5 a4*/ \
1.169 + __asm punpcklbw mm4,mm5 \
1.170 + /*mm6=d7 d6 c7 c6 b7 b6 a7 a6*/ \
1.171 + __asm punpcklbw mm6,mm7 \
1.172 + /*mm5=d5 d4 c5 c4 b5 b4 a5 a4*/ \
1.173 + __asm movq mm5,mm4 \
1.174 + /*mm4=b7 b6 b5 b4 a7 a6 a5 a4*/ \
1.175 + __asm punpcklwd mm4,mm6 \
1.176 + /*mm5=d7 d6 d5 d4 c7 c6 c5 c4*/ \
1.177 + __asm punpckhwd mm5,mm6 \
1.178 + /*mm2=d3 d2 d1 d0 c3 c2 c1 c0*/ \
1.179 + __asm movq mm2,mm3 \
1.180 + /*mm0=a7 a6 a5 a4 a3 a2 a1 a0*/ \
1.181 + __asm punpckldq mm0,mm4 \
1.182 + /*mm1=b7 b6 b5 b4 b3 b2 b1 b0*/ \
1.183 + __asm punpckhdq mm1,mm4 \
1.184 + /*mm2=c7 c6 c5 c4 c3 c2 c1 c0*/ \
1.185 + __asm punpckldq mm2,mm5 \
1.186 + /*mm3=d7 d6 d5 d4 d3 d2 d1 d0*/ \
1.187 + __asm punpckhdq mm3,mm5 \
1.188 + OC_LOOP_FILTER8_MMX \
1.189 + /*mm2={b0+R_0'',...,b7+R_7''}*/ \
1.190 + __asm movq mm0,mm1 \
1.191 + /*mm1={b0+R_0'',c0-R_0'',...,b3+R_3'',c3-R_3''}*/ \
1.192 + __asm punpcklbw mm1,mm2 \
1.193 + /*mm2={b4+R_4'',c4-R_4'',...,b7+R_7'',c7-R_7''}*/ \
1.194 + __asm punpckhbw mm0,mm2 \
1.195 + /*[d]=c1 b1 c0 b0*/ \
1.196 + __asm movd D,mm1 \
1.197 + __asm mov [PIX+1],D_WORD \
1.198 + __asm psrlq mm1,32 \
1.199 + __asm shr D,16 \
1.200 + __asm mov [PIX+YSTRIDE+1],D_WORD \
1.201 + /*[d]=c3 b3 c2 b2*/ \
1.202 + __asm movd D,mm1 \
1.203 + __asm mov [PIX+YSTRIDE*2+1],D_WORD \
1.204 + __asm shr D,16 \
1.205 + __asm mov [PIX+YSTRIDE3+1],D_WORD \
1.206 + __asm lea PIX,[PIX+YSTRIDE*4] \
1.207 + /*[d]=c5 b5 c4 b4*/ \
1.208 + __asm movd D,mm0 \
1.209 + __asm mov [PIX+1],D_WORD \
1.210 + __asm psrlq mm0,32 \
1.211 + __asm shr D,16 \
1.212 + __asm mov [PIX+YSTRIDE+1],D_WORD \
1.213 + /*[d]=c7 b7 c6 b6*/ \
1.214 + __asm movd D,mm0 \
1.215 + __asm mov [PIX+YSTRIDE*2+1],D_WORD \
1.216 + __asm shr D,16 \
1.217 + __asm mov [PIX+YSTRIDE3+1],D_WORD \
1.218 + } \
1.219 + while(0)
1.220 +
1.221 +# endif
1.222 +#endif
