Mercurial > libavcodec.hg

comparison i386/dsputil_mmx.c @ 2217:9492be49de46 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
h264_qpel8_hv_lowpass_mmx2/3dnow
author michael
date Wed, 08 Sep 2004 18:49:12 +0000
parents 1d6835102c51
children 7e0b2e86afa9
comparison
equal deleted inserted replaced
2216:1d6835102c51 2217:9492be49de46
2550 src += 4-13*srcStride;\ 2550 src += 4-13*srcStride;\
2551 dst += 4-8*dstStride;\ 2551 dst += 4-8*dstStride;\
2552 }\ 2552 }\
2553 }\ 2553 }\
2554 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\ 2554 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
2555 OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride);\ 2555 int h=8;\
2556 OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(dst+4, tmp , src+4, dstStride, tmpStride, srcStride);\ 2556 int w=4;\
2557 src += 4*srcStride;\ 2557 src -= 2*srcStride+2;\
2558 dst += 4*dstStride;\ 2558 while(w--){\
2559 OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride);\ 2559 asm volatile(\
2560 OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(dst+4, tmp , src+4, dstStride, tmpStride, srcStride);\ 2560 "pxor %%mm7, %%mm7 \n\t"\
2561 "movd (%0), %%mm0 \n\t"\
2562 "addl %2, %0 \n\t"\
2563 "movd (%0), %%mm1 \n\t"\
2564 "addl %2, %0 \n\t"\
2565 "movd (%0), %%mm2 \n\t"\
2566 "addl %2, %0 \n\t"\
2567 "movd (%0), %%mm3 \n\t"\
2568 "addl %2, %0 \n\t"\
2569 "movd (%0), %%mm4 \n\t"\
2570 "addl %2, %0 \n\t"\
2571 "punpcklbw %%mm7, %%mm0 \n\t"\
2572 "punpcklbw %%mm7, %%mm1 \n\t"\
2573 "punpcklbw %%mm7, %%mm2 \n\t"\
2574 "punpcklbw %%mm7, %%mm3 \n\t"\
2575 "punpcklbw %%mm7, %%mm4 \n\t"\
2576 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*4)\
2577 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*4)\
2578 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*4)\
2579 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*4)\
2580 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*8*4)\
2581 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*8*4)\
2582 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*8*4)\
2583 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*8*4)\
2584 \
2585 : "+a"(src)\
2586 : "c"(tmp), "S"(srcStride), "m"(ff_pw_5)\
2587 : "memory"\
2588 );\
2589 tmp += 4;\
2590 src += 4 - 13*srcStride;\
2591 }\
2592 tmp -= 4*4;\
2593 asm volatile(\
2594 "movq %4, %%mm6 \n\t"\
2595 "1: \n\t"\
2596 "movq (%0), %%mm0 \n\t"\
2597 "movq 8(%0), %%mm3 \n\t"\
2598 "movq 2(%0), %%mm1 \n\t"\
2599 "movq 10(%0), %%mm4 \n\t"\
2600 "paddw %%mm4, %%mm0 \n\t"\
2601 "paddw %%mm3, %%mm1 \n\t"\
2602 "paddw 18(%0), %%mm3 \n\t"\
2603 "paddw 16(%0), %%mm4 \n\t"\
2604 "movq 4(%0), %%mm2 \n\t"\
2605 "movq 12(%0), %%mm5 \n\t"\
2606 "paddw 6(%0), %%mm2 \n\t"\
2607 "paddw 14(%0), %%mm5 \n\t"\
2608 "psubw %%mm1, %%mm0 \n\t"\
2609 "psubw %%mm4, %%mm3 \n\t"\
2610 "psraw $2, %%mm0 \n\t"\
2611 "psraw $2, %%mm3 \n\t"\
2612 "psubw %%mm1, %%mm0 \n\t"\
2613 "psubw %%mm4, %%mm3 \n\t"\
2614 "paddsw %%mm2, %%mm0 \n\t"\
2615 "paddsw %%mm5, %%mm3 \n\t"\
2616 "psraw $2, %%mm0 \n\t"\
2617 "psraw $2, %%mm3 \n\t"\
2618 "paddw %%mm6, %%mm2 \n\t"\
2619 "paddw %%mm6, %%mm5 \n\t"\
2620 "paddw %%mm2, %%mm0 \n\t"\
2621 "paddw %%mm5, %%mm3 \n\t"\
2622 "psraw $6, %%mm0 \n\t"\
2623 "psraw $6, %%mm3 \n\t"\
2624 "packuswb %%mm3, %%mm0 \n\t"\
2625 OP(%%mm0, (%1),%%mm7, q)\
2626 "addl $32, %0 \n\t"\
2627 "addl %3, %1 \n\t"\
2628 "decl %2 \n\t"\
2629 " jnz 1b \n\t"\
2630 : "+a"(tmp), "+c"(dst), "+m"(h)\
2631 : "S"(dstStride), "m"(ff_pw_32)\
2632 : "memory"\
2633 );\
2561 }\ 2634 }\
2562 static void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\ 2635 static void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
2563 OPNAME ## h264_qpel8_v_lowpass_ ## MMX(dst , src , dstStride, srcStride);\ 2636 OPNAME ## h264_qpel8_v_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
2564 OPNAME ## h264_qpel8_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\ 2637 OPNAME ## h264_qpel8_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
2565 src += 8*srcStride;\ 2638 src += 8*srcStride;\