From face578d56c2d1375e40d5e2a28acc122132bc55 Mon Sep 17 00:00:00 2001 From: "Ronald S. Bultje" Date: Fri, 20 Sep 2013 08:01:19 -0400 Subject: [PATCH] Rewrite emu_edge functions to have separate src/dst_stride arguments. This allows supporting files for which the image stride is smaller than the max. block size + number of subpel mc taps, e.g. a 64x64 VP9 file or a 16x16 VP8 file with -fflags +emu_edge. --- libavcodec/cavs.c | 7 +- libavcodec/diracdec.c | 3 +- libavcodec/h264.c | 10 +- libavcodec/mpegvideo.c | 16 +- libavcodec/mpegvideo_enc.c | 6 +- libavcodec/mpegvideo_motion.c | 74 +-- libavcodec/rv34.c | 15 +- libavcodec/snow.c | 4 +- libavcodec/svq3.c | 6 +- libavcodec/vc1dec.c | 47 +- libavcodec/videodsp.h | 20 +- libavcodec/videodsp_template.c | 32 +- libavcodec/vp3.c | 5 +- libavcodec/vp56.c | 2 +- libavcodec/vp8.c | 38 +- libavcodec/vp8.h | 2 +- libavcodec/wmv2.c | 16 +- libavcodec/x86/dsputil_mmx.c | 7 +- libavcodec/x86/videodsp.asm | 860 +++++++++++++-------------------- libavcodec/x86/videodsp_init.c | 179 +++++-- 20 files changed, 666 insertions(+), 683 deletions(-) diff --git a/libavcodec/cavs.c b/libavcodec/cavs.c index 12af0c9a41..28327d7189 100644 --- a/libavcodec/cavs.c +++ b/libavcodec/cavs.c @@ -405,7 +405,8 @@ static inline void mc_dir_part(AVSContext *h, AVFrame *pic, || full_my < 0-extra_height || full_mx + 16/*FIXME*/ > pic_width + extra_width || full_my + 16/*FIXME*/ > pic_height + extra_height){ - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_y - 2 - 2*h->l_stride, h->l_stride, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->l_stride, + src_y - 2 - 2*h->l_stride, h->l_stride, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height); src_y= h->edge_emu_buffer + 2 + 2*h->l_stride; emu=1; @@ -414,14 +415,14 @@ static inline void mc_dir_part(AVSContext *h, AVFrame *pic, qpix_op[luma_xy](dest_y, src_y, h->l_stride); //FIXME try variable height perhaps? if(emu){ - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cb, h->c_stride, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->c_stride, src_cb, h->c_stride, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1); src_cb= h->edge_emu_buffer; } chroma_op(dest_cb, src_cb, h->c_stride, chroma_height, mx&7, my&7); if(emu){ - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cr, h->c_stride, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->c_stride, src_cr, h->c_stride, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1); src_cr= h->edge_emu_buffer; } diff --git a/libavcodec/diracdec.c b/libavcodec/diracdec.c index e5239a77f8..3d70f1723c 100644 --- a/libavcodec/diracdec.c +++ b/libavcodec/diracdec.c @@ -1414,7 +1414,8 @@ static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5], y + p->yblen > p->height+EDGE_WIDTH/2 || x < 0 || y < 0) { for (i = 0; i < nplanes; i++) { - ff_emulated_edge_mc(s->edge_emu_buffer[i], src[i], p->stride, + ff_emulated_edge_mc(s->edge_emu_buffer[i], p->stride, + src[i], p->stride, p->xblen, p->yblen, x, y, p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2); src[i] = s->edge_emu_buffer[i]; diff --git a/libavcodec/h264.c b/libavcodec/h264.c index acaab39f08..f2ee328124 100644 --- a/libavcodec/h264.c +++ b/libavcodec/h264.c @@ -932,7 +932,7 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic, full_my < 0 - extra_height || full_mx + 16 /*FIXME*/ > pic_width + extra_width || full_my + 16 /*FIXME*/ > pic_height + extra_height) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->mb_linesize, src_y - (2 << pixel_shift) - 2 * h->mb_linesize, h->mb_linesize, 16 + 5, 16 + 5 /*FIXME*/, full_mx - 2, @@ -951,7 +951,7 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic, if (chroma_idc == 3 /* yuv444 */) { src_cb = pic->f.data[1] + offset; if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->mb_linesize, src_cb - (2 << pixel_shift) - 2 * h->mb_linesize, h->mb_linesize, 16 + 5, 16 + 5 /*FIXME*/, @@ -965,7 +965,7 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic, src_cr = pic->f.data[2] + offset; if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->mb_linesize, src_cr - (2 << pixel_shift) - 2 * h->mb_linesize, h->mb_linesize, 16 + 5, 16 + 5 /*FIXME*/, @@ -992,7 +992,7 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic, (my >> ysh) * h->mb_uvlinesize; if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cb, h->mb_uvlinesize, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->mb_uvlinesize, src_cb, h->mb_uvlinesize, 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh), pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */)); src_cb = h->edge_emu_buffer; @@ -1002,7 +1002,7 @@ static av_always_inline void mc_dir_part(H264Context *h, Picture *pic, mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7); if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cr, h->mb_uvlinesize, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->mb_uvlinesize, src_cr, h->mb_uvlinesize, 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh), pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */)); src_cr = h->edge_emu_buffer; diff --git a/libavcodec/mpegvideo.c b/libavcodec/mpegvideo.c index 96afca5d7c..32d6fa71f8 100644 --- a/libavcodec/mpegvideo.c +++ b/libavcodec/mpegvideo.c @@ -2204,7 +2204,8 @@ static inline int hpel_motion_lowres(MpegEncContext *s, if ((unsigned)src_x > FFMAX( h_edge_pos - (!!sx) - w, 0) || (unsigned)src_y > FFMAX((v_edge_pos >> field_based) - (!!sy) - h, 0)) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize, w + 1, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + src, s->linesize, w + 1, (h + 1) << field_based, src_x, src_y << field_based, h_edge_pos, @@ -2306,18 +2307,20 @@ static av_always_inline void mpeg_motion_lowres(MpegEncContext *s, if ((unsigned) src_x > FFMAX( h_edge_pos - (!!sx) - 2 * block_s, 0) || uvsrc_y<0 || (unsigned) src_y > FFMAX((v_edge_pos >> field_based) - (!!sy) - h, 0)) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, linesize >> field_based, ptr_y, linesize >> field_based, 17, 17 + field_based, src_x, src_y << field_based, h_edge_pos, v_edge_pos); ptr_y = s->edge_emu_buffer; if (!CONFIG_GRAY || !(s->flags & CODEC_FLAG_GRAY)) { uint8_t *uvbuf = s->edge_emu_buffer + 18 * s->linesize; - s->vdsp.emulated_edge_mc(uvbuf , ptr_cb, uvlinesize >> field_based, 9, + s->vdsp.emulated_edge_mc(uvbuf, uvlinesize >> field_based, + ptr_cb, uvlinesize >> field_based, 9, 9 + field_based, uvsrc_x, uvsrc_y << field_based, h_edge_pos >> 1, v_edge_pos >> 1); - s->vdsp.emulated_edge_mc(uvbuf + 16, ptr_cr, uvlinesize >> field_based, 9, + s->vdsp.emulated_edge_mc(uvbuf + 16, uvlinesize >> field_based, + ptr_cr, uvlinesize >> field_based, 9, 9 + field_based, uvsrc_x, uvsrc_y << field_based, h_edge_pos >> 1, v_edge_pos >> 1); @@ -2390,7 +2393,7 @@ static inline void chroma_4mv_motion_lowres(MpegEncContext *s, ptr = ref_picture[1] + offset; if ((unsigned) src_x > FFMAX(h_edge_pos - (!!sx) - block_s, 0) || (unsigned) src_y > FFMAX(v_edge_pos - (!!sy) - block_s, 0)) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, ptr, s->uvlinesize, 9, 9, src_x, src_y, h_edge_pos, v_edge_pos); ptr = s->edge_emu_buffer; emu = 1; @@ -2401,7 +2404,8 @@ static inline void chroma_4mv_motion_lowres(MpegEncContext *s, ptr = ref_picture[2] + offset; if (emu) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, + ptr, s->uvlinesize, 9, 9, src_x, src_y, h_edge_pos, v_edge_pos); ptr = s->edge_emu_buffer; } diff --git a/libavcodec/mpegvideo_enc.c b/libavcodec/mpegvideo_enc.c index 6f0d96168a..7b69acaa97 100644 --- a/libavcodec/mpegvideo_enc.c +++ b/libavcodec/mpegvideo_enc.c @@ -1819,14 +1819,14 @@ static av_always_inline void encode_mb_internal(MpegEncContext *s, uint8_t *ebuf = s->edge_emu_buffer + 32; int cw = (s->width + s->chroma_x_shift) >> s->chroma_x_shift; int ch = (s->height + s->chroma_y_shift) >> s->chroma_y_shift; - s->vdsp.emulated_edge_mc(ebuf, ptr_y, wrap_y, 16, 16, mb_x * 16, + s->vdsp.emulated_edge_mc(ebuf, wrap_y, ptr_y, wrap_y, 16, 16, mb_x * 16, mb_y * 16, s->width, s->height); ptr_y = ebuf; - s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y, ptr_cb, wrap_c, mb_block_width, + s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y, wrap_c, ptr_cb, wrap_c, mb_block_width, mb_block_height, mb_x * mb_block_width, mb_y * mb_block_height, cw, ch); ptr_cb = ebuf + 18 * wrap_y; - s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y + 16, ptr_cr, wrap_c, mb_block_width, + s->vdsp.emulated_edge_mc(ebuf + 18 * wrap_y + 16, wrap_c, ptr_cr, wrap_c, mb_block_width, mb_block_height, mb_x * mb_block_width, mb_y * mb_block_height, cw, ch); ptr_cr = ebuf + 18 * wrap_y + 16; diff --git a/libavcodec/mpegvideo_motion.c b/libavcodec/mpegvideo_motion.c index e5b7e4648a..09061d8dfe 100644 --- a/libavcodec/mpegvideo_motion.c +++ b/libavcodec/mpegvideo_motion.c @@ -63,7 +63,9 @@ static void gmc1_motion(MpegEncContext *s, if( (unsigned)src_x >= FFMAX(s->h_edge_pos - 17, 0) || (unsigned)src_y >= FFMAX(s->v_edge_pos - 17, 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, linesize, 17, 17, src_x, src_y, s->h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, linesize, + ptr, linesize, 17, 17, src_x, src_y, + s->h_edge_pos, s->v_edge_pos); ptr= s->edge_emu_buffer; } @@ -100,7 +102,9 @@ static void gmc1_motion(MpegEncContext *s, ptr = ref_picture[1] + offset; if( (unsigned)src_x >= FFMAX((s->h_edge_pos>>1) - 9, 0) || (unsigned)src_y >= FFMAX((s->v_edge_pos>>1) - 9, 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, uvlinesize, + ptr, uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; emu=1; } @@ -108,7 +112,9 @@ static void gmc1_motion(MpegEncContext *s, ptr = ref_picture[2] + offset; if(emu){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, uvlinesize, 9, 9, src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, uvlinesize, + ptr, uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; } s->dsp.gmc1(dest_cr, ptr, uvlinesize, 8, motion_x&15, motion_y&15, 128 - s->no_rounding); @@ -196,8 +202,9 @@ static inline int hpel_motion(MpegEncContext *s, if(s->unrestricted_mv && (s->flags&CODEC_FLAG_EMU_EDGE)){ if( (unsigned)src_x > FFMAX(s->h_edge_pos - (motion_x&1) - 8, 0) || (unsigned)src_y > FFMAX(s->v_edge_pos - (motion_y&1) - 8, 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, src, s->linesize, 9, 9, - src_x, src_y, s->h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + src, s->linesize, 9, 9, + src_x, src_y, s->h_edge_pos, s->v_edge_pos); src= s->edge_emu_buffer; emu=1; } @@ -287,19 +294,19 @@ if(s->quarter_sample) "MPEG motion vector out of boundary (%d %d)\n", src_x, src_y); return; } - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize, - 17, 17+field_based, - src_x, src_y<h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + ptr_y, s->linesize, 17, 17+field_based, + src_x, src_y<h_edge_pos, s->v_edge_pos); ptr_y = s->edge_emu_buffer; if(!CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){ uint8_t *uvbuf= s->edge_emu_buffer+18*s->linesize; - s->vdsp.emulated_edge_mc(uvbuf , + s->vdsp.emulated_edge_mc(uvbuf, s->uvlinesize, ptr_cb, s->uvlinesize, 9, 9+field_based, uvsrc_x, uvsrc_y<h_edge_pos>>1, s->v_edge_pos>>1); - s->vdsp.emulated_edge_mc(uvbuf+16, + s->vdsp.emulated_edge_mc(uvbuf+16, s->uvlinesize, ptr_cr, s->uvlinesize, 9, 9+field_based, uvsrc_x, uvsrc_y< FFMAX(s->h_edge_pos - (motion_x&3) - 16, 0) || (unsigned)src_y > FFMAX( v_edge_pos - (motion_y&3) - h , 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr_y, s->linesize, - 17, 17+field_based, src_x, src_y<h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + ptr_y, s->linesize, + 17, 17+field_based, src_x, src_y<h_edge_pos, s->v_edge_pos); ptr_y= s->edge_emu_buffer; if(!CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){ uint8_t *uvbuf= s->edge_emu_buffer + 18*s->linesize; - s->vdsp.emulated_edge_mc(uvbuf, ptr_cb, s->uvlinesize, - 9, 9 + field_based, - uvsrc_x, uvsrc_y<h_edge_pos>>1, s->v_edge_pos>>1); - s->vdsp.emulated_edge_mc(uvbuf + 16, ptr_cr, s->uvlinesize, - 9, 9 + field_based, - uvsrc_x, uvsrc_y<h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(uvbuf, s->uvlinesize, + ptr_cb, s->uvlinesize, + 9, 9 + field_based, + uvsrc_x, uvsrc_y<h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(uvbuf + 16, s->uvlinesize, + ptr_cr, s->uvlinesize, + 9, 9 + field_based, + uvsrc_x, uvsrc_y<h_edge_pos>>1, s->v_edge_pos>>1); ptr_cb= uvbuf; ptr_cr= uvbuf + 16; } @@ -581,9 +591,9 @@ static void chroma_4mv_motion(MpegEncContext *s, if(s->flags&CODEC_FLAG_EMU_EDGE){ if( (unsigned)src_x > FFMAX((s->h_edge_pos>>1) - (dxy &1) - 8, 0) || (unsigned)src_y > FFMAX((s->v_edge_pos>>1) - (dxy>>1) - 8, 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, - 9, 9, src_x, src_y, - s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, + ptr, s->uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; emu=1; } @@ -592,9 +602,9 @@ static void chroma_4mv_motion(MpegEncContext *s, ptr = ref_picture[2] + offset; if(emu){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, - 9, 9, src_x, src_y, - s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, + ptr, s->uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; } pix_op[dxy](dest_cr, ptr, s->uvlinesize, 8); @@ -761,10 +771,10 @@ static av_always_inline void MPV_motion_internal(MpegEncContext *s, if(s->flags&CODEC_FLAG_EMU_EDGE){ if( (unsigned)src_x > FFMAX(s->h_edge_pos - (motion_x&3) - 8, 0) || (unsigned)src_y > FFMAX(s->v_edge_pos - (motion_y&3) - 8, 0)){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, - s->linesize, 9, 9, - src_x, src_y, - s->h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + ptr, s->linesize, 9, 9, + src_x, src_y, + s->h_edge_pos, s->v_edge_pos); ptr= s->edge_emu_buffer; } } diff --git a/libavcodec/rv34.c b/libavcodec/rv34.c index 8834032974..f84fb005df 100644 --- a/libavcodec/rv34.c +++ b/libavcodec/rv34.c @@ -724,13 +724,16 @@ static inline void rv34_mc(RV34DecContext *r, const int block_type, uint8_t *uvbuf = s->edge_emu_buffer + 22 * s->linesize; srcY -= 2 + 2*s->linesize; - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, (width<<3)+6, (height<<3)+6, - src_x - 2, src_y - 2, s->h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, srcY, s->linesize, + (width<<3)+6, (height<<3)+6, src_x - 2, src_y - 2, + s->h_edge_pos, s->v_edge_pos); srcY = s->edge_emu_buffer + 2 + 2*s->linesize; - s->vdsp.emulated_edge_mc(uvbuf , srcU, s->uvlinesize, (width<<2)+1, (height<<2)+1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, s->v_edge_pos >> 1); - s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, (width<<2)+1, (height<<2)+1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, s->v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf, s->uvlinesize, srcU, s->uvlinesize, + (width<<2)+1, (height<<2)+1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, s->v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf + 16, s->uvlinesize, srcV, s->uvlinesize, + (width<<2)+1, (height<<2)+1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, s->v_edge_pos >> 1); srcU = uvbuf; srcV = uvbuf + 16; } diff --git a/libavcodec/snow.c b/libavcodec/snow.c index dc5c78e84a..b54c491f42 100644 --- a/libavcodec/snow.c +++ b/libavcodec/snow.c @@ -349,7 +349,9 @@ void ff_snow_pred_block(SnowContext *s, uint8_t *dst, uint8_t *tmp, ptrdiff_t st src += sx + sy*stride; if( (unsigned)sx >= FFMAX(w - b_w - (HTAPS_MAX-2), 0) || (unsigned)sy >= FFMAX(h - b_h - (HTAPS_MAX-2), 0)){ - s->vdsp.emulated_edge_mc(tmp + MB_SIZE, src, stride, b_w+HTAPS_MAX-1, b_h+HTAPS_MAX-1, sx, sy, w, h); + s->vdsp.emulated_edge_mc(tmp + MB_SIZE, stride, src, stride, + b_w+HTAPS_MAX-1, b_h+HTAPS_MAX-1, + sx, sy, w, h); src= tmp + MB_SIZE; } diff --git a/libavcodec/svq3.c b/libavcodec/svq3.c index 9326165c36..6988a1dd74 100644 --- a/libavcodec/svq3.c +++ b/libavcodec/svq3.c @@ -317,7 +317,8 @@ static inline void svq3_mc_dir_part(SVQ3Context *s, src = pic->f.data[0] + mx + my * h->linesize; if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src, h->linesize, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->linesize, + src, h->linesize, width + 1, height + 1, mx, my, s->h_edge_pos, s->v_edge_pos); src = h->edge_emu_buffer; @@ -343,7 +344,8 @@ static inline void svq3_mc_dir_part(SVQ3Context *s, src = pic->f.data[i] + mx + my * h->uvlinesize; if (emu) { - h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src, h->uvlinesize, + h->vdsp.emulated_edge_mc(h->edge_emu_buffer, h->uvlinesize, + src, h->uvlinesize, width + 1, height + 1, mx, my, (s->h_edge_pos >> 1), s->v_edge_pos >> 1); diff --git a/libavcodec/vc1dec.c b/libavcodec/vc1dec.c index 439d68eee9..399e0d2264 100644 --- a/libavcodec/vc1dec.c +++ b/libavcodec/vc1dec.c @@ -454,15 +454,18 @@ static void vc1_mc_1mv(VC1Context *v, int dir) uint8_t *uvbuf = s->edge_emu_buffer + 19 * s->linesize; srcY -= s->mspel * (1 + s->linesize); - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + srcY, s->linesize, 17 + s->mspel * 2, 17 + s->mspel * 2, src_x - s->mspel, src_y - s->mspel, s->h_edge_pos, v_edge_pos); srcY = s->edge_emu_buffer; - s->vdsp.emulated_edge_mc(uvbuf , srcU, s->uvlinesize, 8 + 1, 8 + 1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); - s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf, s->uvlinesize, srcU, s->uvlinesize, + 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf + 16, s->uvlinesize, srcV, s->uvlinesize, + 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, v_edge_pos >> 1); srcU = uvbuf; srcV = uvbuf + 16; /* if we deal with range reduction we need to scale source blocks */ @@ -697,7 +700,7 @@ static void vc1_mc_4mv_luma(VC1Context *v, int n, int dir, int avg) || (unsigned)(src_y - (s->mspel << fieldmv)) > v_edge_pos - (my & 3) - ((8 + s->mspel * 2) << fieldmv)) { srcY -= s->mspel * (1 + (s->linesize << fieldmv)); /* check emulate edge stride and offset */ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, srcY, s->linesize, 9 + s->mspel * 2, (9 + s->mspel * 2) << fieldmv, src_x - s->mspel, src_y - (s->mspel << fieldmv), s->h_edge_pos, v_edge_pos); @@ -912,11 +915,11 @@ static void vc1_mc_4mv_chroma(VC1Context *v, int dir) || s->h_edge_pos < 18 || v_edge_pos < 18 || (unsigned)uvsrc_x > (s->h_edge_pos >> 1) - 9 || (unsigned)uvsrc_y > (v_edge_pos >> 1) - 9) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer , srcU, s->uvlinesize, - 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, srcU, + s->uvlinesize, 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); - s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize, - 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, s->uvlinesize, srcV, + s->uvlinesize, 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); srcU = s->edge_emu_buffer; srcV = s->edge_emu_buffer + 16; @@ -1033,12 +1036,12 @@ static void vc1_mc_4mv_chroma4(VC1Context *v, int dir, int dir2, int avg) || s->h_edge_pos < 10 || v_edge_pos < (5 << fieldmv) || (unsigned)uvsrc_x > (s->h_edge_pos >> 1) - 5 || (unsigned)uvsrc_y > v_edge_pos - (5 << fieldmv)) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcU, s->uvlinesize, - 5, (5 << fieldmv), uvsrc_x, uvsrc_y, - s->h_edge_pos >> 1, v_edge_pos); - s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, srcV, s->uvlinesize, - 5, (5 << fieldmv), uvsrc_x, uvsrc_y, - s->h_edge_pos >> 1, v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, srcU, + s->uvlinesize, 5, (5 << fieldmv), uvsrc_x, + uvsrc_y, s->h_edge_pos >> 1, v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer + 16, s->uvlinesize, srcV, + s->uvlinesize, 5, (5 << fieldmv), uvsrc_x, + uvsrc_y, s->h_edge_pos >> 1, v_edge_pos); srcU = s->edge_emu_buffer; srcV = s->edge_emu_buffer + 16; @@ -1966,15 +1969,17 @@ static void vc1_interp_mc(VC1Context *v) uint8_t *uvbuf = s->edge_emu_buffer + 19 * s->linesize; srcY -= s->mspel * (1 + s->linesize); - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, srcY, s->linesize, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, srcY, s->linesize, 17 + s->mspel * 2, 17 + s->mspel * 2, src_x - s->mspel, src_y - s->mspel, s->h_edge_pos, v_edge_pos); srcY = s->edge_emu_buffer; - s->vdsp.emulated_edge_mc(uvbuf , srcU, s->uvlinesize, 8 + 1, 8 + 1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); - s->vdsp.emulated_edge_mc(uvbuf + 16, srcV, s->uvlinesize, 8 + 1, 8 + 1, - uvsrc_x, uvsrc_y, s->h_edge_pos >> 1, v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf, s->uvlinesize, srcU, s->uvlinesize, + 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, v_edge_pos >> 1); + s->vdsp.emulated_edge_mc(uvbuf + 16, s->uvlinesize, srcV, s->uvlinesize, + 8 + 1, 8 + 1, uvsrc_x, uvsrc_y, + s->h_edge_pos >> 1, v_edge_pos >> 1); srcU = uvbuf; srcV = uvbuf + 16; /* if we deal with range reduction we need to scale source blocks */ diff --git a/libavcodec/videodsp.h b/libavcodec/videodsp.h index e397720773..e85dff355c 100644 --- a/libavcodec/videodsp.h +++ b/libavcodec/videodsp.h @@ -30,9 +30,10 @@ #include #define EMULATED_EDGE(depth) \ -void ff_emulated_edge_mc_ ## depth (uint8_t *buf, const uint8_t *src, ptrdiff_t linesize,\ - int block_w, int block_h,\ - int src_x, int src_y, int w, int h); +void ff_emulated_edge_mc_ ## depth(uint8_t *dst, ptrdiff_t dst_stride, \ + const uint8_t *src, ptrdiff_t src_stride, \ + int block_w, int block_h,\ + int src_x, int src_y, int w, int h); EMULATED_EDGE(8) EMULATED_EDGE(16) @@ -42,10 +43,12 @@ typedef struct VideoDSPContext { * Copy a rectangular area of samples to a temporary buffer and replicate * the border samples. * - * @param buf destination buffer + * @param dst destination buffer + * @param dst_stride number of bytes between 2 vertically adjacent samples + * in destination buffer * @param src source buffer - * @param linesize number of bytes between 2 vertically adjacent samples - * in both the source and destination buffers + * @param src_stride number of bytes between 2 vertically adjacent samples + * in source buffer * @param block_w width of block * @param block_h height of block * @param src_x x coordinate of the top left sample of the block in the @@ -55,8 +58,9 @@ typedef struct VideoDSPContext { * @param w width of the source buffer * @param h height of the source buffer */ - void (*emulated_edge_mc)(uint8_t *buf, const uint8_t *src, - ptrdiff_t linesize, int block_w, int block_h, + void (*emulated_edge_mc)(uint8_t *dst, ptrdiff_t dst_stride, + const uint8_t *src, ptrdiff_t src_stride, + int block_w, int block_h, int src_x, int src_y, int w, int h); /** diff --git a/libavcodec/videodsp_template.c b/libavcodec/videodsp_template.c index 39aad4deec..f7e7bfdb13 100644 --- a/libavcodec/videodsp_template.c +++ b/libavcodec/videodsp_template.c @@ -20,10 +20,10 @@ */ #include "bit_depth_template.c" -void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src, - ptrdiff_t linesize, - int block_w, int block_h, - int src_x, int src_y, int w, int h) +void FUNC(ff_emulated_edge_mc)(uint8_t *buf, ptrdiff_t buf_stride, + const uint8_t *src, ptrdiff_t src_stride, + int block_w, int block_h, + int src_x, int src_y, int w, int h) { int x, y; int start_y, start_x, end_y, end_x; @@ -32,12 +32,12 @@ void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src, return; if (src_y >= h) { - src -= src_y * linesize; - src += (h - 1) * linesize; + src -= src_y * src_stride; + src += (h - 1) * src_stride; src_y = h - 1; } else if (src_y <= -block_h) { - src -= src_y * linesize; - src += (1 - block_h) * linesize; + src -= src_y * src_stride; + src += (1 - block_h) * src_stride; src_y = 1 - block_h; } if (src_x >= w) { @@ -56,30 +56,30 @@ void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src, av_assert2(start_x < end_x && block_w); w = end_x - start_x; - src += start_y * linesize + start_x * sizeof(pixel); + src += start_y * src_stride + start_x * sizeof(pixel); buf += start_x * sizeof(pixel); // top for (y = 0; y < start_y; y++) { memcpy(buf, src, w * sizeof(pixel)); - buf += linesize; + buf += buf_stride; } // copy existing part for (; y < end_y; y++) { memcpy(buf, src, w * sizeof(pixel)); - src += linesize; - buf += linesize; + src += src_stride; + buf += buf_stride; } // bottom - src -= linesize; + src -= src_stride; for (; y < block_h; y++) { memcpy(buf, src, w * sizeof(pixel)); - buf += linesize; + buf += buf_stride; } - buf -= block_h * linesize + start_x * sizeof(pixel); + buf -= block_h * buf_stride + start_x * sizeof(pixel); while (block_h--) { pixel *bufp = (pixel *) buf; @@ -92,6 +92,6 @@ void FUNC(ff_emulated_edge_mc)(uint8_t *buf, const uint8_t *src, for (x = end_x; x < block_w; x++) { bufp[x] = bufp[end_x - 1]; } - buf += linesize; + buf += buf_stride; } } diff --git a/libavcodec/vp3.c b/libavcodec/vp3.c index 78c094c038..e2e5ab9697 100644 --- a/libavcodec/vp3.c +++ b/libavcodec/vp3.c @@ -1549,7 +1549,10 @@ static void render_slice(Vp3DecodeContext *s, int slice) uint8_t *temp= s->edge_emu_buffer; if(stride<0) temp -= 8*stride; - s->vdsp.emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height); + s->vdsp.emulated_edge_mc(temp, stride, + motion_source, stride, + 9, 9, src_x, src_y, + plane_width, plane_height); motion_source= temp; } } diff --git a/libavcodec/vp56.c b/libavcodec/vp56.c index b5d1884aa5..d7d08f2fd9 100644 --- a/libavcodec/vp56.c +++ b/libavcodec/vp56.c @@ -339,7 +339,7 @@ static void vp56_mc(VP56Context *s, int b, int plane, uint8_t *src, if (x<0 || x+12>=s->plane_width[plane] || y<0 || y+12>=s->plane_height[plane]) { - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, stride, src + s->block_offset[b] + (dy-2)*stride + (dx-2), stride, 12, 12, x, y, s->plane_width[plane], diff --git a/libavcodec/vp8.c b/libavcodec/vp8.c index 48432c17c2..157cebcb1a 100644 --- a/libavcodec/vp8.c +++ b/libavcodec/vp8.c @@ -45,7 +45,6 @@ static void free_buffers(VP8Context *s) pthread_mutex_destroy(&s->thread_data[i].lock); #endif av_freep(&s->thread_data[i].filter_strength); - av_freep(&s->thread_data[i].edge_emu_buffer); } av_freep(&s->thread_data); av_freep(&s->macroblocks_base); @@ -1186,7 +1185,7 @@ void vp8_mc_luma(VP8Context *s, VP8ThreadData *td, uint8_t *dst, uint8_t *src = ref->f->data[0]; if (AV_RN32A(mv)) { - + int src_linesize = linesize; int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx]; int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my]; @@ -1198,12 +1197,15 @@ void vp8_mc_luma(VP8Context *s, VP8ThreadData *td, uint8_t *dst, src += y_off * linesize + x_off; if (x_off < mx_idx || x_off >= width - block_w - subpel_idx[2][mx] || y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) { - s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize, - block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + s->vdsp.emulated_edge_mc(td->edge_emu_buffer, 32, + src - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], + block_h + subpel_idx[1][my], x_off - mx_idx, y_off - my_idx, width, height); - src = td->edge_emu_buffer + mx_idx + linesize * my_idx; + src = td->edge_emu_buffer + mx_idx + 32 * my_idx; + src_linesize = 32; } - mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); + mc_func[my_idx][mx_idx](dst, linesize, src, src_linesize, block_h, mx, my); } else { ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0); mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); @@ -1248,17 +1250,21 @@ void vp8_mc_chroma(VP8Context *s, VP8ThreadData *td, uint8_t *dst1, uint8_t *dst ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0); if (x_off < mx_idx || x_off >= width - block_w - subpel_idx[2][mx] || y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) { - s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize, - block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + s->vdsp.emulated_edge_mc(td->edge_emu_buffer, 32, + src1 - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], + block_h + subpel_idx[1][my], x_off - mx_idx, y_off - my_idx, width, height); - src1 = td->edge_emu_buffer + mx_idx + linesize * my_idx; - mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my); + src1 = td->edge_emu_buffer + mx_idx + 32 * my_idx; + mc_func[my_idx][mx_idx](dst1, linesize, src1, 32, block_h, mx, my); - s->vdsp.emulated_edge_mc(td->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize, - block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + s->vdsp.emulated_edge_mc(td->edge_emu_buffer, 32, + src2 - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], + block_h + subpel_idx[1][my], x_off - mx_idx, y_off - my_idx, width, height); - src2 = td->edge_emu_buffer + mx_idx + linesize * my_idx; - mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my); + src2 = td->edge_emu_buffer + mx_idx + 32 * my_idx; + mc_func[my_idx][mx_idx](dst2, linesize, src2, 32, block_h, mx, my); } else { mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my); mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my); @@ -1944,10 +1950,6 @@ int ff_vp8_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, s->linesize = curframe->tf.f->linesize[0]; s->uvlinesize = curframe->tf.f->linesize[1]; - if (!s->thread_data[0].edge_emu_buffer) - for (i = 0; i < MAX_THREADS; i++) - s->thread_data[i].edge_emu_buffer = av_malloc(21*s->linesize); - memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); /* Zero macroblock structures for top/top-left prediction from outside the frame. */ if (!s->mb_layout) diff --git a/libavcodec/vp8.h b/libavcodec/vp8.h index 6b56e65168..c9a7906df0 100644 --- a/libavcodec/vp8.h +++ b/libavcodec/vp8.h @@ -122,7 +122,7 @@ typedef struct VP8ThreadData { #endif int thread_mb_pos; // (mb_y << 16) | (mb_x & 0xFFFF) int wait_mb_pos; // What the current thread is waiting on. - uint8_t *edge_emu_buffer; + DECLARE_ALIGNED(16, uint8_t, edge_emu_buffer)[21*32]; VP8FilterStrength *filter_strength; } VP8ThreadData; diff --git a/libavcodec/wmv2.c b/libavcodec/wmv2.c index eea819bfdb..8d30259b15 100644 --- a/libavcodec/wmv2.c +++ b/libavcodec/wmv2.c @@ -119,8 +119,10 @@ void ff_mspel_motion(MpegEncContext *s, if(src_x<1 || src_y<1 || src_x + 17 >= s->h_edge_pos || src_y + h+1 >= v_edge_pos){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr - 1 - s->linesize, s->linesize, 19, 19, - src_x-1, src_y-1, s->h_edge_pos, s->v_edge_pos); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->linesize, + ptr - 1 - s->linesize, s->linesize, 19, 19, + src_x-1, src_y-1, + s->h_edge_pos, s->v_edge_pos); ptr= s->edge_emu_buffer + 1 + s->linesize; emu=1; } @@ -159,16 +161,18 @@ void ff_mspel_motion(MpegEncContext *s, offset = (src_y * uvlinesize) + src_x; ptr = ref_picture[1] + offset; if(emu){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9, - src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, + ptr, s->uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; } pix_op[1][dxy](dest_cb, ptr, uvlinesize, h >> 1); ptr = ref_picture[2] + offset; if(emu){ - s->vdsp.emulated_edge_mc(s->edge_emu_buffer, ptr, s->uvlinesize, 9, 9, - src_x, src_y, s->h_edge_pos>>1, s->v_edge_pos>>1); + s->vdsp.emulated_edge_mc(s->edge_emu_buffer, s->uvlinesize, + ptr, s->uvlinesize, 9, 9, src_x, src_y, + s->h_edge_pos>>1, s->v_edge_pos>>1); ptr= s->edge_emu_buffer; } pix_op[1][dxy](dest_cr, ptr, uvlinesize, h >> 1); diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c index 421a6c2090..9cbf58a09c 100644 --- a/libavcodec/x86/dsputil_mmx.c +++ b/libavcodec/x86/dsputil_mmx.c @@ -405,8 +405,9 @@ void ff_draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, } } -typedef void emulated_edge_mc_func(uint8_t *dst, const uint8_t *src, - ptrdiff_t linesize, int block_w, int block_h, +typedef void emulated_edge_mc_func(uint8_t *dst, ptrdiff_t dst_stride, + const uint8_t *src, ptrdiff_t src_linesize, + int block_w, int block_h, int src_x, int src_y, int w, int h); static av_always_inline void gmc(uint8_t *dst, uint8_t *src, @@ -454,7 +455,7 @@ static av_always_inline void gmc(uint8_t *dst, uint8_t *src, src += ix + iy * stride; if (need_emu) { - emu_edge_fn(edge_buf, src, stride, w + 1, h + 1, ix, iy, width, height); + emu_edge_fn(edge_buf, stride, src, stride, w + 1, h + 1, ix, iy, width, height); src = edge_buf; } diff --git a/libavcodec/x86/videodsp.asm b/libavcodec/x86/videodsp.asm index 3f2f01a3c4..cc2105d381 100644 --- a/libavcodec/x86/videodsp.asm +++ b/libavcodec/x86/videodsp.asm @@ -23,576 +23,394 @@ SECTION .text -; void ff_emu_edge_core(uint8_t *buf, const uint8_t *src, x86_reg linesize, -; x86_reg start_y, x86_reg end_y, x86_reg block_h, -; x86_reg start_x, x86_reg end_x, x86_reg block_w); -; -; The actual function itself is below. It basically wraps a very simple -; w = end_x - start_x -; if (w) { -; if (w > 22) { -; jump to the slow loop functions -; } else { -; jump to the fast loop functions -; } -; } -; -; ... and then the same for left/right extend also. See below for loop -; function implementations. Fast are fixed-width, slow is variable-width - -%macro EMU_EDGE_FUNC 0 -%if ARCH_X86_64 -%define w_reg r7 -cglobal emu_edge_core, 6, 9, 1 - mov r8, r5 ; save block_h -%else -%define w_reg r6 -cglobal emu_edge_core, 2, 7, 0 - mov r4, r4m ; end_y - mov r5, r5m ; block_h -%endif - - ; start with vertical extend (top/bottom) and body pixel copy - mov w_reg, r7m - sub w_reg, r6m ; w = start_x - end_x - sub r5, r4 -%if ARCH_X86_64 - sub r4, r3 -%else - sub r4, dword r3m -%endif - cmp w_reg, 22 - jg .slow_v_extend_loop -%if ARCH_X86_32 - mov r2, r2m ; linesize -%endif - sal w_reg, 7 ; w * 128 -%ifdef PIC - lea rax, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)] - add w_reg, rax -%else - lea w_reg, [.emuedge_v_extend_1 - (.emuedge_v_extend_2 - .emuedge_v_extend_1)+w_reg] -%endif - call w_reg ; fast top extend, body copy and bottom extend -.v_extend_end: - - ; horizontal extend (left/right) - mov w_reg, r6m ; start_x - sub r0, w_reg -%if ARCH_X86_64 - mov r3, r0 ; backup of buf+block_h*linesize - mov r5, r8 -%else - mov r0m, r0 ; backup of buf+block_h*linesize - mov r5, r5m -%endif - test w_reg, w_reg - jz .right_extend - cmp w_reg, 22 - jg .slow_left_extend_loop - mov r1, w_reg - dec w_reg - ; FIXME we can do a if size == 1 here if that makes any speed difference, test me - sar w_reg, 1 - sal w_reg, 6 - ; r0=buf+block_h*linesize,r7(64)/r6(32)=start_x offset for funcs - ; r6(rax)/r3(ebx)=val,r2=linesize,r1=start_x,r5=block_h -%ifdef PIC - lea rax, [.emuedge_extend_left_2] - add w_reg, rax -%else - lea w_reg, [.emuedge_extend_left_2+w_reg] -%endif - call w_reg - - ; now r3(64)/r0(32)=buf,r2=linesize,r8/r5=block_h,r6/r3=val, r7/r6=end_x, r1=block_w -.right_extend: -%if ARCH_X86_32 - mov r0, r0m - mov r5, r5m -%endif - mov w_reg, r7m ; end_x - mov r1, r8m ; block_w - mov r4, r1 - sub r1, w_reg - jz .h_extend_end ; if (end_x == block_w) goto h_extend_end - cmp r1, 22 - jg .slow_right_extend_loop - dec r1 - ; FIXME we can do a if size == 1 here if that makes any speed difference, test me - sar r1, 1 - sal r1, 6 -%ifdef PIC - lea rax, [.emuedge_extend_right_2] - add r1, rax -%else - lea r1, [.emuedge_extend_right_2+r1] -%endif - call r1 -.h_extend_end: - RET - -%if ARCH_X86_64 -%define vall al -%define valh ah -%define valw ax -%define valw2 r7w -%define valw3 r3w -%if WIN64 -%define valw4 r7w -%else ; unix64 -%define valw4 r3w -%endif -%define vald eax -%else -%define vall bl -%define valh bh -%define valw bx -%define valw2 r6w -%define valw3 valw2 -%define valw4 valw3 -%define vald ebx -%define stack_offset 0x14 -%endif +; slow vertical extension loop function. Works with variable-width, and +; does per-line reading/writing of source data +%macro V_COPY_ROW 2 ; type (top/body/bottom), h +.%1_y_loop: ; do { + mov wq, r7mp ; initialize w (r7mp = wmp) +.%1_x_loop: ; do { + movu m0, [srcq+wq] ; m0 = read($mmsize) + movu [dstq+wq], m0 ; write(m0, $mmsize) + add wq, mmsize ; w -= $mmsize + cmp wq, -mmsize ; } while (w > $mmsize); + jl .%1_x_loop + movu m0, [srcq-mmsize] ; m0 = read($mmsize) + movu [dstq-mmsize], m0 ; write(m0, $mmsize) +%ifidn %1, body ; if ($type == body) { + add srcq, src_strideq ; src += src_stride +%endif ; } + add dstq, dst_strideq ; dst += dst_stride + dec %2 ; } while (--$h); + jnz .%1_y_loop %endmacro -; macro to read/write a horizontal number of pixels (%2) to/from registers -; on x86-64, - fills xmm0-15 for consecutive sets of 16 pixels -; - if (%2 & 15 == 8) fills the last 8 bytes into rax -; - else if (%2 & 8) fills 8 bytes into mm0 -; - if (%2 & 7 == 4) fills the last 4 bytes into rax -; - else if (%2 & 4) fills 4 bytes into mm0-1 -; - if (%2 & 3 == 3) fills 2 bytes into r7/r3, and 1 into eax -; (note that we're using r3 for body/bottom because it's a shorter -; opcode, and then the loop fits in 128 bytes) -; - else fills remaining bytes into rax -; on x86-32, - fills mm0-7 for consecutive sets of 8 pixels -; - if (%2 & 7 == 4) fills 4 bytes into ebx -; - else if (%2 & 4) fills 4 bytes into mm0-7 -; - if (%2 & 3 == 3) fills 2 bytes into r6, and 1 into ebx -; - else fills remaining bytes into ebx -; writing data out is in the same way -%macro READ_NUM_BYTES 2 -%assign %%src_off 0 ; offset in source buffer -%assign %%smidx 0 ; mmx register idx -%assign %%sxidx 0 ; xmm register idx +%macro vvar_fn 0 +; .----. <- zero +; | | <- top is copied from first line in body of source +; |----| <- start_y +; | | <- body is copied verbatim (line-by-line) from source +; |----| <- end_y +; | | <- bottom is copied from last line in body of source +; '----' <- bh +%if ARCH_X86_64 +cglobal emu_edge_vvar, 7, 8, 1, dst, dst_stride, src, src_stride, \ + start_y, end_y, bh, w +%else ; x86-32 +cglobal emu_edge_vvar, 1, 6, 1, dst, src, start_y, end_y, bh, w +%define src_strideq r3mp +%define dst_strideq r1mp + mov srcq, r2mp + mov start_yq, r4mp + mov end_yq, r5mp + mov bhq, r6mp +%endif + sub bhq, end_yq ; bh -= end_q + sub end_yq, start_yq ; end_q -= start_q + add srcq, r7mp ; (r7mp = wmp) + add dstq, r7mp ; (r7mp = wmp) + neg r7mp ; (r7mp = wmp) + test start_yq, start_yq ; if (start_q) { + jz .body + V_COPY_ROW top, start_yq ; v_copy_row(top, start_yq) +.body: ; } + V_COPY_ROW body, end_yq ; v_copy_row(body, end_yq) + test bhq, bhq ; if (bh) { + jz .end + sub srcq, src_strideq ; src -= src_stride + V_COPY_ROW bottom, bhq ; v_copy_row(bottom, bh) +.end: ; } + RET +%endmacro -%if cpuflag(sse) -%rep %2/16 - movups xmm %+ %%sxidx, [r1+%%src_off] -%assign %%src_off %%src_off+16 -%assign %%sxidx %%sxidx+1 -%endrep ; %2/16 +%if ARCH_X86_32 +INIT_MMX mmx +vvar_fn %endif -%if ARCH_X86_64 -%if (%2-%%src_off) == 8 - mov rax, [r1+%%src_off] -%assign %%src_off %%src_off+8 -%endif ; (%2-%%src_off) == 8 -%endif ; x86-64 +INIT_XMM sse +vvar_fn -%rep (%2-%%src_off)/8 - movq mm %+ %%smidx, [r1+%%src_off] -%assign %%src_off %%src_off+8 -%assign %%smidx %%smidx+1 -%endrep ; (%2-%%dst_off)/8 +%macro hvar_fn 0 +cglobal emu_edge_hvar, 5, 6, 1, dst, dst_stride, start_x, n_words, h, w + lea dstq, [dstq+n_wordsq*2] + neg n_wordsq + lea start_xq, [start_xq+n_wordsq*2] +.y_loop: ; do { + ; FIXME also write a ssse3 version using pshufb + movzx wd, byte [dstq+start_xq] ; w = read(1) + imul wd, 0x01010101 ; w *= 0x01010101 + movd m0, wd + mov wq, n_wordsq ; initialize w +%if cpuflag(sse) + shufps m0, m0, q0000 ; splat +%else ; mmx + punpckldq m0, m0 ; splat +%endif ; mmx/sse +.x_loop: ; do { + movu [dstq+wq*2], m0 ; write($reg, $mmsize) + add wq, mmsize/2 ; w -= $mmsize/2 + cmp wq, -mmsize/2 ; } while (w > $mmsize/2) + jl .x_loop + movu [dstq-mmsize], m0 ; write($reg, $mmsize) + add dstq, dst_strideq ; dst += dst_stride + dec hq ; } while (h--) + jnz .y_loop + RET +%endmacro -%if (%2-%%src_off) == 4 - mov vald, [r1+%%src_off] -%elif (%2-%%src_off) & 4 - movd mm %+ %%smidx, [r1+%%src_off] -%assign %%src_off %%src_off+4 -%endif ; (%2-%%src_off) ==/& 4 +%if ARCH_X86_32 +INIT_MMX mmx +hvar_fn +%endif -%if (%2-%%src_off) == 1 - mov vall, [r1+%%src_off] -%elif (%2-%%src_off) == 2 - mov valw, [r1+%%src_off] -%elif (%2-%%src_off) == 3 -%ifidn %1, top - mov valw2, [r1+%%src_off] +INIT_XMM sse +hvar_fn + +; macro to read/write a horizontal number of pixels (%2) to/from registers +; on sse, - fills xmm0-15 for consecutive sets of 16 pixels +; - if (%2 & 8) fills 8 bytes into xmm$next +; - if (%2 & 4) fills 4 bytes into xmm$next +; - if (%2 & 3) fills 1, 2 or 4 bytes in eax +; on mmx, - fills mm0-7 for consecutive sets of 8 pixels +; - if (%2 & 4) fills 4 bytes into mm$next +; - if (%2 & 3) fills 1, 2 or 4 bytes in eax +; writing data out is in the same way +%macro READ_NUM_BYTES 2 +%assign %%off 0 ; offset in source buffer +%assign %%idx 0 ; mmx/xmm register index + +%rep %2/mmsize + movu m %+ %%idx, [srcq+%%off] +%assign %%off %%off+mmsize +%assign %%idx %%idx+1 +%endrep ; %2/mmsize + +%if mmsize == 16 +%if (%2-%%off) >= 8 +%if %2 > 16 && (%2-%%off) > 8 + movu m %+ %%idx, [srcq+%2-16] +%assign %%off %2 +%else + movq m %+ %%idx, [srcq+%%off] +%assign %%off %%off+8 +%endif +%assign %%idx %%idx+1 +%endif ; (%2-%%off) >= 8 +%endif + +%if (%2-%%off) >= 4 +%if %2 > 8 && (%2-%%off) > 4 + movq m %+ %%idx, [srcq+%2-8] +%assign %%off %2 +%else + movd m %+ %%idx, [srcq+%%off] +%assign %%off %%off+4 +%endif +%assign %%idx %%idx+1 +%endif ; (%2-%%off) >= 4 + +%if (%2-%%off) >= 1 +%if %2 >= 4 + movd m %+ %%idx, [srcq+%2-4] +%elif (%2-%%off) == 1 + mov valb, [srcq+%2-1] +%elif (%2-%%off) == 2 + mov valw, [srcq+%2-2] %elifidn %1, body - mov valw3, [r1+%%src_off] -%elifidn %1, bottom - mov valw4, [r1+%%src_off] -%endif ; %1 ==/!= top - mov vall, [r1+%%src_off+2] -%endif ; (%2-%%src_off) == 1/2/3 + mov vald, [srcq+%2-3] +%else + movd m %+ %%idx, [srcq+%2-3] +%endif +%endif ; (%2-%%off) >= 1 %endmacro ; READ_NUM_BYTES %macro WRITE_NUM_BYTES 2 -%assign %%dst_off 0 ; offset in destination buffer -%assign %%dmidx 0 ; mmx register idx -%assign %%dxidx 0 ; xmm register idx +%assign %%off 0 ; offset in destination buffer +%assign %%idx 0 ; mmx/xmm register index -%if cpuflag(sse) -%rep %2/16 - movups [r0+%%dst_off], xmm %+ %%dxidx -%assign %%dst_off %%dst_off+16 -%assign %%dxidx %%dxidx+1 -%endrep ; %2/16 +%rep %2/mmsize + movu [dstq+%%off], m %+ %%idx +%assign %%off %%off+mmsize +%assign %%idx %%idx+1 +%endrep ; %2/mmsize + +%if mmsize == 16 +%if (%2-%%off) >= 8 +%if %2 > 16 && (%2-%%off) > 8 + movu [dstq+%2-16], m %+ %%idx +%assign %%off %2 +%else + movq [dstq+%%off], m %+ %%idx +%assign %%off %%off+8 +%endif +%assign %%idx %%idx+1 +%endif ; (%2-%%off) >= 8 %endif -%if ARCH_X86_64 -%if (%2-%%dst_off) == 8 - mov [r0+%%dst_off], rax -%assign %%dst_off %%dst_off+8 -%endif ; (%2-%%dst_off) == 8 -%endif ; x86-64 +%if (%2-%%off) >= 4 +%if %2 > 8 && (%2-%%off) > 4 + movq [dstq+%2-8], m %+ %%idx +%assign %%off %2 +%else + movd [dstq+%%off], m %+ %%idx +%assign %%off %%off+4 +%endif +%assign %%idx %%idx+1 +%endif ; (%2-%%off) >= 4 -%rep (%2-%%dst_off)/8 - movq [r0+%%dst_off], mm %+ %%dmidx -%assign %%dst_off %%dst_off+8 -%assign %%dmidx %%dmidx+1 -%endrep ; (%2-%%dst_off)/8 - -%if (%2-%%dst_off) == 4 - mov [r0+%%dst_off], vald -%elif (%2-%%dst_off) & 4 - movd [r0+%%dst_off], mm %+ %%dmidx -%assign %%dst_off %%dst_off+4 -%endif ; (%2-%%dst_off) ==/& 4 - -%if (%2-%%dst_off) == 1 - mov [r0+%%dst_off], vall -%elif (%2-%%dst_off) == 2 - mov [r0+%%dst_off], valw -%elif (%2-%%dst_off) == 3 -%ifidn %1, top - mov [r0+%%dst_off], valw2 +%if (%2-%%off) >= 1 +%if %2 >= 4 + movd [dstq+%2-4], m %+ %%idx +%elif (%2-%%off) == 1 + mov [dstq+%2-1], valb +%elif (%2-%%off) == 2 + mov [dstq+%2-2], valw %elifidn %1, body - mov [r0+%%dst_off], valw3 -%elifidn %1, bottom - mov [r0+%%dst_off], valw4 -%endif ; %1 ==/!= top - mov [r0+%%dst_off+2], vall -%endif ; (%2-%%dst_off) == 1/2/3 + mov [dstq+%2-3], valw + shr vald, 16 + mov [dstq+%2-1], valb +%else + movd vald, m %+ %%idx + mov [dstq+%2-3], valw + shr vald, 16 + mov [dstq+%2-1], valb +%endif +%endif ; (%2-%%off) >= 1 %endmacro ; WRITE_NUM_BYTES ; vertical top/bottom extend and body copy fast loops ; these are function pointers to set-width line copy functions, i.e. ; they read a fixed number of pixels into set registers, and write ; those out into the destination buffer -; r0=buf,r1=src,r2=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h -; r6(eax/64)/r3(ebx/32)=val_reg -%macro VERTICAL_EXTEND 0 -%assign %%n 1 -%rep 22 -ALIGN 128 -.emuedge_v_extend_ %+ %%n: +%macro VERTICAL_EXTEND 2 +%assign %%n %1 +%rep 1+%2-%1 +%if %%n <= 3 +%if ARCH_X86_64 +cglobal emu_edge_vfix %+ %%n, 6, 8, 0, dst, dst_stride, src, src_stride, \ + start_y, end_y, val, bh + mov bhq, r6mp ; r6mp = bhmp +%else ; x86-32 +cglobal emu_edge_vfix %+ %%n, 0, 6, 0, val, dst, src, start_y, end_y, bh + mov dstq, r0mp + mov srcq, r2mp + mov start_yq, r4mp + mov end_yq, r5mp + mov bhq, r6mp +%define dst_strideq r1mp +%define src_strideq r3mp +%endif ; x86-64/32 +%else +%if ARCH_X86_64 +cglobal emu_edge_vfix %+ %%n, 7, 7, 1, dst, dst_stride, src, src_stride, \ + start_y, end_y, bh +%else ; x86-32 +cglobal emu_edge_vfix %+ %%n, 1, 5, 1, dst, src, start_y, end_y, bh + mov srcq, r2mp + mov start_yq, r4mp + mov end_yq, r5mp + mov bhq, r6mp +%define dst_strideq r1mp +%define src_strideq r3mp +%endif ; x86-64/32 +%endif + ; FIXME move this to c wrapper? + sub bhq, end_yq ; bh -= end_y + sub end_yq, start_yq ; end_y -= start_y + ; extend pixels above body -%if ARCH_X86_64 - test r3 , r3 ; if (!start_y) - jz .emuedge_copy_body_ %+ %%n %+ _loop ; goto body -%else ; ARCH_X86_32 - cmp dword r3m, 0 - je .emuedge_copy_body_ %+ %%n %+ _loop -%endif ; ARCH_X86_64/32 - READ_NUM_BYTES top, %%n ; read bytes -.emuedge_extend_top_ %+ %%n %+ _loop: ; do { - WRITE_NUM_BYTES top, %%n ; write bytes - add r0 , r2 ; dst += linesize -%if ARCH_X86_64 - dec r3d -%else ; ARCH_X86_32 - dec dword r3m -%endif ; ARCH_X86_64/32 - jnz .emuedge_extend_top_ %+ %%n %+ _loop ; } while (--start_y) + test start_yq, start_yq ; if (start_y) { + jz .body_loop + READ_NUM_BYTES top, %%n ; $variable_regs = read($n) +.top_loop: ; do { + WRITE_NUM_BYTES top, %%n ; write($variable_regs, $n) + add dstq, dst_strideq ; dst += linesize + dec start_yq ; } while (--start_y) + jnz .top_loop ; } ; copy body pixels -.emuedge_copy_body_ %+ %%n %+ _loop: ; do { - READ_NUM_BYTES body, %%n ; read bytes - WRITE_NUM_BYTES body, %%n ; write bytes - add r0 , r2 ; dst += linesize - add r1 , r2 ; src += linesize - dec r4d - jnz .emuedge_copy_body_ %+ %%n %+ _loop ; } while (--end_y) +.body_loop: ; do { + READ_NUM_BYTES body, %%n ; $variable_regs = read($n) + WRITE_NUM_BYTES body, %%n ; write($variable_regs, $n) + add dstq, dst_strideq ; dst += dst_stride + add srcq, src_strideq ; src += src_stride + dec end_yq ; } while (--end_y) + jnz .body_loop ; copy bottom pixels - test r5 , r5 ; if (!block_h) - jz .emuedge_v_extend_end_ %+ %%n ; goto end - sub r1 , r2 ; src -= linesize - READ_NUM_BYTES bottom, %%n ; read bytes -.emuedge_extend_bottom_ %+ %%n %+ _loop: ; do { - WRITE_NUM_BYTES bottom, %%n ; write bytes - add r0 , r2 ; dst += linesize - dec r5d - jnz .emuedge_extend_bottom_ %+ %%n %+ _loop ; } while (--block_h) + test bhq, bhq ; if (block_h) { + jz .end + sub srcq, src_strideq ; src -= linesize + READ_NUM_BYTES bottom, %%n ; $variable_regs = read($n) +.bottom_loop: ; do { + WRITE_NUM_BYTES bottom, %%n ; write($variable_regs, $n) + add dstq, dst_strideq ; dst += linesize + dec bhq ; } while (--bh) + jnz .bottom_loop ; } -.emuedge_v_extend_end_ %+ %%n: -%if ARCH_X86_64 - ret -%else ; ARCH_X86_32 - rep ret -%endif ; ARCH_X86_64/32 +.end: + RET %assign %%n %%n+1 -%endrep -%endmacro VERTICAL_EXTEND +%endrep ; 1+%2-%1 +%endmacro ; VERTICAL_EXTEND + +INIT_MMX mmx +VERTICAL_EXTEND 1, 15 +%if ARCH_X86_32 +VERTICAL_EXTEND 16, 22 +%endif + +INIT_XMM sse +VERTICAL_EXTEND 16, 22 ; left/right (horizontal) fast extend functions ; these are essentially identical to the vertical extend ones above, ; just left/right separated because number of pixels to extend is ; obviously not the same on both sides. -; for reading, pixels are placed in eax (x86-64) or ebx (x86-64) in the -; lowest two bytes of the register (so val*0x0101), and are splatted -; into each byte of mm0 as well if n_pixels >= 8 %macro READ_V_PIXEL 2 - mov vall, %2 - mov valh, vall +%if %1 == 2 + movzx valw, byte %2 + imul valw, 0x0101 +%else + movzx vald, byte %2 + imul vald, 0x01010101 %if %1 >= 8 - movd mm0, vald -%if cpuflag(mmxext) - pshufw mm0, mm0, 0 -%else ; mmx - punpcklwd mm0, mm0 - punpckldq mm0, mm0 -%endif ; sse + movd m0, vald +%if mmsize == 16 + shufps m0, m0, q0000 +%else + punpckldq m0, m0 +%endif %endif ; %1 >= 8 -%endmacro +%endif +%endmacro ; READ_V_PIXEL %macro WRITE_V_PIXEL 2 -%assign %%dst_off 0 -%rep %1/8 - movq [%2+%%dst_off], mm0 -%assign %%dst_off %%dst_off+8 -%endrep -%if %1 & 4 +%assign %%off 0 +%rep %1/mmsize + movu [%2+%%off], m0 +%assign %%off %%off+mmsize +%endrep ; %1/mmsize + +%if mmsize == 16 +%if %1-%%off >= 8 +%if %1 > 16 && %1-%%off > 8 + movu [%2+%1-16], m0 +%assign %%off %1 +%else + movq [%2+%%off], m0 +%assign %%off %%off+8 +%endif +%endif ; %1-%%off >= 8 +%endif + +%if %1-%%off >= 4 +%if %1 > 8 %% %1-%%off > 4 + movq [%2+%1-8], m0 +%assign %%off %1 +%elif %1 >= 8 && %1-%%off >= 4 + movd [%2+%%off], m0 +%assign %%off %%off+4 +%else + mov [%2+%%off], vald +%assign %%off %%off+4 +%endif +%endif ; %1-%%off >= 4 + +%if %1-%%off >= 2 %if %1 >= 8 - movd [%2+%%dst_off], mm0 -%else ; %1 < 8 - mov [%2+%%dst_off] , valw - mov [%2+%%dst_off+2], valw -%endif ; %1 >=/< 8 -%assign %%dst_off %%dst_off+4 -%endif ; %1 & 4 -%if %1&2 - mov [%2+%%dst_off], valw -%endif ; %1 & 2 -%endmacro + movd [%2+%1-4], m0 +%else + mov [%2+%%off], valw +%endif +%endif ; (%1-%%off)/2 +%endmacro ; WRITE_V_PIXEL -; r0=buf+block_h*linesize, r1=start_x, r2=linesize, r5=block_h, r6/r3=val -%macro LEFT_EXTEND 0 -%assign %%n 2 -%rep 11 -ALIGN 64 -.emuedge_extend_left_ %+ %%n: ; do { - sub r0, r2 ; dst -= linesize - READ_V_PIXEL %%n, [r0+r1] ; read pixels - WRITE_V_PIXEL %%n, r0 ; write pixels - dec r5 - jnz .emuedge_extend_left_ %+ %%n ; } while (--block_h) -%if ARCH_X86_64 - ret -%else ; ARCH_X86_32 - rep ret -%endif ; ARCH_X86_64/32 +%macro H_EXTEND 2 +%assign %%n %1 +%rep 1+(%2-%1)/2 +cglobal emu_edge_hfix %+ %%n, 4, 5, 1, dst, dst_stride, start_x, bh, val +.loop_y: ; do { + READ_V_PIXEL %%n, [dstq+start_xq] ; $variable_regs = read($n) + WRITE_V_PIXEL %%n, dstq ; write($variable_regs, $n) + add dstq, dst_strideq ; dst += dst_stride + dec bhq ; } while (--bh) + jnz .loop_y + RET %assign %%n %%n+2 -%endrep -%endmacro ; LEFT_EXTEND - -; r3/r0=buf+block_h*linesize, r2=linesize, r8/r5=block_h, r0/r6=end_x, r6/r3=val -%macro RIGHT_EXTEND 0 -%assign %%n 2 -%rep 11 -ALIGN 64 -.emuedge_extend_right_ %+ %%n: ; do { -%if ARCH_X86_64 - sub r3, r2 ; dst -= linesize - READ_V_PIXEL %%n, [r3+w_reg-1] ; read pixels - WRITE_V_PIXEL %%n, r3+r4-%%n ; write pixels - dec r8 -%else ; ARCH_X86_32 - sub r0, r2 ; dst -= linesize - READ_V_PIXEL %%n, [r0+w_reg-1] ; read pixels - WRITE_V_PIXEL %%n, r0+r4-%%n ; write pixels - dec r5 -%endif ; ARCH_X86_64/32 - jnz .emuedge_extend_right_ %+ %%n ; } while (--block_h) -%if ARCH_X86_64 - ret -%else ; ARCH_X86_32 - rep ret -%endif ; ARCH_X86_64/32 -%assign %%n %%n+2 -%endrep +%endrep ; 1+(%2-%1)/2 +%endmacro ; H_EXTEND +INIT_MMX mmx +H_EXTEND 2, 14 %if ARCH_X86_32 -%define stack_offset 0x10 -%endif -%endmacro ; RIGHT_EXTEND - -; below follow the "slow" copy/extend functions, these act on a non-fixed -; width specified in a register, and run a loop to copy the full amount -; of bytes. They are optimized for copying of large amounts of pixels per -; line, so they unconditionally splat data into mm registers to copy 8 -; bytes per loop iteration. It could be considered to use xmm for x86-64 -; also, but I haven't optimized this as much (i.e. FIXME) -%macro V_COPY_NPX 4-5 -%if %0 == 4 - test w_reg, %4 - jz .%1_skip_%4_px -%else ; %0 == 5 -.%1_%4_px_loop: -%endif - %3 %2, [r1+cnt_reg] - %3 [r0+cnt_reg], %2 - add cnt_reg, %4 -%if %0 == 5 - sub w_reg, %4 - test w_reg, %5 - jnz .%1_%4_px_loop -%endif -.%1_skip_%4_px: -%endmacro - -%macro V_COPY_ROW 2 -%ifidn %1, bottom - sub r1, linesize -%endif -.%1_copy_loop: - xor cnt_reg, cnt_reg -%if notcpuflag(sse) -%define linesize r2m - V_COPY_NPX %1, mm0, movq, 8, 0xFFFFFFF8 -%else ; sse - V_COPY_NPX %1, xmm0, movups, 16, 0xFFFFFFF0 -%if ARCH_X86_64 -%define linesize r2 - V_COPY_NPX %1, rax , mov, 8 -%else ; ARCH_X86_32 -%define linesize r2m - V_COPY_NPX %1, mm0, movq, 8 -%endif ; ARCH_X86_64/32 -%endif ; sse - V_COPY_NPX %1, vald, mov, 4 - V_COPY_NPX %1, valw, mov, 2 - V_COPY_NPX %1, vall, mov, 1 - mov w_reg, cnt_reg -%ifidn %1, body - add r1, linesize -%endif - add r0, linesize - dec %2 - jnz .%1_copy_loop -%endmacro - -%macro SLOW_V_EXTEND 0 -.slow_v_extend_loop: -; r0=buf,r1=src,r2(64)/r2m(32)=linesize,r3(64)/r3m(32)=start_x,r4=end_y,r5=block_h -; r8(64)/r3(later-64)/r2(32)=cnt_reg,r6(64)/r3(32)=val_reg,r7(64)/r6(32)=w=end_x-start_x -%if ARCH_X86_64 - push r8 ; save old value of block_h - test r3, r3 -%define cnt_reg r8 - jz .do_body_copy ; if (!start_y) goto do_body_copy - V_COPY_ROW top, r3 -%else - cmp dword r3m, 0 -%define cnt_reg r2 - je .do_body_copy ; if (!start_y) goto do_body_copy - V_COPY_ROW top, dword r3m +H_EXTEND 16, 22 %endif -.do_body_copy: - V_COPY_ROW body, r4 - -%if ARCH_X86_64 - pop r8 ; restore old value of block_h -%define cnt_reg r3 -%endif - test r5, r5 -%if ARCH_X86_64 - jz .v_extend_end -%else - jz .skip_bottom_extend -%endif - V_COPY_ROW bottom, r5 -%if ARCH_X86_32 -.skip_bottom_extend: - mov r2, r2m -%endif - jmp .v_extend_end -%endmacro - -%macro SLOW_LEFT_EXTEND 0 -.slow_left_extend_loop: -; r0=buf+block_h*linesize,r2=linesize,r6(64)/r3(32)=val,r5=block_h,r4=cntr,r7/r6=start_x - mov r4, 8 - sub r0, linesize - READ_V_PIXEL 8, [r0+w_reg] -.left_extend_8px_loop: - movq [r0+r4-8], mm0 - add r4, 8 - cmp r4, w_reg - jle .left_extend_8px_loop - sub r4, 8 - cmp r4, w_reg - jge .left_extend_loop_end -.left_extend_2px_loop: - mov [r0+r4], valw - add r4, 2 - cmp r4, w_reg - jl .left_extend_2px_loop -.left_extend_loop_end: - dec r5 - jnz .slow_left_extend_loop -%if ARCH_X86_32 - mov r2, r2m -%endif - jmp .right_extend -%endmacro - -%macro SLOW_RIGHT_EXTEND 0 -.slow_right_extend_loop: -; r3(64)/r0(32)=buf+block_h*linesize,r2=linesize,r4=block_w,r8(64)/r5(32)=block_h, -; r7(64)/r6(32)=end_x,r6/r3=val,r1=cntr -%if ARCH_X86_64 -%define buf_reg r3 -%define bh_reg r8 -%else -%define buf_reg r0 -%define bh_reg r5 -%endif - lea r1, [r4-8] - sub buf_reg, linesize - READ_V_PIXEL 8, [buf_reg+w_reg-1] -.right_extend_8px_loop: - movq [buf_reg+r1], mm0 - sub r1, 8 - cmp r1, w_reg - jge .right_extend_8px_loop - add r1, 8 - cmp r1, w_reg - je .right_extend_loop_end -.right_extend_2px_loop: - sub r1, 2 - mov [buf_reg+r1], valw - cmp r1, w_reg - jg .right_extend_2px_loop -.right_extend_loop_end: - dec bh_reg - jnz .slow_right_extend_loop - jmp .h_extend_end -%endmacro - -%macro emu_edge 1 -INIT_XMM %1 -EMU_EDGE_FUNC -VERTICAL_EXTEND -LEFT_EXTEND -RIGHT_EXTEND -SLOW_V_EXTEND -SLOW_LEFT_EXTEND -SLOW_RIGHT_EXTEND -%endmacro - -emu_edge sse -%if ARCH_X86_32 -emu_edge mmx -%endif +INIT_XMM sse +H_EXTEND 16, 22 %macro PREFETCH_FN 1 cglobal prefetch, 3, 3, 0, buf, stride, h diff --git a/libavcodec/x86/videodsp_init.c b/libavcodec/x86/videodsp_init.c index 17a90115cf..a709d03d26 100644 --- a/libavcodec/x86/videodsp_init.c +++ b/libavcodec/x86/videodsp_init.c @@ -30,32 +30,126 @@ #include "libavcodec/videodsp.h" #if HAVE_YASM -typedef void emu_edge_core_func(uint8_t *buf, const uint8_t *src, - x86_reg linesize, x86_reg start_y, - x86_reg end_y, x86_reg block_h, - x86_reg start_x, x86_reg end_x, - x86_reg block_w); -extern emu_edge_core_func ff_emu_edge_core_mmx; -extern emu_edge_core_func ff_emu_edge_core_sse; +typedef void emu_edge_vfix_func(uint8_t *dst, x86_reg dst_stride, + const uint8_t *src, x86_reg src_stride, + x86_reg start_y, x86_reg end_y, x86_reg bh); +typedef void emu_edge_vvar_func(uint8_t *dst, x86_reg dst_stride, + const uint8_t *src, x86_reg src_stride, + x86_reg start_y, x86_reg end_y, x86_reg bh, + x86_reg w); -static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src, - ptrdiff_t linesize, - int block_w, int block_h, - int src_x, int src_y, - int w, int h, - emu_edge_core_func *core_fn) +extern emu_edge_vfix_func ff_emu_edge_vfix1_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix2_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix3_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix4_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix5_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix6_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix7_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix8_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix9_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix10_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix11_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix12_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix13_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix14_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix15_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix16_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix17_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix18_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix19_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix20_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix21_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix22_mmx; +#if ARCH_X86_32 +static emu_edge_vfix_func *vfixtbl_mmx[22] = { + &ff_emu_edge_vfix1_mmx, &ff_emu_edge_vfix2_mmx, &ff_emu_edge_vfix3_mmx, + &ff_emu_edge_vfix4_mmx, &ff_emu_edge_vfix5_mmx, &ff_emu_edge_vfix6_mmx, + &ff_emu_edge_vfix7_mmx, &ff_emu_edge_vfix8_mmx, &ff_emu_edge_vfix9_mmx, + &ff_emu_edge_vfix10_mmx, &ff_emu_edge_vfix11_mmx, &ff_emu_edge_vfix12_mmx, + &ff_emu_edge_vfix13_mmx, &ff_emu_edge_vfix14_mmx, &ff_emu_edge_vfix15_mmx, + &ff_emu_edge_vfix16_mmx, &ff_emu_edge_vfix17_mmx, &ff_emu_edge_vfix18_mmx, + &ff_emu_edge_vfix19_mmx, &ff_emu_edge_vfix20_mmx, &ff_emu_edge_vfix21_mmx, + &ff_emu_edge_vfix22_mmx +}; +#endif +extern emu_edge_vvar_func ff_emu_edge_vvar_mmx; +extern emu_edge_vfix_func ff_emu_edge_vfix16_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix17_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix18_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix19_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix20_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix21_sse; +extern emu_edge_vfix_func ff_emu_edge_vfix22_sse; +static emu_edge_vfix_func *vfixtbl_sse[22] = { + ff_emu_edge_vfix1_mmx, ff_emu_edge_vfix2_mmx, ff_emu_edge_vfix3_mmx, + ff_emu_edge_vfix4_mmx, ff_emu_edge_vfix5_mmx, ff_emu_edge_vfix6_mmx, + ff_emu_edge_vfix7_mmx, ff_emu_edge_vfix8_mmx, ff_emu_edge_vfix9_mmx, + ff_emu_edge_vfix10_mmx, ff_emu_edge_vfix11_mmx, ff_emu_edge_vfix12_mmx, + ff_emu_edge_vfix13_mmx, ff_emu_edge_vfix14_mmx, ff_emu_edge_vfix15_mmx, + ff_emu_edge_vfix16_sse, ff_emu_edge_vfix17_sse, ff_emu_edge_vfix18_sse, + ff_emu_edge_vfix19_sse, ff_emu_edge_vfix20_sse, ff_emu_edge_vfix21_sse, + ff_emu_edge_vfix22_sse +}; +extern emu_edge_vvar_func ff_emu_edge_vvar_sse; + +typedef void emu_edge_hfix_func(uint8_t *dst, x86_reg dst_stride, + x86_reg start_x, x86_reg bh); +typedef void emu_edge_hvar_func(uint8_t *dst, x86_reg dst_stride, + x86_reg start_x, x86_reg n_words, x86_reg bh); + +extern emu_edge_hfix_func ff_emu_edge_hfix2_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix4_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix6_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix8_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix10_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix12_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix14_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix16_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix18_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix20_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix22_mmx; +#if ARCH_X86_32 +static emu_edge_hfix_func *hfixtbl_mmx[11] = { + ff_emu_edge_hfix2_mmx, ff_emu_edge_hfix4_mmx, ff_emu_edge_hfix6_mmx, + ff_emu_edge_hfix8_mmx, ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx, + ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_mmx, ff_emu_edge_hfix18_mmx, + ff_emu_edge_hfix20_mmx, ff_emu_edge_hfix22_mmx +}; +#endif +extern emu_edge_hvar_func ff_emu_edge_hvar_mmx; +extern emu_edge_hfix_func ff_emu_edge_hfix16_sse; +extern emu_edge_hfix_func ff_emu_edge_hfix18_sse; +extern emu_edge_hfix_func ff_emu_edge_hfix20_sse; +extern emu_edge_hfix_func ff_emu_edge_hfix22_sse; +static emu_edge_hfix_func *hfixtbl_sse[11] = { + ff_emu_edge_hfix2_mmx, ff_emu_edge_hfix4_mmx, ff_emu_edge_hfix6_mmx, + ff_emu_edge_hfix8_mmx, ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx, + ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_sse, ff_emu_edge_hfix18_sse, + ff_emu_edge_hfix20_sse, ff_emu_edge_hfix22_sse +}; +extern emu_edge_hvar_func ff_emu_edge_hvar_sse; + +static av_always_inline void emulated_edge_mc(uint8_t *dst, ptrdiff_t dst_stride, + const uint8_t *src, ptrdiff_t src_stride, + x86_reg block_w, x86_reg block_h, + x86_reg src_x, x86_reg src_y, + x86_reg w, x86_reg h, + emu_edge_vfix_func **vfix_tbl, + emu_edge_vvar_func *v_extend_var, + emu_edge_hfix_func **hfix_tbl, + emu_edge_hvar_func *h_extend_var) { - int start_y, start_x, end_y, end_x, src_y_add = 0; + x86_reg start_y, start_x, end_y, end_x, src_y_add = 0, p; if(!w || !h) return; if (src_y >= h) { - src -= src_y*linesize; + src -= src_y*src_stride; src_y_add = h - 1; src_y = h - 1; } else if (src_y <= -block_h) { - src -= src_y*linesize; + src -= src_y*src_stride; src_y_add = 1 - block_h; src_y = 1 - block_h; } @@ -75,30 +169,59 @@ static av_always_inline void emulated_edge_mc(uint8_t *buf, const uint8_t *src, av_assert2(start_y < end_y && block_h > 0); // fill in the to-be-copied part plus all above/below - src += (src_y_add + start_y) * linesize + start_x; - buf += start_x; - core_fn(buf, src, linesize, start_y, end_y, - block_h, start_x, end_x, block_w); + src += (src_y_add + start_y) * src_stride + start_x; + w = end_x - start_x; + if (w <= 22) { + vfix_tbl[w - 1](dst + start_x, dst_stride, src, src_stride, + start_y, end_y, block_h); + } else { + v_extend_var(dst + start_x, dst_stride, src, src_stride, + start_y, end_y, block_h, w); + } + + // fill left + if (start_x) { + if (start_x <= 22) { + hfix_tbl[(start_x - 1) >> 1](dst, dst_stride, start_x, block_h); + } else { + h_extend_var(dst, dst_stride, + start_x, (start_x + 1) >> 1, block_h); + } + } + + // fill right + p = block_w - end_x; + if (p) { + if (p <= 22) { + hfix_tbl[(p - 1) >> 1](dst + end_x - (p & 1), dst_stride, + -!(p & 1), block_h); + } else { + h_extend_var(dst + end_x - (p & 1), dst_stride, + -!(p & 1), (p + 1) >> 1, block_h); + } + } } #if ARCH_X86_32 -static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src, - ptrdiff_t linesize, +static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, ptrdiff_t buf_stride, + const uint8_t *src, ptrdiff_t src_stride, int block_w, int block_h, int src_x, int src_y, int w, int h) { - emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y, - w, h, &ff_emu_edge_core_mmx); + emulated_edge_mc(buf, buf_stride, src, src_stride, block_w, block_h, + src_x, src_y, w, h, vfixtbl_mmx, &ff_emu_edge_vvar_mmx, + hfixtbl_mmx, &ff_emu_edge_hvar_mmx); } #endif -static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src, - ptrdiff_t linesize, +static av_noinline void emulated_edge_mc_sse(uint8_t *buf, ptrdiff_t buf_stride, + const uint8_t *src, ptrdiff_t src_stride, int block_w, int block_h, int src_x, int src_y, int w, int h) { - emulated_edge_mc(buf, src, linesize, block_w, block_h, src_x, src_y, - w, h, &ff_emu_edge_core_sse); + emulated_edge_mc(buf, buf_stride, src, src_stride, block_w, block_h, + src_x, src_y, w, h, vfixtbl_sse, &ff_emu_edge_vvar_sse, + hfixtbl_sse, &ff_emu_edge_hvar_sse); } #endif /* HAVE_YASM */