Merge commit 'e9a77f4bc084e5174baba3a8bc869ce4d102c4f0'
authorMichael Niedermayer <michaelni@gmx.at>
Sun, 23 Mar 2014 23:46:20 +0000 (00:46 +0100)
committerMichael Niedermayer <michaelni@gmx.at>
Sun, 23 Mar 2014 23:46:20 +0000 (00:46 +0100)
* commit 'e9a77f4bc084e5174baba3a8bc869ce4d102c4f0':
  h264: move macroblock decoding into its own file

Conflicts:
libavcodec/Makefile
libavcodec/h264.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>
1  2 
libavcodec/Makefile
libavcodec/h264.c
libavcodec/h264_mb.c

@@@ -233,9 -196,8 +233,9 @@@ OBJS-$(CONFIG_H263_ENCODER)            
                                            h263.o ituh263enc.o flvenc.o
  OBJS-$(CONFIG_H264_DECODER)            += h264.o h264_cabac.o h264_cavlc.o \
                                            h264_direct.o h264_loopfilter.o  \
-                                           h264_ps.o h264_refs.o h264_sei.o \
-                                           cabac.o
+                                           h264_mb.o h264_ps.o h264_refs.o  \
+                                           h264_sei.o cabac.o
 +OBJS-$(CONFIG_H264_VDA_DECODER)        += vda_h264_dec.o
  OBJS-$(CONFIG_HEVC_DECODER)            += hevc.o hevc_mvs.o hevc_ps.o hevc_sei.o \
                                            hevc_cabac.o hevc_refs.o hevcpred.o    \
                                            hevcdsp.o hevc_filter.o cabac.o
Simple merge
index 0000000,67d9202..ff902be
mode 000000,100644..100644
--- /dev/null
@@@ -1,0 -1,821 +1,820 @@@
 - * This file is part of Libav.
+ /*
+  * H.26L/H.264/AVC/JVT/14496-10/... decoder
+  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
+  *
 - * Libav is free software; you can redistribute it and/or
++ * This file is part of FFmpeg.
+  *
 - * Libav is distributed in the hope that it will be useful,
++ * FFmpeg is free software; you can redistribute it and/or
+  * modify it under the terms of the GNU Lesser General Public
+  * License as published by the Free Software Foundation; either
+  * version 2.1 of the License, or (at your option) any later version.
+  *
 - * License along with Libav; if not, write to the Free Software
++ * FFmpeg is distributed in the hope that it will be useful,
+  * but WITHOUT ANY WARRANTY; without even the implied warranty of
+  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+  * Lesser General Public License for more details.
+  *
+  * You should have received a copy of the GNU Lesser General Public
 -    int filter_height_up   = (raw_my & 3) ? 2 : 0;
++ * License along with FFmpeg; if not, write to the Free Software
+  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+  */
+ /**
+  * @file
+  * H.264 / AVC / MPEG4 part10 macroblock decoding
+  */
+ #include <stdint.h>
+ #include "config.h"
+ #include "libavutil/common.h"
+ #include "libavutil/intreadwrite.h"
+ #include "avcodec.h"
+ #include "h264.h"
+ #include "svq3.h"
+ #include "thread.h"
+ static inline int get_lowest_part_list_y(H264Context *h, H264Picture *pic, int n,
+                                          int height, int y_offset, int list)
+ {
+     int raw_my             = h->mv_cache[list][scan8[n]][1];
 -    int top                = full_my - filter_height_up;
+     int filter_height_down = (raw_my & 3) ? 3 : 0;
+     int full_my            = (raw_my >> 2) + y_offset;
 -    return FFMAX(abs(top), bottom);
+     int bottom             = full_my + filter_height_down + height;
 -        assert(IS_8X8(mb_type));
++    av_assert2(height >= 0);
++
++    return FFMAX(0, bottom);
+ }
+ static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n,
+                                      int height, int y_offset, int list0,
+                                      int list1, int *nrefs)
+ {
+     int my;
+     y_offset += 16 * (h->mb_y >> MB_FIELD(h));
+     if (list0) {
+         int ref_n = h->ref_cache[0][scan8[n]];
+         H264Picture *ref = &h->ref_list[0][ref_n];
+         // Error resilience puts the current picture in the ref list.
+         // Don't try to wait on these as it will cause a deadlock.
+         // Fields can wait on each other, though.
+         if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
+             (ref->reference & 3) != h->picture_structure) {
+             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
+             if (refs[0][ref_n] < 0)
+                 nrefs[0] += 1;
+             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
+         }
+     }
+     if (list1) {
+         int ref_n    = h->ref_cache[1][scan8[n]];
+         H264Picture *ref = &h->ref_list[1][ref_n];
+         if (ref->tf.progress->data != h->cur_pic.tf.progress->data ||
+             (ref->reference & 3) != h->picture_structure) {
+             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
+             if (refs[1][ref_n] < 0)
+                 nrefs[1] += 1;
+             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
+         }
+     }
+ }
+ /**
+  * Wait until all reference frames are available for MC operations.
+  *
+  * @param h the H264 context
+  */
+ static void await_references(H264Context *h)
+ {
+     const int mb_xy   = h->mb_xy;
+     const int mb_type = h->cur_pic.mb_type[mb_xy];
+     int refs[2][48];
+     int nrefs[2] = { 0 };
+     int ref, list;
+     memset(refs, -1, sizeof(refs));
+     if (IS_16X16(mb_type)) {
+         get_lowest_part_y(h, refs, 0, 16, 0,
+                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
+     } else if (IS_16X8(mb_type)) {
+         get_lowest_part_y(h, refs, 0, 8, 0,
+                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
+         get_lowest_part_y(h, refs, 8, 8, 8,
+                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
+     } else if (IS_8X16(mb_type)) {
+         get_lowest_part_y(h, refs, 0, 16, 0,
+                           IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
+         get_lowest_part_y(h, refs, 4, 16, 0,
+                           IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
+     } else {
+         int i;
 -                assert(IS_SUB_4X4(sub_mb_type));
++        av_assert2(IS_8X8(mb_type));
+         for (i = 0; i < 4; i++) {
+             const int sub_mb_type = h->sub_mb_type[i];
+             const int n           = 4 * i;
+             int y_offset          = (i & 2) << 2;
+             if (IS_SUB_8X8(sub_mb_type)) {
+                 get_lowest_part_y(h, refs, n, 8, y_offset,
+                                   IS_DIR(sub_mb_type, 0, 0),
+                                   IS_DIR(sub_mb_type, 0, 1),
+                                   nrefs);
+             } else if (IS_SUB_8X4(sub_mb_type)) {
+                 get_lowest_part_y(h, refs, n, 4, y_offset,
+                                   IS_DIR(sub_mb_type, 0, 0),
+                                   IS_DIR(sub_mb_type, 0, 1),
+                                   nrefs);
+                 get_lowest_part_y(h, refs, n + 2, 4, y_offset + 4,
+                                   IS_DIR(sub_mb_type, 0, 0),
+                                   IS_DIR(sub_mb_type, 0, 1),
+                                   nrefs);
+             } else if (IS_SUB_4X8(sub_mb_type)) {
+                 get_lowest_part_y(h, refs, n, 8, y_offset,
+                                   IS_DIR(sub_mb_type, 0, 0),
+                                   IS_DIR(sub_mb_type, 0, 1),
+                                   nrefs);
+                 get_lowest_part_y(h, refs, n + 1, 8, y_offset,
+                                   IS_DIR(sub_mb_type, 0, 0),
+                                   IS_DIR(sub_mb_type, 0, 1),
+                                   nrefs);
+             } else {
+                 int j;
 -            off = ((mx >> 1) << pixel_shift) +
 -                  ((my >> 1) + (h->mb_x & 7)) * h->uvlinesize +
 -                  (64 << pixel_shift);
++                av_assert2(IS_SUB_4X4(sub_mb_type));
+                 for (j = 0; j < 4; j++) {
+                     int sub_y_offset = y_offset + 2 * (j & 2);
+                     get_lowest_part_y(h, refs, n + j, 4, sub_y_offset,
+                                       IS_DIR(sub_mb_type, 0, 0),
+                                       IS_DIR(sub_mb_type, 0, 1),
+                                       nrefs);
+                 }
+             }
+         }
+     }
+     for (list = h->list_count - 1; list >= 0; list--)
+         for (ref = 0; ref < 48 && nrefs[list]; ref++) {
+             int row = refs[list][ref];
+             if (row >= 0) {
+                 H264Picture *ref_pic  = &h->ref_list[list][ref];
+                 int ref_field         = ref_pic->reference - 1;
+                 int ref_field_picture = ref_pic->field_picture;
+                 int pic_height        = 16 * h->mb_height >> ref_field_picture;
+                 row <<= MB_MBAFF(h);
+                 nrefs[list]--;
+                 if (!FIELD_PICTURE(h) && ref_field_picture) { // frame referencing two fields
+                     ff_thread_await_progress(&ref_pic->tf,
+                                              FFMIN((row >> 1) - !(row & 1),
+                                                    pic_height - 1),
+                                              1);
+                     ff_thread_await_progress(&ref_pic->tf,
+                                              FFMIN((row >> 1), pic_height - 1),
+                                              0);
+                 } else if (FIELD_PICTURE(h) && !ref_field_picture) { // field referencing one field of a frame
+                     ff_thread_await_progress(&ref_pic->tf,
+                                              FFMIN(row * 2 + ref_field,
+                                                    pic_height - 1),
+                                              0);
+                 } else if (FIELD_PICTURE(h)) {
+                     ff_thread_await_progress(&ref_pic->tf,
+                                              FFMIN(row, pic_height - 1),
+                                              ref_field);
+                 } else {
+                     ff_thread_await_progress(&ref_pic->tf,
+                                              FFMIN(row, pic_height - 1),
+                                              0);
+                 }
+             }
+         }
+ }
+ static av_always_inline void mc_dir_part(H264Context *h, H264Picture *pic,
+                                          int n, int square, int height,
+                                          int delta, int list,
+                                          uint8_t *dest_y, uint8_t *dest_cb,
+                                          uint8_t *dest_cr,
+                                          int src_x_offset, int src_y_offset,
+                                          qpel_mc_func *qpix_op,
+                                          h264_chroma_mc_func chroma_op,
+                                          int pixel_shift, int chroma_idc)
+ {
+     const int mx      = h->mv_cache[list][scan8[n]][0] + src_x_offset * 8;
+     int my            = h->mv_cache[list][scan8[n]][1] + src_y_offset * 8;
+     const int luma_xy = (mx & 3) + ((my & 3) << 2);
+     ptrdiff_t offset  = ((mx >> 2) << pixel_shift) + (my >> 2) * h->mb_linesize;
+     uint8_t *src_y    = pic->f.data[0] + offset;
+     uint8_t *src_cb, *src_cr;
+     int extra_width  = 0;
+     int extra_height = 0;
+     int emu = 0;
+     const int full_mx    = mx >> 2;
+     const int full_my    = my >> 2;
+     const int pic_width  = 16 * h->mb_width;
+     const int pic_height = 16 * h->mb_height >> MB_FIELD(h);
+     int ysh;
+     if (mx & 7)
+         extra_width -= 3;
+     if (my & 7)
+         extra_height -= 3;
+     if (full_mx                <          0 - extra_width  ||
+         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 << pixel_shift) - 2 * h->mb_linesize,
+                                  h->mb_linesize, h->mb_linesize,
+                                  16 + 5, 16 + 5 /*FIXME*/, full_mx - 2,
+                                  full_my - 2, pic_width, pic_height);
+         src_y = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
+         emu   = 1;
+     }
+     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); // FIXME try variable height perhaps?
+     if (!square)
+         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
+     if (CONFIG_GRAY && h->flags & CODEC_FLAG_GRAY)
+         return;
+     if (chroma_idc == 3 /* yuv444 */) {
+         src_cb = pic->f.data[1] + offset;
+         if (emu) {
+             h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
+                                      src_cb - (2 << pixel_shift) - 2 * h->mb_linesize,
+                                      h->mb_linesize, h->mb_linesize,
+                                      16 + 5, 16 + 5 /*FIXME*/,
+                                      full_mx - 2, full_my - 2,
+                                      pic_width, pic_height);
+             src_cb = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
+         }
+         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); // FIXME try variable height perhaps?
+         if (!square)
+             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
+         src_cr = pic->f.data[2] + offset;
+         if (emu) {
+             h->vdsp.emulated_edge_mc(h->edge_emu_buffer,
+                                      src_cr - (2 << pixel_shift) - 2 * h->mb_linesize,
+                                      h->mb_linesize, h->mb_linesize,
+                                      16 + 5, 16 + 5 /*FIXME*/,
+                                      full_mx - 2, full_my - 2,
+                                      pic_width, pic_height);
+             src_cr = h->edge_emu_buffer + (2 << pixel_shift) + 2 * h->mb_linesize;
+         }
+         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); // FIXME try variable height perhaps?
+         if (!square)
+             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
+         return;
+     }
+     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
+     if (chroma_idc == 1 /* yuv420 */ && MB_FIELD(h)) {
+         // chroma offset when predicting from a field of opposite parity
+         my  += 2 * ((h->mb_y & 1) - (pic->reference - 1));
+         emu |= (my >> 3) < 0 || (my >> 3) + 8 >= (pic_height >> 1);
+     }
+     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) +
+              (my >> ysh) * h->mb_uvlinesize;
+     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) +
+              (my >> ysh) * h->mb_uvlinesize;
+     if (emu) {
+         h->vdsp.emulated_edge_mc(h->edge_emu_buffer, src_cb,
+                                  h->mb_uvlinesize, 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;
+     }
+     chroma_op(dest_cb, src_cb, h->mb_uvlinesize,
+               height >> (chroma_idc == 1 /* yuv420 */),
+               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->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;
+     }
+     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
+               mx & 7, (my << (chroma_idc == 2 /* yuv422 */)) & 7);
+ }
+ static av_always_inline void mc_part_std(H264Context *h, int n, int square,
+                                          int height, int delta,
+                                          uint8_t *dest_y, uint8_t *dest_cb,
+                                          uint8_t *dest_cr,
+                                          int x_offset, int y_offset,
+                                          qpel_mc_func *qpix_put,
+                                          h264_chroma_mc_func chroma_put,
+                                          qpel_mc_func *qpix_avg,
+                                          h264_chroma_mc_func chroma_avg,
+                                          int list0, int list1,
+                                          int pixel_shift, int chroma_idc)
+ {
+     qpel_mc_func *qpix_op         = qpix_put;
+     h264_chroma_mc_func chroma_op = chroma_put;
+     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+     if (chroma_idc == 3 /* yuv444 */) {
+         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+     } else if (chroma_idc == 2 /* yuv422 */) {
+         dest_cb += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
+         dest_cr += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
+     } else { /* yuv420 */
+         dest_cb += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
+         dest_cr += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
+     }
+     x_offset += 8 * h->mb_x;
+     y_offset += 8 * (h->mb_y >> MB_FIELD(h));
+     if (list0) {
+         H264Picture *ref = &h->ref_list[0][h->ref_cache[0][scan8[n]]];
+         mc_dir_part(h, ref, n, square, height, delta, 0,
+                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
+                     qpix_op, chroma_op, pixel_shift, chroma_idc);
+         qpix_op   = qpix_avg;
+         chroma_op = chroma_avg;
+     }
+     if (list1) {
+         H264Picture *ref = &h->ref_list[1][h->ref_cache[1][scan8[n]]];
+         mc_dir_part(h, ref, n, square, height, delta, 1,
+                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
+                     qpix_op, chroma_op, pixel_shift, chroma_idc);
+     }
+ }
+ static av_always_inline void mc_part_weighted(H264Context *h, int n, int square,
+                                               int height, int delta,
+                                               uint8_t *dest_y, uint8_t *dest_cb,
+                                               uint8_t *dest_cr,
+                                               int x_offset, int y_offset,
+                                               qpel_mc_func *qpix_put,
+                                               h264_chroma_mc_func chroma_put,
+                                               h264_weight_func luma_weight_op,
+                                               h264_weight_func chroma_weight_op,
+                                               h264_biweight_func luma_weight_avg,
+                                               h264_biweight_func chroma_weight_avg,
+                                               int list0, int list1,
+                                               int pixel_shift, int chroma_idc)
+ {
+     int chroma_height;
+     dest_y += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+     if (chroma_idc == 3 /* yuv444 */) {
+         chroma_height     = height;
+         chroma_weight_avg = luma_weight_avg;
+         chroma_weight_op  = luma_weight_op;
+         dest_cb += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+         dest_cr += (2 * x_offset << pixel_shift) + 2 * y_offset * h->mb_linesize;
+     } else if (chroma_idc == 2 /* yuv422 */) {
+         chroma_height = height;
+         dest_cb      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
+         dest_cr      += (x_offset << pixel_shift) + 2 * y_offset * h->mb_uvlinesize;
+     } else { /* yuv420 */
+         chroma_height = height >> 1;
+         dest_cb      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
+         dest_cr      += (x_offset << pixel_shift) + y_offset * h->mb_uvlinesize;
+     }
+     x_offset += 8 * h->mb_x;
+     y_offset += 8 * (h->mb_y >> MB_FIELD(h));
+     if (list0 && list1) {
+         /* don't optimize for luma-only case, since B-frames usually
+          * use implicit weights => chroma too. */
+         uint8_t *tmp_cb = h->bipred_scratchpad;
+         uint8_t *tmp_cr = h->bipred_scratchpad + (16 << pixel_shift);
+         uint8_t *tmp_y  = h->bipred_scratchpad + 16 * h->mb_uvlinesize;
+         int refn0       = h->ref_cache[0][scan8[n]];
+         int refn1       = h->ref_cache[1][scan8[n]];
+         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
+                     dest_y, dest_cb, dest_cr,
+                     x_offset, y_offset, qpix_put, chroma_put,
+                     pixel_shift, chroma_idc);
+         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
+                     tmp_y, tmp_cb, tmp_cr,
+                     x_offset, y_offset, qpix_put, chroma_put,
+                     pixel_shift, chroma_idc);
+         if (h->use_weight == 2) {
+             int weight0 = h->implicit_weight[refn0][refn1][h->mb_y & 1];
+             int weight1 = 64 - weight0;
+             luma_weight_avg(dest_y, tmp_y, h->mb_linesize,
+                             height, 5, weight0, weight1, 0);
+             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
+                               chroma_height, 5, weight0, weight1, 0);
+             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
+                               chroma_height, 5, weight0, weight1, 0);
+         } else {
+             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height,
+                             h->luma_log2_weight_denom,
+                             h->luma_weight[refn0][0][0],
+                             h->luma_weight[refn1][1][0],
+                             h->luma_weight[refn0][0][1] +
+                             h->luma_weight[refn1][1][1]);
+             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height,
+                               h->chroma_log2_weight_denom,
+                               h->chroma_weight[refn0][0][0][0],
+                               h->chroma_weight[refn1][1][0][0],
+                               h->chroma_weight[refn0][0][0][1] +
+                               h->chroma_weight[refn1][1][0][1]);
+             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height,
+                               h->chroma_log2_weight_denom,
+                               h->chroma_weight[refn0][0][1][0],
+                               h->chroma_weight[refn1][1][1][0],
+                               h->chroma_weight[refn0][0][1][1] +
+                               h->chroma_weight[refn1][1][1][1]);
+         }
+     } else {
+         int list     = list1 ? 1 : 0;
+         int refn     = h->ref_cache[list][scan8[n]];
+         H264Picture *ref = &h->ref_list[list][refn];
+         mc_dir_part(h, ref, n, square, height, delta, list,
+                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
+                     qpix_put, chroma_put, pixel_shift, chroma_idc);
+         luma_weight_op(dest_y, h->mb_linesize, height,
+                        h->luma_log2_weight_denom,
+                        h->luma_weight[refn][list][0],
+                        h->luma_weight[refn][list][1]);
+         if (h->use_weight_chroma) {
+             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height,
+                              h->chroma_log2_weight_denom,
+                              h->chroma_weight[refn][list][0][0],
+                              h->chroma_weight[refn][list][0][1]);
+             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height,
+                              h->chroma_log2_weight_denom,
+                              h->chroma_weight[refn][list][1][0],
+                              h->chroma_weight[refn][list][1][1]);
+         }
+     }
+ }
+ static av_always_inline void prefetch_motion(H264Context *h, int list,
+                                              int pixel_shift, int chroma_idc)
+ {
+     /* fetch pixels for estimated mv 4 macroblocks ahead
+      * optimized for 64byte cache lines */
+     const int refn = h->ref_cache[list][scan8[0]];
+     if (refn >= 0) {
+         const int mx  = (h->mv_cache[list][scan8[0]][0] >> 2) + 16 * h->mb_x + 8;
+         const int my  = (h->mv_cache[list][scan8[0]][1] >> 2) + 16 * h->mb_y;
+         uint8_t **src = h->ref_list[list][refn].f.data;
+         int off       = (mx << pixel_shift) +
+                         (my + (h->mb_x & 3) * 4) * h->mb_linesize +
+                         (64 << pixel_shift);
+         h->vdsp.prefetch(src[0] + off, h->linesize, 4);
+         if (chroma_idc == 3 /* yuv444 */) {
+             h->vdsp.prefetch(src[1] + off, h->linesize, 4);
+             h->vdsp.prefetch(src[2] + off, h->linesize, 4);
+         } else {
 -    }
 -    if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
 -        if (chroma444) {
 -            if (deblock_top) {
++            off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (h->mb_x&7))*h->uvlinesize;
+             h->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
+         }
+     }
+ }
+ static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
+                                             uint8_t *src_cb, uint8_t *src_cr,
+                                             int linesize, int uvlinesize,
+                                             int xchg, int chroma444,
+                                             int simple, int pixel_shift)
+ {
+     int deblock_topleft;
+     int deblock_top;
+     int top_idx = 1;
+     uint8_t *top_border_m1;
+     uint8_t *top_border;
+     if (!simple && FRAME_MBAFF(h)) {
+         if (h->mb_y & 1) {
+             if (!MB_MBAFF(h))
+                 return;
+         } else {
+             top_idx = MB_MBAFF(h) ? 0 : 1;
+         }
+     }
+     if (h->deblocking_filter == 2) {
+         deblock_topleft = h->slice_table[h->mb_xy - 1 - h->mb_stride] == h->slice_num;
+         deblock_top     = h->top_type;
+     } else {
+         deblock_topleft = (h->mb_x > 0);
+         deblock_top     = (h->mb_y > !!MB_FIELD(h));
+     }
+     src_y  -= linesize   + 1 + pixel_shift;
+     src_cb -= uvlinesize + 1 + pixel_shift;
+     src_cr -= uvlinesize + 1 + pixel_shift;
+     top_border_m1 = h->top_borders[top_idx][h->mb_x - 1];
+     top_border    = h->top_borders[top_idx][h->mb_x];
+ #define XCHG(a, b, xchg)                        \
+     if (pixel_shift) {                          \
+         if (xchg) {                             \
+             AV_SWAP64(b + 0, a + 0);            \
+             AV_SWAP64(b + 8, a + 8);            \
+         } else {                                \
+             AV_COPY128(b, a);                   \
+         }                                       \
+     } else if (xchg)                            \
+         AV_SWAP64(b, a);                        \
+     else                                        \
+         AV_COPY64(b, a);
+     if (deblock_top) {
+         if (deblock_topleft) {
+             XCHG(top_border_m1 + (8 << pixel_shift),
+                  src_y - (7 << pixel_shift), 1);
+         }
+         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
+         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
+         if (h->mb_x + 1 < h->mb_width) {
+             XCHG(h->top_borders[top_idx][h->mb_x + 1],
+                  src_y + (17 << pixel_shift), 1);
+         }
 -            }
 -        } else {
 -            if (deblock_top) {
++        if (simple || !CONFIG_GRAY || !(h->flags & CODEC_FLAG_GRAY)) {
++            if (chroma444) {
+                 if (deblock_topleft) {
+                     XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
+                     XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
+                 }
+                 XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
+                 XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
+                 XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
+                 XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
+                 if (h->mb_x + 1 < h->mb_width) {
+                     XCHG(h->top_borders[top_idx][h->mb_x + 1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
+                     XCHG(h->top_borders[top_idx][h->mb_x + 1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
+                 }
 -                    h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
++            } else {
+                 if (deblock_topleft) {
+                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
+                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
+                 }
+                 XCHG(top_border + (16 << pixel_shift), src_cb + 1 + pixel_shift, 1);
+                 XCHG(top_border + (24 << pixel_shift), src_cr + 1 + pixel_shift, 1);
+             }
+         }
+     }
+ }
+ static av_always_inline int dctcoef_get(int16_t *mb, int high_bit_depth,
+                                         int index)
+ {
+     if (high_bit_depth) {
+         return AV_RN32A(((int32_t *)mb) + index);
+     } else
+         return AV_RN16A(mb + index);
+ }
+ static av_always_inline void dctcoef_set(int16_t *mb, int high_bit_depth,
+                                          int index, int value)
+ {
+     if (high_bit_depth) {
+         AV_WN32A(((int32_t *)mb) + index, value);
+     } else
+         AV_WN16A(mb + index, value);
+ }
+ static av_always_inline void hl_decode_mb_predict_luma(H264Context *h,
+                                                        int mb_type, int is_h264,
+                                                        int simple,
+                                                        int transform_bypass,
+                                                        int pixel_shift,
+                                                        int *block_offset,
+                                                        int linesize,
+                                                        uint8_t *dest_y, int p)
+ {
+     void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
+     void (*idct_dc_add)(uint8_t *dst, int16_t *block, int stride);
+     int i;
+     int qscale = p == 0 ? h->qscale : h->chroma_qp[p - 1];
+     block_offset += 16 * p;
+     if (IS_INTRA4x4(mb_type)) {
+         if (IS_8x8DCT(mb_type)) {
+             if (transform_bypass) {
+                 idct_dc_add =
+                 idct_add    = h->h264dsp.h264_add_pixels8_clear;
+             } else {
+                 idct_dc_add = h->h264dsp.h264_idct8_dc_add;
+                 idct_add    = h->h264dsp.h264_idct8_add;
+             }
+             for (i = 0; i < 16; i += 4) {
+                 uint8_t *const ptr = dest_y + block_offset[i];
+                 const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
+                 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
 -                        assert(h->mb_y || linesize <= block_offset[i]);
++                    if (h->x264_build != -1) {
++                        h->hpc.pred8x8l_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
++                    } else
++                        h->hpc.pred8x8l_filter_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift),
++                                                        (h-> topleft_samples_available << i) & 0x8000,
++                                                        (h->topright_samples_available << i) & 0x4000, linesize);
+                 } else {
+                     const int nnz = h->non_zero_count_cache[scan8[i + p * 16]];
+                     h->hpc.pred8x8l[dir](ptr, (h->topleft_samples_available << i) & 0x8000,
+                                          (h->topright_samples_available << i) & 0x4000, linesize);
+                     if (nnz) {
+                         if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
+                             idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
+                         else
+                             idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
+                     }
+                 }
+             }
+         } else {
+             if (transform_bypass) {
+                 idct_dc_add  =
+                 idct_add     = h->h264dsp.h264_add_pixels4_clear;
+             } else {
+                 idct_dc_add = h->h264dsp.h264_idct_dc_add;
+                 idct_add    = h->h264dsp.h264_idct_add;
+             }
+             for (i = 0; i < 16; i++) {
+                 uint8_t *const ptr = dest_y + block_offset[i];
+                 const int dir      = h->intra4x4_pred_mode_cache[scan8[i]];
+                 if (transform_bypass && h->sps.profile_idc == 244 && dir <= 1) {
+                     h->hpc.pred4x4_add[dir](ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
+                 } else {
+                     uint8_t *topright;
+                     int nnz, tr;
+                     uint64_t tr_high;
+                     if (dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED) {
+                         const int topright_avail = (h->topright_samples_available << i) & 0x8000;
++                        av_assert2(h->mb_y || linesize <= block_offset[i]);
+                         if (!topright_avail) {
+                             if (pixel_shift) {
+                                 tr_high  = ((uint16_t *)ptr)[3 - linesize / 2] * 0x0001000100010001ULL;
+                                 topright = (uint8_t *)&tr_high;
+                             } else {
+                                 tr       = ptr[3 - linesize] * 0x01010101u;
+                                 topright = (uint8_t *)&tr;
+                             }
+                         } else
+                             topright = ptr + (4 << pixel_shift) - linesize;
+                     } else
+                         topright = NULL;
+                     h->hpc.pred4x4[dir](ptr, topright, linesize);
+                     nnz = h->non_zero_count_cache[scan8[i + p * 16]];
+                     if (nnz) {
+                         if (is_h264) {
+                             if (nnz == 1 && dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
+                                 idct_dc_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
+                             else
+                                 idct_add(ptr, h->mb + (i * 16 + p * 256 << pixel_shift), linesize);
+                         } else if (CONFIG_SVQ3_DECODER)
+                             ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize, qscale, 0);
+                     }
+                 }
+             }
+         }
+     } else {
+         h->hpc.pred16x16[h->intra16x16_pred_mode](dest_y, linesize);
+         if (is_h264) {
+             if (h->non_zero_count_cache[scan8[LUMA_DC_BLOCK_INDEX + p]]) {
+                 if (!transform_bypass)
+                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb + (p * 256 << pixel_shift),
+                                                          h->mb_luma_dc[p],
+                                                          h->dequant4_coeff[p][qscale][0]);
+                 else {
+                     static const uint8_t dc_mapping[16] = {
+                          0 * 16,  1 * 16,  4 * 16,  5 * 16,
+                          2 * 16,  3 * 16,  6 * 16,  7 * 16,
+                          8 * 16,  9 * 16, 12 * 16, 13 * 16,
+                         10 * 16, 11 * 16, 14 * 16, 15 * 16
+                     };
+                     for (i = 0; i < 16; i++)
+                         dctcoef_set(h->mb + (p * 256 << pixel_shift),
+                                     pixel_shift, dc_mapping[i],
+                                     dctcoef_get(h->mb_luma_dc[p],
+                                                 pixel_shift, i));
+                 }
+             }
+         } else if (CONFIG_SVQ3_DECODER)
+             ff_svq3_luma_dc_dequant_idct_c(h->mb + p * 256,
+                                            h->mb_luma_dc[p], qscale);
+     }
+ }
+ static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type,
+                                                     int is_h264, int simple,
+                                                     int transform_bypass,
+                                                     int pixel_shift,
+                                                     int *block_offset,
+                                                     int linesize,
+                                                     uint8_t *dest_y, int p)
+ {
+     void (*idct_add)(uint8_t *dst, int16_t *block, int stride);
+     int i;
+     block_offset += 16 * p;
+     if (!IS_INTRA4x4(mb_type)) {
+         if (is_h264) {
+             if (IS_INTRA16x16(mb_type)) {
+                 if (transform_bypass) {
+                     if (h->sps.profile_idc == 244 &&
+                         (h->intra16x16_pred_mode == VERT_PRED8x8 ||
+                          h->intra16x16_pred_mode == HOR_PRED8x8)) {
+                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset,
+                                                                       h->mb + (p * 256 << pixel_shift),
+                                                                       linesize);
+                     } else {
+                         for (i = 0; i < 16; i++)
+                             if (h->non_zero_count_cache[scan8[i + p * 16]] ||
+                                 dctcoef_get(h->mb, pixel_shift, i * 16 + p * 256))
+                                 h->h264dsp.h264_add_pixels4_clear(dest_y + block_offset[i],
+                                                                   h->mb + (i * 16 + p * 256 << pixel_shift),
+                                                                   linesize);
+                     }
+                 } else {
+                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset,
+                                                     h->mb + (p * 256 << pixel_shift),
+                                                     linesize,
+                                                     h->non_zero_count_cache + p * 5 * 8);
+                 }
+             } else if (h->cbp & 15) {
+                 if (transform_bypass) {
+                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
+                     idct_add = IS_8x8DCT(mb_type) ? h->h264dsp.h264_add_pixels8_clear
+                                                   : h->h264dsp.h264_add_pixels4_clear;
+                     for (i = 0; i < 16; i += di)
+                         if (h->non_zero_count_cache[scan8[i + p * 16]])
+                             idct_add(dest_y + block_offset[i],
+                                      h->mb + (i * 16 + p * 256 << pixel_shift),
+                                      linesize);
+                 } else {
+                     if (IS_8x8DCT(mb_type))
+                         h->h264dsp.h264_idct8_add4(dest_y, block_offset,
+                                                    h->mb + (p * 256 << pixel_shift),
+                                                    linesize,
+                                                    h->non_zero_count_cache + p * 5 * 8);
+                     else
+                         h->h264dsp.h264_idct_add16(dest_y, block_offset,
+                                                    h->mb + (p * 256 << pixel_shift),
+                                                    linesize,
+                                                    h->non_zero_count_cache + p * 5 * 8);
+                 }
+             }
+         } else if (CONFIG_SVQ3_DECODER) {
+             for (i = 0; i < 16; i++)
+                 if (h->non_zero_count_cache[scan8[i + p * 16]] || h->mb[i * 16 + p * 256]) {
+                     // FIXME benchmark weird rule, & below
+                     uint8_t *const ptr = dest_y + block_offset[i];
+                     ff_svq3_add_idct_c(ptr, h->mb + i * 16 + p * 256, linesize,
+                                        h->qscale, IS_INTRA(mb_type) ? 1 : 0);
+                 }
+         }
+     }
+ }
+ #define BITS   8
+ #define SIMPLE 1
+ #include "h264_mb_template.c"
+ #undef  BITS
+ #define BITS   16
+ #include "h264_mb_template.c"
+ #undef  SIMPLE
+ #define SIMPLE 0
+ #include "h264_mb_template.c"
+ void ff_h264_hl_decode_mb(H264Context *h)
+ {
+     const int mb_xy   = h->mb_xy;
+     const int mb_type = h->cur_pic.mb_type[mb_xy];
+     int is_complex    = CONFIG_SMALL || h->is_complex ||
+                         IS_INTRA_PCM(mb_type) || h->qscale == 0;
+     if (CHROMA444(h)) {
+         if (is_complex || h->pixel_shift)
+             hl_decode_mb_444_complex(h);
+         else
+             hl_decode_mb_444_simple_8(h);
+     } else if (is_complex) {
+         hl_decode_mb_complex(h);
+     } else if (h->pixel_shift) {
+         hl_decode_mb_simple_16(h);
+     } else
+         hl_decode_mb_simple_8(h);
+ }