src/gallium/drivers/radeon/radeon_uvd.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2011 Advanced Micro Devices, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 /*
  29  * Authors:
  30  *      Christian König <christian.koenig@amd.com>
  31  *
  32  */
  33
  34 #include <sys/types.h>
  35 #include <assert.h>
  36 #include <errno.h>
  37 #include <unistd.h>
  38 #include <stdio.h>
  39
  40 #include "pipe/p_video_codec.h"
  41
  42 #include "util/u_memory.h"
  43 #include "util/u_video.h"
  44
  45 #include "vl/vl_defines.h"
  46 #include "vl/vl_mpeg12_decoder.h"
  47
  48 #include "r600_pipe_common.h"
  49 #include "radeon_video.h"
  50 #include "radeon_uvd.h"
  51
  52 #define NUM_BUFFERS 4
  53
  54 #define NUM_MPEG2_REFS 6
  55 #define NUM_H264_REFS 17
  56 #define NUM_VC1_REFS 5
  57
  58 #define FB_BUFFER_OFFSET 0x1000
  59 #define FB_BUFFER_SIZE 2048
  60 #define IT_SCALING_TABLE_SIZE 992
  61
  62 /* UVD decoder representation */
  63 struct ruvd_decoder {
  64         struct pipe_video_codec         base;
  65
  66         ruvd_set_dtb                    set_dtb;
  67
  68         unsigned                        stream_handle;
  69         unsigned                        stream_type;
  70         unsigned                        frame_number;
  71
  72         struct pipe_screen              *screen;
  73         struct radeon_winsys*           ws;
  74         struct radeon_winsys_cs*        cs;
  75
  76         unsigned                        cur_buffer;
  77
  78         struct rvid_buffer              msg_fb_it_buffers[NUM_BUFFERS];
  79         struct ruvd_msg                 *msg;
  80         uint32_t                        *fb;
  81         uint8_t                         *it;
  82
  83         struct rvid_buffer              bs_buffers[NUM_BUFFERS];
  84         void*                           bs_ptr;
  85         unsigned                        bs_size;
  86
  87         struct rvid_buffer              dpb;
  88         bool                            use_legacy;
  89         struct rvid_buffer              ctx;
  90 };
  91
  92 /* flush IB to the hardware */
  93 static void flush(struct ruvd_decoder *dec)
  94 {
  95         dec->ws->cs_flush(dec->cs, RADEON_FLUSH_ASYNC, NULL, 0);
  96 }
  97
  98 /* add a new set register command to the IB */
  99 static void set_reg(struct ruvd_decoder *dec, unsigned reg, uint32_t val)
 100 {
 101         uint32_t *pm4 = dec->cs->buf;
 102         pm4[dec->cs->cdw++] = RUVD_PKT0(reg >> 2, 0);
 103         pm4[dec->cs->cdw++] = val;
 104 }
 105
 106 /* send a command to the VCPU through the GPCOM registers */
 107 static void send_cmd(struct ruvd_decoder *dec, unsigned cmd,
 108                      struct pb_buffer* buf, uint32_t off,
 109                      enum radeon_bo_usage usage, enum radeon_bo_domain domain)
 110 {
 111         int reloc_idx;
 112
 113         reloc_idx = dec->ws->cs_add_buffer(dec->cs, buf, usage, domain,
 114                                           RADEON_PRIO_UVD);
 115         if (!dec->use_legacy) {
 116                 uint64_t addr;
 117                 addr = dec->ws->buffer_get_virtual_address(buf);
 118                 addr = addr + off;
 119                 set_reg(dec, RUVD_GPCOM_VCPU_DATA0, addr);
 120                 set_reg(dec, RUVD_GPCOM_VCPU_DATA1, addr >> 32);
 121         } else {
 122                 set_reg(dec, RUVD_GPCOM_VCPU_DATA0, off);
 123                 set_reg(dec, RUVD_GPCOM_VCPU_DATA1, reloc_idx * 4);
 124         }
 125         set_reg(dec, RUVD_GPCOM_VCPU_CMD, cmd << 1);
 126 }
 127
 128 /* do the codec needs an IT buffer ?*/
 129 static bool have_it(struct ruvd_decoder *dec)
 130 {
 131         return dec->stream_type == RUVD_CODEC_H264_PERF ||
 132                 dec->stream_type == RUVD_CODEC_H265;
 133 }
 134
 135 /* map the next available message/feedback/itscaling buffer */
 136 static void map_msg_fb_it_buf(struct ruvd_decoder *dec)
 137 {
 138         struct rvid_buffer* buf;
 139         uint8_t *ptr;
 140
 141         /* grab the current message/feedback buffer */
 142         buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
 143
 144         /* and map it for CPU access */
 145         ptr = dec->ws->buffer_map(buf->res->buf, dec->cs, PIPE_TRANSFER_WRITE);
 146
 147         /* calc buffer offsets */
 148         dec->msg = (struct ruvd_msg *)ptr;
 149         dec->fb = (uint32_t *)(ptr + FB_BUFFER_OFFSET);
 150         if (have_it(dec))
 151                 dec->it = (uint8_t *)(ptr + FB_BUFFER_OFFSET + FB_BUFFER_SIZE);
 152 }
 153
 154 /* unmap and send a message command to the VCPU */
 155 static void send_msg_buf(struct ruvd_decoder *dec)
 156 {
 157         struct rvid_buffer* buf;
 158
 159         /* ignore the request if message/feedback buffer isn't mapped */
 160         if (!dec->msg || !dec->fb)
 161                 return;
 162
 163         /* grab the current message buffer */
 164         buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
 165
 166         /* unmap the buffer */
 167         dec->ws->buffer_unmap(buf->res->buf);
 168         dec->msg = NULL;
 169         dec->fb = NULL;
 170         dec->it = NULL;
 171
 172         /* and send it to the hardware */
 173         send_cmd(dec, RUVD_CMD_MSG_BUFFER, buf->res->buf, 0,
 174                  RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
 175 }
 176
 177 /* cycle to the next set of buffers */
 178 static void next_buffer(struct ruvd_decoder *dec)
 179 {
 180         ++dec->cur_buffer;
 181         dec->cur_buffer %= NUM_BUFFERS;
 182 }
 183
 184 /* convert the profile into something UVD understands */
 185 static uint32_t profile2stream_type(struct ruvd_decoder *dec, unsigned family)
 186 {
 187         switch (u_reduce_video_profile(dec->base.profile)) {
 188         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 189                 return (family >= CHIP_TONGA) ?
 190                         RUVD_CODEC_H264_PERF : RUVD_CODEC_H264;
 191
 192         case PIPE_VIDEO_FORMAT_VC1:
 193                 return RUVD_CODEC_VC1;
 194
 195         case PIPE_VIDEO_FORMAT_MPEG12:
 196                 return RUVD_CODEC_MPEG2;
 197
 198         case PIPE_VIDEO_FORMAT_MPEG4:
 199                 return RUVD_CODEC_MPEG4;
 200
 201         case PIPE_VIDEO_FORMAT_HEVC:
 202                 return RUVD_CODEC_H265;
 203
 204         default:
 205                 assert(0);
 206                 return 0;
 207         }
 208 }
 209
 210 static unsigned calc_ctx_size(struct ruvd_decoder *dec)
 211 {
 212         unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
 213         unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
 214
 215         unsigned max_references = dec->base.max_references + 1;
 216
 217         if (dec->base.width * dec->base.height >= 4096*2000)
 218                 max_references = MAX2(max_references, 8);
 219         else
 220                 max_references = MAX2(max_references, 17);
 221
 222         width = align (width, 16);
 223         height = align (height, 16);
 224         return ((width + 255) / 16) * ((height + 255) / 16) * 16 * max_references + 52 * 1024;
 225 }
 226
 227 /* calculate size of reference picture buffer */
 228 static unsigned calc_dpb_size(struct ruvd_decoder *dec)
 229 {
 230         unsigned width_in_mb, height_in_mb, image_size, dpb_size;
 231
 232         // always align them to MB size for dpb calculation
 233         unsigned width = align(dec->base.width, VL_MACROBLOCK_WIDTH);
 234         unsigned height = align(dec->base.height, VL_MACROBLOCK_HEIGHT);
 235
 236         // always one more for currently decoded picture
 237         unsigned max_references = dec->base.max_references + 1;
 238
 239         // aligned size of a single frame
 240         image_size = width * height;
 241         image_size += image_size / 2;
 242         image_size = align(image_size, 1024);
 243
 244         // picture width & height in 16 pixel units
 245         width_in_mb = width / VL_MACROBLOCK_WIDTH;
 246         height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
 247
 248         switch (u_reduce_video_profile(dec->base.profile)) {
 249         case PIPE_VIDEO_FORMAT_MPEG4_AVC: {
 250                 if (!dec->use_legacy) {
 251                         unsigned fs_in_mb = width_in_mb * height_in_mb;
 252                         unsigned alignment = 64, num_dpb_buffer;
 253
 254                         if (dec->stream_type == RUVD_CODEC_H264_PERF)
 255                                 alignment = 256;
 256                         switch(dec->base.level) {
 257                         case 30:
 258                                 num_dpb_buffer = 8100 / fs_in_mb;
 259                                 break;
 260                         case 31:
 261                                 num_dpb_buffer = 18000 / fs_in_mb;
 262                                 break;
 263                         case 32:
 264                                 num_dpb_buffer = 20480 / fs_in_mb;
 265                                 break;
 266                         case 41:
 267                                 num_dpb_buffer = 32768 / fs_in_mb;
 268                                 break;
 269                         case 42:
 270                                 num_dpb_buffer = 34816 / fs_in_mb;
 271                                 break;
 272                         case 50:
 273                                 num_dpb_buffer = 110400 / fs_in_mb;
 274                                 break;
 275                         case 51:
 276                                 num_dpb_buffer = 184320 / fs_in_mb;
 277                                 break;
 278                         default:
 279                                 num_dpb_buffer = 184320 / fs_in_mb;
 280                                 break;
 281                         }
 282                         num_dpb_buffer++;
 283                         max_references = MAX2(MIN2(NUM_H264_REFS, num_dpb_buffer), max_references);
 284                         dpb_size = image_size * max_references;
 285                         dpb_size += max_references * align(width_in_mb * height_in_mb  * 192, alignment);
 286                         dpb_size += align(width_in_mb * height_in_mb * 32, alignment);
 287                 } else {
 288                         // the firmware seems to allways assume a minimum of ref frames
 289                         max_references = MAX2(NUM_H264_REFS, max_references);
 290                         // reference picture buffer
 291                         dpb_size = image_size * max_references;
 292                         // macroblock context buffer
 293                         dpb_size += width_in_mb * height_in_mb * max_references * 192;
 294                         // IT surface buffer
 295                         dpb_size += width_in_mb * height_in_mb * 32;
 296                 }
 297                 break;
 298         }
 299
 300         case PIPE_VIDEO_FORMAT_HEVC:
 301                 if (dec->base.width * dec->base.height >= 4096*2000)
 302                         max_references = MAX2(max_references, 8);
 303                 else
 304                         max_references = MAX2(max_references, 17);
 305
 306                 width = align (width, 16);
 307                 height = align (height, 16);
 308                 dpb_size = align((width * height * 3) / 2, 256) * max_references;
 309                 break;
 310
 311         case PIPE_VIDEO_FORMAT_VC1:
 312                 // the firmware seems to allways assume a minimum of ref frames
 313                 max_references = MAX2(NUM_VC1_REFS, max_references);
 314
 315                 // reference picture buffer
 316                 dpb_size = image_size * max_references;
 317
 318                 // CONTEXT_BUFFER
 319                 dpb_size += width_in_mb * height_in_mb * 128;
 320
 321                 // IT surface buffer
 322                 dpb_size += width_in_mb * 64;
 323
 324                 // DB surface buffer
 325                 dpb_size += width_in_mb * 128;
 326
 327                 // BP
 328                 dpb_size += align(MAX2(width_in_mb, height_in_mb) * 7 * 16, 64);
 329                 break;
 330
 331         case PIPE_VIDEO_FORMAT_MPEG12:
 332                 // reference picture buffer, must be big enough for all frames
 333                 dpb_size = image_size * NUM_MPEG2_REFS;
 334                 break;
 335
 336         case PIPE_VIDEO_FORMAT_MPEG4:
 337                 // reference picture buffer
 338                 dpb_size = image_size * max_references;
 339
 340                 // CM
 341                 dpb_size += width_in_mb * height_in_mb * 64;
 342
 343                 // IT surface buffer
 344                 dpb_size += align(width_in_mb * height_in_mb * 32, 64);
 345
 346                 dpb_size = MAX2(dpb_size, 30 * 1024 * 1024);
 347                 break;
 348
 349         default:
 350                 // something is missing here
 351                 assert(0);
 352
 353                 // at least use a sane default value
 354                 dpb_size = 32 * 1024 * 1024;
 355                 break;
 356         }
 357         return dpb_size;
 358 }
 359
 360 /* free associated data in the video buffer callback */
 361 static void ruvd_destroy_associated_data(void *data)
 362 {
 363         /* NOOP, since we only use an intptr */
 364 }
 365
 366 /* get h264 specific message bits */
 367 static struct ruvd_h264 get_h264_msg(struct ruvd_decoder *dec, struct pipe_h264_picture_desc *pic)
 368 {
 369         struct ruvd_h264 result;
 370
 371         memset(&result, 0, sizeof(result));
 372         switch (pic->base.profile) {
 373         case PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE:
 374                 result.profile = RUVD_H264_PROFILE_BASELINE;
 375                 break;
 376
 377         case PIPE_VIDEO_PROFILE_MPEG4_AVC_MAIN:
 378                 result.profile = RUVD_H264_PROFILE_MAIN;
 379                 break;
 380
 381         case PIPE_VIDEO_PROFILE_MPEG4_AVC_HIGH:
 382                 result.profile = RUVD_H264_PROFILE_HIGH;
 383                 break;
 384
 385         default:
 386                 assert(0);
 387                 break;
 388         }
 389
 390         result.level = dec->base.level;
 391
 392         result.sps_info_flags = 0;
 393         result.sps_info_flags |= pic->pps->sps->direct_8x8_inference_flag << 0;
 394         result.sps_info_flags |= pic->pps->sps->mb_adaptive_frame_field_flag << 1;
 395         result.sps_info_flags |= pic->pps->sps->frame_mbs_only_flag << 2;
 396         result.sps_info_flags |= pic->pps->sps->delta_pic_order_always_zero_flag << 3;
 397
 398         result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
 399         result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
 400         result.log2_max_frame_num_minus4 = pic->pps->sps->log2_max_frame_num_minus4;
 401         result.pic_order_cnt_type = pic->pps->sps->pic_order_cnt_type;
 402         result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
 403
 404         switch (dec->base.chroma_format) {
 405         case PIPE_VIDEO_CHROMA_FORMAT_400:
 406                 result.chroma_format = 0;
 407                 break;
 408         case PIPE_VIDEO_CHROMA_FORMAT_420:
 409                 result.chroma_format = 1;
 410                 break;
 411         case PIPE_VIDEO_CHROMA_FORMAT_422:
 412                 result.chroma_format = 2;
 413                 break;
 414         case PIPE_VIDEO_CHROMA_FORMAT_444:
 415                 result.chroma_format = 3;
 416                 break;
 417         }
 418
 419         result.pps_info_flags = 0;
 420         result.pps_info_flags |= pic->pps->transform_8x8_mode_flag << 0;
 421         result.pps_info_flags |= pic->pps->redundant_pic_cnt_present_flag << 1;
 422         result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 2;
 423         result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag << 3;
 424         result.pps_info_flags |= pic->pps->weighted_bipred_idc << 4;
 425         result.pps_info_flags |= pic->pps->weighted_pred_flag << 6;
 426         result.pps_info_flags |= pic->pps->bottom_field_pic_order_in_frame_present_flag << 7;
 427         result.pps_info_flags |= pic->pps->entropy_coding_mode_flag << 8;
 428
 429         result.num_slice_groups_minus1 = pic->pps->num_slice_groups_minus1;
 430         result.slice_group_map_type = pic->pps->slice_group_map_type;
 431         result.slice_group_change_rate_minus1 = pic->pps->slice_group_change_rate_minus1;
 432         result.pic_init_qp_minus26 = pic->pps->pic_init_qp_minus26;
 433         result.chroma_qp_index_offset = pic->pps->chroma_qp_index_offset;
 434         result.second_chroma_qp_index_offset = pic->pps->second_chroma_qp_index_offset;
 435
 436         memcpy(result.scaling_list_4x4, pic->pps->ScalingList4x4, 6*16);
 437         memcpy(result.scaling_list_8x8, pic->pps->ScalingList8x8, 2*64);
 438
 439         if (dec->stream_type == RUVD_CODEC_H264_PERF) {
 440                 memcpy(dec->it, result.scaling_list_4x4, 6*16);
 441                 memcpy((dec->it + 96), result.scaling_list_8x8, 2*64);
 442         }
 443
 444         result.num_ref_frames = pic->num_ref_frames;
 445
 446         result.num_ref_idx_l0_active_minus1 = pic->num_ref_idx_l0_active_minus1;
 447         result.num_ref_idx_l1_active_minus1 = pic->num_ref_idx_l1_active_minus1;
 448
 449         result.frame_num = pic->frame_num;
 450         memcpy(result.frame_num_list, pic->frame_num_list, 4*16);
 451         result.curr_field_order_cnt_list[0] = pic->field_order_cnt[0];
 452         result.curr_field_order_cnt_list[1] = pic->field_order_cnt[1];
 453         memcpy(result.field_order_cnt_list, pic->field_order_cnt_list, 4*16*2);
 454
 455         result.decoded_pic_idx = pic->frame_num;
 456
 457         return result;
 458 }
 459
 460 /* get h265 specific message bits */
 461 static struct ruvd_h265 get_h265_msg(struct ruvd_decoder *dec, struct pipe_video_buffer *target,
 462                                      struct pipe_h265_picture_desc *pic)
 463 {
 464         struct ruvd_h265 result;
 465         unsigned i;
 466
 467         memset(&result, 0, sizeof(result));
 468
 469         result.sps_info_flags = 0;
 470         result.sps_info_flags |= pic->pps->sps->scaling_list_enabled_flag << 0;
 471         result.sps_info_flags |= pic->pps->sps->amp_enabled_flag << 1;
 472         result.sps_info_flags |= pic->pps->sps->sample_adaptive_offset_enabled_flag << 2;
 473         result.sps_info_flags |= pic->pps->sps->pcm_enabled_flag << 3;
 474         result.sps_info_flags |= pic->pps->sps->pcm_loop_filter_disabled_flag << 4;
 475         result.sps_info_flags |= pic->pps->sps->long_term_ref_pics_present_flag << 5;
 476         result.sps_info_flags |= pic->pps->sps->sps_temporal_mvp_enabled_flag << 6;
 477         result.sps_info_flags |= pic->pps->sps->strong_intra_smoothing_enabled_flag << 7;
 478         result.sps_info_flags |= pic->pps->sps->separate_colour_plane_flag << 8;
 479         if (((struct r600_common_screen*)dec->screen)->family == CHIP_CARRIZO)
 480                 result.sps_info_flags |= 1 << 9;
 481         if (pic->UseRefPicList == true)
 482                 result.sps_info_flags |= 1 << 10;
 483
 484         result.chroma_format = pic->pps->sps->chroma_format_idc;
 485         result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
 486         result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
 487         result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
 488         result.sps_max_dec_pic_buffering_minus1 = pic->pps->sps->sps_max_dec_pic_buffering_minus1;
 489         result.log2_min_luma_coding_block_size_minus3 = pic->pps->sps->log2_min_luma_coding_block_size_minus3;
 490         result.log2_diff_max_min_luma_coding_block_size = pic->pps->sps->log2_diff_max_min_luma_coding_block_size;
 491         result.log2_min_transform_block_size_minus2 = pic->pps->sps->log2_min_transform_block_size_minus2;
 492         result.log2_diff_max_min_transform_block_size = pic->pps->sps->log2_diff_max_min_transform_block_size;
 493         result.max_transform_hierarchy_depth_inter = pic->pps->sps->max_transform_hierarchy_depth_inter;
 494         result.max_transform_hierarchy_depth_intra = pic->pps->sps->max_transform_hierarchy_depth_intra;
 495         result.pcm_sample_bit_depth_luma_minus1 = pic->pps->sps->pcm_sample_bit_depth_luma_minus1;
 496         result.pcm_sample_bit_depth_chroma_minus1 = pic->pps->sps->pcm_sample_bit_depth_chroma_minus1;
 497         result.log2_min_pcm_luma_coding_block_size_minus3 = pic->pps->sps->log2_min_pcm_luma_coding_block_size_minus3;
 498         result.log2_diff_max_min_pcm_luma_coding_block_size = pic->pps->sps->log2_diff_max_min_pcm_luma_coding_block_size;
 499         result.num_short_term_ref_pic_sets = pic->pps->sps->num_short_term_ref_pic_sets;
 500
 501         result.pps_info_flags = 0;
 502         result.pps_info_flags |= pic->pps->dependent_slice_segments_enabled_flag << 0;
 503         result.pps_info_flags |= pic->pps->output_flag_present_flag << 1;
 504         result.pps_info_flags |= pic->pps->sign_data_hiding_enabled_flag << 2;
 505         result.pps_info_flags |= pic->pps->cabac_init_present_flag << 3;
 506         result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 4;
 507         result.pps_info_flags |= pic->pps->transform_skip_enabled_flag << 5;
 508         result.pps_info_flags |= pic->pps->cu_qp_delta_enabled_flag << 6;
 509         result.pps_info_flags |= pic->pps->pps_slice_chroma_qp_offsets_present_flag << 7;
 510         result.pps_info_flags |= pic->pps->weighted_pred_flag << 8;
 511         result.pps_info_flags |= pic->pps->weighted_bipred_flag << 9;
 512         result.pps_info_flags |= pic->pps->transquant_bypass_enabled_flag << 10;
 513         result.pps_info_flags |= pic->pps->tiles_enabled_flag << 11;
 514         result.pps_info_flags |= pic->pps->entropy_coding_sync_enabled_flag << 12;
 515         result.pps_info_flags |= pic->pps->uniform_spacing_flag << 13;
 516         result.pps_info_flags |= pic->pps->loop_filter_across_tiles_enabled_flag << 14;
 517         result.pps_info_flags |= pic->pps->pps_loop_filter_across_slices_enabled_flag << 15;
 518         result.pps_info_flags |= pic->pps->deblocking_filter_override_enabled_flag << 16;
 519         result.pps_info_flags |= pic->pps->pps_deblocking_filter_disabled_flag << 17;
 520         result.pps_info_flags |= pic->pps->lists_modification_present_flag << 18;
 521         result.pps_info_flags |= pic->pps->slice_segment_header_extension_present_flag << 19;
 522         //result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag; ???
 523
 524         result.num_extra_slice_header_bits = pic->pps->num_extra_slice_header_bits;
 525         result.num_long_term_ref_pic_sps = pic->pps->sps->num_long_term_ref_pics_sps;
 526         result.num_ref_idx_l0_default_active_minus1 = pic->pps->num_ref_idx_l0_default_active_minus1;
 527         result.num_ref_idx_l1_default_active_minus1 = pic->pps->num_ref_idx_l1_default_active_minus1;
 528         result.pps_cb_qp_offset = pic->pps->pps_cb_qp_offset;
 529         result.pps_cr_qp_offset = pic->pps->pps_cr_qp_offset;
 530         result.pps_beta_offset_div2 = pic->pps->pps_beta_offset_div2;
 531         result.pps_tc_offset_div2 = pic->pps->pps_tc_offset_div2;
 532         result.diff_cu_qp_delta_depth = pic->pps->diff_cu_qp_delta_depth;
 533         result.num_tile_columns_minus1 = pic->pps->num_tile_columns_minus1;
 534         result.num_tile_rows_minus1 = pic->pps->num_tile_rows_minus1;
 535         result.log2_parallel_merge_level_minus2 = pic->pps->log2_parallel_merge_level_minus2;
 536         result.init_qp_minus26 = pic->pps->init_qp_minus26;
 537
 538         for (i = 0; i < 19; ++i)
 539                 result.column_width_minus1[i] = pic->pps->column_width_minus1[i];
 540
 541         for (i = 0; i < 21; ++i)
 542                 result.row_height_minus1[i] = pic->pps->row_height_minus1[i];
 543
 544         result.num_delta_pocs_ref_rps_idx = pic->NumDeltaPocsOfRefRpsIdx;
 545         result.curr_idx = pic->CurrPicOrderCntVal;
 546         result.curr_poc = pic->CurrPicOrderCntVal;
 547
 548         vl_video_buffer_set_associated_data(target, &dec->base,
 549                                             (void *)(uintptr_t)pic->CurrPicOrderCntVal,
 550                                             &ruvd_destroy_associated_data);
 551
 552         for (i = 0; i < 16; ++i) {
 553                 struct pipe_video_buffer *ref = pic->ref[i];
 554                 uintptr_t ref_pic = 0;
 555
 556                 result.poc_list[i] = pic->PicOrderCntVal[i];
 557
 558                 if (ref)
 559                         ref_pic = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
 560                 else
 561                         ref_pic = 0x7F;
 562                 result.ref_pic_list[i] = ref_pic;
 563         }
 564
 565         for (i = 0; i < 8; ++i) {
 566                 result.ref_pic_set_st_curr_before[i] = 0xFF;
 567                 result.ref_pic_set_st_curr_after[i] = 0xFF;
 568                 result.ref_pic_set_lt_curr[i] = 0xFF;
 569         }
 570
 571         for (i = 0; i < pic->NumPocStCurrBefore; ++i)
 572                 result.ref_pic_set_st_curr_before[i] = pic->RefPicSetStCurrBefore[i];
 573
 574         for (i = 0; i < pic->NumPocStCurrAfter; ++i)
 575                 result.ref_pic_set_st_curr_after[i] = pic->RefPicSetStCurrAfter[i];
 576
 577         for (i = 0; i < pic->NumPocLtCurr; ++i)
 578                 result.ref_pic_set_lt_curr[i] = pic->RefPicSetLtCurr[i];
 579
 580         for (i = 0; i < 6; ++i)
 581                 result.ucScalingListDCCoefSizeID2[i] = pic->pps->sps->ScalingListDCCoeff16x16[i];
 582
 583         for (i = 0; i < 2; ++i)
 584                 result.ucScalingListDCCoefSizeID3[i] = pic->pps->sps->ScalingListDCCoeff32x32[i];
 585
 586         memcpy(dec->it, pic->pps->sps->ScalingList4x4, 6 * 16);
 587         memcpy(dec->it + 96, pic->pps->sps->ScalingList8x8, 6 * 64);
 588         memcpy(dec->it + 480, pic->pps->sps->ScalingList16x16, 6 * 64);
 589         memcpy(dec->it + 864, pic->pps->sps->ScalingList32x32, 2 * 64);
 590
 591         for (i = 0 ; i < 2 ; i++) {
 592                 for (int j = 0 ; j < 15 ; j++)
 593                         result.direct_reflist[i][j] = pic->RefPicList[i][j];
 594         }
 595
 596         /* TODO
 597         result.highestTid;
 598         result.isNonRef;
 599
 600         IDRPicFlag;
 601         RAPPicFlag;
 602         NumPocTotalCurr;
 603         NumShortTermPictureSliceHeaderBits;
 604         NumLongTermPictureSliceHeaderBits;
 605
 606         IsLongTerm[16];
 607         */
 608
 609         return result;
 610 }
 611
 612 /* get vc1 specific message bits */
 613 static struct ruvd_vc1 get_vc1_msg(struct pipe_vc1_picture_desc *pic)
 614 {
 615         struct ruvd_vc1 result;
 616
 617         memset(&result, 0, sizeof(result));
 618
 619         switch(pic->base.profile) {
 620         case PIPE_VIDEO_PROFILE_VC1_SIMPLE:
 621                 result.profile = RUVD_VC1_PROFILE_SIMPLE;
 622                 result.level = 1;
 623                 break;
 624
 625         case PIPE_VIDEO_PROFILE_VC1_MAIN:
 626                 result.profile = RUVD_VC1_PROFILE_MAIN;
 627                 result.level = 2;
 628                 break;
 629
 630         case PIPE_VIDEO_PROFILE_VC1_ADVANCED:
 631                 result.profile = RUVD_VC1_PROFILE_ADVANCED;
 632                 result.level = 4;
 633                 break;
 634
 635         default:
 636                 assert(0);
 637         }
 638
 639         /* fields common for all profiles */
 640         result.sps_info_flags |= pic->postprocflag << 7;
 641         result.sps_info_flags |= pic->pulldown << 6;
 642         result.sps_info_flags |= pic->interlace << 5;
 643         result.sps_info_flags |= pic->tfcntrflag << 4;
 644         result.sps_info_flags |= pic->finterpflag << 3;
 645         result.sps_info_flags |= pic->psf << 1;
 646
 647         result.pps_info_flags |= pic->range_mapy_flag << 31;
 648         result.pps_info_flags |= pic->range_mapy << 28;
 649         result.pps_info_flags |= pic->range_mapuv_flag << 27;
 650         result.pps_info_flags |= pic->range_mapuv << 24;
 651         result.pps_info_flags |= pic->multires << 21;
 652         result.pps_info_flags |= pic->maxbframes << 16;
 653         result.pps_info_flags |= pic->overlap << 11;
 654         result.pps_info_flags |= pic->quantizer << 9;
 655         result.pps_info_flags |= pic->panscan_flag << 7;
 656         result.pps_info_flags |= pic->refdist_flag << 6;
 657         result.pps_info_flags |= pic->vstransform << 0;
 658
 659         /* some fields only apply to main/advanced profile */
 660         if (pic->base.profile != PIPE_VIDEO_PROFILE_VC1_SIMPLE) {
 661                 result.pps_info_flags |= pic->syncmarker << 20;
 662                 result.pps_info_flags |= pic->rangered << 19;
 663                 result.pps_info_flags |= pic->loopfilter << 5;
 664                 result.pps_info_flags |= pic->fastuvmc << 4;
 665                 result.pps_info_flags |= pic->extended_mv << 3;
 666                 result.pps_info_flags |= pic->extended_dmv << 8;
 667                 result.pps_info_flags |= pic->dquant << 1;
 668         }
 669
 670         result.chroma_format = 1;
 671
 672 #if 0
 673 //(((unsigned int)(pPicParams->advance.reserved1))        << SPS_INFO_VC1_RESERVED_SHIFT)
 674 uint32_t        slice_count
 675 uint8_t         picture_type
 676 uint8_t         frame_coding_mode
 677 uint8_t         deblockEnable
 678 uint8_t         pquant
 679 #endif
 680
 681         return result;
 682 }
 683
 684 /* extract the frame number from a referenced video buffer */
 685 static uint32_t get_ref_pic_idx(struct ruvd_decoder *dec, struct pipe_video_buffer *ref)
 686 {
 687         uint32_t min = MAX2(dec->frame_number, NUM_MPEG2_REFS) - NUM_MPEG2_REFS;
 688         uint32_t max = MAX2(dec->frame_number, 1) - 1;
 689         uintptr_t frame;
 690
 691         /* seems to be the most sane fallback */
 692         if (!ref)
 693                 return max;
 694
 695         /* get the frame number from the associated data */
 696         frame = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
 697
 698         /* limit the frame number to a valid range */
 699         return MAX2(MIN2(frame, max), min);
 700 }
 701
 702 /* get mpeg2 specific msg bits */
 703 static struct ruvd_mpeg2 get_mpeg2_msg(struct ruvd_decoder *dec,
 704                                        struct pipe_mpeg12_picture_desc *pic)
 705 {
 706         const int *zscan = pic->alternate_scan ? vl_zscan_alternate : vl_zscan_normal;
 707         struct ruvd_mpeg2 result;
 708         unsigned i;
 709
 710         memset(&result, 0, sizeof(result));
 711         result.decoded_pic_idx = dec->frame_number;
 712         for (i = 0; i < 2; ++i)
 713                 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
 714
 715         result.load_intra_quantiser_matrix = 1;
 716         result.load_nonintra_quantiser_matrix = 1;
 717
 718         for (i = 0; i < 64; ++i) {
 719                 result.intra_quantiser_matrix[i] = pic->intra_matrix[zscan[i]];
 720                 result.nonintra_quantiser_matrix[i] = pic->non_intra_matrix[zscan[i]];
 721         }
 722
 723         result.profile_and_level_indication = 0;
 724         result.chroma_format = 0x1;
 725
 726         result.picture_coding_type = pic->picture_coding_type;
 727         result.f_code[0][0] = pic->f_code[0][0] + 1;
 728         result.f_code[0][1] = pic->f_code[0][1] + 1;
 729         result.f_code[1][0] = pic->f_code[1][0] + 1;
 730         result.f_code[1][1] = pic->f_code[1][1] + 1;
 731         result.intra_dc_precision = pic->intra_dc_precision;
 732         result.pic_structure = pic->picture_structure;
 733         result.top_field_first = pic->top_field_first;
 734         result.frame_pred_frame_dct = pic->frame_pred_frame_dct;
 735         result.concealment_motion_vectors = pic->concealment_motion_vectors;
 736         result.q_scale_type = pic->q_scale_type;
 737         result.intra_vlc_format = pic->intra_vlc_format;
 738         result.alternate_scan = pic->alternate_scan;
 739
 740         return result;
 741 }
 742
 743 /* get mpeg4 specific msg bits */
 744 static struct ruvd_mpeg4 get_mpeg4_msg(struct ruvd_decoder *dec,
 745                                        struct pipe_mpeg4_picture_desc *pic)
 746 {
 747         struct ruvd_mpeg4 result;
 748         unsigned i;
 749
 750         memset(&result, 0, sizeof(result));
 751         result.decoded_pic_idx = dec->frame_number;
 752         for (i = 0; i < 2; ++i)
 753                 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
 754
 755         result.variant_type = 0;
 756         result.profile_and_level_indication = 0xF0; // ASP Level0
 757
 758         result.video_object_layer_verid = 0x5; // advanced simple
 759         result.video_object_layer_shape = 0x0; // rectangular
 760
 761         result.video_object_layer_width = dec->base.width;
 762         result.video_object_layer_height = dec->base.height;
 763
 764         result.vop_time_increment_resolution = pic->vop_time_increment_resolution;
 765
 766         result.flags |= pic->short_video_header << 0;
 767         //result.flags |= obmc_disable << 1;
 768         result.flags |= pic->interlaced << 2;
 769         result.flags |= 1 << 3; // load_intra_quant_mat
 770         result.flags |= 1 << 4; // load_nonintra_quant_mat
 771         result.flags |= pic->quarter_sample << 5;
 772         result.flags |= 1 << 6; // complexity_estimation_disable
 773         result.flags |= pic->resync_marker_disable << 7;
 774         //result.flags |= data_partitioned << 8;
 775         //result.flags |= reversible_vlc << 9;
 776         result.flags |= 0 << 10; // newpred_enable
 777         result.flags |= 0 << 11; // reduced_resolution_vop_enable
 778         //result.flags |= scalability << 12;
 779         //result.flags |= is_object_layer_identifier << 13;
 780         //result.flags |= fixed_vop_rate << 14;
 781         //result.flags |= newpred_segment_type << 15;
 782
 783         result.quant_type = pic->quant_type;
 784
 785         for (i = 0; i < 64; ++i) {
 786                 result.intra_quant_mat[i] = pic->intra_matrix[vl_zscan_normal[i]];
 787                 result.nonintra_quant_mat[i] = pic->non_intra_matrix[vl_zscan_normal[i]];
 788         }
 789
 790         /*
 791         int32_t         trd [2]
 792         int32_t         trb [2]
 793         uint8_t         vop_coding_type
 794         uint8_t         vop_fcode_forward
 795         uint8_t         vop_fcode_backward
 796         uint8_t         rounding_control
 797         uint8_t         alternate_vertical_scan_flag
 798         uint8_t         top_field_first
 799         */
 800
 801         return result;
 802 }
 803
 804 /**
 805  * destroy this video decoder
 806  */
 807 static void ruvd_destroy(struct pipe_video_codec *decoder)
 808 {
 809         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 810         unsigned i;
 811
 812         assert(decoder);
 813
 814         map_msg_fb_it_buf(dec);
 815         memset(dec->msg, 0, sizeof(*dec->msg));
 816         dec->msg->size = sizeof(*dec->msg);
 817         dec->msg->msg_type = RUVD_MSG_DESTROY;
 818         dec->msg->stream_handle = dec->stream_handle;
 819         send_msg_buf(dec);
 820
 821         flush(dec);
 822
 823         dec->ws->cs_destroy(dec->cs);
 824
 825         for (i = 0; i < NUM_BUFFERS; ++i) {
 826                 rvid_destroy_buffer(&dec->msg_fb_it_buffers[i]);
 827                 rvid_destroy_buffer(&dec->bs_buffers[i]);
 828         }
 829
 830         rvid_destroy_buffer(&dec->dpb);
 831         if (u_reduce_video_profile(dec->base.profile) == PIPE_VIDEO_FORMAT_HEVC)
 832                 rvid_destroy_buffer(&dec->ctx);
 833
 834         FREE(dec);
 835 }
 836
 837 /**
 838  * start decoding of a new frame
 839  */
 840 static void ruvd_begin_frame(struct pipe_video_codec *decoder,
 841                              struct pipe_video_buffer *target,
 842                              struct pipe_picture_desc *picture)
 843 {
 844         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 845         uintptr_t frame;
 846
 847         assert(decoder);
 848
 849         frame = ++dec->frame_number;
 850         vl_video_buffer_set_associated_data(target, decoder, (void *)frame,
 851                                             &ruvd_destroy_associated_data);
 852
 853         dec->bs_size = 0;
 854         dec->bs_ptr = dec->ws->buffer_map(
 855                 dec->bs_buffers[dec->cur_buffer].res->buf,
 856                 dec->cs, PIPE_TRANSFER_WRITE);
 857 }
 858
 859 /**
 860  * decode a macroblock
 861  */
 862 static void ruvd_decode_macroblock(struct pipe_video_codec *decoder,
 863                                    struct pipe_video_buffer *target,
 864                                    struct pipe_picture_desc *picture,
 865                                    const struct pipe_macroblock *macroblocks,
 866                                    unsigned num_macroblocks)
 867 {
 868         /* not supported (yet) */
 869         assert(0);
 870 }
 871
 872 /**
 873  * decode a bitstream
 874  */
 875 static void ruvd_decode_bitstream(struct pipe_video_codec *decoder,
 876                                   struct pipe_video_buffer *target,
 877                                   struct pipe_picture_desc *picture,
 878                                   unsigned num_buffers,
 879                                   const void * const *buffers,
 880                                   const unsigned *sizes)
 881 {
 882         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 883         unsigned i;
 884
 885         assert(decoder);
 886
 887         if (!dec->bs_ptr)
 888                 return;
 889
 890         for (i = 0; i < num_buffers; ++i) {
 891                 struct rvid_buffer *buf = &dec->bs_buffers[dec->cur_buffer];
 892                 unsigned new_size = dec->bs_size + sizes[i];
 893
 894                 if (new_size > buf->res->buf->size) {
 895                         dec->ws->buffer_unmap(buf->res->buf);
 896                         if (!rvid_resize_buffer(dec->screen, dec->cs, buf, new_size)) {
 897                                 RVID_ERR("Can't resize bitstream buffer!");
 898                                 return;
 899                         }
 900
 901                         dec->bs_ptr = dec->ws->buffer_map(buf->res->buf, dec->cs,
 902                                                           PIPE_TRANSFER_WRITE);
 903                         if (!dec->bs_ptr)
 904                                 return;
 905
 906                         dec->bs_ptr += dec->bs_size;
 907                 }
 908
 909                 memcpy(dec->bs_ptr, buffers[i], sizes[i]);
 910                 dec->bs_size += sizes[i];
 911                 dec->bs_ptr += sizes[i];
 912         }
 913 }
 914
 915 /**
 916  * end decoding of the current frame
 917  */
 918 static void ruvd_end_frame(struct pipe_video_codec *decoder,
 919                            struct pipe_video_buffer *target,
 920                            struct pipe_picture_desc *picture)
 921 {
 922         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 923         struct pb_buffer *dt;
 924         struct rvid_buffer *msg_fb_it_buf, *bs_buf;
 925         unsigned bs_size;
 926
 927         assert(decoder);
 928
 929         if (!dec->bs_ptr)
 930                 return;
 931
 932         msg_fb_it_buf = &dec->msg_fb_it_buffers[dec->cur_buffer];
 933         bs_buf = &dec->bs_buffers[dec->cur_buffer];
 934
 935         bs_size = align(dec->bs_size, 128);
 936         memset(dec->bs_ptr, 0, bs_size - dec->bs_size);
 937         dec->ws->buffer_unmap(bs_buf->res->buf);
 938
 939         map_msg_fb_it_buf(dec);
 940         dec->msg->size = sizeof(*dec->msg);
 941         dec->msg->msg_type = RUVD_MSG_DECODE;
 942         dec->msg->stream_handle = dec->stream_handle;
 943         dec->msg->status_report_feedback_number = dec->frame_number;
 944
 945         dec->msg->body.decode.stream_type = dec->stream_type;
 946         dec->msg->body.decode.decode_flags = 0x1;
 947         dec->msg->body.decode.width_in_samples = dec->base.width;
 948         dec->msg->body.decode.height_in_samples = dec->base.height;
 949
 950         if ((picture->profile == PIPE_VIDEO_PROFILE_VC1_SIMPLE) ||
 951             (picture->profile == PIPE_VIDEO_PROFILE_VC1_MAIN)) {
 952                 dec->msg->body.decode.width_in_samples = align(dec->msg->body.decode.width_in_samples, 16) / 16;
 953                 dec->msg->body.decode.height_in_samples = align(dec->msg->body.decode.height_in_samples, 16) / 16;
 954         }
 955
 956         dec->msg->body.decode.dpb_size = dec->dpb.res->buf->size;
 957         dec->msg->body.decode.bsd_size = bs_size;
 958         dec->msg->body.decode.db_pitch = dec->base.width;
 959
 960         dt = dec->set_dtb(dec->msg, (struct vl_video_buffer *)target);
 961         if (((struct r600_common_screen*)dec->screen)->family >= CHIP_STONEY)
 962                 dec->msg->body.decode.dt_wa_chroma_top_offset = dec->msg->body.decode.dt_pitch / 2;
 963
 964         switch (u_reduce_video_profile(picture->profile)) {
 965         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 966                 dec->msg->body.decode.codec.h264 = get_h264_msg(dec, (struct pipe_h264_picture_desc*)picture);
 967                 break;
 968
 969         case PIPE_VIDEO_FORMAT_HEVC:
 970                 dec->msg->body.decode.codec.h265 = get_h265_msg(dec, target, (struct pipe_h265_picture_desc*)picture);
 971                 break;
 972
 973         case PIPE_VIDEO_FORMAT_VC1:
 974                 dec->msg->body.decode.codec.vc1 = get_vc1_msg((struct pipe_vc1_picture_desc*)picture);
 975                 break;
 976
 977         case PIPE_VIDEO_FORMAT_MPEG12:
 978                 dec->msg->body.decode.codec.mpeg2 = get_mpeg2_msg(dec, (struct pipe_mpeg12_picture_desc*)picture);
 979                 break;
 980
 981         case PIPE_VIDEO_FORMAT_MPEG4:
 982                 dec->msg->body.decode.codec.mpeg4 = get_mpeg4_msg(dec, (struct pipe_mpeg4_picture_desc*)picture);
 983                 break;
 984
 985         default:
 986                 assert(0);
 987                 return;
 988         }
 989
 990         dec->msg->body.decode.db_surf_tile_config = dec->msg->body.decode.dt_surf_tile_config;
 991         dec->msg->body.decode.extension_support = 0x1;
 992
 993         /* set at least the feedback buffer size */
 994         dec->fb[0] = FB_BUFFER_SIZE;
 995
 996         send_msg_buf(dec);
 997
 998         send_cmd(dec, RUVD_CMD_DPB_BUFFER, dec->dpb.res->buf, 0,
 999                  RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
1000         if (u_reduce_video_profile(picture->profile) == PIPE_VIDEO_FORMAT_HEVC) {
1001                 send_cmd(dec, RUVD_CMD_CONTEXT_BUFFER, dec->ctx.res->buf, 0,
1002                         RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
1003         }
1004         send_cmd(dec, RUVD_CMD_BITSTREAM_BUFFER, bs_buf->res->buf,
1005                  0, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
1006         send_cmd(dec, RUVD_CMD_DECODING_TARGET_BUFFER, dt, 0,
1007                  RADEON_USAGE_WRITE, RADEON_DOMAIN_VRAM);
1008         send_cmd(dec, RUVD_CMD_FEEDBACK_BUFFER, msg_fb_it_buf->res->buf,
1009                  FB_BUFFER_OFFSET, RADEON_USAGE_WRITE, RADEON_DOMAIN_GTT);
1010         if (have_it(dec))
1011                 send_cmd(dec, RUVD_CMD_ITSCALING_TABLE_BUFFER, msg_fb_it_buf->res->buf,
1012                          FB_BUFFER_OFFSET + FB_BUFFER_SIZE, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
1013         set_reg(dec, RUVD_ENGINE_CNTL, 1);
1014
1015         flush(dec);
1016         next_buffer(dec);
1017 }
1018
1019 /**
1020  * flush any outstanding command buffers to the hardware
1021  */
1022 static void ruvd_flush(struct pipe_video_codec *decoder)
1023 {
1024 }
1025
1026 /**
1027  * create and UVD decoder
1028  */
1029 struct pipe_video_codec *ruvd_create_decoder(struct pipe_context *context,
1030                                              const struct pipe_video_codec *templ,
1031                                              ruvd_set_dtb set_dtb)
1032 {
1033         struct radeon_winsys* ws = ((struct r600_common_context *)context)->ws;
1034         struct r600_common_context *rctx = (struct r600_common_context*)context;
1035         unsigned dpb_size;
1036         unsigned width = templ->width, height = templ->height;
1037         unsigned bs_buf_size;
1038         struct radeon_info info;
1039         struct ruvd_decoder *dec;
1040         int i;
1041
1042         ws->query_info(ws, &info);
1043
1044         switch(u_reduce_video_profile(templ->profile)) {
1045         case PIPE_VIDEO_FORMAT_MPEG12:
1046                 if (templ->entrypoint > PIPE_VIDEO_ENTRYPOINT_BITSTREAM || info.family < CHIP_PALM)
1047                         return vl_create_mpeg12_decoder(context, templ);
1048
1049                 /* fall through */
1050         case PIPE_VIDEO_FORMAT_MPEG4:
1051         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
1052                 width = align(width, VL_MACROBLOCK_WIDTH);
1053                 height = align(height, VL_MACROBLOCK_HEIGHT);
1054                 break;
1055
1056         default:
1057                 break;
1058         }
1059
1060
1061         dec = CALLOC_STRUCT(ruvd_decoder);
1062
1063         if (!dec)
1064                 return NULL;
1065
1066         if (info.drm_major < 3)
1067                 dec->use_legacy = TRUE;
1068
1069         dec->base = *templ;
1070         dec->base.context = context;
1071         dec->base.width = width;
1072         dec->base.height = height;
1073
1074         dec->base.destroy = ruvd_destroy;
1075         dec->base.begin_frame = ruvd_begin_frame;
1076         dec->base.decode_macroblock = ruvd_decode_macroblock;
1077         dec->base.decode_bitstream = ruvd_decode_bitstream;
1078         dec->base.end_frame = ruvd_end_frame;
1079         dec->base.flush = ruvd_flush;
1080
1081         dec->stream_type = profile2stream_type(dec, info.family);
1082         dec->set_dtb = set_dtb;
1083         dec->stream_handle = rvid_alloc_stream_handle();
1084         dec->screen = context->screen;
1085         dec->ws = ws;
1086         dec->cs = ws->cs_create(rctx->ctx, RING_UVD, NULL, NULL, NULL);
1087         if (!dec->cs) {
1088                 RVID_ERR("Can't get command submission context.\n");
1089                 goto error;
1090         }
1091
1092         bs_buf_size = width * height * 512 / (16 * 16);
1093         for (i = 0; i < NUM_BUFFERS; ++i) {
1094                 unsigned msg_fb_it_size = FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
1095                 STATIC_ASSERT(sizeof(struct ruvd_msg) <= FB_BUFFER_OFFSET);
1096                 if (have_it(dec))
1097                         msg_fb_it_size += IT_SCALING_TABLE_SIZE;
1098                 if (!rvid_create_buffer(dec->screen, &dec->msg_fb_it_buffers[i],
1099                                         msg_fb_it_size, PIPE_USAGE_STAGING)) {
1100                         RVID_ERR("Can't allocated message buffers.\n");
1101                         goto error;
1102                 }
1103
1104                 if (!rvid_create_buffer(dec->screen, &dec->bs_buffers[i],
1105                                         bs_buf_size, PIPE_USAGE_STAGING)) {
1106                         RVID_ERR("Can't allocated bitstream buffers.\n");
1107                         goto error;
1108                 }
1109
1110                 rvid_clear_buffer(context, &dec->msg_fb_it_buffers[i]);
1111                 rvid_clear_buffer(context, &dec->bs_buffers[i]);
1112         }
1113
1114         dpb_size = calc_dpb_size(dec);
1115
1116         if (!rvid_create_buffer(dec->screen, &dec->dpb, dpb_size, PIPE_USAGE_DEFAULT)) {
1117                 RVID_ERR("Can't allocated dpb.\n");
1118                 goto error;
1119         }
1120
1121         rvid_clear_buffer(context, &dec->dpb);
1122
1123         if (u_reduce_video_profile(dec->base.profile) == PIPE_VIDEO_FORMAT_HEVC) {
1124                 unsigned ctx_size = calc_ctx_size(dec);
1125                 if (!rvid_create_buffer(dec->screen, &dec->ctx, ctx_size, PIPE_USAGE_DEFAULT)) {
1126                         RVID_ERR("Can't allocated context buffer.\n");
1127                         goto error;
1128                 }
1129                 rvid_clear_buffer(context, &dec->ctx);
1130         }
1131
1132         map_msg_fb_it_buf(dec);
1133         dec->msg->size = sizeof(*dec->msg);
1134         dec->msg->msg_type = RUVD_MSG_CREATE;
1135         dec->msg->stream_handle = dec->stream_handle;
1136         dec->msg->body.create.stream_type = dec->stream_type;
1137         dec->msg->body.create.width_in_samples = dec->base.width;
1138         dec->msg->body.create.height_in_samples = dec->base.height;
1139         dec->msg->body.create.dpb_size = dpb_size;
1140         send_msg_buf(dec);
1141         flush(dec);
1142         next_buffer(dec);
1143
1144         return &dec->base;
1145
1146 error:
1147         if (dec->cs) dec->ws->cs_destroy(dec->cs);
1148
1149         for (i = 0; i < NUM_BUFFERS; ++i) {
1150                 rvid_destroy_buffer(&dec->msg_fb_it_buffers[i]);
1151                 rvid_destroy_buffer(&dec->bs_buffers[i]);
1152         }
1153
1154         rvid_destroy_buffer(&dec->dpb);
1155         if (u_reduce_video_profile(dec->base.profile) == PIPE_VIDEO_FORMAT_HEVC)
1156                 rvid_destroy_buffer(&dec->ctx);
1157
1158         FREE(dec);
1159
1160         return NULL;
1161 }
1162
1163 /* calculate top/bottom offset */
1164 static unsigned texture_offset(struct radeon_surf *surface, unsigned layer)
1165 {
1166         return surface->level[0].offset +
1167                 layer * surface->level[0].slice_size;
1168 }
1169
1170 /* hw encode the aspect of macro tiles */
1171 static unsigned macro_tile_aspect(unsigned macro_tile_aspect)
1172 {
1173         switch (macro_tile_aspect) {
1174         default:
1175         case 1: macro_tile_aspect = 0;  break;
1176         case 2: macro_tile_aspect = 1;  break;
1177         case 4: macro_tile_aspect = 2;  break;
1178         case 8: macro_tile_aspect = 3;  break;
1179         }
1180         return macro_tile_aspect;
1181 }
1182
1183 /* hw encode the bank width and height */
1184 static unsigned bank_wh(unsigned bankwh)
1185 {
1186         switch (bankwh) {
1187         default:
1188         case 1: bankwh = 0;     break;
1189         case 2: bankwh = 1;     break;
1190         case 4: bankwh = 2;     break;
1191         case 8: bankwh = 3;     break;
1192         }
1193         return bankwh;
1194 }
1195
1196 /**
1197  * fill decoding target field from the luma and chroma surfaces
1198  */
1199 void ruvd_set_dt_surfaces(struct ruvd_msg *msg, struct radeon_surf *luma,
1200                           struct radeon_surf *chroma)
1201 {
1202         msg->body.decode.dt_pitch = luma->level[0].pitch_bytes;
1203         switch (luma->level[0].mode) {
1204         case RADEON_SURF_MODE_LINEAR_ALIGNED:
1205                 msg->body.decode.dt_tiling_mode = RUVD_TILE_LINEAR;
1206                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_LINEAR;
1207                 break;
1208         case RADEON_SURF_MODE_1D:
1209                 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
1210                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_1D_THIN;
1211                 break;
1212         case RADEON_SURF_MODE_2D:
1213                 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
1214                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_2D_THIN;
1215                 break;
1216         default:
1217                 assert(0);
1218                 break;
1219         }
1220
1221         msg->body.decode.dt_luma_top_offset = texture_offset(luma, 0);
1222         msg->body.decode.dt_chroma_top_offset = texture_offset(chroma, 0);
1223         if (msg->body.decode.dt_field_mode) {
1224                 msg->body.decode.dt_luma_bottom_offset = texture_offset(luma, 1);
1225                 msg->body.decode.dt_chroma_bottom_offset = texture_offset(chroma, 1);
1226         } else {
1227                 msg->body.decode.dt_luma_bottom_offset = msg->body.decode.dt_luma_top_offset;
1228                 msg->body.decode.dt_chroma_bottom_offset = msg->body.decode.dt_chroma_top_offset;
1229         }
1230
1231         assert(luma->bankw == chroma->bankw);
1232         assert(luma->bankh == chroma->bankh);
1233         assert(luma->mtilea == chroma->mtilea);
1234
1235         msg->body.decode.dt_surf_tile_config |= RUVD_BANK_WIDTH(bank_wh(luma->bankw));
1236         msg->body.decode.dt_surf_tile_config |= RUVD_BANK_HEIGHT(bank_wh(luma->bankh));
1237         msg->body.decode.dt_surf_tile_config |= RUVD_MACRO_TILE_ASPECT_RATIO(macro_tile_aspect(luma->mtilea));
1238 }