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 "radeon/drm/radeon_winsys.h"
  49 #include "r600_pipe_common.h"
  50 #include "radeon_video.h"
  51 #include "radeon_uvd.h"
  52
  53 #define NUM_BUFFERS 4
  54
  55 #define NUM_MPEG2_REFS 6
  56 #define NUM_H264_REFS 17
  57 #define NUM_VC1_REFS 5
  58
  59 #define FB_BUFFER_OFFSET 0x1000
  60 #define FB_BUFFER_SIZE 2048
  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                        frame_number;
  70
  71         struct pipe_screen              *screen;
  72         struct radeon_winsys*           ws;
  73         struct radeon_winsys_cs*        cs;
  74
  75         unsigned                        cur_buffer;
  76
  77         struct rvid_buffer              msg_fb_buffers[NUM_BUFFERS];
  78         struct ruvd_msg                 *msg;
  79         uint32_t                        *fb;
  80
  81         struct rvid_buffer              bs_buffers[NUM_BUFFERS];
  82         void*                           bs_ptr;
  83         unsigned                        bs_size;
  84
  85         struct rvid_buffer              dpb;
  86 };
  87
  88 /* flush IB to the hardware */
  89 static void flush(struct ruvd_decoder *dec)
  90 {
  91         dec->ws->cs_flush(dec->cs, RADEON_FLUSH_ASYNC, NULL, 0);
  92 }
  93
  94 /* add a new set register command to the IB */
  95 static void set_reg(struct ruvd_decoder *dec, unsigned reg, uint32_t val)
  96 {
  97         uint32_t *pm4 = dec->cs->buf;
  98         pm4[dec->cs->cdw++] = RUVD_PKT0(reg >> 2, 0);
  99         pm4[dec->cs->cdw++] = val;
 100 }
 101
 102 /* send a command to the VCPU through the GPCOM registers */
 103 static void send_cmd(struct ruvd_decoder *dec, unsigned cmd,
 104                      struct radeon_winsys_cs_handle* cs_buf, uint32_t off,
 105                      enum radeon_bo_usage usage, enum radeon_bo_domain domain)
 106 {
 107         int reloc_idx;
 108
 109         reloc_idx = dec->ws->cs_add_reloc(dec->cs, cs_buf, usage, domain,
 110                                           RADEON_PRIO_MIN);
 111         set_reg(dec, RUVD_GPCOM_VCPU_DATA0, off);
 112         set_reg(dec, RUVD_GPCOM_VCPU_DATA1, reloc_idx * 4);
 113         set_reg(dec, RUVD_GPCOM_VCPU_CMD, cmd << 1);
 114 }
 115
 116 /* map the next available message/feedback buffer */
 117 static void map_msg_fb_buf(struct ruvd_decoder *dec)
 118 {
 119         struct rvid_buffer* buf;
 120         uint8_t *ptr;
 121
 122         /* grab the current message/feedback buffer */
 123         buf = &dec->msg_fb_buffers[dec->cur_buffer];
 124
 125         /* and map it for CPU access */
 126         ptr = dec->ws->buffer_map(buf->res->cs_buf, dec->cs, PIPE_TRANSFER_WRITE);
 127
 128         /* calc buffer offsets */
 129         dec->msg = (struct ruvd_msg *)ptr;
 130         dec->fb = (uint32_t *)(ptr + FB_BUFFER_OFFSET);
 131 }
 132
 133 /* unmap and send a message command to the VCPU */
 134 static void send_msg_buf(struct ruvd_decoder *dec)
 135 {
 136         struct rvid_buffer* buf;
 137
 138         /* ignore the request if message/feedback buffer isn't mapped */
 139         if (!dec->msg || !dec->fb)
 140                 return;
 141
 142         /* grab the current message buffer */
 143         buf = &dec->msg_fb_buffers[dec->cur_buffer];
 144
 145         /* unmap the buffer */
 146         dec->ws->buffer_unmap(buf->res->cs_buf);
 147         dec->msg = NULL;
 148         dec->fb = NULL;
 149
 150         /* and send it to the hardware */
 151         send_cmd(dec, RUVD_CMD_MSG_BUFFER, buf->res->cs_buf, 0,
 152                  RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
 153 }
 154
 155 /* cycle to the next set of buffers */
 156 static void next_buffer(struct ruvd_decoder *dec)
 157 {
 158         ++dec->cur_buffer;
 159         dec->cur_buffer %= NUM_BUFFERS;
 160 }
 161
 162 /* convert the profile into something UVD understands */
 163 static uint32_t profile2stream_type(enum pipe_video_profile profile)
 164 {
 165         switch (u_reduce_video_profile(profile)) {
 166         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 167                 return RUVD_CODEC_H264;
 168
 169         case PIPE_VIDEO_FORMAT_VC1:
 170                 return RUVD_CODEC_VC1;
 171
 172         case PIPE_VIDEO_FORMAT_MPEG12:
 173                 return RUVD_CODEC_MPEG2;
 174
 175         case PIPE_VIDEO_FORMAT_MPEG4:
 176                 return RUVD_CODEC_MPEG4;
 177
 178         default:
 179                 assert(0);
 180                 return 0;
 181         }
 182 }
 183
 184 /* calculate size of reference picture buffer */
 185 static unsigned calc_dpb_size(const struct pipe_video_codec *templ)
 186 {
 187         unsigned width_in_mb, height_in_mb, image_size, dpb_size;
 188
 189         // always align them to MB size for dpb calculation
 190         unsigned width = align(templ->width, VL_MACROBLOCK_WIDTH);
 191         unsigned height = align(templ->height, VL_MACROBLOCK_HEIGHT);
 192
 193         // always one more for currently decoded picture
 194         unsigned max_references = templ->max_references + 1;
 195
 196         // aligned size of a single frame
 197         image_size = width * height;
 198         image_size += image_size / 2;
 199         image_size = align(image_size, 1024);
 200
 201         // picture width & height in 16 pixel units
 202         width_in_mb = width / VL_MACROBLOCK_WIDTH;
 203         height_in_mb = align(height / VL_MACROBLOCK_HEIGHT, 2);
 204
 205         switch (u_reduce_video_profile(templ->profile)) {
 206         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 207                 // the firmware seems to allways assume a minimum of ref frames
 208                 max_references = MAX2(NUM_H264_REFS, max_references);
 209
 210                 // reference picture buffer
 211                 dpb_size = image_size * max_references;
 212
 213                 // macroblock context buffer
 214                 dpb_size += width_in_mb * height_in_mb * max_references * 192;
 215
 216                 // IT surface buffer
 217                 dpb_size += width_in_mb * height_in_mb * 32;
 218                 break;
 219
 220         case PIPE_VIDEO_FORMAT_VC1:
 221                 // the firmware seems to allways assume a minimum of ref frames
 222                 max_references = MAX2(NUM_VC1_REFS, max_references);
 223
 224                 // reference picture buffer
 225                 dpb_size = image_size * max_references;
 226
 227                 // CONTEXT_BUFFER
 228                 dpb_size += width_in_mb * height_in_mb * 128;
 229
 230                 // IT surface buffer
 231                 dpb_size += width_in_mb * 64;
 232
 233                 // DB surface buffer
 234                 dpb_size += width_in_mb * 128;
 235
 236                 // BP
 237                 dpb_size += align(MAX2(width_in_mb, height_in_mb) * 7 * 16, 64);
 238                 break;
 239
 240         case PIPE_VIDEO_FORMAT_MPEG12:
 241                 // reference picture buffer, must be big enough for all frames
 242                 dpb_size = image_size * NUM_MPEG2_REFS;
 243                 break;
 244
 245         case PIPE_VIDEO_FORMAT_MPEG4:
 246                 // reference picture buffer
 247                 dpb_size = image_size * max_references;
 248
 249                 // CM
 250                 dpb_size += width_in_mb * height_in_mb * 64;
 251
 252                 // IT surface buffer
 253                 dpb_size += align(width_in_mb * height_in_mb * 32, 64);
 254                 break;
 255
 256         default:
 257                 // something is missing here
 258                 assert(0);
 259
 260                 // at least use a sane default value
 261                 dpb_size = 32 * 1024 * 1024;
 262                 break;
 263         }
 264         return dpb_size;
 265 }
 266
 267 /* get h264 specific message bits */
 268 static struct ruvd_h264 get_h264_msg(struct ruvd_decoder *dec, struct pipe_h264_picture_desc *pic)
 269 {
 270         struct ruvd_h264 result;
 271
 272         memset(&result, 0, sizeof(result));
 273         switch (pic->base.profile) {
 274         case PIPE_VIDEO_PROFILE_MPEG4_AVC_BASELINE:
 275                 result.profile = RUVD_H264_PROFILE_BASELINE;
 276                 break;
 277
 278         case PIPE_VIDEO_PROFILE_MPEG4_AVC_MAIN:
 279                 result.profile = RUVD_H264_PROFILE_MAIN;
 280                 break;
 281
 282         case PIPE_VIDEO_PROFILE_MPEG4_AVC_HIGH:
 283                 result.profile = RUVD_H264_PROFILE_HIGH;
 284                 break;
 285
 286         default:
 287                 assert(0);
 288                 break;
 289         }
 290         if (((dec->base.width * dec->base.height) >> 8) <= 1620)
 291                 result.level = 30;
 292         else
 293                 result.level = 41;
 294
 295         result.sps_info_flags = 0;
 296         result.sps_info_flags |= pic->pps->sps->direct_8x8_inference_flag << 0;
 297         result.sps_info_flags |= pic->pps->sps->mb_adaptive_frame_field_flag << 1;
 298         result.sps_info_flags |= pic->pps->sps->frame_mbs_only_flag << 2;
 299         result.sps_info_flags |= pic->pps->sps->delta_pic_order_always_zero_flag << 3;
 300
 301         result.bit_depth_luma_minus8 = pic->pps->sps->bit_depth_luma_minus8;
 302         result.bit_depth_chroma_minus8 = pic->pps->sps->bit_depth_chroma_minus8;
 303         result.log2_max_frame_num_minus4 = pic->pps->sps->log2_max_frame_num_minus4;
 304         result.pic_order_cnt_type = pic->pps->sps->pic_order_cnt_type;
 305         result.log2_max_pic_order_cnt_lsb_minus4 = pic->pps->sps->log2_max_pic_order_cnt_lsb_minus4;
 306
 307         switch (dec->base.chroma_format) {
 308         case PIPE_VIDEO_CHROMA_FORMAT_400:
 309                 result.chroma_format = 0;
 310                 break;
 311         case PIPE_VIDEO_CHROMA_FORMAT_420:
 312                 result.chroma_format = 1;
 313                 break;
 314         case PIPE_VIDEO_CHROMA_FORMAT_422:
 315                 result.chroma_format = 2;
 316                 break;
 317         case PIPE_VIDEO_CHROMA_FORMAT_444:
 318                 result.chroma_format = 3;
 319                 break;
 320         }
 321
 322         result.pps_info_flags = 0;
 323         result.pps_info_flags |= pic->pps->transform_8x8_mode_flag << 0;
 324         result.pps_info_flags |= pic->pps->redundant_pic_cnt_present_flag << 1;
 325         result.pps_info_flags |= pic->pps->constrained_intra_pred_flag << 2;
 326         result.pps_info_flags |= pic->pps->deblocking_filter_control_present_flag << 3;
 327         result.pps_info_flags |= pic->pps->weighted_bipred_idc << 4;
 328         result.pps_info_flags |= pic->pps->weighted_pred_flag << 6;
 329         result.pps_info_flags |= pic->pps->bottom_field_pic_order_in_frame_present_flag << 7;
 330         result.pps_info_flags |= pic->pps->entropy_coding_mode_flag << 8;
 331
 332         result.num_slice_groups_minus1 = pic->pps->num_slice_groups_minus1;
 333         result.slice_group_map_type = pic->pps->slice_group_map_type;
 334         result.slice_group_change_rate_minus1 = pic->pps->slice_group_change_rate_minus1;
 335         result.pic_init_qp_minus26 = pic->pps->pic_init_qp_minus26;
 336         result.chroma_qp_index_offset = pic->pps->chroma_qp_index_offset;
 337         result.second_chroma_qp_index_offset = pic->pps->second_chroma_qp_index_offset;
 338
 339         memcpy(result.scaling_list_4x4, pic->pps->ScalingList4x4, 6*16);
 340         memcpy(result.scaling_list_8x8, pic->pps->ScalingList8x8, 2*64);
 341
 342         result.num_ref_frames = pic->num_ref_frames;
 343
 344         result.num_ref_idx_l0_active_minus1 = pic->num_ref_idx_l0_active_minus1;
 345         result.num_ref_idx_l1_active_minus1 = pic->num_ref_idx_l1_active_minus1;
 346
 347         result.frame_num = pic->frame_num;
 348         memcpy(result.frame_num_list, pic->frame_num_list, 4*16);
 349         result.curr_field_order_cnt_list[0] = pic->field_order_cnt[0];
 350         result.curr_field_order_cnt_list[1] = pic->field_order_cnt[1];
 351         memcpy(result.field_order_cnt_list, pic->field_order_cnt_list, 4*16*2);
 352
 353         result.decoded_pic_idx = pic->frame_num;
 354
 355         return result;
 356 }
 357
 358 /* get vc1 specific message bits */
 359 static struct ruvd_vc1 get_vc1_msg(struct pipe_vc1_picture_desc *pic)
 360 {
 361         struct ruvd_vc1 result;
 362
 363         memset(&result, 0, sizeof(result));
 364
 365         switch(pic->base.profile) {
 366         case PIPE_VIDEO_PROFILE_VC1_SIMPLE:
 367                 result.profile = RUVD_VC1_PROFILE_SIMPLE;
 368                 result.level = 1;
 369                 break;
 370
 371         case PIPE_VIDEO_PROFILE_VC1_MAIN:
 372                 result.profile = RUVD_VC1_PROFILE_MAIN;
 373                 result.level = 2;
 374                 break;
 375
 376         case PIPE_VIDEO_PROFILE_VC1_ADVANCED:
 377                 result.profile = RUVD_VC1_PROFILE_ADVANCED;
 378                 result.level = 4;
 379                 break;
 380
 381         default:
 382                 assert(0);
 383         }
 384
 385         /* fields common for all profiles */
 386         result.sps_info_flags |= pic->postprocflag << 7;
 387         result.sps_info_flags |= pic->pulldown << 6;
 388         result.sps_info_flags |= pic->interlace << 5;
 389         result.sps_info_flags |= pic->tfcntrflag << 4;
 390         result.sps_info_flags |= pic->finterpflag << 3;
 391         result.sps_info_flags |= pic->psf << 1;
 392
 393         result.pps_info_flags |= pic->range_mapy_flag << 31;
 394         result.pps_info_flags |= pic->range_mapy << 28;
 395         result.pps_info_flags |= pic->range_mapuv_flag << 27;
 396         result.pps_info_flags |= pic->range_mapuv << 24;
 397         result.pps_info_flags |= pic->multires << 21;
 398         result.pps_info_flags |= pic->maxbframes << 16;
 399         result.pps_info_flags |= pic->overlap << 11;
 400         result.pps_info_flags |= pic->quantizer << 9;
 401         result.pps_info_flags |= pic->panscan_flag << 7;
 402         result.pps_info_flags |= pic->refdist_flag << 6;
 403         result.pps_info_flags |= pic->vstransform << 0;
 404
 405         /* some fields only apply to main/advanced profile */
 406         if (pic->base.profile != PIPE_VIDEO_PROFILE_VC1_SIMPLE) {
 407                 result.pps_info_flags |= pic->syncmarker << 20;
 408                 result.pps_info_flags |= pic->rangered << 19;
 409                 result.pps_info_flags |= pic->loopfilter << 5;
 410                 result.pps_info_flags |= pic->fastuvmc << 4;
 411                 result.pps_info_flags |= pic->extended_mv << 3;
 412                 result.pps_info_flags |= pic->extended_dmv << 8;
 413                 result.pps_info_flags |= pic->dquant << 1;
 414         }
 415
 416         result.chroma_format = 1;
 417
 418 #if 0
 419 //(((unsigned int)(pPicParams->advance.reserved1))        << SPS_INFO_VC1_RESERVED_SHIFT)
 420 uint32_t        slice_count
 421 uint8_t         picture_type
 422 uint8_t         frame_coding_mode
 423 uint8_t         deblockEnable
 424 uint8_t         pquant
 425 #endif
 426
 427         return result;
 428 }
 429
 430 /* extract the frame number from a referenced video buffer */
 431 static uint32_t get_ref_pic_idx(struct ruvd_decoder *dec, struct pipe_video_buffer *ref)
 432 {
 433         uint32_t min = MAX2(dec->frame_number, NUM_MPEG2_REFS) - NUM_MPEG2_REFS;
 434         uint32_t max = MAX2(dec->frame_number, 1) - 1;
 435         uintptr_t frame;
 436
 437         /* seems to be the most sane fallback */
 438         if (!ref)
 439                 return max;
 440
 441         /* get the frame number from the associated data */
 442         frame = (uintptr_t)vl_video_buffer_get_associated_data(ref, &dec->base);
 443
 444         /* limit the frame number to a valid range */
 445         return MAX2(MIN2(frame, max), min);
 446 }
 447
 448 /* get mpeg2 specific msg bits */
 449 static struct ruvd_mpeg2 get_mpeg2_msg(struct ruvd_decoder *dec,
 450                                        struct pipe_mpeg12_picture_desc *pic)
 451 {
 452         const int *zscan = pic->alternate_scan ? vl_zscan_alternate : vl_zscan_normal;
 453         struct ruvd_mpeg2 result;
 454         unsigned i;
 455
 456         memset(&result, 0, sizeof(result));
 457         result.decoded_pic_idx = dec->frame_number;
 458         for (i = 0; i < 2; ++i)
 459                 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
 460
 461         result.load_intra_quantiser_matrix = 1;
 462         result.load_nonintra_quantiser_matrix = 1;
 463
 464         for (i = 0; i < 64; ++i) {
 465                 result.intra_quantiser_matrix[i] = pic->intra_matrix[zscan[i]];
 466                 result.nonintra_quantiser_matrix[i] = pic->non_intra_matrix[zscan[i]];
 467         }
 468
 469         result.profile_and_level_indication = 0;
 470         result.chroma_format = 0x1;
 471
 472         result.picture_coding_type = pic->picture_coding_type;
 473         result.f_code[0][0] = pic->f_code[0][0] + 1;
 474         result.f_code[0][1] = pic->f_code[0][1] + 1;
 475         result.f_code[1][0] = pic->f_code[1][0] + 1;
 476         result.f_code[1][1] = pic->f_code[1][1] + 1;
 477         result.intra_dc_precision = pic->intra_dc_precision;
 478         result.pic_structure = pic->picture_structure;
 479         result.top_field_first = pic->top_field_first;
 480         result.frame_pred_frame_dct = pic->frame_pred_frame_dct;
 481         result.concealment_motion_vectors = pic->concealment_motion_vectors;
 482         result.q_scale_type = pic->q_scale_type;
 483         result.intra_vlc_format = pic->intra_vlc_format;
 484         result.alternate_scan = pic->alternate_scan;
 485
 486         return result;
 487 }
 488
 489 /* get mpeg4 specific msg bits */
 490 static struct ruvd_mpeg4 get_mpeg4_msg(struct ruvd_decoder *dec,
 491                                        struct pipe_mpeg4_picture_desc *pic)
 492 {
 493         struct ruvd_mpeg4 result;
 494         unsigned i;
 495
 496         memset(&result, 0, sizeof(result));
 497         result.decoded_pic_idx = dec->frame_number;
 498         for (i = 0; i < 2; ++i)
 499                 result.ref_pic_idx[i] = get_ref_pic_idx(dec, pic->ref[i]);
 500
 501         result.variant_type = 0;
 502         result.profile_and_level_indication = 0xF0; // ASP Level0
 503
 504         result.video_object_layer_verid = 0x5; // advanced simple
 505         result.video_object_layer_shape = 0x0; // rectangular
 506
 507         result.video_object_layer_width = dec->base.width;
 508         result.video_object_layer_height = dec->base.height;
 509
 510         result.vop_time_increment_resolution = pic->vop_time_increment_resolution;
 511
 512         result.flags |= pic->short_video_header << 0;
 513         //result.flags |= obmc_disable << 1;
 514         result.flags |= pic->interlaced << 2;
 515         result.flags |= 1 << 3; // load_intra_quant_mat
 516         result.flags |= 1 << 4; // load_nonintra_quant_mat
 517         result.flags |= pic->quarter_sample << 5;
 518         result.flags |= 1 << 6; // complexity_estimation_disable
 519         result.flags |= pic->resync_marker_disable << 7;
 520         //result.flags |= data_partitioned << 8;
 521         //result.flags |= reversible_vlc << 9;
 522         result.flags |= 0 << 10; // newpred_enable
 523         result.flags |= 0 << 11; // reduced_resolution_vop_enable
 524         //result.flags |= scalability << 12;
 525         //result.flags |= is_object_layer_identifier << 13;
 526         //result.flags |= fixed_vop_rate << 14;
 527         //result.flags |= newpred_segment_type << 15;
 528
 529         result.quant_type = pic->quant_type;
 530
 531         for (i = 0; i < 64; ++i) {
 532                 result.intra_quant_mat[i] = pic->intra_matrix[vl_zscan_normal[i]];
 533                 result.nonintra_quant_mat[i] = pic->non_intra_matrix[vl_zscan_normal[i]];
 534         }
 535
 536         /*
 537         int32_t         trd [2]
 538         int32_t         trb [2]
 539         uint8_t         vop_coding_type
 540         uint8_t         vop_fcode_forward
 541         uint8_t         vop_fcode_backward
 542         uint8_t         rounding_control
 543         uint8_t         alternate_vertical_scan_flag
 544         uint8_t         top_field_first
 545         */
 546
 547         return result;
 548 }
 549
 550 /**
 551  * destroy this video decoder
 552  */
 553 static void ruvd_destroy(struct pipe_video_codec *decoder)
 554 {
 555         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 556         unsigned i;
 557
 558         assert(decoder);
 559
 560         map_msg_fb_buf(dec);
 561         memset(dec->msg, 0, sizeof(*dec->msg));
 562         dec->msg->size = sizeof(*dec->msg);
 563         dec->msg->msg_type = RUVD_MSG_DESTROY;
 564         dec->msg->stream_handle = dec->stream_handle;
 565         send_msg_buf(dec);
 566
 567         flush(dec);
 568
 569         dec->ws->cs_destroy(dec->cs);
 570
 571         for (i = 0; i < NUM_BUFFERS; ++i) {
 572                 rvid_destroy_buffer(&dec->msg_fb_buffers[i]);
 573                 rvid_destroy_buffer(&dec->bs_buffers[i]);
 574         }
 575
 576         rvid_destroy_buffer(&dec->dpb);
 577
 578         FREE(dec);
 579 }
 580
 581 /* free associated data in the video buffer callback */
 582 static void ruvd_destroy_associated_data(void *data)
 583 {
 584         /* NOOP, since we only use an intptr */
 585 }
 586
 587 /**
 588  * start decoding of a new frame
 589  */
 590 static void ruvd_begin_frame(struct pipe_video_codec *decoder,
 591                              struct pipe_video_buffer *target,
 592                              struct pipe_picture_desc *picture)
 593 {
 594         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 595         uintptr_t frame;
 596
 597         assert(decoder);
 598
 599         frame = ++dec->frame_number;
 600         vl_video_buffer_set_associated_data(target, decoder, (void *)frame,
 601                                             &ruvd_destroy_associated_data);
 602
 603         dec->bs_size = 0;
 604         dec->bs_ptr = dec->ws->buffer_map(
 605                 dec->bs_buffers[dec->cur_buffer].res->cs_buf,
 606                 dec->cs, PIPE_TRANSFER_WRITE);
 607 }
 608
 609 /**
 610  * decode a macroblock
 611  */
 612 static void ruvd_decode_macroblock(struct pipe_video_codec *decoder,
 613                                    struct pipe_video_buffer *target,
 614                                    struct pipe_picture_desc *picture,
 615                                    const struct pipe_macroblock *macroblocks,
 616                                    unsigned num_macroblocks)
 617 {
 618         /* not supported (yet) */
 619         assert(0);
 620 }
 621
 622 /**
 623  * decode a bitstream
 624  */
 625 static void ruvd_decode_bitstream(struct pipe_video_codec *decoder,
 626                                   struct pipe_video_buffer *target,
 627                                   struct pipe_picture_desc *picture,
 628                                   unsigned num_buffers,
 629                                   const void * const *buffers,
 630                                   const unsigned *sizes)
 631 {
 632         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 633         unsigned i;
 634
 635         assert(decoder);
 636
 637         if (!dec->bs_ptr)
 638                 return;
 639
 640         for (i = 0; i < num_buffers; ++i) {
 641                 struct rvid_buffer *buf = &dec->bs_buffers[dec->cur_buffer];
 642                 unsigned new_size = dec->bs_size + sizes[i];
 643
 644                 if (new_size > buf->res->buf->size) {
 645                         dec->ws->buffer_unmap(buf->res->cs_buf);
 646                         if (!rvid_resize_buffer(dec->screen, dec->cs, buf, new_size)) {
 647                                 RVID_ERR("Can't resize bitstream buffer!");
 648                                 return;
 649                         }
 650
 651                         dec->bs_ptr = dec->ws->buffer_map(buf->res->cs_buf, dec->cs,
 652                                                           PIPE_TRANSFER_WRITE);
 653                         if (!dec->bs_ptr)
 654                                 return;
 655
 656                         dec->bs_ptr += dec->bs_size;
 657                 }
 658
 659                 memcpy(dec->bs_ptr, buffers[i], sizes[i]);
 660                 dec->bs_size += sizes[i];
 661                 dec->bs_ptr += sizes[i];
 662         }
 663 }
 664
 665 /**
 666  * end decoding of the current frame
 667  */
 668 static void ruvd_end_frame(struct pipe_video_codec *decoder,
 669                            struct pipe_video_buffer *target,
 670                            struct pipe_picture_desc *picture)
 671 {
 672         struct ruvd_decoder *dec = (struct ruvd_decoder*)decoder;
 673         struct radeon_winsys_cs_handle *dt;
 674         struct rvid_buffer *msg_fb_buf, *bs_buf;
 675         unsigned bs_size;
 676
 677         assert(decoder);
 678
 679         if (!dec->bs_ptr)
 680                 return;
 681
 682         msg_fb_buf = &dec->msg_fb_buffers[dec->cur_buffer];
 683         bs_buf = &dec->bs_buffers[dec->cur_buffer];
 684
 685         bs_size = align(dec->bs_size, 128);
 686         memset(dec->bs_ptr, 0, bs_size - dec->bs_size);
 687         dec->ws->buffer_unmap(bs_buf->res->cs_buf);
 688
 689         map_msg_fb_buf(dec);
 690         dec->msg->size = sizeof(*dec->msg);
 691         dec->msg->msg_type = RUVD_MSG_DECODE;
 692         dec->msg->stream_handle = dec->stream_handle;
 693         dec->msg->status_report_feedback_number = dec->frame_number;
 694
 695         dec->msg->body.decode.stream_type = profile2stream_type(dec->base.profile);
 696         dec->msg->body.decode.decode_flags = 0x1;
 697         dec->msg->body.decode.width_in_samples = dec->base.width;
 698         dec->msg->body.decode.height_in_samples = dec->base.height;
 699
 700         dec->msg->body.decode.dpb_size = dec->dpb.res->buf->size;
 701         dec->msg->body.decode.bsd_size = bs_size;
 702
 703         dt = dec->set_dtb(dec->msg, (struct vl_video_buffer *)target);
 704
 705         switch (u_reduce_video_profile(picture->profile)) {
 706         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 707                 dec->msg->body.decode.codec.h264 = get_h264_msg(dec, (struct pipe_h264_picture_desc*)picture);
 708                 break;
 709
 710         case PIPE_VIDEO_FORMAT_VC1:
 711                 dec->msg->body.decode.codec.vc1 = get_vc1_msg((struct pipe_vc1_picture_desc*)picture);
 712                 break;
 713
 714         case PIPE_VIDEO_FORMAT_MPEG12:
 715                 dec->msg->body.decode.codec.mpeg2 = get_mpeg2_msg(dec, (struct pipe_mpeg12_picture_desc*)picture);
 716                 break;
 717
 718         case PIPE_VIDEO_FORMAT_MPEG4:
 719                 dec->msg->body.decode.codec.mpeg4 = get_mpeg4_msg(dec, (struct pipe_mpeg4_picture_desc*)picture);
 720                 break;
 721
 722         default:
 723                 assert(0);
 724                 return;
 725         }
 726
 727         dec->msg->body.decode.db_surf_tile_config = dec->msg->body.decode.dt_surf_tile_config;
 728         dec->msg->body.decode.extension_support = 0x1;
 729
 730         /* set at least the feedback buffer size */
 731         dec->fb[0] = FB_BUFFER_SIZE;
 732
 733         send_msg_buf(dec);
 734
 735         send_cmd(dec, RUVD_CMD_DPB_BUFFER, dec->dpb.res->cs_buf, 0,
 736                  RADEON_USAGE_READWRITE, RADEON_DOMAIN_VRAM);
 737         send_cmd(dec, RUVD_CMD_BITSTREAM_BUFFER, bs_buf->res->cs_buf,
 738                  0, RADEON_USAGE_READ, RADEON_DOMAIN_GTT);
 739         send_cmd(dec, RUVD_CMD_DECODING_TARGET_BUFFER, dt, 0,
 740                  RADEON_USAGE_WRITE, RADEON_DOMAIN_VRAM);
 741         send_cmd(dec, RUVD_CMD_FEEDBACK_BUFFER, msg_fb_buf->res->cs_buf,
 742                  FB_BUFFER_OFFSET, RADEON_USAGE_WRITE, RADEON_DOMAIN_GTT);
 743         set_reg(dec, RUVD_ENGINE_CNTL, 1);
 744
 745         flush(dec);
 746         next_buffer(dec);
 747 }
 748
 749 /**
 750  * flush any outstanding command buffers to the hardware
 751  */
 752 static void ruvd_flush(struct pipe_video_codec *decoder)
 753 {
 754 }
 755
 756 /**
 757  * create and UVD decoder
 758  */
 759 struct pipe_video_codec *ruvd_create_decoder(struct pipe_context *context,
 760                                              const struct pipe_video_codec *templ,
 761                                              ruvd_set_dtb set_dtb)
 762 {
 763         struct radeon_winsys* ws = ((struct r600_common_context *)context)->ws;
 764         unsigned dpb_size = calc_dpb_size(templ);
 765         unsigned width = templ->width, height = templ->height;
 766         unsigned bs_buf_size;
 767         struct radeon_info info;
 768         struct ruvd_decoder *dec;
 769         int i;
 770
 771         ws->query_info(ws, &info);
 772
 773         switch(u_reduce_video_profile(templ->profile)) {
 774         case PIPE_VIDEO_FORMAT_MPEG12:
 775                 if (templ->entrypoint > PIPE_VIDEO_ENTRYPOINT_BITSTREAM || info.family < CHIP_PALM)
 776                         return vl_create_mpeg12_decoder(context, templ);
 777
 778                 /* fall through */
 779         case PIPE_VIDEO_FORMAT_MPEG4:
 780         case PIPE_VIDEO_FORMAT_MPEG4_AVC:
 781                 width = align(width, VL_MACROBLOCK_WIDTH);
 782                 height = align(height, VL_MACROBLOCK_HEIGHT);
 783                 break;
 784
 785         default:
 786                 break;
 787         }
 788
 789
 790         dec = CALLOC_STRUCT(ruvd_decoder);
 791
 792         if (!dec)
 793                 return NULL;
 794
 795         dec->base = *templ;
 796         dec->base.context = context;
 797         dec->base.width = width;
 798         dec->base.height = height;
 799
 800         dec->base.destroy = ruvd_destroy;
 801         dec->base.begin_frame = ruvd_begin_frame;
 802         dec->base.decode_macroblock = ruvd_decode_macroblock;
 803         dec->base.decode_bitstream = ruvd_decode_bitstream;
 804         dec->base.end_frame = ruvd_end_frame;
 805         dec->base.flush = ruvd_flush;
 806
 807         dec->set_dtb = set_dtb;
 808         dec->stream_handle = rvid_alloc_stream_handle();
 809         dec->screen = context->screen;
 810         dec->ws = ws;
 811         dec->cs = ws->cs_create(ws, RING_UVD, NULL, NULL, NULL);
 812         if (!dec->cs) {
 813                 RVID_ERR("Can't get command submission context.\n");
 814                 goto error;
 815         }
 816
 817         bs_buf_size = width * height * 512 / (16 * 16);
 818         for (i = 0; i < NUM_BUFFERS; ++i) {
 819                 unsigned msg_fb_size = FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
 820                 STATIC_ASSERT(sizeof(struct ruvd_msg) <= FB_BUFFER_OFFSET);
 821                 if (!rvid_create_buffer(dec->screen, &dec->msg_fb_buffers[i],
 822                                         msg_fb_size, PIPE_USAGE_DEFAULT)) {
 823                         RVID_ERR("Can't allocated message buffers.\n");
 824                         goto error;
 825                 }
 826
 827                 if (!rvid_create_buffer(dec->screen, &dec->bs_buffers[i],
 828                                         bs_buf_size, PIPE_USAGE_STAGING)) {
 829                         RVID_ERR("Can't allocated bitstream buffers.\n");
 830                         goto error;
 831                 }
 832
 833                 rvid_clear_buffer(dec->ws, dec->cs, &dec->msg_fb_buffers[i]);
 834                 rvid_clear_buffer(dec->ws, dec->cs, &dec->bs_buffers[i]);
 835         }
 836
 837         if (!rvid_create_buffer(dec->screen, &dec->dpb, dpb_size, PIPE_USAGE_DEFAULT)) {
 838                 RVID_ERR("Can't allocated dpb.\n");
 839                 goto error;
 840         }
 841
 842         rvid_clear_buffer(dec->ws, dec->cs, &dec->dpb);
 843
 844         map_msg_fb_buf(dec);
 845         dec->msg->size = sizeof(*dec->msg);
 846         dec->msg->msg_type = RUVD_MSG_CREATE;
 847         dec->msg->stream_handle = dec->stream_handle;
 848         dec->msg->body.create.stream_type = profile2stream_type(dec->base.profile);
 849         dec->msg->body.create.width_in_samples = dec->base.width;
 850         dec->msg->body.create.height_in_samples = dec->base.height;
 851         dec->msg->body.create.dpb_size = dec->dpb.res->buf->size;
 852         send_msg_buf(dec);
 853         flush(dec);
 854         next_buffer(dec);
 855
 856         return &dec->base;
 857
 858 error:
 859         if (dec->cs) dec->ws->cs_destroy(dec->cs);
 860
 861         for (i = 0; i < NUM_BUFFERS; ++i) {
 862                 rvid_destroy_buffer(&dec->msg_fb_buffers[i]);
 863                 rvid_destroy_buffer(&dec->bs_buffers[i]);
 864         }
 865
 866         rvid_destroy_buffer(&dec->dpb);
 867
 868         FREE(dec);
 869
 870         return NULL;
 871 }
 872
 873 /* calculate top/bottom offset */
 874 static unsigned texture_offset(struct radeon_surface *surface, unsigned layer)
 875 {
 876         return surface->level[0].offset +
 877                 layer * surface->level[0].slice_size;
 878 }
 879
 880 /* hw encode the aspect of macro tiles */
 881 static unsigned macro_tile_aspect(unsigned macro_tile_aspect)
 882 {
 883         switch (macro_tile_aspect) {
 884         default:
 885         case 1: macro_tile_aspect = 0;  break;
 886         case 2: macro_tile_aspect = 1;  break;
 887         case 4: macro_tile_aspect = 2;  break;
 888         case 8: macro_tile_aspect = 3;  break;
 889         }
 890         return macro_tile_aspect;
 891 }
 892
 893 /* hw encode the bank width and height */
 894 static unsigned bank_wh(unsigned bankwh)
 895 {
 896         switch (bankwh) {
 897         default:
 898         case 1: bankwh = 0;     break;
 899         case 2: bankwh = 1;     break;
 900         case 4: bankwh = 2;     break;
 901         case 8: bankwh = 3;     break;
 902         }
 903         return bankwh;
 904 }
 905
 906 /**
 907  * fill decoding target field from the luma and chroma surfaces
 908  */
 909 void ruvd_set_dt_surfaces(struct ruvd_msg *msg, struct radeon_surface *luma,
 910                           struct radeon_surface *chroma)
 911 {
 912         msg->body.decode.dt_pitch = luma->level[0].pitch_bytes;
 913         switch (luma->level[0].mode) {
 914         case RADEON_SURF_MODE_LINEAR_ALIGNED:
 915                 msg->body.decode.dt_tiling_mode = RUVD_TILE_LINEAR;
 916                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_LINEAR;
 917                 break;
 918         case RADEON_SURF_MODE_1D:
 919                 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
 920                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_1D_THIN;
 921                 break;
 922         case RADEON_SURF_MODE_2D:
 923                 msg->body.decode.dt_tiling_mode = RUVD_TILE_8X8;
 924                 msg->body.decode.dt_array_mode = RUVD_ARRAY_MODE_2D_THIN;
 925                 break;
 926         default:
 927                 assert(0);
 928                 break;
 929         }
 930
 931         msg->body.decode.dt_luma_top_offset = texture_offset(luma, 0);
 932         msg->body.decode.dt_chroma_top_offset = texture_offset(chroma, 0);
 933         if (msg->body.decode.dt_field_mode) {
 934                 msg->body.decode.dt_luma_bottom_offset = texture_offset(luma, 1);
 935                 msg->body.decode.dt_chroma_bottom_offset = texture_offset(chroma, 1);
 936         } else {
 937                 msg->body.decode.dt_luma_bottom_offset = msg->body.decode.dt_luma_top_offset;
 938                 msg->body.decode.dt_chroma_bottom_offset = msg->body.decode.dt_chroma_top_offset;
 939         }
 940
 941         assert(luma->bankw == chroma->bankw);
 942         assert(luma->bankh == chroma->bankh);
 943         assert(luma->mtilea == chroma->mtilea);
 944
 945         msg->body.decode.dt_surf_tile_config |= RUVD_BANK_WIDTH(bank_wh(luma->bankw));
 946         msg->body.decode.dt_surf_tile_config |= RUVD_BANK_HEIGHT(bank_wh(luma->bankh));
 947         msg->body.decode.dt_surf_tile_config |= RUVD_MACRO_TILE_ASPECT_RATIO(macro_tile_aspect(luma->mtilea));
 948 }