src/gallium/drivers/radeon/radeon_vce.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2013 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 #include <stdio.h>
  29
  30 #include "pipe/p_video_codec.h"
  31
  32 #include "util/u_video.h"
  33 #include "util/u_memory.h"
  34
  35 #include "vl/vl_video_buffer.h"
  36
  37 #include "radeonsi/si_pipe.h"
  38 #include "radeon_video.h"
  39 #include "radeon_vce.h"
  40
  41 #define FW_40_2_2 ((40 << 24) | (2 << 16) | (2 << 8))
  42 #define FW_50_0_1 ((50 << 24) | (0 << 16) | (1 << 8))
  43 #define FW_50_1_2 ((50 << 24) | (1 << 16) | (2 << 8))
  44 #define FW_50_10_2 ((50 << 24) | (10 << 16) | (2 << 8))
  45 #define FW_50_17_3 ((50 << 24) | (17 << 16) | (3 << 8))
  46 #define FW_52_0_3 ((52 << 24) | (0 << 16) | (3 << 8))
  47 #define FW_52_4_3 ((52 << 24) | (4 << 16) | (3 << 8))
  48 #define FW_52_8_3 ((52 << 24) | (8 << 16) | (3 << 8))
  49 #define FW_53 (53 << 24)
  50
  51 /**
  52  * flush commands to the hardware
  53  */
  54 static void flush(struct rvce_encoder *enc)
  55 {
  56         enc->ws->cs_flush(enc->cs, PIPE_FLUSH_ASYNC, NULL);
  57         enc->task_info_idx = 0;
  58         enc->bs_idx = 0;
  59 }
  60
  61 #if 0
  62 static void dump_feedback(struct rvce_encoder *enc, struct rvid_buffer *fb)
  63 {
  64         uint32_t *ptr = enc->ws->buffer_map(fb->res->buf, enc->cs, PIPE_TRANSFER_READ_WRITE);
  65         unsigned i = 0;
  66         fprintf(stderr, "\n");
  67         fprintf(stderr, "encStatus:\t\t\t%08x\n", ptr[i++]);
  68         fprintf(stderr, "encHasBitstream:\t\t%08x\n", ptr[i++]);
  69         fprintf(stderr, "encHasAudioBitstream:\t\t%08x\n", ptr[i++]);
  70         fprintf(stderr, "encBitstreamOffset:\t\t%08x\n", ptr[i++]);
  71         fprintf(stderr, "encBitstreamSize:\t\t%08x\n", ptr[i++]);
  72         fprintf(stderr, "encAudioBitstreamOffset:\t%08x\n", ptr[i++]);
  73         fprintf(stderr, "encAudioBitstreamSize:\t\t%08x\n", ptr[i++]);
  74         fprintf(stderr, "encExtrabytes:\t\t\t%08x\n", ptr[i++]);
  75         fprintf(stderr, "encAudioExtrabytes:\t\t%08x\n", ptr[i++]);
  76         fprintf(stderr, "videoTimeStamp:\t\t\t%08x\n", ptr[i++]);
  77         fprintf(stderr, "audioTimeStamp:\t\t\t%08x\n", ptr[i++]);
  78         fprintf(stderr, "videoOutputType:\t\t%08x\n", ptr[i++]);
  79         fprintf(stderr, "attributeFlags:\t\t\t%08x\n", ptr[i++]);
  80         fprintf(stderr, "seiPrivatePackageOffset:\t%08x\n", ptr[i++]);
  81         fprintf(stderr, "seiPrivatePackageSize:\t\t%08x\n", ptr[i++]);
  82         fprintf(stderr, "\n");
  83         enc->ws->buffer_unmap(fb->res->buf);
  84 }
  85 #endif
  86
  87 /**
  88  * reset the CPB handling
  89  */
  90 static void reset_cpb(struct rvce_encoder *enc)
  91 {
  92         unsigned i;
  93
  94         LIST_INITHEAD(&enc->cpb_slots);
  95         for (i = 0; i < enc->cpb_num; ++i) {
  96                 struct rvce_cpb_slot *slot = &enc->cpb_array[i];
  97                 slot->index = i;
  98                 slot->picture_type = PIPE_H264_ENC_PICTURE_TYPE_SKIP;
  99                 slot->frame_num = 0;
 100                 slot->pic_order_cnt = 0;
 101                 LIST_ADDTAIL(&slot->list, &enc->cpb_slots);
 102         }
 103 }
 104
 105 /**
 106  * sort l0 and l1 to the top of the list
 107  */
 108 static void sort_cpb(struct rvce_encoder *enc)
 109 {
 110         struct rvce_cpb_slot *i, *l0 = NULL, *l1 = NULL;
 111
 112         LIST_FOR_EACH_ENTRY(i, &enc->cpb_slots, list) {
 113                 if (i->frame_num == enc->pic.ref_idx_l0)
 114                         l0 = i;
 115
 116                 if (i->frame_num == enc->pic.ref_idx_l1)
 117                         l1 = i;
 118
 119                 if (enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_P && l0)
 120                         break;
 121
 122                 if (enc->pic.picture_type == PIPE_H264_ENC_PICTURE_TYPE_B &&
 123                     l0 && l1)
 124                         break;
 125         }
 126
 127         if (l1) {
 128                 LIST_DEL(&l1->list);
 129                 LIST_ADD(&l1->list, &enc->cpb_slots);
 130         }
 131
 132         if (l0) {
 133                 LIST_DEL(&l0->list);
 134                 LIST_ADD(&l0->list, &enc->cpb_slots);
 135         }
 136 }
 137
 138 /**
 139  * get number of cpbs based on dpb
 140  */
 141 static unsigned get_cpb_num(struct rvce_encoder *enc)
 142 {
 143         unsigned w = align(enc->base.width, 16) / 16;
 144         unsigned h = align(enc->base.height, 16) / 16;
 145         unsigned dpb;
 146
 147         switch (enc->base.level) {
 148         case 10:
 149                 dpb = 396;
 150                 break;
 151         case 11:
 152                 dpb = 900;
 153                 break;
 154         case 12:
 155         case 13:
 156         case 20:
 157                 dpb = 2376;
 158                 break;
 159         case 21:
 160                 dpb = 4752;
 161                 break;
 162         case 22:
 163         case 30:
 164                 dpb = 8100;
 165                 break;
 166         case 31:
 167                 dpb = 18000;
 168                 break;
 169         case 32:
 170                 dpb = 20480;
 171                 break;
 172         case 40:
 173         case 41:
 174                 dpb = 32768;
 175                 break;
 176         case 42:
 177                 dpb = 34816;
 178                 break;
 179         case 50:
 180                 dpb = 110400;
 181                 break;
 182         default:
 183         case 51:
 184         case 52:
 185                 dpb = 184320;
 186                 break;
 187         }
 188
 189         return MIN2(dpb / (w * h), 16);
 190 }
 191
 192 /**
 193  * Get the slot for the currently encoded frame
 194  */
 195 struct rvce_cpb_slot *si_current_slot(struct rvce_encoder *enc)
 196 {
 197         return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.prev, list);
 198 }
 199
 200 /**
 201  * Get the slot for L0
 202  */
 203 struct rvce_cpb_slot *si_l0_slot(struct rvce_encoder *enc)
 204 {
 205         return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.next, list);
 206 }
 207
 208 /**
 209  * Get the slot for L1
 210  */
 211 struct rvce_cpb_slot *si_l1_slot(struct rvce_encoder *enc)
 212 {
 213         return LIST_ENTRY(struct rvce_cpb_slot, enc->cpb_slots.next->next, list);
 214 }
 215
 216 /**
 217  * Calculate the offsets into the CPB
 218  */
 219 void si_vce_frame_offset(struct rvce_encoder *enc, struct rvce_cpb_slot *slot,
 220                          signed *luma_offset, signed *chroma_offset)
 221 {
 222         struct si_screen *sscreen = (struct si_screen *)enc->screen;
 223         unsigned pitch, vpitch, fsize;
 224
 225         if (sscreen->info.chip_class < GFX9) {
 226                 pitch = align(enc->luma->u.legacy.level[0].nblk_x * enc->luma->bpe, 128);
 227                 vpitch = align(enc->luma->u.legacy.level[0].nblk_y, 16);
 228         } else {
 229                 pitch = align(enc->luma->u.gfx9.surf_pitch * enc->luma->bpe, 256);
 230                 vpitch = align(enc->luma->u.gfx9.surf_height, 16);
 231         }
 232         fsize = pitch * (vpitch + vpitch / 2);
 233
 234         *luma_offset = slot->index * fsize;
 235         *chroma_offset = *luma_offset + pitch * vpitch;
 236 }
 237
 238 /**
 239  * destroy this video encoder
 240  */
 241 static void rvce_destroy(struct pipe_video_codec *encoder)
 242 {
 243         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 244         if (enc->stream_handle) {
 245                 struct rvid_buffer fb;
 246                 si_vid_create_buffer(enc->screen, &fb, 512, PIPE_USAGE_STAGING);
 247                 enc->fb = &fb;
 248                 enc->session(enc);
 249                 enc->destroy(enc);
 250                 flush(enc);
 251                 si_vid_destroy_buffer(&fb);
 252         }
 253         si_vid_destroy_buffer(&enc->cpb);
 254         enc->ws->cs_destroy(enc->cs);
 255         FREE(enc->cpb_array);
 256         FREE(enc);
 257 }
 258
 259 static void rvce_begin_frame(struct pipe_video_codec *encoder,
 260                              struct pipe_video_buffer *source,
 261                              struct pipe_picture_desc *picture)
 262 {
 263         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 264         struct vl_video_buffer *vid_buf = (struct vl_video_buffer *)source;
 265         struct pipe_h264_enc_picture_desc *pic = (struct pipe_h264_enc_picture_desc *)picture;
 266
 267         bool need_rate_control =
 268                 enc->pic.rate_ctrl.rate_ctrl_method != pic->rate_ctrl.rate_ctrl_method ||
 269                 enc->pic.quant_i_frames != pic->quant_i_frames ||
 270                 enc->pic.quant_p_frames != pic->quant_p_frames ||
 271                 enc->pic.quant_b_frames != pic->quant_b_frames;
 272
 273         enc->pic = *pic;
 274         si_get_pic_param(enc, pic);
 275
 276         enc->get_buffer(vid_buf->resources[0], &enc->handle, &enc->luma);
 277         enc->get_buffer(vid_buf->resources[1], NULL, &enc->chroma);
 278
 279         if (pic->picture_type == PIPE_H264_ENC_PICTURE_TYPE_IDR)
 280                 reset_cpb(enc);
 281         else if (pic->picture_type == PIPE_H264_ENC_PICTURE_TYPE_P ||
 282                  pic->picture_type == PIPE_H264_ENC_PICTURE_TYPE_B)
 283                 sort_cpb(enc);
 284
 285         if (!enc->stream_handle) {
 286                 struct rvid_buffer fb;
 287                 enc->stream_handle = si_vid_alloc_stream_handle();
 288                 si_vid_create_buffer(enc->screen, &fb, 512, PIPE_USAGE_STAGING);
 289                 enc->fb = &fb;
 290                 enc->session(enc);
 291                 enc->create(enc);
 292                 enc->config(enc);
 293                 enc->feedback(enc);
 294                 flush(enc);
 295                 //dump_feedback(enc, &fb);
 296                 si_vid_destroy_buffer(&fb);
 297                 need_rate_control = false;
 298         }
 299
 300         if (need_rate_control) {
 301                 enc->session(enc);
 302                 enc->config(enc);
 303                 flush(enc);
 304         }
 305 }
 306
 307 static void rvce_encode_bitstream(struct pipe_video_codec *encoder,
 308                                   struct pipe_video_buffer *source,
 309                                   struct pipe_resource *destination,
 310                                   void **fb)
 311 {
 312         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 313         enc->get_buffer(destination, &enc->bs_handle, NULL);
 314         enc->bs_size = destination->width0;
 315
 316         *fb = enc->fb = CALLOC_STRUCT(rvid_buffer);
 317         if (!si_vid_create_buffer(enc->screen, enc->fb, 512, PIPE_USAGE_STAGING)) {
 318                 RVID_ERR("Can't create feedback buffer.\n");
 319                 return;
 320         }
 321         if (!radeon_emitted(enc->cs, 0))
 322                 enc->session(enc);
 323         enc->encode(enc);
 324         enc->feedback(enc);
 325 }
 326
 327 static void rvce_end_frame(struct pipe_video_codec *encoder,
 328                            struct pipe_video_buffer *source,
 329                            struct pipe_picture_desc *picture)
 330 {
 331         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 332         struct rvce_cpb_slot *slot = LIST_ENTRY(
 333                 struct rvce_cpb_slot, enc->cpb_slots.prev, list);
 334
 335         if (!enc->dual_inst || enc->bs_idx > 1)
 336                 flush(enc);
 337
 338         /* update the CPB backtrack with the just encoded frame */
 339         slot->picture_type = enc->pic.picture_type;
 340         slot->frame_num = enc->pic.frame_num;
 341         slot->pic_order_cnt = enc->pic.pic_order_cnt;
 342         if (!enc->pic.not_referenced) {
 343                 LIST_DEL(&slot->list);
 344                 LIST_ADD(&slot->list, &enc->cpb_slots);
 345         }
 346 }
 347
 348 static void rvce_get_feedback(struct pipe_video_codec *encoder,
 349                               void *feedback, unsigned *size)
 350 {
 351         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 352         struct rvid_buffer *fb = feedback;
 353
 354         if (size) {
 355                 uint32_t *ptr = enc->ws->buffer_map(fb->res->buf, enc->cs, PIPE_TRANSFER_READ_WRITE);
 356
 357                 if (ptr[1]) {
 358                         *size = ptr[4] - ptr[9];
 359                 } else {
 360                         *size = 0;
 361                 }
 362
 363                 enc->ws->buffer_unmap(fb->res->buf);
 364         }
 365         //dump_feedback(enc, fb);
 366         si_vid_destroy_buffer(fb);
 367         FREE(fb);
 368 }
 369
 370 /**
 371  * flush any outstanding command buffers to the hardware
 372  */
 373 static void rvce_flush(struct pipe_video_codec *encoder)
 374 {
 375         struct rvce_encoder *enc = (struct rvce_encoder*)encoder;
 376
 377         flush(enc);
 378 }
 379
 380 static void rvce_cs_flush(void *ctx, unsigned flags,
 381                           struct pipe_fence_handle **fence)
 382 {
 383         // just ignored
 384 }
 385
 386 struct pipe_video_codec *si_vce_create_encoder(struct pipe_context *context,
 387                                                const struct pipe_video_codec *templ,
 388                                                struct radeon_winsys* ws,
 389                                                rvce_get_buffer get_buffer)
 390 {
 391         struct si_screen *sscreen = (struct si_screen *)context->screen;
 392         struct si_context *sctx = (struct si_context*)context;
 393         struct rvce_encoder *enc;
 394         struct pipe_video_buffer *tmp_buf, templat = {};
 395         struct radeon_surf *tmp_surf;
 396         unsigned cpb_size;
 397
 398         if (!sscreen->info.vce_fw_version) {
 399                 RVID_ERR("Kernel doesn't supports VCE!\n");
 400                 return NULL;
 401
 402         } else if (!si_vce_is_fw_version_supported(sscreen)) {
 403                 RVID_ERR("Unsupported VCE fw version loaded!\n");
 404                 return NULL;
 405         }
 406
 407         enc = CALLOC_STRUCT(rvce_encoder);
 408         if (!enc)
 409                 return NULL;
 410
 411         if (sscreen->info.drm_major == 3)
 412                 enc->use_vm = true;
 413         if ((sscreen->info.drm_major == 2 && sscreen->info.drm_minor >= 42) ||
 414             sscreen->info.drm_major == 3)
 415                 enc->use_vui = true;
 416         if (sscreen->info.family >= CHIP_TONGA &&
 417             sscreen->info.family != CHIP_STONEY &&
 418             sscreen->info.family != CHIP_POLARIS11 &&
 419             sscreen->info.family != CHIP_POLARIS12 &&
 420             sscreen->info.family != CHIP_VEGAM)
 421                 enc->dual_pipe = true;
 422         /* TODO enable B frame with dual instance */
 423         if ((sscreen->info.family >= CHIP_TONGA) &&
 424                 (templ->max_references == 1) &&
 425                 (sscreen->info.vce_harvest_config == 0))
 426                 enc->dual_inst = true;
 427
 428         enc->base = *templ;
 429         enc->base.context = context;
 430
 431         enc->base.destroy = rvce_destroy;
 432         enc->base.begin_frame = rvce_begin_frame;
 433         enc->base.encode_bitstream = rvce_encode_bitstream;
 434         enc->base.end_frame = rvce_end_frame;
 435         enc->base.flush = rvce_flush;
 436         enc->base.get_feedback = rvce_get_feedback;
 437         enc->get_buffer = get_buffer;
 438
 439         enc->screen = context->screen;
 440         enc->ws = ws;
 441         enc->cs = ws->cs_create(sctx->ctx, RING_VCE, rvce_cs_flush, enc);
 442         if (!enc->cs) {
 443                 RVID_ERR("Can't get command submission context.\n");
 444                 goto error;
 445         }
 446
 447         templat.buffer_format = PIPE_FORMAT_NV12;
 448         templat.chroma_format = PIPE_VIDEO_CHROMA_FORMAT_420;
 449         templat.width = enc->base.width;
 450         templat.height = enc->base.height;
 451         templat.interlaced = false;
 452         if (!(tmp_buf = context->create_video_buffer(context, &templat))) {
 453                 RVID_ERR("Can't create video buffer.\n");
 454                 goto error;
 455         }
 456
 457         enc->cpb_num = get_cpb_num(enc);
 458         if (!enc->cpb_num)
 459                 goto error;
 460
 461         get_buffer(((struct vl_video_buffer *)tmp_buf)->resources[0], NULL, &tmp_surf);
 462
 463         cpb_size = (sscreen->info.chip_class < GFX9) ?
 464                 align(tmp_surf->u.legacy.level[0].nblk_x * tmp_surf->bpe, 128) *
 465                 align(tmp_surf->u.legacy.level[0].nblk_y, 32) :
 466
 467                 align(tmp_surf->u.gfx9.surf_pitch * tmp_surf->bpe, 256) *
 468                 align(tmp_surf->u.gfx9.surf_height, 32);
 469
 470         cpb_size = cpb_size * 3 / 2;
 471         cpb_size = cpb_size * enc->cpb_num;
 472         if (enc->dual_pipe)
 473                 cpb_size +=  RVCE_MAX_AUX_BUFFER_NUM *
 474                         RVCE_MAX_BITSTREAM_OUTPUT_ROW_SIZE * 2;
 475         tmp_buf->destroy(tmp_buf);
 476         if (!si_vid_create_buffer(enc->screen, &enc->cpb, cpb_size, PIPE_USAGE_DEFAULT)) {
 477                 RVID_ERR("Can't create CPB buffer.\n");
 478                 goto error;
 479         }
 480
 481         enc->cpb_array = CALLOC(enc->cpb_num, sizeof(struct rvce_cpb_slot));
 482         if (!enc->cpb_array)
 483                 goto error;
 484
 485         reset_cpb(enc);
 486
 487         switch (sscreen->info.vce_fw_version) {
 488         case FW_40_2_2:
 489                 si_vce_40_2_2_init(enc);
 490                 si_get_pic_param = si_vce_40_2_2_get_param;
 491                 break;
 492
 493         case FW_50_0_1:
 494         case FW_50_1_2:
 495         case FW_50_10_2:
 496         case FW_50_17_3:
 497                 si_vce_50_init(enc);
 498                 si_get_pic_param = si_vce_50_get_param;
 499                 break;
 500
 501         case FW_52_0_3:
 502         case FW_52_4_3:
 503         case FW_52_8_3:
 504                 si_vce_52_init(enc);
 505                 si_get_pic_param = si_vce_52_get_param;
 506                 break;
 507
 508         default:
 509                 if ((sscreen->info.vce_fw_version & (0xff << 24)) == FW_53) {
 510                         si_vce_52_init(enc);
 511                         si_get_pic_param = si_vce_52_get_param;
 512                 } else
 513                         goto error;
 514         }
 515
 516         return &enc->base;
 517
 518 error:
 519         if (enc->cs)
 520                 enc->ws->cs_destroy(enc->cs);
 521
 522         si_vid_destroy_buffer(&enc->cpb);
 523
 524         FREE(enc->cpb_array);
 525         FREE(enc);
 526         return NULL;
 527 }
 528
 529 /**
 530  * check if kernel has the right fw version loaded
 531  */
 532 bool si_vce_is_fw_version_supported(struct si_screen *sscreen)
 533 {
 534         switch (sscreen->info.vce_fw_version) {
 535         case FW_40_2_2:
 536         case FW_50_0_1:
 537         case FW_50_1_2:
 538         case FW_50_10_2:
 539         case FW_50_17_3:
 540         case FW_52_0_3:
 541         case FW_52_4_3:
 542         case FW_52_8_3:
 543                 return true;
 544         default:
 545                 if ((sscreen->info.vce_fw_version & (0xff << 24)) == FW_53)
 546                         return true;
 547                 else
 548                         return false;
 549         }
 550 }
 551
 552 /**
 553  * Add the buffer as relocation to the current command submission
 554  */
 555 void si_vce_add_buffer(struct rvce_encoder *enc, struct pb_buffer *buf,
 556                        enum radeon_bo_usage usage, enum radeon_bo_domain domain,
 557                        signed offset)
 558 {
 559         int reloc_idx;
 560
 561         reloc_idx = enc->ws->cs_add_buffer(enc->cs, buf, usage | RADEON_USAGE_SYNCHRONIZED,
 562                                            domain, RADEON_PRIO_VCE);
 563         if (enc->use_vm) {
 564                 uint64_t addr;
 565                 addr = enc->ws->buffer_get_virtual_address(buf);
 566                 addr = addr + offset;
 567                 RVCE_CS(addr >> 32);
 568                 RVCE_CS(addr);
 569         } else {
 570                 offset += enc->ws->buffer_get_reloc_offset(buf);
 571                 RVCE_CS(reloc_idx * 4);
 572                 RVCE_CS(offset);
 573         }
 574 }