1 /* 2 * Copyright 2014 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24 /** 25 * \file amdgpu.h 26 * 27 * Declare public libdrm_amdgpu API 28 * 29 * This file define API exposed by libdrm_amdgpu library. 30 * User wanted to use libdrm_amdgpu functionality must include 31 * this file. 32 * 33 */ 34 #ifndef _AMDGPU_H_ 35 #define _AMDGPU_H_ 36 37 #include <stdint.h> 38 #include <stdbool.h> 39 40 struct drm_amdgpu_info_hw_ip; 41 42 /*--------------------------------------------------------------------------*/ 43 /* --------------------------- Defines ------------------------------------ */ 44 /*--------------------------------------------------------------------------*/ 45 46 /** 47 * Define max. number of Command Buffers (IB) which could be sent to the single 48 * hardware IP to accommodate CE/DE requirements 49 * 50 * \sa amdgpu_cs_ib_info 51 */ 52 #define AMDGPU_CS_MAX_IBS_PER_SUBMIT 4 53 54 /** 55 * Special timeout value meaning that the timeout is infinite. 56 */ 57 #define AMDGPU_TIMEOUT_INFINITE 0xffffffffffffffffull 58 59 /** 60 * Used in amdgpu_cs_query_fence_status(), meaning that the given timeout 61 * is absolute. 62 */ 63 #define AMDGPU_QUERY_FENCE_TIMEOUT_IS_ABSOLUTE (1 << 0) 64 65 /*--------------------------------------------------------------------------*/ 66 /* ----------------------------- Enums ------------------------------------ */ 67 /*--------------------------------------------------------------------------*/ 68 69 /** 70 * Enum describing possible handle types 71 * 72 * \sa amdgpu_bo_import, amdgpu_bo_export 73 * 74 */ 75 enum amdgpu_bo_handle_type { 76 /** GEM flink name (needs DRM authentication, used by DRI2) */ 77 amdgpu_bo_handle_type_gem_flink_name = 0, 78 79 /** KMS handle which is used by all driver ioctls */ 80 amdgpu_bo_handle_type_kms = 1, 81 82 /** DMA-buf fd handle */ 83 amdgpu_bo_handle_type_dma_buf_fd = 2 84 }; 85 86 /** Define known types of GPU VM VA ranges */ 87 enum amdgpu_gpu_va_range 88 { 89 /** Allocate from "normal"/general range */ 90 amdgpu_gpu_va_range_general = 0 91 }; 92 93 /*--------------------------------------------------------------------------*/ 94 /* -------------------------- Datatypes ----------------------------------- */ 95 /*--------------------------------------------------------------------------*/ 96 97 /** 98 * Define opaque pointer to context associated with fd. 99 * This context will be returned as the result of 100 * "initialize" function and should be pass as the first 101 * parameter to any API call 102 */ 103 typedef struct amdgpu_device *amdgpu_device_handle; 104 105 /** 106 * Define GPU Context type as pointer to opaque structure 107 * Example of GPU Context is the "rendering" context associated 108 * with OpenGL context (glCreateContext) 109 */ 110 typedef struct amdgpu_context *amdgpu_context_handle; 111 112 /** 113 * Define handle for amdgpu resources: buffer, GDS, etc. 114 */ 115 typedef struct amdgpu_bo *amdgpu_bo_handle; 116 117 /** 118 * Define handle for list of BOs 119 */ 120 typedef struct amdgpu_bo_list *amdgpu_bo_list_handle; 121 122 /** 123 * Define handle to be used to work with VA allocated ranges 124 */ 125 typedef struct amdgpu_va *amdgpu_va_handle; 126 127 /*--------------------------------------------------------------------------*/ 128 /* -------------------------- Structures ---------------------------------- */ 129 /*--------------------------------------------------------------------------*/ 130 131 /** 132 * Structure describing memory allocation request 133 * 134 * \sa amdgpu_bo_alloc() 135 * 136 */ 137 struct amdgpu_bo_alloc_request { 138 /** Allocation request. It must be aligned correctly. */ 139 uint64_t alloc_size; 140 141 /** 142 * It may be required to have some specific alignment requirements 143 * for physical back-up storage (e.g. for displayable surface). 144 * If 0 there is no special alignment requirement 145 */ 146 uint64_t phys_alignment; 147 148 /** 149 * UMD should specify where to allocate memory and how it 150 * will be accessed by the CPU. 151 */ 152 uint32_t preferred_heap; 153 154 /** Additional flags passed on allocation */ 155 uint64_t flags; 156 }; 157 158 /** 159 * Special UMD specific information associated with buffer. 160 * 161 * It may be need to pass some buffer charactersitic as part 162 * of buffer sharing. Such information are defined UMD and 163 * opaque for libdrm_amdgpu as well for kernel driver. 164 * 165 * \sa amdgpu_bo_set_metadata(), amdgpu_bo_query_info, 166 * amdgpu_bo_import(), amdgpu_bo_export 167 * 168 */ 169 struct amdgpu_bo_metadata { 170 /** Special flag associated with surface */ 171 uint64_t flags; 172 173 /** 174 * ASIC-specific tiling information (also used by DCE). 175 * The encoding is defined by the AMDGPU_TILING_* definitions. 176 */ 177 uint64_t tiling_info; 178 179 /** Size of metadata associated with the buffer, in bytes. */ 180 uint32_t size_metadata; 181 182 /** UMD specific metadata. Opaque for kernel */ 183 uint32_t umd_metadata[64]; 184 }; 185 186 /** 187 * Structure describing allocated buffer. Client may need 188 * to query such information as part of 'sharing' buffers mechanism 189 * 190 * \sa amdgpu_bo_set_metadata(), amdgpu_bo_query_info(), 191 * amdgpu_bo_import(), amdgpu_bo_export() 192 */ 193 struct amdgpu_bo_info { 194 /** Allocated memory size */ 195 uint64_t alloc_size; 196 197 /** 198 * It may be required to have some specific alignment requirements 199 * for physical back-up storage. 200 */ 201 uint64_t phys_alignment; 202 203 /** Heap where to allocate memory. */ 204 uint32_t preferred_heap; 205 206 /** Additional allocation flags. */ 207 uint64_t alloc_flags; 208 209 /** Metadata associated with buffer if any. */ 210 struct amdgpu_bo_metadata metadata; 211 }; 212 213 /** 214 * Structure with information about "imported" buffer 215 * 216 * \sa amdgpu_bo_import() 217 * 218 */ 219 struct amdgpu_bo_import_result { 220 /** Handle of memory/buffer to use */ 221 amdgpu_bo_handle buf_handle; 222 223 /** Buffer size */ 224 uint64_t alloc_size; 225 }; 226 227 /** 228 * 229 * Structure to describe GDS partitioning information. 230 * \note OA and GWS resources are asscoiated with GDS partition 231 * 232 * \sa amdgpu_gpu_resource_query_gds_info 233 * 234 */ 235 struct amdgpu_gds_resource_info { 236 uint32_t gds_gfx_partition_size; 237 uint32_t compute_partition_size; 238 uint32_t gds_total_size; 239 uint32_t gws_per_gfx_partition; 240 uint32_t gws_per_compute_partition; 241 uint32_t oa_per_gfx_partition; 242 uint32_t oa_per_compute_partition; 243 }; 244 245 /** 246 * Structure describing CS fence 247 * 248 * \sa amdgpu_cs_query_fence_status(), amdgpu_cs_request, amdgpu_cs_submit() 249 * 250 */ 251 struct amdgpu_cs_fence { 252 253 /** In which context IB was sent to execution */ 254 amdgpu_context_handle context; 255 256 /** To which HW IP type the fence belongs */ 257 uint32_t ip_type; 258 259 /** IP instance index if there are several IPs of the same type. */ 260 uint32_t ip_instance; 261 262 /** Ring index of the HW IP */ 263 uint32_t ring; 264 265 /** Specify fence for which we need to check submission status.*/ 266 uint64_t fence; 267 }; 268 269 /** 270 * Structure describing IB 271 * 272 * \sa amdgpu_cs_request, amdgpu_cs_submit() 273 * 274 */ 275 struct amdgpu_cs_ib_info { 276 /** Special flags */ 277 uint64_t flags; 278 279 /** Virtual MC address of the command buffer */ 280 uint64_t ib_mc_address; 281 282 /** 283 * Size of Command Buffer to be submitted. 284 * - The size is in units of dwords (4 bytes). 285 * - Could be 0 286 */ 287 uint32_t size; 288 }; 289 290 /** 291 * Structure describing fence information 292 * 293 * \sa amdgpu_cs_request, amdgpu_cs_query_fence, 294 * amdgpu_cs_submit(), amdgpu_cs_query_fence_status() 295 */ 296 struct amdgpu_cs_fence_info { 297 /** buffer object for the fence */ 298 amdgpu_bo_handle handle; 299 300 /** fence offset in the unit of sizeof(uint64_t) */ 301 uint64_t offset; 302 }; 303 304 /** 305 * Structure describing submission request 306 * 307 * \note We could have several IBs as packet. e.g. CE, CE, DE case for gfx 308 * 309 * \sa amdgpu_cs_submit() 310 */ 311 struct amdgpu_cs_request { 312 /** Specify flags with additional information */ 313 uint64_t flags; 314 315 /** Specify HW IP block type to which to send the IB. */ 316 unsigned ip_type; 317 318 /** IP instance index if there are several IPs of the same type. */ 319 unsigned ip_instance; 320 321 /** 322 * Specify ring index of the IP. We could have several rings 323 * in the same IP. E.g. 0 for SDMA0 and 1 for SDMA1. 324 */ 325 uint32_t ring; 326 327 /** 328 * List handle with resources used by this request. 329 */ 330 amdgpu_bo_list_handle resources; 331 332 /** 333 * Number of dependencies this Command submission needs to 334 * wait for before starting execution. 335 */ 336 uint32_t number_of_dependencies; 337 338 /** 339 * Array of dependencies which need to be met before 340 * execution can start. 341 */ 342 struct amdgpu_cs_fence *dependencies; 343 344 /** Number of IBs to submit in the field ibs. */ 345 uint32_t number_of_ibs; 346 347 /** 348 * IBs to submit. Those IBs will be submit together as single entity 349 */ 350 struct amdgpu_cs_ib_info *ibs; 351 352 /** 353 * The returned sequence number for the command submission 354 */ 355 uint64_t seq_no; 356 357 /** 358 * The fence information 359 */ 360 struct amdgpu_cs_fence_info fence_info; 361 }; 362 363 /** 364 * Structure which provide information about GPU VM MC Address space 365 * alignments requirements 366 * 367 * \sa amdgpu_query_buffer_size_alignment 368 */ 369 struct amdgpu_buffer_size_alignments { 370 /** Size alignment requirement for allocation in 371 * local memory */ 372 uint64_t size_local; 373 374 /** 375 * Size alignment requirement for allocation in remote memory 376 */ 377 uint64_t size_remote; 378 }; 379 380 /** 381 * Structure which provide information about heap 382 * 383 * \sa amdgpu_query_heap_info() 384 * 385 */ 386 struct amdgpu_heap_info { 387 /** Theoretical max. available memory in the given heap */ 388 uint64_t heap_size; 389 390 /** 391 * Number of bytes allocated in the heap. This includes all processes 392 * and private allocations in the kernel. It changes when new buffers 393 * are allocated, freed, and moved. It cannot be larger than 394 * heap_size. 395 */ 396 uint64_t heap_usage; 397 398 /** 399 * Theoretical possible max. size of buffer which 400 * could be allocated in the given heap 401 */ 402 uint64_t max_allocation; 403 }; 404 405 /** 406 * Describe GPU h/w info needed for UMD correct initialization 407 * 408 * \sa amdgpu_query_gpu_info() 409 */ 410 struct amdgpu_gpu_info { 411 /** Asic id */ 412 uint32_t asic_id; 413 /** Chip revision */ 414 uint32_t chip_rev; 415 /** Chip external revision */ 416 uint32_t chip_external_rev; 417 /** Family ID */ 418 uint32_t family_id; 419 /** Special flags */ 420 uint64_t ids_flags; 421 /** max engine clock*/ 422 uint64_t max_engine_clk; 423 /** max memory clock */ 424 uint64_t max_memory_clk; 425 /** number of shader engines */ 426 uint32_t num_shader_engines; 427 /** number of shader arrays per engine */ 428 uint32_t num_shader_arrays_per_engine; 429 /** Number of available good shader pipes */ 430 uint32_t avail_quad_shader_pipes; 431 /** Max. number of shader pipes.(including good and bad pipes */ 432 uint32_t max_quad_shader_pipes; 433 /** Number of parameter cache entries per shader quad pipe */ 434 uint32_t cache_entries_per_quad_pipe; 435 /** Number of available graphics context */ 436 uint32_t num_hw_gfx_contexts; 437 /** Number of render backend pipes */ 438 uint32_t rb_pipes; 439 /** Enabled render backend pipe mask */ 440 uint32_t enabled_rb_pipes_mask; 441 /** Frequency of GPU Counter */ 442 uint32_t gpu_counter_freq; 443 /** CC_RB_BACKEND_DISABLE.BACKEND_DISABLE per SE */ 444 uint32_t backend_disable[4]; 445 /** Value of MC_ARB_RAMCFG register*/ 446 uint32_t mc_arb_ramcfg; 447 /** Value of GB_ADDR_CONFIG */ 448 uint32_t gb_addr_cfg; 449 /** Values of the GB_TILE_MODE0..31 registers */ 450 uint32_t gb_tile_mode[32]; 451 /** Values of GB_MACROTILE_MODE0..15 registers */ 452 uint32_t gb_macro_tile_mode[16]; 453 /** Value of PA_SC_RASTER_CONFIG register per SE */ 454 uint32_t pa_sc_raster_cfg[4]; 455 /** Value of PA_SC_RASTER_CONFIG_1 register per SE */ 456 uint32_t pa_sc_raster_cfg1[4]; 457 /* CU info */ 458 uint32_t cu_active_number; 459 uint32_t cu_ao_mask; 460 uint32_t cu_bitmap[4][4]; 461 /* video memory type info*/ 462 uint32_t vram_type; 463 /* video memory bit width*/ 464 uint32_t vram_bit_width; 465 /** constant engine ram size*/ 466 uint32_t ce_ram_size; 467 /* vce harvesting instance */ 468 uint32_t vce_harvest_config; 469 /* PCI revision ID */ 470 uint32_t pci_rev_id; 471 }; 472 473 474 /*--------------------------------------------------------------------------*/ 475 /*------------------------- Functions --------------------------------------*/ 476 /*--------------------------------------------------------------------------*/ 477 478 /* 479 * Initialization / Cleanup 480 * 481 */ 482 483 /** 484 * 485 * \param fd - \c [in] File descriptor for AMD GPU device 486 * received previously as the result of 487 * e.g. drmOpen() call. 488 * For legacy fd type, the DRI2/DRI3 489 * authentication should be done before 490 * calling this function. 491 * \param major_version - \c [out] Major version of library. It is assumed 492 * that adding new functionality will cause 493 * increase in major version 494 * \param minor_version - \c [out] Minor version of library 495 * \param device_handle - \c [out] Pointer to opaque context which should 496 * be passed as the first parameter on each 497 * API call 498 * 499 * 500 * \return 0 on success\n 501 * <0 - Negative POSIX Error code 502 * 503 * 504 * \sa amdgpu_device_deinitialize() 505 */ 506 int amdgpu_device_initialize(int fd, 507 uint32_t *major_version, 508 uint32_t *minor_version, 509 amdgpu_device_handle *device_handle); 510 511 /** 512 * 513 * When access to such library does not needed any more the special 514 * function must be call giving opportunity to clean up any 515 * resources if needed. 516 * 517 * \param device_handle - \c [in] Context associated with file 518 * descriptor for AMD GPU device 519 * received previously as the 520 * result e.g. of drmOpen() call. 521 * 522 * \return 0 on success\n 523 * <0 - Negative POSIX Error code 524 * 525 * \sa amdgpu_device_initialize() 526 * 527 */ 528 int amdgpu_device_deinitialize(amdgpu_device_handle device_handle); 529 530 /* 531 * Memory Management 532 * 533 */ 534 535 /** 536 * Allocate memory to be used by UMD for GPU related operations 537 * 538 * \param dev - \c [in] Device handle. 539 * See #amdgpu_device_initialize() 540 * \param alloc_buffer - \c [in] Pointer to the structure describing an 541 * allocation request 542 * \param buf_handle - \c [out] Allocated buffer handle 543 * 544 * \return 0 on success\n 545 * <0 - Negative POSIX Error code 546 * 547 * \sa amdgpu_bo_free() 548 */ 549 int amdgpu_bo_alloc(amdgpu_device_handle dev, 550 struct amdgpu_bo_alloc_request *alloc_buffer, 551 amdgpu_bo_handle *buf_handle); 552 553 /** 554 * Associate opaque data with buffer to be queried by another UMD 555 * 556 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 557 * \param buf_handle - \c [in] Buffer handle 558 * \param info - \c [in] Metadata to associated with buffer 559 * 560 * \return 0 on success\n 561 * <0 - Negative POSIX Error code 562 */ 563 int amdgpu_bo_set_metadata(amdgpu_bo_handle buf_handle, 564 struct amdgpu_bo_metadata *info); 565 566 /** 567 * Query buffer information including metadata previusly associated with 568 * buffer. 569 * 570 * \param dev - \c [in] Device handle. 571 * See #amdgpu_device_initialize() 572 * \param buf_handle - \c [in] Buffer handle 573 * \param info - \c [out] Structure describing buffer 574 * 575 * \return 0 on success\n 576 * <0 - Negative POSIX Error code 577 * 578 * \sa amdgpu_bo_set_metadata(), amdgpu_bo_alloc() 579 */ 580 int amdgpu_bo_query_info(amdgpu_bo_handle buf_handle, 581 struct amdgpu_bo_info *info); 582 583 /** 584 * Allow others to get access to buffer 585 * 586 * \param dev - \c [in] Device handle. 587 * See #amdgpu_device_initialize() 588 * \param buf_handle - \c [in] Buffer handle 589 * \param type - \c [in] Type of handle requested 590 * \param shared_handle - \c [out] Special "shared" handle 591 * 592 * \return 0 on success\n 593 * <0 - Negative POSIX Error code 594 * 595 * \sa amdgpu_bo_import() 596 * 597 */ 598 int amdgpu_bo_export(amdgpu_bo_handle buf_handle, 599 enum amdgpu_bo_handle_type type, 600 uint32_t *shared_handle); 601 602 /** 603 * Request access to "shared" buffer 604 * 605 * \param dev - \c [in] Device handle. 606 * See #amdgpu_device_initialize() 607 * \param type - \c [in] Type of handle requested 608 * \param shared_handle - \c [in] Shared handle received as result "import" 609 * operation 610 * \param output - \c [out] Pointer to structure with information 611 * about imported buffer 612 * 613 * \return 0 on success\n 614 * <0 - Negative POSIX Error code 615 * 616 * \note Buffer must be "imported" only using new "fd" (different from 617 * one used by "exporter"). 618 * 619 * \sa amdgpu_bo_export() 620 * 621 */ 622 int amdgpu_bo_import(amdgpu_device_handle dev, 623 enum amdgpu_bo_handle_type type, 624 uint32_t shared_handle, 625 struct amdgpu_bo_import_result *output); 626 627 /** 628 * Request GPU access to user allocated memory e.g. via "malloc" 629 * 630 * \param dev - [in] Device handle. See #amdgpu_device_initialize() 631 * \param cpu - [in] CPU address of user allocated memory which we 632 * want to map to GPU address space (make GPU accessible) 633 * (This address must be correctly aligned). 634 * \param size - [in] Size of allocation (must be correctly aligned) 635 * \param buf_handle - [out] Buffer handle for the userptr memory 636 * resource on submission and be used in other operations. 637 * 638 * 639 * \return 0 on success\n 640 * <0 - Negative POSIX Error code 641 * 642 * \note 643 * This call doesn't guarantee that such memory will be persistently 644 * "locked" / make non-pageable. The purpose of this call is to provide 645 * opportunity for GPU get access to this resource during submission. 646 * 647 * The maximum amount of memory which could be mapped in this call depends 648 * if overcommit is disabled or not. If overcommit is disabled than the max. 649 * amount of memory to be pinned will be limited by left "free" size in total 650 * amount of memory which could be locked simultaneously ("GART" size). 651 * 652 * Supported (theoretical) max. size of mapping is restricted only by 653 * "GART" size. 654 * 655 * It is responsibility of caller to correctly specify access rights 656 * on VA assignment. 657 */ 658 int amdgpu_create_bo_from_user_mem(amdgpu_device_handle dev, 659 void *cpu, uint64_t size, 660 amdgpu_bo_handle *buf_handle); 661 662 /** 663 * Free previosuly allocated memory 664 * 665 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 666 * \param buf_handle - \c [in] Buffer handle to free 667 * 668 * \return 0 on success\n 669 * <0 - Negative POSIX Error code 670 * 671 * \note In the case of memory shared between different applications all 672 * resources will be “physically” freed only all such applications 673 * will be terminated 674 * \note If is UMD responsibility to ‘free’ buffer only when there is no 675 * more GPU access 676 * 677 * \sa amdgpu_bo_set_metadata(), amdgpu_bo_alloc() 678 * 679 */ 680 int amdgpu_bo_free(amdgpu_bo_handle buf_handle); 681 682 /** 683 * Request CPU access to GPU accessable memory 684 * 685 * \param buf_handle - \c [in] Buffer handle 686 * \param cpu - \c [out] CPU address to be used for access 687 * 688 * \return 0 on success\n 689 * <0 - Negative POSIX Error code 690 * 691 * \sa amdgpu_bo_cpu_unmap() 692 * 693 */ 694 int amdgpu_bo_cpu_map(amdgpu_bo_handle buf_handle, void **cpu); 695 696 /** 697 * Release CPU access to GPU memory 698 * 699 * \param buf_handle - \c [in] Buffer handle 700 * 701 * \return 0 on success\n 702 * <0 - Negative POSIX Error code 703 * 704 * \sa amdgpu_bo_cpu_map() 705 * 706 */ 707 int amdgpu_bo_cpu_unmap(amdgpu_bo_handle buf_handle); 708 709 /** 710 * Wait until a buffer is not used by the device. 711 * 712 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 713 * \param buf_handle - \c [in] Buffer handle. 714 * \param timeout_ns - Timeout in nanoseconds. 715 * \param buffer_busy - 0 if buffer is idle, all GPU access was completed 716 * and no GPU access is scheduled. 717 * 1 GPU access is in fly or scheduled 718 * 719 * \return 0 - on success 720 * <0 - Negative POSIX Error code 721 */ 722 int amdgpu_bo_wait_for_idle(amdgpu_bo_handle buf_handle, 723 uint64_t timeout_ns, 724 bool *buffer_busy); 725 726 /** 727 * Creates a BO list handle for command submission. 728 * 729 * \param dev - \c [in] Device handle. 730 * See #amdgpu_device_initialize() 731 * \param number_of_resources - \c [in] Number of BOs in the list 732 * \param resources - \c [in] List of BO handles 733 * \param resource_prios - \c [in] Optional priority for each handle 734 * \param result - \c [out] Created BO list handle 735 * 736 * \return 0 on success\n 737 * <0 - Negative POSIX Error code 738 * 739 * \sa amdgpu_bo_list_destroy() 740 */ 741 int amdgpu_bo_list_create(amdgpu_device_handle dev, 742 uint32_t number_of_resources, 743 amdgpu_bo_handle *resources, 744 uint8_t *resource_prios, 745 amdgpu_bo_list_handle *result); 746 747 /** 748 * Destroys a BO list handle. 749 * 750 * \param handle - \c [in] BO list handle. 751 * 752 * \return 0 on success\n 753 * <0 - Negative POSIX Error code 754 * 755 * \sa amdgpu_bo_list_create() 756 */ 757 int amdgpu_bo_list_destroy(amdgpu_bo_list_handle handle); 758 759 /** 760 * Update resources for existing BO list 761 * 762 * \param handle - \c [in] BO list handle 763 * \param number_of_resources - \c [in] Number of BOs in the list 764 * \param resources - \c [in] List of BO handles 765 * \param resource_prios - \c [in] Optional priority for each handle 766 * 767 * \return 0 on success\n 768 * <0 - Negative POSIX Error code 769 * 770 * \sa amdgpu_bo_list_update() 771 */ 772 int amdgpu_bo_list_update(amdgpu_bo_list_handle handle, 773 uint32_t number_of_resources, 774 amdgpu_bo_handle *resources, 775 uint8_t *resource_prios); 776 777 /* 778 * GPU Execution context 779 * 780 */ 781 782 /** 783 * Create GPU execution Context 784 * 785 * For the purpose of GPU Scheduler and GPU Robustness extensions it is 786 * necessary to have information/identify rendering/compute contexts. 787 * It also may be needed to associate some specific requirements with such 788 * contexts. Kernel driver will guarantee that submission from the same 789 * context will always be executed in order (first come, first serve). 790 * 791 * 792 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 793 * \param context - \c [out] GPU Context handle 794 * 795 * \return 0 on success\n 796 * <0 - Negative POSIX Error code 797 * 798 * \sa amdgpu_cs_ctx_free() 799 * 800 */ 801 int amdgpu_cs_ctx_create(amdgpu_device_handle dev, 802 amdgpu_context_handle *context); 803 804 /** 805 * 806 * Destroy GPU execution context when not needed any more 807 * 808 * \param context - \c [in] GPU Context handle 809 * 810 * \return 0 on success\n 811 * <0 - Negative POSIX Error code 812 * 813 * \sa amdgpu_cs_ctx_create() 814 * 815 */ 816 int amdgpu_cs_ctx_free(amdgpu_context_handle context); 817 818 /** 819 * Query reset state for the specific GPU Context 820 * 821 * \param context - \c [in] GPU Context handle 822 * \param state - \c [out] One of AMDGPU_CTX_*_RESET 823 * \param hangs - \c [out] Number of hangs caused by the context. 824 * 825 * \return 0 on success\n 826 * <0 - Negative POSIX Error code 827 * 828 * \sa amdgpu_cs_ctx_create() 829 * 830 */ 831 int amdgpu_cs_query_reset_state(amdgpu_context_handle context, 832 uint32_t *state, uint32_t *hangs); 833 834 /* 835 * Command Buffers Management 836 * 837 */ 838 839 /** 840 * Send request to submit command buffers to hardware. 841 * 842 * Kernel driver could use GPU Scheduler to make decision when physically 843 * sent this request to the hardware. Accordingly this request could be put 844 * in queue and sent for execution later. The only guarantee is that request 845 * from the same GPU context to the same ip:ip_instance:ring will be executed in 846 * order. 847 * 848 * The caller can specify the user fence buffer/location with the fence_info in the 849 * cs_request.The sequence number is returned via the 'seq_no' paramter 850 * in ibs_request structure. 851 * 852 * 853 * \param dev - \c [in] Device handle. 854 * See #amdgpu_device_initialize() 855 * \param context - \c [in] GPU Context 856 * \param flags - \c [in] Global submission flags 857 * \param ibs_request - \c [in/out] Pointer to submission requests. 858 * We could submit to the several 859 * engines/rings simulteniously as 860 * 'atomic' operation 861 * \param number_of_requests - \c [in] Number of submission requests 862 * 863 * \return 0 on success\n 864 * <0 - Negative POSIX Error code 865 * 866 * \note It is required to pass correct resource list with buffer handles 867 * which will be accessible by command buffers from submission 868 * This will allow kernel driver to correctly implement "paging". 869 * Failure to do so will have unpredictable results. 870 * 871 * \sa amdgpu_command_buffer_alloc(), amdgpu_command_buffer_free(), 872 * amdgpu_cs_query_fence_status() 873 * 874 */ 875 int amdgpu_cs_submit(amdgpu_context_handle context, 876 uint64_t flags, 877 struct amdgpu_cs_request *ibs_request, 878 uint32_t number_of_requests); 879 880 /** 881 * Query status of Command Buffer Submission 882 * 883 * \param fence - \c [in] Structure describing fence to query 884 * \param timeout_ns - \c [in] Timeout value to wait 885 * \param flags - \c [in] Flags for the query 886 * \param expired - \c [out] If fence expired or not.\n 887 * 0 – if fence is not expired\n 888 * !0 - otherwise 889 * 890 * \return 0 on success\n 891 * <0 - Negative POSIX Error code 892 * 893 * \note If UMD wants only to check operation status and returned immediately 894 * then timeout value as 0 must be passed. In this case success will be 895 * returned in the case if submission was completed or timeout error 896 * code. 897 * 898 * \sa amdgpu_cs_submit() 899 */ 900 int amdgpu_cs_query_fence_status(struct amdgpu_cs_fence *fence, 901 uint64_t timeout_ns, 902 uint64_t flags, 903 uint32_t *expired); 904 905 /* 906 * Query / Info API 907 * 908 */ 909 910 /** 911 * Query allocation size alignments 912 * 913 * UMD should query information about GPU VM MC size alignments requirements 914 * to be able correctly choose required allocation size and implement 915 * internal optimization if needed. 916 * 917 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 918 * \param info - \c [out] Pointer to structure to get size alignment 919 * requirements 920 * 921 * \return 0 on success\n 922 * <0 - Negative POSIX Error code 923 * 924 */ 925 int amdgpu_query_buffer_size_alignment(amdgpu_device_handle dev, 926 struct amdgpu_buffer_size_alignments 927 *info); 928 929 /** 930 * Query firmware versions 931 * 932 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 933 * \param fw_type - \c [in] AMDGPU_INFO_FW_* 934 * \param ip_instance - \c [in] Index of the IP block of the same type. 935 * \param index - \c [in] Index of the engine. (for SDMA and MEC) 936 * \param version - \c [out] Pointer to to the "version" return value 937 * \param feature - \c [out] Pointer to to the "feature" return value 938 * 939 * \return 0 on success\n 940 * <0 - Negative POSIX Error code 941 * 942 */ 943 int amdgpu_query_firmware_version(amdgpu_device_handle dev, unsigned fw_type, 944 unsigned ip_instance, unsigned index, 945 uint32_t *version, uint32_t *feature); 946 947 /** 948 * Query the number of HW IP instances of a certain type. 949 * 950 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 951 * \param type - \c [in] Hardware IP block type = AMDGPU_HW_IP_* 952 * \param count - \c [out] Pointer to structure to get information 953 * 954 * \return 0 on success\n 955 * <0 - Negative POSIX Error code 956 */ 957 int amdgpu_query_hw_ip_count(amdgpu_device_handle dev, unsigned type, 958 uint32_t *count); 959 960 /** 961 * Query engine information 962 * 963 * This query allows UMD to query information different engines and their 964 * capabilities. 965 * 966 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 967 * \param type - \c [in] Hardware IP block type = AMDGPU_HW_IP_* 968 * \param ip_instance - \c [in] Index of the IP block of the same type. 969 * \param info - \c [out] Pointer to structure to get information 970 * 971 * \return 0 on success\n 972 * <0 - Negative POSIX Error code 973 */ 974 int amdgpu_query_hw_ip_info(amdgpu_device_handle dev, unsigned type, 975 unsigned ip_instance, 976 struct drm_amdgpu_info_hw_ip *info); 977 978 /** 979 * Query heap information 980 * 981 * This query allows UMD to query potentially available memory resources and 982 * adjust their logic if necessary. 983 * 984 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 985 * \param heap - \c [in] Heap type 986 * \param info - \c [in] Pointer to structure to get needed information 987 * 988 * \return 0 on success\n 989 * <0 - Negative POSIX Error code 990 * 991 */ 992 int amdgpu_query_heap_info(amdgpu_device_handle dev, uint32_t heap, 993 uint32_t flags, struct amdgpu_heap_info *info); 994 995 /** 996 * Get the CRTC ID from the mode object ID 997 * 998 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 999 * \param id - \c [in] Mode object ID 1000 * \param result - \c [in] Pointer to the CRTC ID 1001 * 1002 * \return 0 on success\n 1003 * <0 - Negative POSIX Error code 1004 * 1005 */ 1006 int amdgpu_query_crtc_from_id(amdgpu_device_handle dev, unsigned id, 1007 int32_t *result); 1008 1009 /** 1010 * Query GPU H/w Info 1011 * 1012 * Query hardware specific information 1013 * 1014 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 1015 * \param heap - \c [in] Heap type 1016 * \param info - \c [in] Pointer to structure to get needed information 1017 * 1018 * \return 0 on success\n 1019 * <0 - Negative POSIX Error code 1020 * 1021 */ 1022 int amdgpu_query_gpu_info(amdgpu_device_handle dev, 1023 struct amdgpu_gpu_info *info); 1024 1025 /** 1026 * Query hardware or driver information. 1027 * 1028 * The return size is query-specific and depends on the "info_id" parameter. 1029 * No more than "size" bytes is returned. 1030 * 1031 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 1032 * \param info_id - \c [in] AMDGPU_INFO_* 1033 * \param size - \c [in] Size of the returned value. 1034 * \param value - \c [out] Pointer to the return value. 1035 * 1036 * \return 0 on success\n 1037 * <0 - Negative POSIX error code 1038 * 1039 */ 1040 int amdgpu_query_info(amdgpu_device_handle dev, unsigned info_id, 1041 unsigned size, void *value); 1042 1043 /** 1044 * Query information about GDS 1045 * 1046 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize() 1047 * \param gds_info - \c [out] Pointer to structure to get GDS information 1048 * 1049 * \return 0 on success\n 1050 * <0 - Negative POSIX Error code 1051 * 1052 */ 1053 int amdgpu_query_gds_info(amdgpu_device_handle dev, 1054 struct amdgpu_gds_resource_info *gds_info); 1055 1056 /** 1057 * Read a set of consecutive memory-mapped registers. 1058 * Not all registers are allowed to be read by userspace. 1059 * 1060 * \param dev - \c [in] Device handle. See #amdgpu_device_initialize( 1061 * \param dword_offset - \c [in] Register offset in dwords 1062 * \param count - \c [in] The number of registers to read starting 1063 * from the offset 1064 * \param instance - \c [in] GRBM_GFX_INDEX selector. It may have other 1065 * uses. Set it to 0xffffffff if unsure. 1066 * \param flags - \c [in] Flags with additional information. 1067 * \param values - \c [out] The pointer to return values. 1068 * 1069 * \return 0 on success\n 1070 * <0 - Negative POSIX error code 1071 * 1072 */ 1073 int amdgpu_read_mm_registers(amdgpu_device_handle dev, unsigned dword_offset, 1074 unsigned count, uint32_t instance, uint32_t flags, 1075 uint32_t *values); 1076 1077 /** 1078 * Flag to request VA address range in the 32bit address space 1079 */ 1080 #define AMDGPU_VA_RANGE_32_BIT 0x1 1081 1082 /** 1083 * Allocate virtual address range 1084 * 1085 * \param dev - [in] Device handle. See #amdgpu_device_initialize() 1086 * \param va_range_type - \c [in] Type of MC va range from which to allocate 1087 * \param size - \c [in] Size of range. Size must be correctly* aligned. 1088 * It is client responsibility to correctly aligned size based on the future 1089 * usage of allocated range. 1090 * \param va_base_alignment - \c [in] Overwrite base address alignment 1091 * requirement for GPU VM MC virtual 1092 * address assignment. Must be multiple of size alignments received as 1093 * 'amdgpu_buffer_size_alignments'. 1094 * If 0 use the default one. 1095 * \param va_base_required - \c [in] Specified required va base address. 1096 * If 0 then library choose available one. 1097 * If !0 value will be passed and those value already "in use" then 1098 * corresponding error status will be returned. 1099 * \param va_base_allocated - \c [out] On return: Allocated VA base to be used 1100 * by client. 1101 * \param va_range_handle - \c [out] On return: Handle assigned to allocation 1102 * \param flags - \c [in] flags for special VA range 1103 * 1104 * \return 0 on success\n 1105 * >0 - AMD specific error code\n 1106 * <0 - Negative POSIX Error code 1107 * 1108 * \notes \n 1109 * It is client responsibility to correctly handle VA assignments and usage. 1110 * Neither kernel driver nor libdrm_amdpgu are able to prevent and 1111 * detect wrong va assignemnt. 1112 * 1113 * It is client responsibility to correctly handle multi-GPU cases and to pass 1114 * the corresponding arrays of all devices handles where corresponding VA will 1115 * be used. 1116 * 1117 */ 1118 int amdgpu_va_range_alloc(amdgpu_device_handle dev, 1119 enum amdgpu_gpu_va_range va_range_type, 1120 uint64_t size, 1121 uint64_t va_base_alignment, 1122 uint64_t va_base_required, 1123 uint64_t *va_base_allocated, 1124 amdgpu_va_handle *va_range_handle, 1125 uint64_t flags); 1126 1127 /** 1128 * Free previously allocated virtual address range 1129 * 1130 * 1131 * \param va_range_handle - \c [in] Handle assigned to VA allocation 1132 * 1133 * \return 0 on success\n 1134 * >0 - AMD specific error code\n 1135 * <0 - Negative POSIX Error code 1136 * 1137 */ 1138 int amdgpu_va_range_free(amdgpu_va_handle va_range_handle); 1139 1140 /** 1141 * Query virtual address range 1142 * 1143 * UMD can query GPU VM range supported by each device 1144 * to initialize its own VAM accordingly. 1145 * 1146 * \param dev - [in] Device handle. See #amdgpu_device_initialize() 1147 * \param type - \c [in] Type of virtual address range 1148 * \param offset - \c [out] Start offset of virtual address range 1149 * \param size - \c [out] Size of virtual address range 1150 * 1151 * \return 0 on success\n 1152 * <0 - Negative POSIX Error code 1153 * 1154 */ 1155 1156 int amdgpu_va_range_query(amdgpu_device_handle dev, 1157 enum amdgpu_gpu_va_range type, 1158 uint64_t *start, 1159 uint64_t *end); 1160 1161 /** 1162 * VA mapping/unmapping for the buffer object 1163 * 1164 * \param bo - \c [in] BO handle 1165 * \param offset - \c [in] Start offset to map 1166 * \param size - \c [in] Size to map 1167 * \param addr - \c [in] Start virtual address. 1168 * \param flags - \c [in] Supported flags for mapping/unmapping 1169 * \param ops - \c [in] AMDGPU_VA_OP_MAP or AMDGPU_VA_OP_UNMAP 1170 * 1171 * \return 0 on success\n 1172 * <0 - Negative POSIX Error code 1173 * 1174 */ 1175 1176 int amdgpu_bo_va_op(amdgpu_bo_handle bo, 1177 uint64_t offset, 1178 uint64_t size, 1179 uint64_t addr, 1180 uint64_t flags, 1181 uint32_t ops); 1182 1183 #endif /* #ifdef _AMDGPU_H_ */ 1184