projects/linux·blob master · refs

564 lines · 16.0 KiB

1/* SPDX-License-Identifier: GPL-2.0-only WITH Linux-syscall-note */
2/*
3 * Copyright (C) 2020-2025 Intel Corporation
4 */
5
6#ifndef __UAPI_IVPU_DRM_H__
7#define __UAPI_IVPU_DRM_H__
8
9#include "drm.h"
10
11#if defined(__cplusplus)
12extern "C" {
13#endif
14
15#define DRM_IVPU_GET_PARAM 0x00
16#define DRM_IVPU_SET_PARAM 0x01
17#define DRM_IVPU_BO_CREATE 0x02
18#define DRM_IVPU_BO_INFO 0x03
19#define DRM_IVPU_SUBMIT 0x05
20#define DRM_IVPU_BO_WAIT 0x06
21#define DRM_IVPU_METRIC_STREAMER_START 0x07
22#define DRM_IVPU_METRIC_STREAMER_STOP 0x08
23#define DRM_IVPU_METRIC_STREAMER_GET_DATA 0x09
24#define DRM_IVPU_METRIC_STREAMER_GET_INFO 0x0a
25#define DRM_IVPU_CMDQ_CREATE 0x0b
26#define DRM_IVPU_CMDQ_DESTROY 0x0c
27#define DRM_IVPU_CMDQ_SUBMIT 0x0d
28#define DRM_IVPU_BO_CREATE_FROM_USERPTR 0x0e
29
30#define DRM_IOCTL_IVPU_GET_PARAM \
31 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_GET_PARAM, struct drm_ivpu_param)
32
33#define DRM_IOCTL_IVPU_SET_PARAM \
34 DRM_IOW(DRM_COMMAND_BASE + DRM_IVPU_SET_PARAM, struct drm_ivpu_param)
35
36#define DRM_IOCTL_IVPU_BO_CREATE \
37 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_BO_CREATE, struct drm_ivpu_bo_create)
38
39#define DRM_IOCTL_IVPU_BO_INFO \
40 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_BO_INFO, struct drm_ivpu_bo_info)
41
42#define DRM_IOCTL_IVPU_SUBMIT \
43 DRM_IOW(DRM_COMMAND_BASE + DRM_IVPU_SUBMIT, struct drm_ivpu_submit)
44
45#define DRM_IOCTL_IVPU_BO_WAIT \
46 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_BO_WAIT, struct drm_ivpu_bo_wait)
47
48#define DRM_IOCTL_IVPU_METRIC_STREAMER_START \
49 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_METRIC_STREAMER_START, \
50 struct drm_ivpu_metric_streamer_start)
51
52#define DRM_IOCTL_IVPU_METRIC_STREAMER_STOP \
53 DRM_IOW(DRM_COMMAND_BASE + DRM_IVPU_METRIC_STREAMER_STOP, \
54 struct drm_ivpu_metric_streamer_stop)
55
56#define DRM_IOCTL_IVPU_METRIC_STREAMER_GET_DATA \
57 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_METRIC_STREAMER_GET_DATA, \
58 struct drm_ivpu_metric_streamer_get_data)
59
60#define DRM_IOCTL_IVPU_METRIC_STREAMER_GET_INFO \
61 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_METRIC_STREAMER_GET_INFO, \
62 struct drm_ivpu_metric_streamer_get_data)
63
64#define DRM_IOCTL_IVPU_CMDQ_CREATE \
65 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_CMDQ_CREATE, struct drm_ivpu_cmdq_create)
66
67#define DRM_IOCTL_IVPU_CMDQ_DESTROY \
68 DRM_IOW(DRM_COMMAND_BASE + DRM_IVPU_CMDQ_DESTROY, struct drm_ivpu_cmdq_destroy)
69
70#define DRM_IOCTL_IVPU_CMDQ_SUBMIT \
71 DRM_IOW(DRM_COMMAND_BASE + DRM_IVPU_CMDQ_SUBMIT, struct drm_ivpu_cmdq_submit)
72
73#define DRM_IOCTL_IVPU_BO_CREATE_FROM_USERPTR \
74 DRM_IOWR(DRM_COMMAND_BASE + DRM_IVPU_BO_CREATE_FROM_USERPTR, \
75 struct drm_ivpu_bo_create_from_userptr)
76
77/**
78 * DOC: contexts
79 *
80 * VPU contexts have private virtual address space, job queues and priority.
81 * Each context is identified by an unique ID. Context is created on open().
82 */
83
84#define DRM_IVPU_PARAM_DEVICE_ID 0
85#define DRM_IVPU_PARAM_DEVICE_REVISION 1
86#define DRM_IVPU_PARAM_PLATFORM_TYPE 2
87#define DRM_IVPU_PARAM_CORE_CLOCK_RATE 3
88#define DRM_IVPU_PARAM_NUM_CONTEXTS 4
89#define DRM_IVPU_PARAM_CONTEXT_BASE_ADDRESS 5
90#define DRM_IVPU_PARAM_CONTEXT_PRIORITY 6 /* Deprecated */
91#define DRM_IVPU_PARAM_CONTEXT_ID 7
92#define DRM_IVPU_PARAM_FW_API_VERSION 8
93#define DRM_IVPU_PARAM_ENGINE_HEARTBEAT 9
94#define DRM_IVPU_PARAM_UNIQUE_INFERENCE_ID 10
95#define DRM_IVPU_PARAM_TILE_CONFIG 11
96#define DRM_IVPU_PARAM_SKU 12
97#define DRM_IVPU_PARAM_CAPABILITIES 13
98#define DRM_IVPU_PARAM_PREEMPT_BUFFER_SIZE 14
99
100#define DRM_IVPU_PLATFORM_TYPE_SILICON 0
101
102/* Deprecated, use DRM_IVPU_JOB_PRIORITY */
103#define DRM_IVPU_CONTEXT_PRIORITY_IDLE 0
104#define DRM_IVPU_CONTEXT_PRIORITY_NORMAL 1
105#define DRM_IVPU_CONTEXT_PRIORITY_FOCUS 2
106#define DRM_IVPU_CONTEXT_PRIORITY_REALTIME 3
107
108#define DRM_IVPU_JOB_PRIORITY_DEFAULT 0
109#define DRM_IVPU_JOB_PRIORITY_IDLE 1
110#define DRM_IVPU_JOB_PRIORITY_NORMAL 2
111#define DRM_IVPU_JOB_PRIORITY_FOCUS 3
112#define DRM_IVPU_JOB_PRIORITY_REALTIME 4
113
114/**
115 * DRM_IVPU_CAP_METRIC_STREAMER
116 *
117 * Metric streamer support. Provides sampling of various hardware performance
118 * metrics like DMA bandwidth and cache miss/hits. Can be used for profiling.
119 */
120#define DRM_IVPU_CAP_METRIC_STREAMER 1
121/**
122 * DRM_IVPU_CAP_DMA_MEMORY_RANGE
123 *
124 * Driver has capability to allocate separate memory range
125 * accessible by hardware DMA.
126 */
127#define DRM_IVPU_CAP_DMA_MEMORY_RANGE 2
128/**
129 * DRM_IVPU_CAP_MANAGE_CMDQ
130 *
131 * Driver supports explicit command queue operations like command queue create,
132 * command queue destroy and submit job on specific command queue.
133 */
134#define DRM_IVPU_CAP_MANAGE_CMDQ 3
135/**
136 * DRM_IVPU_CAP_BO_CREATE_FROM_USERPTR
137 *
138 * Driver supports creating buffer objects from user space memory pointers.
139 * This allows creating GEM buffers from existing user memory regions.
140 */
141#define DRM_IVPU_CAP_BO_CREATE_FROM_USERPTR 4
142
143/**
144 * struct drm_ivpu_param - Get/Set VPU parameters
145 */
146struct drm_ivpu_param {
147 /**
148 * @param:
149 *
150 * Supported params:
151 *
152 * %DRM_IVPU_PARAM_DEVICE_ID:
153 * PCI Device ID of the VPU device (read-only)
154 *
155 * %DRM_IVPU_PARAM_DEVICE_REVISION:
156 * VPU device revision (read-only)
157 *
158 * %DRM_IVPU_PARAM_PLATFORM_TYPE:
159 * Returns %DRM_IVPU_PLATFORM_TYPE_SILICON on real hardware or device specific
160 * platform type when executing on a simulator or emulator (read-only)
161 *
162 * %DRM_IVPU_PARAM_CORE_CLOCK_RATE:
163 * Maximum frequency of the NPU data processing unit clock (read-only)
164 *
165 * %DRM_IVPU_PARAM_NUM_CONTEXTS:
166 * Maximum number of simultaneously existing contexts (read-only)
167 *
168 * %DRM_IVPU_PARAM_CONTEXT_BASE_ADDRESS:
169 * Lowest VPU virtual address available in the current context (read-only)
170 *
171 * %DRM_IVPU_PARAM_CONTEXT_ID:
172 * Current context ID, always greater than 0 (read-only)
173 *
174 * %DRM_IVPU_PARAM_FW_API_VERSION:
175 * Firmware API version array (read-only)
176 *
177 * %DRM_IVPU_PARAM_ENGINE_HEARTBEAT:
178 * Heartbeat value from an engine (read-only).
179 * Engine ID (i.e. DRM_IVPU_ENGINE_COMPUTE) is given via index.
180 *
181 * %DRM_IVPU_PARAM_UNIQUE_INFERENCE_ID:
182 * Device-unique inference ID (read-only)
183 *
184 * %DRM_IVPU_PARAM_TILE_CONFIG:
185 * VPU tile configuration (read-only)
186 *
187 * %DRM_IVPU_PARAM_SKU:
188 * VPU SKU ID (read-only)
189 *
190 * %DRM_IVPU_PARAM_CAPABILITIES:
191 * Supported capabilities (read-only)
192 *
193 * %DRM_IVPU_PARAM_PREEMPT_BUFFER_SIZE:
194 * Size of the preemption buffer (read-only)
195 */
196 __u32 param;
197
198 /** @index: Index for params that have multiple instances */
199 __u32 index;
200
201 /** @value: Param value */
202 __u64 value;
203};
204
205#define DRM_IVPU_BO_SHAVE_MEM 0x00000001
206#define DRM_IVPU_BO_HIGH_MEM DRM_IVPU_BO_SHAVE_MEM
207#define DRM_IVPU_BO_MAPPABLE 0x00000002
208#define DRM_IVPU_BO_DMA_MEM 0x00000004
209#define DRM_IVPU_BO_READ_ONLY 0x00000008
210
211#define DRM_IVPU_BO_CACHED 0x00000000
212#define DRM_IVPU_BO_UNCACHED 0x00010000
213#define DRM_IVPU_BO_WC 0x00020000
214#define DRM_IVPU_BO_CACHE_MASK 0x00030000
215
216#define DRM_IVPU_BO_FLAGS \
217 (DRM_IVPU_BO_HIGH_MEM | \
218 DRM_IVPU_BO_MAPPABLE | \
219 DRM_IVPU_BO_DMA_MEM | \
220 DRM_IVPU_BO_READ_ONLY | \
221 DRM_IVPU_BO_CACHE_MASK)
222
223/**
224 * struct drm_ivpu_bo_create - Create BO backed by SHMEM
225 *
226 * Create GEM buffer object allocated in SHMEM memory.
227 */
228struct drm_ivpu_bo_create {
229 /** @size: The size in bytes of the allocated memory */
230 __u64 size;
231
232 /**
233 * @flags:
234 *
235 * Supported flags:
236 *
237 * %DRM_IVPU_BO_HIGH_MEM:
238 *
239 * Allocate VPU address from >4GB range.
240 * Buffer object with vpu address >4GB can be always accessed by the
241 * VPU DMA engine, but some HW generation may not be able to access
242 * this memory from then firmware running on the VPU management processor.
243 * Suitable for input, output and some scratch buffers.
244 *
245 * %DRM_IVPU_BO_MAPPABLE:
246 *
247 * Buffer object can be mapped using mmap().
248 *
249 * %DRM_IVPU_BO_CACHED:
250 *
251 * Allocated BO will be cached on host side (WB) and snooped on the VPU side.
252 * This is the default caching mode.
253 *
254 * %DRM_IVPU_BO_UNCACHED:
255 *
256 * Not supported. Use DRM_IVPU_BO_WC instead.
257 *
258 * %DRM_IVPU_BO_WC:
259 *
260 * Allocated BO will use write combining buffer for writes but reads will be
261 * uncached.
262 */
263 __u32 flags;
264
265 /** @handle: Returned GEM object handle */
266 __u32 handle;
267
268 /** @vpu_addr: Returned VPU virtual address */
269 __u64 vpu_addr;
270};
271
272/**
273 * struct drm_ivpu_bo_create_from_userptr - Create dma-buf from user pointer
274 *
275 * Create a GEM buffer object from a user pointer to a memory region.
276 */
277struct drm_ivpu_bo_create_from_userptr {
278 /** @user_ptr: User pointer to memory region (must be page aligned) */
279 __u64 user_ptr;
280
281 /** @size: Size of the memory region in bytes (must be page aligned) */
282 __u64 size;
283
284 /**
285 * @flags:
286 *
287 * Supported flags:
288 *
289 * %DRM_IVPU_BO_HIGH_MEM:
290 *
291 * Allocate VPU address from >4GB range.
292 *
293 * %DRM_IVPU_BO_DMA_MEM:
294 *
295 * Allocate from DMA memory range accessible by hardware DMA.
296 *
297 * %DRM_IVPU_BO_READ_ONLY:
298 *
299 * Allocate as a read-only buffer object.
300 */
301 __u32 flags;
302
303 /** @handle: Returned GEM object handle */
304 __u32 handle;
305
306 /** @vpu_addr: Returned VPU virtual address */
307 __u64 vpu_addr;
308};
309
310/**
311 * struct drm_ivpu_bo_info - Query buffer object info
312 */
313struct drm_ivpu_bo_info {
314 /** @handle: Handle of the queried BO */
315 __u32 handle;
316
317 /** @flags: Returned flags used to create the BO */
318 __u32 flags;
319
320 /** @vpu_addr: Returned VPU virtual address */
321 __u64 vpu_addr;
322
323 /**
324 * @mmap_offset:
325 *
326 * Returned offset to be used in mmap(). 0 in case the BO is not mappable.
327 */
328 __u64 mmap_offset;
329
330 /** @size: Returned GEM object size, aligned to PAGE_SIZE */
331 __u64 size;
332};
333
334/* drm_ivpu_submit engines */
335#define DRM_IVPU_ENGINE_COMPUTE 0
336#define DRM_IVPU_ENGINE_COPY 1 /* Deprecated */
337
338/**
339 * struct drm_ivpu_submit - Submit commands to the VPU
340 *
341 * Execute a single command buffer on a given VPU engine.
342 * Handles to all referenced buffer objects have to be provided in @buffers_ptr.
343 *
344 * User space may wait on job completion using %DRM_IVPU_BO_WAIT ioctl.
345 */
346struct drm_ivpu_submit {
347 /**
348 * @buffers_ptr:
349 *
350 * A pointer to an u32 array of GEM handles of the BOs required for this job.
351 * The number of elements in the array must be equal to the value given by @buffer_count.
352 *
353 * The first BO is the command buffer. The rest of array has to contain all
354 * BOs referenced from the command buffer.
355 */
356 __u64 buffers_ptr;
357
358 /** @buffer_count: Number of elements in the @buffers_ptr */
359 __u32 buffer_count;
360
361 /**
362 * @engine: Select the engine this job should be executed on
363 *
364 * %DRM_IVPU_ENGINE_COMPUTE:
365 *
366 * Performs Deep Learning Neural Compute Inference Operations
367 */
368 __u32 engine;
369
370 /** @flags: Reserved for future use - must be zero */
371 __u32 flags;
372
373 /**
374 * @commands_offset:
375 *
376 * Offset inside the first buffer in @buffers_ptr containing commands
377 * to be executed. The offset has to be 8-byte aligned.
378 */
379 __u32 commands_offset;
380
381 /**
382 * @priority:
383 *
384 * Priority to be set for related job command queue, can be one of the following:
385 * %DRM_IVPU_JOB_PRIORITY_DEFAULT
386 * %DRM_IVPU_JOB_PRIORITY_IDLE
387 * %DRM_IVPU_JOB_PRIORITY_NORMAL
388 * %DRM_IVPU_JOB_PRIORITY_FOCUS
389 * %DRM_IVPU_JOB_PRIORITY_REALTIME
390 */
391 __u32 priority;
392};
393
394/**
395 * struct drm_ivpu_cmdq_submit - Submit commands to the VPU using explicit command queue
396 *
397 * Execute a single command buffer on a given command queue.
398 * Handles to all referenced buffer objects have to be provided in @buffers_ptr.
399 *
400 * User space may wait on job completion using %DRM_IVPU_BO_WAIT ioctl.
401 */
402struct drm_ivpu_cmdq_submit {
403 /**
404 * @buffers_ptr:
405 *
406 * A pointer to an u32 array of GEM handles of the BOs required for this job.
407 * The number of elements in the array must be equal to the value given by @buffer_count.
408 *
409 * The first BO is the command buffer. The rest of array has to contain all
410 * BOs referenced from the command buffer.
411 */
412 __u64 buffers_ptr;
413
414 /** @buffer_count: Number of elements in the @buffers_ptr */
415 __u32 buffer_count;
416
417 /** @cmdq_id: ID for the command queue where job will be submitted */
418 __u32 cmdq_id;
419
420 /** @flags: Reserved for future use - must be zero */
421 __u32 flags;
422
423 /**
424 * @commands_offset:
425 *
426 * Offset inside the first buffer in @buffers_ptr containing commands
427 * to be executed. The offset has to be 8-byte aligned.
428 */
429 __u32 commands_offset;
430 /**
431 * @preempt_buffer_index:
432 *
433 * Index of the preemption buffer in the buffers_ptr array.
434 */
435 __u32 preempt_buffer_index;
436 __u32 reserved;
437};
438
439/* drm_ivpu_bo_wait job status codes */
440#define DRM_IVPU_JOB_STATUS_SUCCESS 0
441#define DRM_IVPU_JOB_STATUS_ABORTED 256
442
443/**
444 * struct drm_ivpu_bo_wait - Wait for BO to become inactive
445 *
446 * Blocks until a given buffer object becomes inactive.
447 * With @timeout_ms set to 0 returns immediately.
448 */
449struct drm_ivpu_bo_wait {
450 /** @handle: Handle to the buffer object to be waited on */
451 __u32 handle;
452
453 /** @flags: Reserved for future use - must be zero */
454 __u32 flags;
455
456 /** @timeout_ns: Absolute timeout in nanoseconds (may be zero) */
457 __s64 timeout_ns;
458
459 /**
460 * @job_status:
461 *
462 * Job status code which is updated after the job is completed.
463 * &DRM_IVPU_JOB_STATUS_SUCCESS or device specific error otherwise.
464 * Valid only if @handle points to a command buffer.
465 */
466 __u32 job_status;
467
468 /** @pad: Padding - must be zero */
469 __u32 pad;
470};
471
472/**
473 * struct drm_ivpu_metric_streamer_start - Start collecting metric data
474 */
475struct drm_ivpu_metric_streamer_start {
476 /** @metric_group_mask: Indicates metric streamer instance */
477 __u64 metric_group_mask;
478 /** @sampling_period_ns: Sampling period in nanoseconds */
479 __u64 sampling_period_ns;
480 /**
481 * @read_period_samples:
482 *
483 * Number of samples after which user space will try to read the data.
484 * Reading the data after significantly longer period may cause data loss.
485 */
486 __u32 read_period_samples;
487 /** @sample_size: Returned size of a single sample in bytes */
488 __u32 sample_size;
489 /** @max_data_size: Returned max @data_size from %DRM_IOCTL_IVPU_METRIC_STREAMER_GET_DATA */
490 __u32 max_data_size;
491};
492
493/**
494 * struct drm_ivpu_metric_streamer_get_data - Copy collected metric data
495 */
496struct drm_ivpu_metric_streamer_get_data {
497 /** @metric_group_mask: Indicates metric streamer instance */
498 __u64 metric_group_mask;
499 /** @buffer_ptr: A pointer to a destination for the copied data */
500 __u64 buffer_ptr;
501 /** @buffer_size: Size of the destination buffer */
502 __u64 buffer_size;
503 /**
504 * @data_size: Returned size of copied metric data
505 *
506 * If the @buffer_size is zero, returns the amount of data ready to be copied.
507 */
508 __u64 data_size;
509};
510
511/* Command queue flags */
512#define DRM_IVPU_CMDQ_FLAG_TURBO 0x00000001
513
514/**
515 * struct drm_ivpu_cmdq_create - Create command queue for job submission
516 */
517struct drm_ivpu_cmdq_create {
518 /** @cmdq_id: Returned ID of created command queue */
519 __u32 cmdq_id;
520 /**
521 * @priority:
522 *
523 * Priority to be set for related job command queue, can be one of the following:
524 * %DRM_IVPU_JOB_PRIORITY_DEFAULT
525 * %DRM_IVPU_JOB_PRIORITY_IDLE
526 * %DRM_IVPU_JOB_PRIORITY_NORMAL
527 * %DRM_IVPU_JOB_PRIORITY_FOCUS
528 * %DRM_IVPU_JOB_PRIORITY_REALTIME
529 */
530 __u32 priority;
531 /**
532 * @flags:
533 *
534 * Supported flags:
535 *
536 * %DRM_IVPU_CMDQ_FLAG_TURBO
537 *
538 * Enable low-latency mode for the command queue. The NPU will maximize performance
539 * when executing jobs from such queue at the cost of increased power usage.
540 */
541 __u32 flags;
542};
543
544/**
545 * struct drm_ivpu_cmdq_destroy - Destroy a command queue
546 */
547struct drm_ivpu_cmdq_destroy {
548 /** @cmdq_id: ID of command queue to destroy */
549 __u32 cmdq_id;
550};
551
552/**
553 * struct drm_ivpu_metric_streamer_stop - Stop collecting metric data
554 */
555struct drm_ivpu_metric_streamer_stop {
556 /** @metric_group_mask: Indicates metric streamer instance */
557 __u64 metric_group_mask;
558};
559
560#if defined(__cplusplus)
561}
562#endif
563
564#endif /* __UAPI_IVPU_DRM_H__ */