2 * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
3 * Copyright 2010 Marek Olšák <maraeo@gmail.com>
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE. */
24 #ifndef RADEON_WINSYS_H
25 #define RADEON_WINSYS_H
27 /* The public winsys interface header for the radeon driver. */
29 #include "pipebuffer/pb_buffer.h"
31 #include "amd/common/ac_gpu_info.h"
32 #include "amd/common/ac_surface.h"
34 #define RADEON_FLUSH_ASYNC (1 << 0)
35 #define RADEON_FLUSH_END_OF_FRAME (1 << 1)
38 enum radeon_bo_layout
{
39 RADEON_LAYOUT_LINEAR
= 0,
41 RADEON_LAYOUT_SQUARETILED
,
46 enum radeon_bo_domain
{ /* bitfield */
47 RADEON_DOMAIN_GTT
= 2,
48 RADEON_DOMAIN_VRAM
= 4,
49 RADEON_DOMAIN_VRAM_GTT
= RADEON_DOMAIN_VRAM
| RADEON_DOMAIN_GTT
52 enum radeon_bo_flag
{ /* bitfield */
53 RADEON_FLAG_GTT_WC
= (1 << 0),
54 RADEON_FLAG_NO_CPU_ACCESS
= (1 << 1),
55 RADEON_FLAG_NO_SUBALLOC
= (1 << 2),
56 RADEON_FLAG_SPARSE
= (1 << 3),
59 enum radeon_bo_usage
{ /* bitfield */
60 RADEON_USAGE_READ
= 2,
61 RADEON_USAGE_WRITE
= 4,
62 RADEON_USAGE_READWRITE
= RADEON_USAGE_READ
| RADEON_USAGE_WRITE
,
64 /* The winsys ensures that the CS submission will be scheduled after
65 * previously flushed CSs referencing this BO in a conflicting way.
67 RADEON_USAGE_SYNCHRONIZED
= 8
70 #define RADEON_SPARSE_PAGE_SIZE (64 * 1024)
83 enum radeon_value_id
{
84 RADEON_REQUESTED_VRAM_MEMORY
,
85 RADEON_REQUESTED_GTT_MEMORY
,
88 RADEON_BUFFER_WAIT_TIME_NS
,
89 RADEON_NUM_MAPPED_BUFFERS
,
93 RADEON_GFX_BO_LIST_COUNTER
, /* number of BOs submitted in gfx IBs */
94 RADEON_NUM_BYTES_MOVED
,
96 RADEON_NUM_VRAM_CPU_PAGE_FAULTS
,
98 RADEON_VRAM_VIS_USAGE
,
100 RADEON_GPU_TEMPERATURE
, /* DRM 2.42.0 */
103 RADEON_GPU_RESET_COUNTER
, /* DRM 2.43.0 */
104 RADEON_CS_THREAD_TIME
,
107 /* Each group of four has the same priority. */
108 enum radeon_bo_priority
{
109 RADEON_PRIO_FENCE
= 0,
111 RADEON_PRIO_SO_FILLED_SIZE
,
114 RADEON_PRIO_IB1
= 4, /* main IB submitted to the kernel */
115 RADEON_PRIO_IB2
, /* IB executed with INDIRECT_BUFFER */
116 RADEON_PRIO_DRAW_INDIRECT
,
117 RADEON_PRIO_INDEX_BUFFER
,
121 RADEON_PRIO_SDMA_BUFFER
,
122 RADEON_PRIO_SDMA_TEXTURE
,
124 RADEON_PRIO_CP_DMA
= 12,
126 RADEON_PRIO_CONST_BUFFER
= 16,
127 RADEON_PRIO_DESCRIPTORS
,
128 RADEON_PRIO_BORDER_COLORS
,
130 RADEON_PRIO_SAMPLER_BUFFER
= 20,
131 RADEON_PRIO_VERTEX_BUFFER
,
133 RADEON_PRIO_SHADER_RW_BUFFER
= 24,
134 RADEON_PRIO_COMPUTE_GLOBAL
,
136 RADEON_PRIO_SAMPLER_TEXTURE
= 28,
137 RADEON_PRIO_SHADER_RW_IMAGE
,
139 RADEON_PRIO_SAMPLER_TEXTURE_MSAA
= 32,
141 RADEON_PRIO_COLOR_BUFFER
= 36,
143 RADEON_PRIO_DEPTH_BUFFER
= 40,
145 RADEON_PRIO_COLOR_BUFFER_MSAA
= 44,
147 RADEON_PRIO_DEPTH_BUFFER_MSAA
= 48,
149 RADEON_PRIO_CMASK
= 52,
152 RADEON_PRIO_SHADER_BINARY
, /* the hw can't hide instruction cache misses */
154 RADEON_PRIO_SHADER_RINGS
= 56,
156 RADEON_PRIO_SCRATCH_BUFFER
= 60,
157 /* 63 is the maximum value */
160 struct winsys_handle
;
161 struct radeon_winsys_ctx
;
163 struct radeon_winsys_cs_chunk
{
164 unsigned cdw
; /* Number of used dwords. */
165 unsigned max_dw
; /* Maximum number of dwords. */
166 uint32_t *buf
; /* The base pointer of the chunk. */
169 struct radeon_winsys_cs
{
170 struct radeon_winsys_cs_chunk current
;
171 struct radeon_winsys_cs_chunk
*prev
;
172 unsigned num_prev
; /* Number of previous chunks. */
173 unsigned max_prev
; /* Space in array pointed to by prev. */
174 unsigned prev_dw
; /* Total number of dwords in previous chunks. */
176 /* Memory usage of the buffer list. These are always 0 for preamble IBs. */
181 /* Tiling info for display code, DRI sharing, and other data. */
182 struct radeon_bo_metadata
{
183 /* Tiling flags describing the texture layout for display code
188 enum radeon_bo_layout microtile
;
189 enum radeon_bo_layout macrotile
;
190 unsigned pipe_config
;
202 unsigned swizzle_mode
:5;
206 /* Additional metadata associated with the buffer, in bytes.
207 * The maximum size is 64 * 4. This is opaque for the winsys & kernel.
208 * Supported by amdgpu only.
210 uint32_t size_metadata
;
211 uint32_t metadata
[64];
214 enum radeon_feature_id
{
215 RADEON_FID_R300_HYPERZ_ACCESS
, /* ZMask + HiZ */
216 RADEON_FID_R300_CMASK_ACCESS
,
219 struct radeon_bo_list_item
{
222 uint64_t priority_usage
; /* mask of (1 << RADEON_PRIO_*) */
225 struct radeon_winsys
{
227 * The screen object this winsys was created for
229 struct pipe_screen
*screen
;
232 * Decrement the winsys reference count.
234 * \param ws The winsys this function is called for.
235 * \return True if the winsys and screen should be destroyed.
237 bool (*unref
)(struct radeon_winsys
*ws
);
240 * Destroy this winsys.
242 * \param ws The winsys this function is called from.
244 void (*destroy
)(struct radeon_winsys
*ws
);
247 * Query an info structure from winsys.
249 * \param ws The winsys this function is called from.
250 * \param info Return structure
252 void (*query_info
)(struct radeon_winsys
*ws
,
253 struct radeon_info
*info
);
255 /**************************************************************************
256 * Buffer management. Buffer attributes are mostly fixed over its lifetime.
258 * Remember that gallium gets to choose the interface it needs, and the
259 * window systems must then implement that interface (rather than the
260 * other way around...).
261 *************************************************************************/
264 * Create a buffer object.
266 * \param ws The winsys this function is called from.
267 * \param size The size to allocate.
268 * \param alignment An alignment of the buffer in memory.
269 * \param use_reusable_pool Whether the cache buffer manager should be used.
270 * \param domain A bitmask of the RADEON_DOMAIN_* flags.
271 * \return The created buffer object.
273 struct pb_buffer
*(*buffer_create
)(struct radeon_winsys
*ws
,
276 enum radeon_bo_domain domain
,
277 enum radeon_bo_flag flags
);
280 * Map the entire data store of a buffer object into the client's address
283 * \param buf A winsys buffer object to map.
284 * \param cs A command stream to flush if the buffer is referenced by it.
285 * \param usage A bitmask of the PIPE_TRANSFER_* flags.
286 * \return The pointer at the beginning of the buffer.
288 void *(*buffer_map
)(struct pb_buffer
*buf
,
289 struct radeon_winsys_cs
*cs
,
290 enum pipe_transfer_usage usage
);
293 * Unmap a buffer object from the client's address space.
295 * \param buf A winsys buffer object to unmap.
297 void (*buffer_unmap
)(struct pb_buffer
*buf
);
300 * Wait for the buffer and return true if the buffer is not used
303 * The timeout of 0 will only return the status.
304 * The timeout of PIPE_TIMEOUT_INFINITE will always wait until the buffer
307 bool (*buffer_wait
)(struct pb_buffer
*buf
, uint64_t timeout
,
308 enum radeon_bo_usage usage
);
311 * Return buffer metadata.
312 * (tiling info for display code, DRI sharing, and other data)
314 * \param buf A winsys buffer object to get the flags from.
317 void (*buffer_get_metadata
)(struct pb_buffer
*buf
,
318 struct radeon_bo_metadata
*md
);
321 * Set buffer metadata.
322 * (tiling info for display code, DRI sharing, and other data)
324 * \param buf A winsys buffer object to set the flags for.
327 void (*buffer_set_metadata
)(struct pb_buffer
*buf
,
328 struct radeon_bo_metadata
*md
);
331 * Get a winsys buffer from a winsys handle. The internal structure
332 * of the handle is platform-specific and only a winsys should access it.
334 * \param ws The winsys this function is called from.
335 * \param whandle A winsys handle pointer as was received from a state
337 * \param stride The returned buffer stride in bytes.
339 struct pb_buffer
*(*buffer_from_handle
)(struct radeon_winsys
*ws
,
340 struct winsys_handle
*whandle
,
341 unsigned *stride
, unsigned *offset
);
344 * Get a winsys buffer from a user pointer. The resulting buffer can't
345 * be exported. Both pointer and size must be page aligned.
347 * \param ws The winsys this function is called from.
348 * \param pointer User pointer to turn into a buffer object.
349 * \param Size Size in bytes for the new buffer.
351 struct pb_buffer
*(*buffer_from_ptr
)(struct radeon_winsys
*ws
,
352 void *pointer
, uint64_t size
);
355 * Whether the buffer was created from a user pointer.
357 * \param buf A winsys buffer object
358 * \return whether \p buf was created via buffer_from_ptr
360 bool (*buffer_is_user_ptr
)(struct pb_buffer
*buf
);
362 /** Whether the buffer was suballocated. */
363 bool (*buffer_is_suballocated
)(struct pb_buffer
*buf
);
366 * Get a winsys handle from a winsys buffer. The internal structure
367 * of the handle is platform-specific and only a winsys should access it.
369 * \param buf A winsys buffer object to get the handle from.
370 * \param whandle A winsys handle pointer.
371 * \param stride A stride of the buffer in bytes, for texturing.
372 * \return true on success.
374 bool (*buffer_get_handle
)(struct pb_buffer
*buf
,
375 unsigned stride
, unsigned offset
,
377 struct winsys_handle
*whandle
);
380 * Change the commitment of a (64KB-page aligned) region of the given
383 * \warning There is no automatic synchronization with command submission.
385 * \note Only implemented by the amdgpu winsys.
387 * \return false on out of memory or other failure, true on success.
389 bool (*buffer_commit
)(struct pb_buffer
*buf
,
390 uint64_t offset
, uint64_t size
,
394 * Return the virtual address of a buffer.
396 * When virtual memory is not in use, this is the offset relative to the
397 * relocation base (non-zero for sub-allocated buffers).
399 * \param buf A winsys buffer object
400 * \return virtual address
402 uint64_t (*buffer_get_virtual_address
)(struct pb_buffer
*buf
);
405 * Return the offset of this buffer relative to the relocation base.
406 * This is only non-zero for sub-allocated buffers.
408 * This is only supported in the radeon winsys, since amdgpu uses virtual
409 * addresses in submissions even for the video engines.
411 * \param buf A winsys buffer object
412 * \return the offset for relocations
414 unsigned (*buffer_get_reloc_offset
)(struct pb_buffer
*buf
);
417 * Query the initial placement of the buffer from the kernel driver.
419 enum radeon_bo_domain (*buffer_get_initial_domain
)(struct pb_buffer
*buf
);
421 /**************************************************************************
422 * Command submission.
424 * Each pipe context should create its own command stream and submit
425 * commands independently of other contexts.
426 *************************************************************************/
429 * Create a command submission context.
430 * Various command streams can be submitted to the same context.
432 struct radeon_winsys_ctx
*(*ctx_create
)(struct radeon_winsys
*ws
);
437 void (*ctx_destroy
)(struct radeon_winsys_ctx
*ctx
);
440 * Query a GPU reset status.
442 enum pipe_reset_status (*ctx_query_reset_status
)(struct radeon_winsys_ctx
*ctx
);
445 * Create a command stream.
447 * \param ctx The submission context
448 * \param ring_type The ring type (GFX, DMA, UVD)
449 * \param flush Flush callback function associated with the command stream.
450 * \param user User pointer that will be passed to the flush callback.
452 struct radeon_winsys_cs
*(*cs_create
)(struct radeon_winsys_ctx
*ctx
,
453 enum ring_type ring_type
,
454 void (*flush
)(void *ctx
, unsigned flags
,
455 struct pipe_fence_handle
**fence
),
459 * Destroy a command stream.
461 * \param cs A command stream to destroy.
463 void (*cs_destroy
)(struct radeon_winsys_cs
*cs
);
466 * Add a buffer. Each buffer used by a CS must be added using this function.
468 * \param cs Command stream
470 * \param usage Whether the buffer is used for read and/or write.
471 * \param domain Bitmask of the RADEON_DOMAIN_* flags.
472 * \param priority A higher number means a greater chance of being
473 * placed in the requested domain. 15 is the maximum.
474 * \return Buffer index.
476 unsigned (*cs_add_buffer
)(struct radeon_winsys_cs
*cs
,
477 struct pb_buffer
*buf
,
478 enum radeon_bo_usage usage
,
479 enum radeon_bo_domain domain
,
480 enum radeon_bo_priority priority
);
483 * Return the index of an already-added buffer.
485 * Not supported on amdgpu. Drivers with GPUVM should not care about
488 * \param cs Command stream
490 * \return The buffer index, or -1 if the buffer has not been added.
492 int (*cs_lookup_buffer
)(struct radeon_winsys_cs
*cs
,
493 struct pb_buffer
*buf
);
496 * Return true if there is enough memory in VRAM and GTT for the buffers
497 * added so far. If the validation fails, all buffers which have
498 * been added since the last call of cs_validate will be removed and
499 * the CS will be flushed (provided there are still any buffers).
501 * \param cs A command stream to validate.
503 bool (*cs_validate
)(struct radeon_winsys_cs
*cs
);
506 * Check whether the given number of dwords is available in the IB.
507 * Optionally chain a new chunk of the IB if necessary and supported.
509 * \param cs A command stream.
510 * \param dw Number of CS dwords requested by the caller.
512 bool (*cs_check_space
)(struct radeon_winsys_cs
*cs
, unsigned dw
);
515 * Return the buffer list.
517 * This is the buffer list as passed to the kernel, i.e. it only contains
518 * the parent buffers of sub-allocated buffers.
520 * \param cs Command stream
521 * \param list Returned buffer list. Set to NULL to query the count only.
522 * \return The buffer count.
524 unsigned (*cs_get_buffer_list
)(struct radeon_winsys_cs
*cs
,
525 struct radeon_bo_list_item
*list
);
528 * Flush a command stream.
530 * \param cs A command stream to flush.
531 * \param flags, RADEON_FLUSH_ASYNC or 0.
532 * \param fence Pointer to a fence. If non-NULL, a fence is inserted
533 * after the CS and is returned through this parameter.
534 * \return Negative POSIX error code or 0 for success.
535 * Asynchronous submissions never return an error.
537 int (*cs_flush
)(struct radeon_winsys_cs
*cs
,
539 struct pipe_fence_handle
**fence
);
542 * Create a fence before the CS is flushed.
543 * The user must flush manually to complete the initializaton of the fence.
544 * The fence must not be used before the flush.
546 struct pipe_fence_handle
*(*cs_get_next_fence
)(struct radeon_winsys_cs
*cs
);
549 * Return true if a buffer is referenced by a command stream.
551 * \param cs A command stream.
552 * \param buf A winsys buffer.
554 bool (*cs_is_buffer_referenced
)(struct radeon_winsys_cs
*cs
,
555 struct pb_buffer
*buf
,
556 enum radeon_bo_usage usage
);
559 * Request access to a feature for a command stream.
561 * \param cs A command stream.
562 * \param fid Feature ID, one of RADEON_FID_*
563 * \param enable Whether to enable or disable the feature.
565 bool (*cs_request_feature
)(struct radeon_winsys_cs
*cs
,
566 enum radeon_feature_id fid
,
569 * Make sure all asynchronous flush of the cs have completed
571 * \param cs A command stream.
573 void (*cs_sync_flush
)(struct radeon_winsys_cs
*cs
);
576 * Wait for the fence and return true if the fence has been signalled.
577 * The timeout of 0 will only return the status.
578 * The timeout of PIPE_TIMEOUT_INFINITE will always wait until the fence
581 bool (*fence_wait
)(struct radeon_winsys
*ws
,
582 struct pipe_fence_handle
*fence
,
586 * Reference counting for fences.
588 void (*fence_reference
)(struct pipe_fence_handle
**dst
,
589 struct pipe_fence_handle
*src
);
594 * \param ws The winsys this function is called from.
595 * \param tex Input texture description
596 * \param flags Bitmask of RADEON_SURF_* flags
597 * \param bpe Bytes per pixel, it can be different for Z buffers.
598 * \param mode Preferred tile mode. (linear, 1D, or 2D)
599 * \param surf Output structure
601 int (*surface_init
)(struct radeon_winsys
*ws
,
602 const struct pipe_resource
*tex
,
603 unsigned flags
, unsigned bpe
,
604 enum radeon_surf_mode mode
,
605 struct radeon_surf
*surf
);
607 uint64_t (*query_value
)(struct radeon_winsys
*ws
,
608 enum radeon_value_id value
);
610 bool (*read_registers
)(struct radeon_winsys
*ws
, unsigned reg_offset
,
611 unsigned num_registers
, uint32_t *out
);
613 const char* (*get_chip_name
)(struct radeon_winsys
*ws
);
616 static inline bool radeon_emitted(struct radeon_winsys_cs
*cs
, unsigned num_dw
)
618 return cs
&& (cs
->prev_dw
+ cs
->current
.cdw
> num_dw
);
621 static inline void radeon_emit(struct radeon_winsys_cs
*cs
, uint32_t value
)
623 cs
->current
.buf
[cs
->current
.cdw
++] = value
;
626 static inline void radeon_emit_array(struct radeon_winsys_cs
*cs
,
627 const uint32_t *values
, unsigned count
)
629 memcpy(cs
->current
.buf
+ cs
->current
.cdw
, values
, count
* 4);
630 cs
->current
.cdw
+= count
;
634 RADEON_HEAP_VRAM_NO_CPU_ACCESS
,
636 RADEON_HEAP_VRAM_GTT
, /* combined heaps */
639 RADEON_MAX_SLAB_HEAPS
,
640 RADEON_MAX_CACHED_HEAPS
= RADEON_MAX_SLAB_HEAPS
,
643 static inline enum radeon_bo_domain
radeon_domain_from_heap(enum radeon_heap heap
)
646 case RADEON_HEAP_VRAM_NO_CPU_ACCESS
:
647 case RADEON_HEAP_VRAM
:
648 return RADEON_DOMAIN_VRAM
;
649 case RADEON_HEAP_VRAM_GTT
:
650 return RADEON_DOMAIN_VRAM_GTT
;
651 case RADEON_HEAP_GTT_WC
:
652 case RADEON_HEAP_GTT
:
653 return RADEON_DOMAIN_GTT
;
656 return (enum radeon_bo_domain
)0;
660 static inline unsigned radeon_flags_from_heap(enum radeon_heap heap
)
663 case RADEON_HEAP_VRAM_NO_CPU_ACCESS
:
664 return RADEON_FLAG_GTT_WC
| RADEON_FLAG_NO_CPU_ACCESS
;
665 case RADEON_HEAP_VRAM
:
666 case RADEON_HEAP_VRAM_GTT
:
667 case RADEON_HEAP_GTT_WC
:
668 return RADEON_FLAG_GTT_WC
;
669 case RADEON_HEAP_GTT
:
675 /* The pb cache bucket is chosen to minimize pb_cache misses.
676 * It must be between 0 and 3 inclusive.
678 static inline unsigned radeon_get_pb_cache_bucket_index(enum radeon_heap heap
)
681 case RADEON_HEAP_VRAM_NO_CPU_ACCESS
:
683 case RADEON_HEAP_VRAM
:
684 case RADEON_HEAP_VRAM_GTT
:
686 case RADEON_HEAP_GTT_WC
:
688 case RADEON_HEAP_GTT
:
694 /* Return the heap index for winsys allocators, or -1 on failure. */
695 static inline int radeon_get_heap_index(enum radeon_bo_domain domain
,
696 enum radeon_bo_flag flags
)
698 /* VRAM implies WC (write combining) */
699 assert(!(domain
& RADEON_DOMAIN_VRAM
) || flags
& RADEON_FLAG_GTT_WC
);
700 /* NO_CPU_ACCESS implies VRAM only. */
701 assert(!(flags
& RADEON_FLAG_NO_CPU_ACCESS
) || domain
== RADEON_DOMAIN_VRAM
);
703 /* Unsupported flags: NO_SUBALLOC, SPARSE. */
704 if (flags
& ~(RADEON_FLAG_GTT_WC
| RADEON_FLAG_NO_CPU_ACCESS
))
708 case RADEON_DOMAIN_VRAM
:
709 if (flags
& RADEON_FLAG_NO_CPU_ACCESS
)
710 return RADEON_HEAP_VRAM_NO_CPU_ACCESS
;
712 return RADEON_HEAP_VRAM
;
713 case RADEON_DOMAIN_VRAM_GTT
:
714 return RADEON_HEAP_VRAM_GTT
;
715 case RADEON_DOMAIN_GTT
:
716 if (flags
& RADEON_FLAG_GTT_WC
)
717 return RADEON_HEAP_GTT_WC
;
719 return RADEON_HEAP_GTT
;