2 #include "util/u_inlines.h"
3 #include "util/u_memory.h"
4 #include "util/u_math.h"
5 #include "util/u_surface.h"
7 #include "nouveau_screen.h"
8 #include "nouveau_context.h"
9 #include "nouveau_winsys.h"
10 #include "nouveau_fence.h"
11 #include "nouveau_buffer.h"
12 #include "nouveau_mm.h"
14 #define NOUVEAU_TRANSFER_PUSHBUF_THRESHOLD 192
16 struct nouveau_transfer
{
17 struct pipe_transfer base
;
20 struct nouveau_bo
*bo
;
21 struct nouveau_mm_allocation
*mm
;
25 static INLINE
struct nouveau_transfer
*
26 nouveau_transfer(struct pipe_transfer
*transfer
)
28 return (struct nouveau_transfer
*)transfer
;
32 nouveau_buffer_malloc(struct nv04_resource
*buf
)
35 buf
->data
= align_malloc(buf
->base
.width0
, NOUVEAU_MIN_BUFFER_MAP_ALIGN
);
40 nouveau_buffer_allocate(struct nouveau_screen
*screen
,
41 struct nv04_resource
*buf
, unsigned domain
)
43 uint32_t size
= buf
->base
.width0
;
45 if (buf
->base
.bind
& (PIPE_BIND_CONSTANT_BUFFER
|
46 PIPE_BIND_COMPUTE_RESOURCE
|
47 PIPE_BIND_SHADER_BUFFER
|
48 PIPE_BIND_SHADER_IMAGE
))
49 size
= align(size
, 0x100);
51 if (domain
== NOUVEAU_BO_VRAM
) {
52 buf
->mm
= nouveau_mm_allocate(screen
->mm_VRAM
, size
,
53 &buf
->bo
, &buf
->offset
);
55 return nouveau_buffer_allocate(screen
, buf
, NOUVEAU_BO_GART
);
56 NOUVEAU_DRV_STAT(screen
, buf_obj_current_bytes_vid
, buf
->base
.width0
);
58 if (domain
== NOUVEAU_BO_GART
) {
59 buf
->mm
= nouveau_mm_allocate(screen
->mm_GART
, size
,
60 &buf
->bo
, &buf
->offset
);
63 NOUVEAU_DRV_STAT(screen
, buf_obj_current_bytes_sys
, buf
->base
.width0
);
66 if (!nouveau_buffer_malloc(buf
))
71 buf
->address
= buf
->bo
->offset
+ buf
->offset
;
73 util_range_set_empty(&buf
->valid_buffer_range
);
79 release_allocation(struct nouveau_mm_allocation
**mm
,
80 struct nouveau_fence
*fence
)
82 nouveau_fence_work(fence
, nouveau_mm_free_work
, *mm
);
87 nouveau_buffer_release_gpu_storage(struct nv04_resource
*buf
)
89 nouveau_bo_ref(NULL
, &buf
->bo
);
92 release_allocation(&buf
->mm
, buf
->fence
);
94 if (buf
->domain
== NOUVEAU_BO_VRAM
)
95 NOUVEAU_DRV_STAT_RES(buf
, buf_obj_current_bytes_vid
, -(uint64_t)buf
->base
.width0
);
96 if (buf
->domain
== NOUVEAU_BO_GART
)
97 NOUVEAU_DRV_STAT_RES(buf
, buf_obj_current_bytes_sys
, -(uint64_t)buf
->base
.width0
);
102 static INLINE boolean
103 nouveau_buffer_reallocate(struct nouveau_screen
*screen
,
104 struct nv04_resource
*buf
, unsigned domain
)
106 nouveau_buffer_release_gpu_storage(buf
);
108 nouveau_fence_ref(NULL
, &buf
->fence
);
109 nouveau_fence_ref(NULL
, &buf
->fence_wr
);
111 buf
->status
&= NOUVEAU_BUFFER_STATUS_REALLOC_MASK
;
113 return nouveau_buffer_allocate(screen
, buf
, domain
);
117 nouveau_buffer_destroy(struct pipe_screen
*pscreen
,
118 struct pipe_resource
*presource
)
120 struct nv04_resource
*res
= nv04_resource(presource
);
122 nouveau_buffer_release_gpu_storage(res
);
124 if (res
->data
&& !(res
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
))
125 align_free(res
->data
);
127 nouveau_fence_ref(NULL
, &res
->fence
);
128 nouveau_fence_ref(NULL
, &res
->fence_wr
);
130 util_range_destroy(&res
->valid_buffer_range
);
134 NOUVEAU_DRV_STAT(nouveau_screen(pscreen
), buf_obj_current_count
, -1);
137 /* Set up a staging area for the transfer. This is either done in "regular"
138 * system memory if the driver supports push_data (nv50+) and the data is
139 * small enough (and permit_pb == true), or in GART memory.
142 nouveau_transfer_staging(struct nouveau_context
*nv
,
143 struct nouveau_transfer
*tx
, boolean permit_pb
)
145 const unsigned adj
= tx
->base
.box
.x
& NOUVEAU_MIN_BUFFER_MAP_ALIGN_MASK
;
146 const unsigned size
= align(tx
->base
.box
.width
, 4) + adj
;
151 if ((size
<= NOUVEAU_TRANSFER_PUSHBUF_THRESHOLD
) && permit_pb
) {
152 tx
->map
= align_malloc(size
, NOUVEAU_MIN_BUFFER_MAP_ALIGN
);
157 nouveau_mm_allocate(nv
->screen
->mm_GART
, size
, &tx
->bo
, &tx
->offset
);
160 if (!nouveau_bo_map(tx
->bo
, 0, NULL
))
161 tx
->map
= (uint8_t *)tx
->bo
->map
+ tx
->offset
;
167 /* Copies data from the resource into the the transfer's temporary GART
168 * buffer. Also updates buf->data if present.
170 * Maybe just migrate to GART right away if we actually need to do this. */
172 nouveau_transfer_read(struct nouveau_context
*nv
, struct nouveau_transfer
*tx
)
174 struct nv04_resource
*buf
= nv04_resource(tx
->base
.resource
);
175 const unsigned base
= tx
->base
.box
.x
;
176 const unsigned size
= tx
->base
.box
.width
;
178 NOUVEAU_DRV_STAT(nv
->screen
, buf_read_bytes_staging_vid
, size
);
180 nv
->copy_data(nv
, tx
->bo
, tx
->offset
, NOUVEAU_BO_GART
,
181 buf
->bo
, buf
->offset
+ base
, buf
->domain
, size
);
183 if (nouveau_bo_wait(tx
->bo
, NOUVEAU_BO_RD
, nv
->client
))
187 memcpy(buf
->data
+ base
, tx
->map
, size
);
193 nouveau_transfer_write(struct nouveau_context
*nv
, struct nouveau_transfer
*tx
,
194 unsigned offset
, unsigned size
)
196 struct nv04_resource
*buf
= nv04_resource(tx
->base
.resource
);
197 uint8_t *data
= tx
->map
+ offset
;
198 const unsigned base
= tx
->base
.box
.x
+ offset
;
199 const boolean can_cb
= !((base
| size
) & 3);
202 memcpy(data
, buf
->data
+ base
, size
);
204 buf
->status
|= NOUVEAU_BUFFER_STATUS_DIRTY
;
206 if (buf
->domain
== NOUVEAU_BO_VRAM
)
207 NOUVEAU_DRV_STAT(nv
->screen
, buf_write_bytes_staging_vid
, size
);
208 if (buf
->domain
== NOUVEAU_BO_GART
)
209 NOUVEAU_DRV_STAT(nv
->screen
, buf_write_bytes_staging_sys
, size
);
212 nv
->copy_data(nv
, buf
->bo
, buf
->offset
+ base
, buf
->domain
,
213 tx
->bo
, tx
->offset
+ offset
, NOUVEAU_BO_GART
, size
);
215 if ((buf
->base
.bind
& PIPE_BIND_CONSTANT_BUFFER
) && nv
->push_cb
&& can_cb
)
216 nv
->push_cb(nv
, buf
->bo
, buf
->domain
, buf
->offset
, buf
->base
.width0
,
217 base
, size
/ 4, (const uint32_t *)data
);
219 nv
->push_data(nv
, buf
->bo
, buf
->offset
+ base
, buf
->domain
, size
, data
);
221 nouveau_fence_ref(nv
->screen
->fence
.current
, &buf
->fence
);
222 nouveau_fence_ref(nv
->screen
->fence
.current
, &buf
->fence_wr
);
225 /* Does a CPU wait for the buffer's backing data to become reliably accessible
226 * for write/read by waiting on the buffer's relevant fences.
228 static INLINE boolean
229 nouveau_buffer_sync(struct nv04_resource
*buf
, unsigned rw
)
231 if (rw
== PIPE_TRANSFER_READ
) {
234 NOUVEAU_DRV_STAT_RES(buf
, buf_non_kernel_fence_sync_count
,
235 !nouveau_fence_signalled(buf
->fence_wr
));
236 if (!nouveau_fence_wait(buf
->fence_wr
))
241 NOUVEAU_DRV_STAT_RES(buf
, buf_non_kernel_fence_sync_count
,
242 !nouveau_fence_signalled(buf
->fence
));
243 if (!nouveau_fence_wait(buf
->fence
))
246 nouveau_fence_ref(NULL
, &buf
->fence
);
248 nouveau_fence_ref(NULL
, &buf
->fence_wr
);
253 static INLINE boolean
254 nouveau_buffer_busy(struct nv04_resource
*buf
, unsigned rw
)
256 if (rw
== PIPE_TRANSFER_READ
)
257 return (buf
->fence_wr
&& !nouveau_fence_signalled(buf
->fence_wr
));
259 return (buf
->fence
&& !nouveau_fence_signalled(buf
->fence
));
263 nouveau_buffer_transfer_init(struct nouveau_transfer
*tx
,
264 struct pipe_resource
*resource
,
265 const struct pipe_box
*box
,
268 tx
->base
.resource
= resource
;
270 tx
->base
.usage
= usage
;
271 tx
->base
.box
.x
= box
->x
;
274 tx
->base
.box
.width
= box
->width
;
275 tx
->base
.box
.height
= 1;
276 tx
->base
.box
.depth
= 1;
278 tx
->base
.layer_stride
= 0;
285 nouveau_buffer_transfer_del(struct nouveau_context
*nv
,
286 struct nouveau_transfer
*tx
)
289 if (likely(tx
->bo
)) {
290 nouveau_bo_ref(NULL
, &tx
->bo
);
292 release_allocation(&tx
->mm
, nv
->screen
->fence
.current
);
295 (tx
->base
.box
.x
& NOUVEAU_MIN_BUFFER_MAP_ALIGN_MASK
));
300 /* Creates a cache in system memory of the buffer data. */
302 nouveau_buffer_cache(struct nouveau_context
*nv
, struct nv04_resource
*buf
)
304 struct nouveau_transfer tx
;
306 tx
.base
.resource
= &buf
->base
;
308 tx
.base
.box
.width
= buf
->base
.width0
;
313 if (!nouveau_buffer_malloc(buf
))
315 if (!(buf
->status
& NOUVEAU_BUFFER_STATUS_DIRTY
))
317 nv
->stats
.buf_cache_count
++;
319 if (!nouveau_transfer_staging(nv
, &tx
, FALSE
))
322 ret
= nouveau_transfer_read(nv
, &tx
);
324 buf
->status
&= ~NOUVEAU_BUFFER_STATUS_DIRTY
;
325 memcpy(buf
->data
, tx
.map
, buf
->base
.width0
);
327 nouveau_buffer_transfer_del(nv
, &tx
);
332 #define NOUVEAU_TRANSFER_DISCARD \
333 (PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)
335 /* Checks whether it is possible to completely discard the memory backing this
336 * resource. This can be useful if we would otherwise have to wait for a read
337 * operation to complete on this data.
339 static INLINE boolean
340 nouveau_buffer_should_discard(struct nv04_resource
*buf
, unsigned usage
)
342 if (!(usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
))
344 if (unlikely(buf
->base
.bind
& PIPE_BIND_SHARED
))
346 if (unlikely(usage
& PIPE_TRANSFER_PERSISTENT
))
348 return buf
->mm
&& nouveau_buffer_busy(buf
, PIPE_TRANSFER_WRITE
);
351 /* Returns a pointer to a memory area representing a window into the
354 * This may or may not be the _actual_ memory area of the resource. However
355 * when calling nouveau_buffer_transfer_unmap, if it wasn't the actual memory
356 * area, the contents of the returned map are copied over to the resource.
358 * The usage indicates what the caller plans to do with the map:
360 * WRITE means that the user plans to write to it
362 * READ means that the user plans on reading from it
364 * DISCARD_WHOLE_RESOURCE means that the whole resource is going to be
365 * potentially overwritten, and even if it isn't, the bits that aren't don't
366 * need to be maintained.
368 * DISCARD_RANGE means that all the data in the specified range is going to
371 * The strategy for determining what kind of memory area to return is complex,
372 * see comments inside of the function.
375 nouveau_buffer_transfer_map(struct pipe_context
*pipe
,
376 struct pipe_resource
*resource
,
377 unsigned level
, unsigned usage
,
378 const struct pipe_box
*box
,
379 struct pipe_transfer
**ptransfer
)
381 struct nouveau_context
*nv
= nouveau_context(pipe
);
382 struct nv04_resource
*buf
= nv04_resource(resource
);
383 struct nouveau_transfer
*tx
= MALLOC_STRUCT(nouveau_transfer
);
389 nouveau_buffer_transfer_init(tx
, resource
, box
, usage
);
390 *ptransfer
= &tx
->base
;
392 if (usage
& PIPE_TRANSFER_READ
)
393 NOUVEAU_DRV_STAT(nv
->screen
, buf_transfers_rd
, 1);
394 if (usage
& PIPE_TRANSFER_WRITE
)
395 NOUVEAU_DRV_STAT(nv
->screen
, buf_transfers_wr
, 1);
397 /* If we are trying to write to an uninitialized range, the user shouldn't
398 * care what was there before. So we can treat the write as if the target
399 * range were being discarded. Furthermore, since we know that even if this
400 * buffer is busy due to GPU activity, because the contents were
401 * uninitialized, the GPU can't care what was there, and so we can treat
402 * the write as being unsynchronized.
404 if ((usage
& PIPE_TRANSFER_WRITE
) &&
405 !util_ranges_intersect(&buf
->valid_buffer_range
, box
->x
, box
->x
+ box
->width
))
406 usage
|= PIPE_TRANSFER_DISCARD_RANGE
| PIPE_TRANSFER_UNSYNCHRONIZED
;
408 if (usage
& PIPE_TRANSFER_PERSISTENT
)
409 usage
|= PIPE_TRANSFER_UNSYNCHRONIZED
;
411 if (buf
->domain
== NOUVEAU_BO_VRAM
) {
412 if (usage
& NOUVEAU_TRANSFER_DISCARD
) {
413 /* Set up a staging area for the user to write to. It will be copied
414 * back into VRAM on unmap. */
415 if (usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
)
416 buf
->status
&= NOUVEAU_BUFFER_STATUS_REALLOC_MASK
;
417 nouveau_transfer_staging(nv
, tx
, TRUE
);
419 if (buf
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
) {
420 /* The GPU is currently writing to this buffer. Copy its current
421 * contents to a staging area in the GART. This is necessary since
422 * not the whole area being mapped is being discarded.
425 align_free(buf
->data
);
428 nouveau_transfer_staging(nv
, tx
, FALSE
);
429 nouveau_transfer_read(nv
, tx
);
431 /* The buffer is currently idle. Create a staging area for writes,
432 * and make sure that the cached data is up-to-date. */
433 if (usage
& PIPE_TRANSFER_WRITE
)
434 nouveau_transfer_staging(nv
, tx
, TRUE
);
436 nouveau_buffer_cache(nv
, buf
);
439 return buf
->data
? (buf
->data
+ box
->x
) : tx
->map
;
441 if (unlikely(buf
->domain
== 0)) {
442 return buf
->data
+ box
->x
;
445 /* At this point, buf->domain == GART */
447 if (nouveau_buffer_should_discard(buf
, usage
)) {
448 int ref
= buf
->base
.reference
.count
- 1;
449 nouveau_buffer_reallocate(nv
->screen
, buf
, buf
->domain
);
450 if (ref
> 0) /* any references inside context possible ? */
451 nv
->invalidate_resource_storage(nv
, &buf
->base
, ref
);
454 /* Note that nouveau_bo_map ends up doing a nouveau_bo_wait with the
455 * relevant flags. If buf->mm is set, that means this resource is part of a
456 * larger slab bo that holds multiple resources. So in that case, don't
457 * wait on the whole slab and instead use the logic below to return a
458 * reasonable buffer for that case.
460 ret
= nouveau_bo_map(buf
->bo
,
461 buf
->mm
? 0 : nouveau_screen_transfer_flags(usage
),
467 map
= (uint8_t *)buf
->bo
->map
+ buf
->offset
+ box
->x
;
469 /* using kernel fences only if !buf->mm */
470 if ((usage
& PIPE_TRANSFER_UNSYNCHRONIZED
) || !buf
->mm
)
473 /* If the GPU is currently reading/writing this buffer, we shouldn't
474 * interfere with its progress. So instead we either wait for the GPU to
475 * complete its operation, or set up a staging area to perform our work in.
477 if (nouveau_buffer_busy(buf
, usage
& PIPE_TRANSFER_READ_WRITE
)) {
478 if (unlikely(usage
& PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE
)) {
479 /* Discarding was not possible, must sync because
480 * subsequent transfers might use UNSYNCHRONIZED. */
481 nouveau_buffer_sync(buf
, usage
& PIPE_TRANSFER_READ_WRITE
);
483 if (usage
& PIPE_TRANSFER_DISCARD_RANGE
) {
484 /* The whole range is being discarded, so it doesn't matter what was
485 * there before. No need to copy anything over. */
486 nouveau_transfer_staging(nv
, tx
, TRUE
);
489 if (nouveau_buffer_busy(buf
, PIPE_TRANSFER_READ
)) {
490 if (usage
& PIPE_TRANSFER_DONTBLOCK
)
493 nouveau_buffer_sync(buf
, usage
& PIPE_TRANSFER_READ_WRITE
);
495 /* It is expected that the returned buffer be a representation of the
496 * data in question, so we must copy it over from the buffer. */
497 nouveau_transfer_staging(nv
, tx
, TRUE
);
499 memcpy(tx
->map
, map
, box
->width
);
511 nouveau_buffer_transfer_flush_region(struct pipe_context
*pipe
,
512 struct pipe_transfer
*transfer
,
513 const struct pipe_box
*box
)
515 struct nouveau_transfer
*tx
= nouveau_transfer(transfer
);
516 struct nv04_resource
*buf
= nv04_resource(transfer
->resource
);
519 nouveau_transfer_write(nouveau_context(pipe
), tx
, box
->x
, box
->width
);
521 util_range_add(&buf
->valid_buffer_range
,
522 tx
->base
.box
.x
+ box
->x
,
523 tx
->base
.box
.x
+ box
->x
+ box
->width
);
526 /* Unmap stage of the transfer. If it was a WRITE transfer and the map that
527 * was returned was not the real resource's data, this needs to transfer the
528 * data back to the resource.
530 * Also marks vbo dirty based on the buffer's binding
533 nouveau_buffer_transfer_unmap(struct pipe_context
*pipe
,
534 struct pipe_transfer
*transfer
)
536 struct nouveau_context
*nv
= nouveau_context(pipe
);
537 struct nouveau_transfer
*tx
= nouveau_transfer(transfer
);
538 struct nv04_resource
*buf
= nv04_resource(transfer
->resource
);
540 if (tx
->base
.usage
& PIPE_TRANSFER_WRITE
) {
541 if (!(tx
->base
.usage
& PIPE_TRANSFER_FLUSH_EXPLICIT
) && tx
->map
)
542 nouveau_transfer_write(nv
, tx
, 0, tx
->base
.box
.width
);
544 if (likely(buf
->domain
)) {
545 const uint8_t bind
= buf
->base
.bind
;
546 /* make sure we invalidate dedicated caches */
547 if (bind
& (PIPE_BIND_VERTEX_BUFFER
| PIPE_BIND_INDEX_BUFFER
))
548 nv
->vbo_dirty
= TRUE
;
551 util_range_add(&buf
->valid_buffer_range
,
552 tx
->base
.box
.x
, tx
->base
.box
.x
+ tx
->base
.box
.width
);
555 if (!tx
->bo
&& (tx
->base
.usage
& PIPE_TRANSFER_WRITE
))
556 NOUVEAU_DRV_STAT(nv
->screen
, buf_write_bytes_direct
, tx
->base
.box
.width
);
558 nouveau_buffer_transfer_del(nv
, tx
);
564 nouveau_copy_buffer(struct nouveau_context
*nv
,
565 struct nv04_resource
*dst
, unsigned dstx
,
566 struct nv04_resource
*src
, unsigned srcx
, unsigned size
)
568 assert(dst
->base
.target
== PIPE_BUFFER
&& src
->base
.target
== PIPE_BUFFER
);
570 if (likely(dst
->domain
) && likely(src
->domain
)) {
572 dst
->bo
, dst
->offset
+ dstx
, dst
->domain
,
573 src
->bo
, src
->offset
+ srcx
, src
->domain
, size
);
575 dst
->status
|= NOUVEAU_BUFFER_STATUS_GPU_WRITING
;
576 nouveau_fence_ref(nv
->screen
->fence
.current
, &dst
->fence
);
577 nouveau_fence_ref(nv
->screen
->fence
.current
, &dst
->fence_wr
);
579 src
->status
|= NOUVEAU_BUFFER_STATUS_GPU_READING
;
580 nouveau_fence_ref(nv
->screen
->fence
.current
, &src
->fence
);
582 struct pipe_box src_box
;
586 src_box
.width
= size
;
589 util_resource_copy_region(&nv
->pipe
,
590 &dst
->base
, 0, dstx
, 0, 0,
591 &src
->base
, 0, &src_box
);
594 util_range_add(&dst
->valid_buffer_range
, dstx
, dstx
+ size
);
599 nouveau_resource_map_offset(struct nouveau_context
*nv
,
600 struct nv04_resource
*res
, uint32_t offset
,
603 if (unlikely(res
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
))
604 return res
->data
+ offset
;
606 if (res
->domain
== NOUVEAU_BO_VRAM
) {
607 if (!res
->data
|| (res
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
))
608 nouveau_buffer_cache(nv
, res
);
610 if (res
->domain
!= NOUVEAU_BO_GART
)
611 return res
->data
+ offset
;
615 rw
= (flags
& NOUVEAU_BO_WR
) ? PIPE_TRANSFER_WRITE
: PIPE_TRANSFER_READ
;
616 nouveau_buffer_sync(res
, rw
);
617 if (nouveau_bo_map(res
->bo
, 0, NULL
))
620 if (nouveau_bo_map(res
->bo
, flags
, nv
->client
))
623 return (uint8_t *)res
->bo
->map
+ res
->offset
+ offset
;
627 const struct u_resource_vtbl nouveau_buffer_vtbl
=
629 u_default_resource_get_handle
, /* get_handle */
630 nouveau_buffer_destroy
, /* resource_destroy */
631 nouveau_buffer_transfer_map
, /* transfer_map */
632 nouveau_buffer_transfer_flush_region
, /* transfer_flush_region */
633 nouveau_buffer_transfer_unmap
, /* transfer_unmap */
634 u_default_transfer_inline_write
/* transfer_inline_write */
637 struct pipe_resource
*
638 nouveau_buffer_create(struct pipe_screen
*pscreen
,
639 const struct pipe_resource
*templ
)
641 struct nouveau_screen
*screen
= nouveau_screen(pscreen
);
642 struct nv04_resource
*buffer
;
645 buffer
= CALLOC_STRUCT(nv04_resource
);
649 buffer
->base
= *templ
;
650 buffer
->vtbl
= &nouveau_buffer_vtbl
;
651 pipe_reference_init(&buffer
->base
.reference
, 1);
652 buffer
->base
.screen
= pscreen
;
654 if (buffer
->base
.flags
& (PIPE_RESOURCE_FLAG_MAP_PERSISTENT
|
655 PIPE_RESOURCE_FLAG_MAP_COHERENT
)) {
656 buffer
->domain
= NOUVEAU_BO_GART
;
657 } else if (buffer
->base
.bind
&
658 (screen
->vidmem_bindings
& screen
->sysmem_bindings
)) {
659 switch (buffer
->base
.usage
) {
660 case PIPE_USAGE_DEFAULT
:
661 case PIPE_USAGE_IMMUTABLE
:
662 buffer
->domain
= NV_VRAM_DOMAIN(screen
);
664 case PIPE_USAGE_DYNAMIC
:
665 /* For most apps, we'd have to do staging transfers to avoid sync
666 * with this usage, and GART -> GART copies would be suboptimal.
668 buffer
->domain
= NV_VRAM_DOMAIN(screen
);
670 case PIPE_USAGE_STAGING
:
671 case PIPE_USAGE_STREAM
:
672 buffer
->domain
= NOUVEAU_BO_GART
;
679 if (buffer
->base
.bind
& screen
->vidmem_bindings
)
680 buffer
->domain
= NV_VRAM_DOMAIN(screen
);
682 if (buffer
->base
.bind
& screen
->sysmem_bindings
)
683 buffer
->domain
= NOUVEAU_BO_GART
;
685 ret
= nouveau_buffer_allocate(screen
, buffer
, buffer
->domain
);
690 if (buffer
->domain
== NOUVEAU_BO_VRAM
&& screen
->hint_buf_keep_sysmem_copy
)
691 nouveau_buffer_cache(NULL
, buffer
);
693 NOUVEAU_DRV_STAT(screen
, buf_obj_current_count
, 1);
695 util_range_init(&buffer
->valid_buffer_range
);
697 return &buffer
->base
;
705 struct pipe_resource
*
706 nouveau_user_buffer_create(struct pipe_screen
*pscreen
, void *ptr
,
707 unsigned bytes
, unsigned bind
)
709 struct nv04_resource
*buffer
;
711 buffer
= CALLOC_STRUCT(nv04_resource
);
715 pipe_reference_init(&buffer
->base
.reference
, 1);
716 buffer
->vtbl
= &nouveau_buffer_vtbl
;
717 buffer
->base
.screen
= pscreen
;
718 buffer
->base
.format
= PIPE_FORMAT_R8_UNORM
;
719 buffer
->base
.usage
= PIPE_USAGE_IMMUTABLE
;
720 buffer
->base
.bind
= bind
;
721 buffer
->base
.width0
= bytes
;
722 buffer
->base
.height0
= 1;
723 buffer
->base
.depth0
= 1;
726 buffer
->status
= NOUVEAU_BUFFER_STATUS_USER_MEMORY
;
728 util_range_init(&buffer
->valid_buffer_range
);
729 util_range_add(&buffer
->valid_buffer_range
, 0, bytes
);
731 return &buffer
->base
;
734 static INLINE boolean
735 nouveau_buffer_data_fetch(struct nouveau_context
*nv
, struct nv04_resource
*buf
,
736 struct nouveau_bo
*bo
, unsigned offset
, unsigned size
)
738 if (!nouveau_buffer_malloc(buf
))
740 if (nouveau_bo_map(bo
, NOUVEAU_BO_RD
, nv
->client
))
742 memcpy(buf
->data
, (uint8_t *)bo
->map
+ offset
, size
);
746 /* Migrate a linear buffer (vertex, index, constants) USER -> GART -> VRAM. */
748 nouveau_buffer_migrate(struct nouveau_context
*nv
,
749 struct nv04_resource
*buf
, const unsigned new_domain
)
751 struct nouveau_screen
*screen
= nv
->screen
;
752 struct nouveau_bo
*bo
;
753 const unsigned old_domain
= buf
->domain
;
754 unsigned size
= buf
->base
.width0
;
758 assert(new_domain
!= old_domain
);
760 if (new_domain
== NOUVEAU_BO_GART
&& old_domain
== 0) {
761 if (!nouveau_buffer_allocate(screen
, buf
, new_domain
))
763 ret
= nouveau_bo_map(buf
->bo
, 0, nv
->client
);
766 memcpy((uint8_t *)buf
->bo
->map
+ buf
->offset
, buf
->data
, size
);
767 align_free(buf
->data
);
769 if (old_domain
!= 0 && new_domain
!= 0) {
770 struct nouveau_mm_allocation
*mm
= buf
->mm
;
772 if (new_domain
== NOUVEAU_BO_VRAM
) {
773 /* keep a system memory copy of our data in case we hit a fallback */
774 if (!nouveau_buffer_data_fetch(nv
, buf
, buf
->bo
, buf
->offset
, size
))
776 if (nouveau_mesa_debug
)
777 debug_printf("migrating %u KiB to VRAM\n", size
/ 1024);
780 offset
= buf
->offset
;
784 nouveau_buffer_allocate(screen
, buf
, new_domain
);
786 nv
->copy_data(nv
, buf
->bo
, buf
->offset
, new_domain
,
787 bo
, offset
, old_domain
, buf
->base
.width0
);
789 nouveau_bo_ref(NULL
, &bo
);
791 release_allocation(&mm
, screen
->fence
.current
);
793 if (new_domain
== NOUVEAU_BO_VRAM
&& old_domain
== 0) {
794 struct nouveau_transfer tx
;
795 if (!nouveau_buffer_allocate(screen
, buf
, NOUVEAU_BO_VRAM
))
797 tx
.base
.resource
= &buf
->base
;
799 tx
.base
.box
.width
= buf
->base
.width0
;
802 if (!nouveau_transfer_staging(nv
, &tx
, FALSE
))
804 nouveau_transfer_write(nv
, &tx
, 0, tx
.base
.box
.width
);
805 nouveau_buffer_transfer_del(nv
, &tx
);
809 assert(buf
->domain
== new_domain
);
813 /* Migrate data from glVertexAttribPointer(non-VBO) user buffers to GART.
814 * We'd like to only allocate @size bytes here, but then we'd have to rebase
815 * the vertex indices ...
818 nouveau_user_buffer_upload(struct nouveau_context
*nv
,
819 struct nv04_resource
*buf
,
820 unsigned base
, unsigned size
)
822 struct nouveau_screen
*screen
= nouveau_screen(buf
->base
.screen
);
825 assert(buf
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
);
827 buf
->base
.width0
= base
+ size
;
828 if (!nouveau_buffer_reallocate(screen
, buf
, NOUVEAU_BO_GART
))
831 ret
= nouveau_bo_map(buf
->bo
, 0, nv
->client
);
834 memcpy((uint8_t *)buf
->bo
->map
+ buf
->offset
+ base
, buf
->data
+ base
, size
);
840 /* Scratch data allocation. */
843 nouveau_scratch_bo_alloc(struct nouveau_context
*nv
, struct nouveau_bo
**pbo
,
846 return nouveau_bo_new(nv
->screen
->device
, NOUVEAU_BO_GART
| NOUVEAU_BO_MAP
,
847 4096, size
, NULL
, pbo
);
851 nouveau_scratch_unref_bos(void *d
)
853 struct runout
*b
= d
;
856 for (i
= 0; i
< b
->nr
; ++i
)
857 nouveau_bo_ref(NULL
, &b
->bo
[i
]);
863 nouveau_scratch_runout_release(struct nouveau_context
*nv
)
865 if (!nv
->scratch
.runout
)
868 if (!nouveau_fence_work(nv
->screen
->fence
.current
, nouveau_scratch_unref_bos
,
873 nv
->scratch
.runout
= NULL
;
876 /* Allocate an extra bo if we can't fit everything we need simultaneously.
877 * (Could happen for very large user arrays.)
879 static INLINE boolean
880 nouveau_scratch_runout(struct nouveau_context
*nv
, unsigned size
)
885 if (nv
->scratch
.runout
)
886 n
= nv
->scratch
.runout
->nr
;
889 nv
->scratch
.runout
= REALLOC(nv
->scratch
.runout
, n
== 0 ? 0 :
890 (sizeof(*nv
->scratch
.runout
) + (n
+ 0) * sizeof(void *)),
891 sizeof(*nv
->scratch
.runout
) + (n
+ 1) * sizeof(void *));
892 nv
->scratch
.runout
->nr
= n
+ 1;
893 nv
->scratch
.runout
->bo
[n
] = NULL
;
895 ret
= nouveau_scratch_bo_alloc(nv
, &nv
->scratch
.runout
->bo
[n
], size
);
897 ret
= nouveau_bo_map(nv
->scratch
.runout
->bo
[n
], 0, NULL
);
899 nouveau_bo_ref(NULL
, &nv
->scratch
.runout
->bo
[--nv
->scratch
.runout
->nr
]);
902 nv
->scratch
.current
= nv
->scratch
.runout
->bo
[n
];
903 nv
->scratch
.offset
= 0;
904 nv
->scratch
.end
= size
;
905 nv
->scratch
.map
= nv
->scratch
.current
->map
;
910 /* Continue to next scratch buffer, if available (no wrapping, large enough).
911 * Allocate it if it has not yet been created.
913 static INLINE boolean
914 nouveau_scratch_next(struct nouveau_context
*nv
, unsigned size
)
916 struct nouveau_bo
*bo
;
918 const unsigned i
= (nv
->scratch
.id
+ 1) % NOUVEAU_MAX_SCRATCH_BUFS
;
920 if ((size
> nv
->scratch
.bo_size
) || (i
== nv
->scratch
.wrap
))
924 bo
= nv
->scratch
.bo
[i
];
926 ret
= nouveau_scratch_bo_alloc(nv
, &bo
, nv
->scratch
.bo_size
);
929 nv
->scratch
.bo
[i
] = bo
;
931 nv
->scratch
.current
= bo
;
932 nv
->scratch
.offset
= 0;
933 nv
->scratch
.end
= nv
->scratch
.bo_size
;
935 ret
= nouveau_bo_map(bo
, NOUVEAU_BO_WR
, nv
->client
);
937 nv
->scratch
.map
= bo
->map
;
942 nouveau_scratch_more(struct nouveau_context
*nv
, unsigned min_size
)
946 ret
= nouveau_scratch_next(nv
, min_size
);
948 ret
= nouveau_scratch_runout(nv
, min_size
);
953 /* Copy data to a scratch buffer and return address & bo the data resides in. */
955 nouveau_scratch_data(struct nouveau_context
*nv
,
956 const void *data
, unsigned base
, unsigned size
,
957 struct nouveau_bo
**bo
)
959 unsigned bgn
= MAX2(base
, nv
->scratch
.offset
);
960 unsigned end
= bgn
+ size
;
962 if (end
>= nv
->scratch
.end
) {
964 if (!nouveau_scratch_more(nv
, end
))
968 nv
->scratch
.offset
= align(end
, 4);
970 memcpy(nv
->scratch
.map
+ bgn
, (const uint8_t *)data
+ base
, size
);
972 *bo
= nv
->scratch
.current
;
973 return (*bo
)->offset
+ (bgn
- base
);
977 nouveau_scratch_get(struct nouveau_context
*nv
,
978 unsigned size
, uint64_t *gpu_addr
, struct nouveau_bo
**pbo
)
980 unsigned bgn
= nv
->scratch
.offset
;
981 unsigned end
= nv
->scratch
.offset
+ size
;
983 if (end
>= nv
->scratch
.end
) {
985 if (!nouveau_scratch_more(nv
, end
))
989 nv
->scratch
.offset
= align(end
, 4);
991 *pbo
= nv
->scratch
.current
;
992 *gpu_addr
= nv
->scratch
.current
->offset
+ bgn
;
993 return nv
->scratch
.map
+ bgn
;