2 #include "util/u_inlines.h"
3 #include "util/u_memory.h"
4 #include "util/u_math.h"
6 #include "nouveau_screen.h"
7 #include "nouveau_context.h"
8 #include "nouveau_winsys.h"
9 #include "nouveau_fence.h"
10 #include "nouveau_buffer.h"
11 #include "nouveau_mm.h"
13 struct nouveau_transfer
{
14 struct pipe_transfer base
;
17 static INLINE
struct nouveau_transfer
*
18 nouveau_transfer(struct pipe_transfer
*transfer
)
20 return (struct nouveau_transfer
*)transfer
;
24 nouveau_buffer_allocate(struct nouveau_screen
*screen
,
25 struct nv04_resource
*buf
, unsigned domain
)
27 uint32_t size
= buf
->base
.width0
;
29 if (buf
->base
.bind
& PIPE_BIND_CONSTANT_BUFFER
)
30 size
= align(size
, 0x100);
32 if (domain
== NOUVEAU_BO_VRAM
) {
33 buf
->mm
= nouveau_mm_allocate(screen
->mm_VRAM
, size
,
34 &buf
->bo
, &buf
->offset
);
36 return nouveau_buffer_allocate(screen
, buf
, NOUVEAU_BO_GART
);
38 if (domain
== NOUVEAU_BO_GART
) {
39 buf
->mm
= nouveau_mm_allocate(screen
->mm_GART
, size
,
40 &buf
->bo
, &buf
->offset
);
44 if (domain
!= NOUVEAU_BO_GART
) {
46 buf
->data
= MALLOC(buf
->base
.width0
);
53 buf
->address
= buf
->bo
->offset
+ buf
->offset
;
59 release_allocation(struct nouveau_mm_allocation
**mm
,
60 struct nouveau_fence
*fence
)
62 nouveau_fence_work(fence
, nouveau_mm_free_work
, *mm
);
67 nouveau_buffer_release_gpu_storage(struct nv04_resource
*buf
)
69 nouveau_bo_ref(NULL
, &buf
->bo
);
72 release_allocation(&buf
->mm
, buf
->fence
);
78 nouveau_buffer_reallocate(struct nouveau_screen
*screen
,
79 struct nv04_resource
*buf
, unsigned domain
)
81 nouveau_buffer_release_gpu_storage(buf
);
83 return nouveau_buffer_allocate(screen
, buf
, domain
);
87 nouveau_buffer_destroy(struct pipe_screen
*pscreen
,
88 struct pipe_resource
*presource
)
90 struct nv04_resource
*res
= nv04_resource(presource
);
92 nouveau_buffer_release_gpu_storage(res
);
94 if (res
->data
&& !(res
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
))
100 /* Maybe just migrate to GART right away if we actually need to do this. */
102 nouveau_buffer_download(struct nouveau_context
*nv
, struct nv04_resource
*buf
,
103 unsigned start
, unsigned size
)
105 struct nouveau_mm_allocation
*mm
;
106 struct nouveau_bo
*bounce
= NULL
;
109 assert(buf
->domain
== NOUVEAU_BO_VRAM
);
111 mm
= nouveau_mm_allocate(nv
->screen
->mm_GART
, size
, &bounce
, &offset
);
115 nv
->copy_data(nv
, bounce
, offset
, NOUVEAU_BO_GART
,
116 buf
->bo
, buf
->offset
+ start
, NOUVEAU_BO_VRAM
, size
);
118 if (nouveau_bo_map(bounce
, NOUVEAU_BO_RD
, nv
->screen
->client
))
120 memcpy(buf
->data
+ start
, (uint8_t *)bounce
->map
+ offset
, size
);
122 buf
->status
&= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING
;
124 nouveau_bo_ref(NULL
, &bounce
);
131 nouveau_buffer_upload(struct nouveau_context
*nv
, struct nv04_resource
*buf
,
132 unsigned start
, unsigned size
)
134 struct nouveau_mm_allocation
*mm
;
135 struct nouveau_bo
*bounce
= NULL
;
138 if (size
<= 192 && (nv
->push_data
|| nv
->push_cb
)) {
139 if (buf
->base
.bind
& PIPE_BIND_CONSTANT_BUFFER
)
140 nv
->push_cb(nv
, buf
->bo
, buf
->domain
, buf
->offset
, buf
->base
.width0
,
141 start
, size
/ 4, (const uint32_t *)(buf
->data
+ start
));
143 nv
->push_data(nv
, buf
->bo
, buf
->offset
+ start
, buf
->domain
,
144 size
, buf
->data
+ start
);
148 mm
= nouveau_mm_allocate(nv
->screen
->mm_GART
, size
, &bounce
, &offset
);
152 nouveau_bo_map(bounce
, 0, nv
->screen
->client
);
153 memcpy((uint8_t *)bounce
->map
+ offset
, buf
->data
+ start
, size
);
155 nv
->copy_data(nv
, buf
->bo
, buf
->offset
+ start
, NOUVEAU_BO_VRAM
,
156 bounce
, offset
, NOUVEAU_BO_GART
, size
);
158 nouveau_bo_ref(NULL
, &bounce
);
160 release_allocation(&mm
, nv
->screen
->fence
.current
);
162 if (start
== 0 && size
== buf
->base
.width0
)
163 buf
->status
&= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING
;
167 static struct pipe_transfer
*
168 nouveau_buffer_transfer_get(struct pipe_context
*pipe
,
169 struct pipe_resource
*resource
,
170 unsigned level
, unsigned usage
,
171 const struct pipe_box
*box
)
173 struct nv04_resource
*buf
= nv04_resource(resource
);
174 struct nouveau_context
*nv
= nouveau_context(pipe
);
175 struct nouveau_transfer
*xfr
= CALLOC_STRUCT(nouveau_transfer
);
179 xfr
->base
.resource
= resource
;
180 xfr
->base
.box
.x
= box
->x
;
181 xfr
->base
.box
.width
= box
->width
;
182 xfr
->base
.usage
= usage
;
184 if (buf
->domain
== NOUVEAU_BO_VRAM
) {
185 if (usage
& PIPE_TRANSFER_READ
) {
186 if (buf
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
)
187 nouveau_buffer_download(nv
, buf
, 0, buf
->base
.width0
);
195 nouveau_buffer_transfer_destroy(struct pipe_context
*pipe
,
196 struct pipe_transfer
*transfer
)
198 struct nv04_resource
*buf
= nv04_resource(transfer
->resource
);
199 struct nouveau_transfer
*xfr
= nouveau_transfer(transfer
);
200 struct nouveau_context
*nv
= nouveau_context(pipe
);
202 if (xfr
->base
.usage
& PIPE_TRANSFER_WRITE
) {
203 if (buf
->domain
== NOUVEAU_BO_VRAM
) {
204 nouveau_buffer_upload(nv
, buf
, transfer
->box
.x
, transfer
->box
.width
);
207 if (buf
->domain
!= 0 && (buf
->base
.bind
& (PIPE_BIND_VERTEX_BUFFER
|
208 PIPE_BIND_INDEX_BUFFER
)))
209 nouveau_context(pipe
)->vbo_dirty
= TRUE
;
215 static INLINE boolean
216 nouveau_buffer_sync(struct nv04_resource
*buf
, unsigned rw
)
218 if (rw
== PIPE_TRANSFER_READ
) {
221 if (!nouveau_fence_wait(buf
->fence_wr
))
226 if (!nouveau_fence_wait(buf
->fence
))
229 nouveau_fence_ref(NULL
, &buf
->fence
);
231 nouveau_fence_ref(NULL
, &buf
->fence_wr
);
236 static INLINE boolean
237 nouveau_buffer_busy(struct nv04_resource
*buf
, unsigned rw
)
239 if (rw
== PIPE_TRANSFER_READ
)
240 return (buf
->fence_wr
&& !nouveau_fence_signalled(buf
->fence_wr
));
242 return (buf
->fence
&& !nouveau_fence_signalled(buf
->fence
));
246 nouveau_buffer_transfer_map(struct pipe_context
*pipe
,
247 struct pipe_transfer
*transfer
)
249 struct nouveau_context
*nv
= nouveau_context(pipe
);
250 struct nouveau_transfer
*xfr
= nouveau_transfer(transfer
);
251 struct nv04_resource
*buf
= nv04_resource(transfer
->resource
);
252 struct nouveau_bo
*bo
= buf
->bo
;
255 uint32_t offset
= xfr
->base
.box
.x
;
258 if (buf
->domain
!= NOUVEAU_BO_GART
)
259 return buf
->data
+ offset
;
262 flags
= nouveau_screen_transfer_flags(xfr
->base
.usage
);
264 offset
+= buf
->offset
;
266 ret
= nouveau_bo_map(buf
->bo
, flags
, nv
->screen
->client
);
269 map
= (uint8_t *)bo
->map
+ offset
;
272 if (xfr
->base
.usage
& PIPE_TRANSFER_DONTBLOCK
) {
273 if (nouveau_buffer_busy(buf
, xfr
->base
.usage
& PIPE_TRANSFER_READ_WRITE
))
276 if (!(xfr
->base
.usage
& PIPE_TRANSFER_UNSYNCHRONIZED
)) {
277 nouveau_buffer_sync(buf
, xfr
->base
.usage
& PIPE_TRANSFER_READ_WRITE
);
286 nouveau_buffer_transfer_flush_region(struct pipe_context
*pipe
,
287 struct pipe_transfer
*transfer
,
288 const struct pipe_box
*box
)
291 struct nv04_resource
*res
= nv04_resource(transfer
->resource
);
292 struct nouveau_bo
*bo
= res
->bo
;
293 unsigned offset
= res
->offset
+ transfer
->box
.x
+ box
->x
;
295 /* not using non-snoop system memory yet, no need for cflush */
299 /* XXX: maybe need to upload for VRAM buffers here */
304 nouveau_buffer_transfer_unmap(struct pipe_context
*pipe
,
305 struct pipe_transfer
*transfer
)
311 nouveau_resource_map_offset(struct nouveau_context
*nv
,
312 struct nv04_resource
*res
, uint32_t offset
,
315 if ((res
->domain
== NOUVEAU_BO_VRAM
) &&
316 (res
->status
& NOUVEAU_BUFFER_STATUS_GPU_WRITING
))
317 nouveau_buffer_download(nv
, res
, 0, res
->base
.width0
);
319 if ((res
->domain
!= NOUVEAU_BO_GART
) ||
320 (res
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
))
321 return res
->data
+ offset
;
325 rw
= (flags
& NOUVEAU_BO_WR
) ? PIPE_TRANSFER_WRITE
: PIPE_TRANSFER_READ
;
326 nouveau_buffer_sync(res
, rw
);
327 if (nouveau_bo_map(res
->bo
, 0, NULL
))
330 if (nouveau_bo_map(res
->bo
, flags
, nv
->screen
->client
))
333 return (uint8_t *)res
->bo
->map
+ res
->offset
+ offset
;
337 const struct u_resource_vtbl nouveau_buffer_vtbl
=
339 u_default_resource_get_handle
, /* get_handle */
340 nouveau_buffer_destroy
, /* resource_destroy */
341 nouveau_buffer_transfer_get
, /* get_transfer */
342 nouveau_buffer_transfer_destroy
, /* transfer_destroy */
343 nouveau_buffer_transfer_map
, /* transfer_map */
344 nouveau_buffer_transfer_flush_region
, /* transfer_flush_region */
345 nouveau_buffer_transfer_unmap
, /* transfer_unmap */
346 u_default_transfer_inline_write
/* transfer_inline_write */
349 struct pipe_resource
*
350 nouveau_buffer_create(struct pipe_screen
*pscreen
,
351 const struct pipe_resource
*templ
)
353 struct nouveau_screen
*screen
= nouveau_screen(pscreen
);
354 struct nv04_resource
*buffer
;
357 buffer
= CALLOC_STRUCT(nv04_resource
);
361 buffer
->base
= *templ
;
362 buffer
->vtbl
= &nouveau_buffer_vtbl
;
363 pipe_reference_init(&buffer
->base
.reference
, 1);
364 buffer
->base
.screen
= pscreen
;
366 if ((buffer
->base
.bind
& screen
->sysmem_bindings
) == screen
->sysmem_bindings
)
367 ret
= nouveau_buffer_allocate(screen
, buffer
, 0);
369 ret
= nouveau_buffer_allocate(screen
, buffer
, NOUVEAU_BO_GART
);
374 return &buffer
->base
;
382 struct pipe_resource
*
383 nouveau_user_buffer_create(struct pipe_screen
*pscreen
, void *ptr
,
384 unsigned bytes
, unsigned bind
)
386 struct nv04_resource
*buffer
;
388 buffer
= CALLOC_STRUCT(nv04_resource
);
392 pipe_reference_init(&buffer
->base
.reference
, 1);
393 buffer
->vtbl
= &nouveau_buffer_vtbl
;
394 buffer
->base
.screen
= pscreen
;
395 buffer
->base
.format
= PIPE_FORMAT_R8_UNORM
;
396 buffer
->base
.usage
= PIPE_USAGE_IMMUTABLE
;
397 buffer
->base
.bind
= bind
;
398 buffer
->base
.width0
= bytes
;
399 buffer
->base
.height0
= 1;
400 buffer
->base
.depth0
= 1;
403 buffer
->status
= NOUVEAU_BUFFER_STATUS_USER_MEMORY
;
405 return &buffer
->base
;
408 /* Like download, but for GART buffers. Merge ? */
409 static INLINE boolean
410 nouveau_buffer_data_fetch(struct nouveau_context
*nv
, struct nv04_resource
*buf
,
411 struct nouveau_bo
*bo
, unsigned offset
, unsigned size
)
414 buf
->data
= MALLOC(size
);
418 if (nouveau_bo_map(bo
, NOUVEAU_BO_RD
, nv
->screen
->client
))
420 memcpy(buf
->data
, (uint8_t *)bo
->map
+ offset
, size
);
425 /* Migrate a linear buffer (vertex, index, constants) USER -> GART -> VRAM. */
427 nouveau_buffer_migrate(struct nouveau_context
*nv
,
428 struct nv04_resource
*buf
, const unsigned new_domain
)
430 struct nouveau_screen
*screen
= nv
->screen
;
431 struct nouveau_bo
*bo
;
432 const unsigned old_domain
= buf
->domain
;
433 unsigned size
= buf
->base
.width0
;
437 assert(new_domain
!= old_domain
);
439 if (new_domain
== NOUVEAU_BO_GART
&& old_domain
== 0) {
440 if (!nouveau_buffer_allocate(screen
, buf
, new_domain
))
442 ret
= nouveau_bo_map(buf
->bo
, 0, nv
->screen
->client
);
445 memcpy((uint8_t *)buf
->bo
->map
+ buf
->offset
, buf
->data
, size
);
448 if (old_domain
!= 0 && new_domain
!= 0) {
449 struct nouveau_mm_allocation
*mm
= buf
->mm
;
451 if (new_domain
== NOUVEAU_BO_VRAM
) {
452 /* keep a system memory copy of our data in case we hit a fallback */
453 if (!nouveau_buffer_data_fetch(nv
, buf
, buf
->bo
, buf
->offset
, size
))
455 if (nouveau_mesa_debug
)
456 debug_printf("migrating %u KiB to VRAM\n", size
/ 1024);
459 offset
= buf
->offset
;
463 nouveau_buffer_allocate(screen
, buf
, new_domain
);
465 nv
->copy_data(nv
, buf
->bo
, buf
->offset
, new_domain
,
466 bo
, offset
, old_domain
, buf
->base
.width0
);
468 nouveau_bo_ref(NULL
, &bo
);
470 release_allocation(&mm
, screen
->fence
.current
);
472 if (new_domain
== NOUVEAU_BO_VRAM
&& old_domain
== 0) {
473 if (!nouveau_buffer_allocate(screen
, buf
, NOUVEAU_BO_VRAM
))
475 if (!nouveau_buffer_upload(nv
, buf
, 0, buf
->base
.width0
))
480 assert(buf
->domain
== new_domain
);
484 /* Migrate data from glVertexAttribPointer(non-VBO) user buffers to GART.
485 * We'd like to only allocate @size bytes here, but then we'd have to rebase
486 * the vertex indices ...
489 nouveau_user_buffer_upload(struct nouveau_context
*nv
,
490 struct nv04_resource
*buf
,
491 unsigned base
, unsigned size
)
493 struct nouveau_screen
*screen
= nouveau_screen(buf
->base
.screen
);
496 assert(buf
->status
& NOUVEAU_BUFFER_STATUS_USER_MEMORY
);
498 buf
->base
.width0
= base
+ size
;
499 if (!nouveau_buffer_reallocate(screen
, buf
, NOUVEAU_BO_GART
))
502 ret
= nouveau_bo_map(buf
->bo
, 0, nv
->screen
->client
);
505 memcpy((uint8_t *)buf
->bo
->map
+ buf
->offset
+ base
, buf
->data
+ base
, size
);
511 /* Scratch data allocation. */
514 nouveau_scratch_bo_alloc(struct nouveau_context
*nv
, struct nouveau_bo
**pbo
,
517 return nouveau_bo_new(nv
->screen
->device
, NOUVEAU_BO_GART
| NOUVEAU_BO_MAP
,
518 4096, size
, NULL
, pbo
);
522 nouveau_scratch_runout_release(struct nouveau_context
*nv
)
524 if (!nv
->scratch
.nr_runout
)
526 while (nv
->scratch
.nr_runout
--)
527 nouveau_bo_ref(NULL
, &nv
->scratch
.runout
[nv
->scratch
.nr_runout
]);
529 FREE(nv
->scratch
.runout
);
531 nv
->scratch
.runout
= NULL
;
534 /* Allocate an extra bo if we can't fit everything we need simultaneously.
535 * (Could happen for very large user arrays.)
537 static INLINE boolean
538 nouveau_scratch_runout(struct nouveau_context
*nv
, unsigned size
)
541 const unsigned n
= nv
->scratch
.nr_runout
++;
543 nv
->scratch
.runout
= REALLOC(nv
->scratch
.runout
,
544 (n
+ 0) * sizeof(*nv
->scratch
.runout
),
545 (n
+ 1) * sizeof(*nv
->scratch
.runout
));
546 nv
->scratch
.runout
[n
] = NULL
;
548 ret
= nouveau_scratch_bo_alloc(nv
, &nv
->scratch
.runout
[n
], size
);
550 ret
= nouveau_bo_map(nv
->scratch
.runout
[n
], 0, NULL
);
552 nouveau_bo_ref(NULL
, &nv
->scratch
.runout
[--nv
->scratch
.nr_runout
]);
555 nv
->scratch
.current
= nv
->scratch
.runout
[n
];
556 nv
->scratch
.offset
= 0;
557 nv
->scratch
.end
= size
;
558 nv
->scratch
.map
= nv
->scratch
.current
->map
;
563 /* Continue to next scratch buffer, if available (no wrapping, large enough).
564 * Allocate it if it has not yet been created.
566 static INLINE boolean
567 nouveau_scratch_next(struct nouveau_context
*nv
, unsigned size
)
569 struct nouveau_bo
*bo
;
571 const unsigned i
= (nv
->scratch
.id
+ 1) % NOUVEAU_MAX_SCRATCH_BUFS
;
573 if ((size
> nv
->scratch
.bo_size
) || (i
== nv
->scratch
.wrap
))
577 bo
= nv
->scratch
.bo
[i
];
579 ret
= nouveau_scratch_bo_alloc(nv
, &bo
, nv
->scratch
.bo_size
);
582 nv
->scratch
.bo
[i
] = bo
;
584 nv
->scratch
.current
= bo
;
585 nv
->scratch
.offset
= 0;
586 nv
->scratch
.end
= nv
->scratch
.bo_size
;
588 ret
= nouveau_bo_map(bo
, NOUVEAU_BO_WR
, nv
->screen
->client
);
590 nv
->scratch
.map
= bo
->map
;
595 nouveau_scratch_more(struct nouveau_context
*nv
, unsigned min_size
)
599 ret
= nouveau_scratch_next(nv
, min_size
);
601 ret
= nouveau_scratch_runout(nv
, min_size
);
605 /* Upload data to scratch memory and update buffer address.
606 * Returns the bo the data resides in, if successful.
609 nouveau_scratch_data(struct nouveau_context
*nv
,
610 struct nv04_resource
*buf
, unsigned base
, unsigned size
)
612 struct nouveau_bo
*bo
;
613 unsigned bgn
= MAX2(base
, nv
->scratch
.offset
);
614 unsigned end
= bgn
+ size
;
616 if (end
>= nv
->scratch
.end
) {
618 if (!nouveau_scratch_more(nv
, end
))
622 nv
->scratch
.offset
= align(end
, 4);
624 memcpy(nv
->scratch
.map
+ bgn
, buf
->data
+ base
, size
);
626 bo
= nv
->scratch
.current
;
627 buf
->address
= bo
->offset
+ (bgn
- base
);
632 nouveau_scratch_get(struct nouveau_context
*nv
,
633 unsigned size
, uint64_t *gpu_addr
, struct nouveau_bo
**pbo
)
635 unsigned bgn
= nv
->scratch
.offset
;
636 unsigned end
= nv
->scratch
.offset
+ size
;
638 if (end
>= nv
->scratch
.end
) {
640 if (!nouveau_scratch_more(nv
, end
))
644 nv
->scratch
.offset
= align(end
, 4);
646 *pbo
= nv
->scratch
.current
;
647 *gpu_addr
= nv
->scratch
.current
->offset
+ bgn
;
648 return nv
->scratch
.map
+ bgn
;