2 * Permission is hereby granted, free of charge, to any person obtaining a
3 * copy of this software and associated documentation files (the "Software"),
4 * to deal in the Software without restriction, including without limitation
5 * on the rights to use, copy, modify, merge, publish, distribute, sub
6 * license, and/or sell copies of the Software, and to permit persons to whom
7 * the Software is furnished to do so, subject to the following conditions:
9 * The above copyright notice and this permission notice (including the next
10 * paragraph) shall be included in all copies or substantial portions of the
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * Adam Rak <adam.rak@streamnovation.com>
25 #include "pipe/p_defines.h"
26 #include "pipe/p_state.h"
27 #include "pipe/p_context.h"
28 #include "util/u_blitter.h"
29 #include "util/u_double_list.h"
30 #include "util/u_transfer.h"
31 #include "util/u_surface.h"
32 #include "util/u_pack_color.h"
33 #include "util/u_memory.h"
34 #include "util/u_inlines.h"
35 #include "util/u_framebuffer.h"
37 #include "r600_resource.h"
38 #include "r600_shader.h"
39 #include "r600_pipe.h"
40 #include "r600_formats.h"
41 #include "compute_memory_pool.h"
42 #include "evergreen_compute_internal.h"
48 struct compute_memory_pool
* compute_memory_pool_new(
49 struct r600_screen
* rscreen
)
51 struct compute_memory_pool
* pool
= (struct compute_memory_pool
*)
52 CALLOC(sizeof(struct compute_memory_pool
), 1);
54 COMPUTE_DBG("* compute_memory_pool_new()\n");
56 pool
->screen
= rscreen
;
60 static void compute_memory_pool_init(struct compute_memory_pool
* pool
,
61 unsigned initial_size_in_dw
)
64 COMPUTE_DBG("* compute_memory_pool_init() initial_size_in_dw = %ld\n",
67 /* XXX: pool->shadow is used when the buffer needs to be resized, but
68 * resizing does not work at the moment.
69 * pool->shadow = (uint32_t*)CALLOC(4, pool->size_in_dw);
72 pool
->size_in_dw
= initial_size_in_dw
;
73 pool
->bo
= (struct r600_resource
*)r600_compute_buffer_alloc_vram(pool
->screen
,
74 pool
->size_in_dw
* 4);
78 * Frees all stuff in the pool and the pool struct itself too
80 void compute_memory_pool_delete(struct compute_memory_pool
* pool
)
82 COMPUTE_DBG("* compute_memory_pool_delete()\n");
85 pool
->screen
->screen
.resource_destroy((struct pipe_screen
*)
86 pool
->screen
, (struct pipe_resource
*)pool
->bo
);
92 * Searches for an empty space in the pool, return with the pointer to the
93 * allocatable space in the pool, returns -1 on failure.
95 int64_t compute_memory_prealloc_chunk(
96 struct compute_memory_pool
* pool
,
99 assert(size_in_dw
<= pool
->size_in_dw
);
101 struct compute_memory_item
*item
;
105 COMPUTE_DBG("* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
108 for (item
= pool
->item_list
; item
; item
= item
->next
) {
109 if (item
->start_in_dw
> -1) {
110 if (item
->start_in_dw
-last_end
> size_in_dw
) {
114 last_end
= item
->start_in_dw
+ item
->size_in_dw
;
115 last_end
+= (1024 - last_end
% 1024);
119 if (pool
->size_in_dw
- last_end
< size_in_dw
) {
127 * Search for the chunk where we can link our new chunk after it.
129 struct compute_memory_item
* compute_memory_postalloc_chunk(
130 struct compute_memory_pool
* pool
,
133 struct compute_memory_item
* item
;
135 COMPUTE_DBG("* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
138 /* Check if we can insert it in the front of the list */
139 if (pool
->item_list
&& pool
->item_list
->start_in_dw
> start_in_dw
) {
143 for (item
= pool
->item_list
; item
; item
= item
->next
) {
145 if (item
->start_in_dw
< start_in_dw
146 && item
->next
->start_in_dw
> start_in_dw
) {
152 assert(item
->start_in_dw
< start_in_dw
);
157 assert(0 && "unreachable");
162 * Reallocates pool, conserves data
164 void compute_memory_grow_pool(struct compute_memory_pool
* pool
,
165 struct pipe_context
* pipe
, int new_size_in_dw
)
167 COMPUTE_DBG("* compute_memory_grow_pool() new_size_in_dw = %d\n",
170 assert(new_size_in_dw
>= pool
->size_in_dw
);
172 assert(!pool
->bo
&& "Growing the global memory pool is not yet "
173 "supported. You will see this message if you are trying to"
174 "use more than 64 kb of memory");
177 compute_memory_pool_init(pool
, MAX2(new_size_in_dw
, 1024 * 16));
179 /* XXX: Growing memory pools does not work at the moment. I think
180 * it is because we are using fragment shaders to copy data to
181 * the new texture and some of the compute registers are being
182 * included in the 3D command stream. */
183 fprintf(stderr
, "Warning: growing the global memory pool to"
184 "more than 64 kb is not yet supported\n");
185 new_size_in_dw
+= 1024 - (new_size_in_dw
% 1024);
187 COMPUTE_DBG(" Aligned size = %d\n", new_size_in_dw
);
189 compute_memory_shadow(pool
, pipe
, 1);
190 pool
->shadow
= realloc(pool
->shadow
, new_size_in_dw
*4);
191 pool
->size_in_dw
= new_size_in_dw
;
192 pool
->screen
->screen
.resource_destroy(
193 (struct pipe_screen
*)pool
->screen
,
194 (struct pipe_resource
*)pool
->bo
);
195 pool
->bo
= (struct r600_resource
*)r600_compute_buffer_alloc_vram(
197 pool
->size_in_dw
* 4);
198 compute_memory_shadow(pool
, pipe
, 0);
203 * Copy pool from device to host, or host to device.
205 void compute_memory_shadow(struct compute_memory_pool
* pool
,
206 struct pipe_context
* pipe
, int device_to_host
)
208 struct compute_memory_item chunk
;
210 COMPUTE_DBG("* compute_memory_shadow() device_to_host = %d\n",
214 chunk
.start_in_dw
= 0;
215 chunk
.size_in_dw
= pool
->size_in_dw
;
216 chunk
.prev
= chunk
.next
= NULL
;
217 compute_memory_transfer(pool
, pipe
, device_to_host
, &chunk
,
218 pool
->shadow
, 0, pool
->size_in_dw
*4);
222 * Allocates pending allocations in the pool
224 void compute_memory_finalize_pending(struct compute_memory_pool
* pool
,
225 struct pipe_context
* pipe
)
227 struct compute_memory_item
*pending_list
= NULL
, *end_p
= NULL
;
228 struct compute_memory_item
*item
, *next
;
230 int64_t allocated
= 0;
231 int64_t unallocated
= 0;
233 COMPUTE_DBG("* compute_memory_finalize_pending()\n");
235 for (item
= pool
->item_list
; item
; item
= item
->next
) {
236 COMPUTE_DBG(" + list: offset = %i id = %i size = %i "
237 "(%i bytes)\n",item
->start_in_dw
, item
->id
,
238 item
->size_in_dw
, item
->size_in_dw
* 4);
241 /* Search through the list of memory items in the pool */
242 for (item
= pool
->item_list
; item
; item
= next
) {
245 /* Check if the item is pending. */
246 if (item
->start_in_dw
== -1) {
247 /* It is pending, so add it to the pending_list... */
255 /* ... and then remove it from the item list. */
257 item
->prev
->next
= next
;
260 pool
->item_list
= next
;
264 next
->prev
= item
->prev
;
267 /* This sequence makes the item be at the end of the list */
272 /* Update the amount of space we will need to allocate. */
273 unallocated
+= item
->size_in_dw
+1024;
276 /* The item is not pendng, so update the amount of space
277 * that has already been allocated. */
278 allocated
+= item
->size_in_dw
;
282 /* If we require more space than the size of the pool, then grow the
285 * XXX: I'm pretty sure this won't work. Imagine this scenario:
293 * Allocated size = 150
294 * Pending Item D Size = 200
296 * In this case, there are 300 units of free space in the pool, but
297 * they aren't contiguous, so it will be impossible to allocate Item D.
299 if (pool
->size_in_dw
< allocated
+unallocated
) {
300 compute_memory_grow_pool(pool
, pipe
, allocated
+unallocated
);
303 /* Loop through all the pending items, allocate space for them and
304 * add them back to the item_list. */
305 for (item
= pending_list
; item
; item
= next
) {
310 /* Search for free space in the pool for this item. */
311 while ((start_in_dw
=compute_memory_prealloc_chunk(pool
,
312 item
->size_in_dw
)) == -1) {
313 int64_t need
= item
->size_in_dw
+2048 -
314 (pool
->size_in_dw
- allocated
);
316 need
+= 1024 - (need
% 1024);
319 compute_memory_grow_pool(pool
,
321 pool
->size_in_dw
+ need
);
324 need
= pool
->size_in_dw
/ 10;
325 need
+= 1024 - (need
% 1024);
326 compute_memory_grow_pool(pool
,
328 pool
->size_in_dw
+ need
);
331 COMPUTE_DBG(" + Found space for Item %p id = %u "
332 "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
333 item
, item
->id
, start_in_dw
, start_in_dw
* 4,
334 item
->size_in_dw
, item
->size_in_dw
* 4);
336 item
->start_in_dw
= start_in_dw
;
340 if (pool
->item_list
) {
341 struct compute_memory_item
*pos
;
343 pos
= compute_memory_postalloc_chunk(pool
, start_in_dw
);
346 item
->next
= pos
->next
;
349 item
->next
->prev
= item
;
352 /* Add item to the front of the list */
353 item
->next
= pool
->item_list
->next
;
354 if (pool
->item_list
->next
) {
355 pool
->item_list
->next
->prev
= item
;
357 item
->prev
= pool
->item_list
->prev
;
358 if (pool
->item_list
->prev
) {
359 pool
->item_list
->prev
->next
= item
;
361 pool
->item_list
= item
;
365 pool
->item_list
= item
;
368 allocated
+= item
->size_in_dw
;
373 void compute_memory_free(struct compute_memory_pool
* pool
, int64_t id
)
375 struct compute_memory_item
*item
, *next
;
377 COMPUTE_DBG("* compute_memory_free() id + %ld \n", id
);
379 for (item
= pool
->item_list
; item
; item
= next
) {
382 if (item
->id
== id
) {
384 item
->prev
->next
= item
->next
;
387 pool
->item_list
= item
->next
;
391 item
->next
->prev
= item
->prev
;
400 fprintf(stderr
, "Internal error, invalid id %"PRIi64
" "
401 "for compute_memory_free\n", id
);
403 assert(0 && "error");
407 * Creates pending allocations
409 struct compute_memory_item
* compute_memory_alloc(
410 struct compute_memory_pool
* pool
,
413 struct compute_memory_item
*new_item
;
415 COMPUTE_DBG("* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
416 size_in_dw
, 4 * size_in_dw
);
418 new_item
= (struct compute_memory_item
*)
419 CALLOC(sizeof(struct compute_memory_item
), 1);
420 new_item
->size_in_dw
= size_in_dw
;
421 new_item
->start_in_dw
= -1; /* mark pending */
422 new_item
->id
= pool
->next_id
++;
423 new_item
->pool
= pool
;
425 struct compute_memory_item
*last_item
;
427 if (pool
->item_list
) {
428 for (last_item
= pool
->item_list
; last_item
->next
;
429 last_item
= last_item
->next
);
431 last_item
->next
= new_item
;
432 new_item
->prev
= last_item
;
435 pool
->item_list
= new_item
;
438 COMPUTE_DBG(" + Adding item %p id = %u size = %u (%u bytes)\n",
439 new_item
, new_item
->id
, new_item
->size_in_dw
,
440 new_item
->size_in_dw
* 4);
445 * Transfer data host<->device, offset and size is in bytes
447 void compute_memory_transfer(
448 struct compute_memory_pool
* pool
,
449 struct pipe_context
* pipe
,
451 struct compute_memory_item
* chunk
,
456 int64_t aligned_size
= pool
->size_in_dw
;
457 struct pipe_resource
* gart
= (struct pipe_resource
*)pool
->bo
;
458 int64_t internal_offset
= chunk
->start_in_dw
*4 + offset_in_chunk
;
460 struct pipe_transfer
*xfer
;
465 COMPUTE_DBG("* compute_memory_transfer() device_to_host = %d, "
466 "offset_in_chunk = %d, size = %d\n", device_to_host
,
467 offset_in_chunk
, size
);
471 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_READ
,
472 &(struct pipe_box
) { .width
= aligned_size
,
473 .height
= 1, .depth
= 1 });
475 map
= pipe
->transfer_map(pipe
, xfer
);
477 memcpy(data
, map
+ internal_offset
, size
);
478 pipe
->transfer_unmap(pipe
, xfer
);
479 pipe
->transfer_destroy(pipe
, xfer
);
481 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_WRITE
,
482 &(struct pipe_box
) { .width
= aligned_size
,
483 .height
= 1, .depth
= 1 });
485 map
= pipe
->transfer_map(pipe
, xfer
);
487 memcpy(map
+ internal_offset
, data
, size
);
488 pipe
->transfer_unmap(pipe
, xfer
);
489 pipe
->transfer_destroy(pipe
, xfer
);
494 * Transfer data between chunk<->data, it is for VRAM<->GART transfers
496 void compute_memory_transfer_direct(
497 struct compute_memory_pool
* pool
,
499 struct compute_memory_item
* chunk
,
500 struct r600_resource
* data
,