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 for (item
= pool
->item_list
; item
; item
= item
->next
) {
140 if (item
->start_in_dw
< start_in_dw
141 && item
->next
->start_in_dw
> start_in_dw
) {
147 assert(item
->start_in_dw
< start_in_dw
);
152 assert(0 && "unreachable");
157 * Reallocates pool, conserves data
159 void compute_memory_grow_pool(struct compute_memory_pool
* pool
,
160 struct pipe_context
* pipe
, int new_size_in_dw
)
162 COMPUTE_DBG("* compute_memory_grow_pool() new_size_in_dw = %d\n",
165 assert(new_size_in_dw
>= pool
->size_in_dw
);
167 assert(!pool
->bo
&& "Growing the global memory pool is not yet "
168 "supported. You will see this message if you are trying to"
169 "use more than 64 kb of memory");
172 compute_memory_pool_init(pool
, MAX2(new_size_in_dw
, 1024 * 16));
174 /* XXX: Growing memory pools does not work at the moment. I think
175 * it is because we are using fragment shaders to copy data to
176 * the new texture and some of the compute registers are being
177 * included in the 3D command stream. */
178 fprintf(stderr
, "Warning: growing the global memory pool to"
179 "more than 64 kb is not yet supported\n");
180 new_size_in_dw
+= 1024 - (new_size_in_dw
% 1024);
182 COMPUTE_DBG(" Aligned size = %d\n", new_size_in_dw
);
184 compute_memory_shadow(pool
, pipe
, 1);
185 pool
->shadow
= realloc(pool
->shadow
, new_size_in_dw
*4);
186 pool
->size_in_dw
= new_size_in_dw
;
187 pool
->screen
->screen
.resource_destroy(
188 (struct pipe_screen
*)pool
->screen
,
189 (struct pipe_resource
*)pool
->bo
);
190 pool
->bo
= (struct r600_resource
*)r600_compute_buffer_alloc_vram(
192 pool
->size_in_dw
* 4);
193 compute_memory_shadow(pool
, pipe
, 0);
198 * Copy pool from device to host, or host to device.
200 void compute_memory_shadow(struct compute_memory_pool
* pool
,
201 struct pipe_context
* pipe
, int device_to_host
)
203 struct compute_memory_item chunk
;
205 COMPUTE_DBG("* compute_memory_shadow() device_to_host = %d\n",
209 chunk
.start_in_dw
= 0;
210 chunk
.size_in_dw
= pool
->size_in_dw
;
211 chunk
.prev
= chunk
.next
= NULL
;
212 compute_memory_transfer(pool
, pipe
, device_to_host
, &chunk
,
213 pool
->shadow
, 0, pool
->size_in_dw
*4);
217 * Allocates pending allocations in the pool
219 void compute_memory_finalize_pending(struct compute_memory_pool
* pool
,
220 struct pipe_context
* pipe
)
222 struct compute_memory_item
*pending_list
= NULL
, *end_p
= NULL
;
223 struct compute_memory_item
*item
, *next
;
225 int64_t allocated
= 0;
226 int64_t unallocated
= 0;
228 COMPUTE_DBG("* compute_memory_finalize_pending()\n");
230 for (item
= pool
->item_list
; item
; item
= item
->next
) {
231 COMPUTE_DBG(" + list: offset = %i id = %i size = %i "
232 "(%i bytes)\n",item
->start_in_dw
, item
->id
,
233 item
->size_in_dw
, item
->size_in_dw
* 4);
236 /* Search through the list of memory items in the pool */
237 for (item
= pool
->item_list
; item
; item
= next
) {
240 /* Check if the item is pending. */
241 if (item
->start_in_dw
== -1) {
242 /* It is pending, so add it to the pending_list... */
250 /* ... and then remove it from the item list. */
252 item
->prev
->next
= next
;
255 pool
->item_list
= next
;
259 next
->prev
= item
->prev
;
262 /* This sequence makes the item be at the end of the list */
267 /* Update the amount of space we will need to allocate. */
268 unallocated
+= item
->size_in_dw
+1024;
271 /* The item is not pendng, so update the amount of space
272 * that has already been allocated. */
273 allocated
+= item
->size_in_dw
;
277 /* If we require more space than the size of the pool, then grow the
280 * XXX: I'm pretty sure this won't work. Imagine this scenario:
288 * Allocated size = 150
289 * Pending Item D Size = 200
291 * In this case, there are 300 units of free space in the pool, but
292 * they aren't contiguous, so it will be impossible to allocate Item D.
294 if (pool
->size_in_dw
< allocated
+unallocated
) {
295 compute_memory_grow_pool(pool
, pipe
, allocated
+unallocated
);
298 /* Loop through all the pending items, allocate space for them and
299 * add them back to the item_list. */
300 for (item
= pending_list
; item
; item
= next
) {
305 /* Search for free space in the pool for this item. */
306 while ((start_in_dw
=compute_memory_prealloc_chunk(pool
,
307 item
->size_in_dw
)) == -1) {
308 int64_t need
= item
->size_in_dw
+2048 -
309 (pool
->size_in_dw
- allocated
);
311 need
+= 1024 - (need
% 1024);
314 compute_memory_grow_pool(pool
,
316 pool
->size_in_dw
+ need
);
319 need
= pool
->size_in_dw
/ 10;
320 need
+= 1024 - (need
% 1024);
321 compute_memory_grow_pool(pool
,
323 pool
->size_in_dw
+ need
);
326 COMPUTE_DBG(" + Found space for Item %p id = %u "
327 "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
328 item
, item
->id
, start_in_dw
, start_in_dw
* 4,
329 item
->size_in_dw
, item
->size_in_dw
* 4);
331 item
->start_in_dw
= start_in_dw
;
335 if (pool
->item_list
) {
336 struct compute_memory_item
*pos
;
338 pos
= compute_memory_postalloc_chunk(pool
, start_in_dw
);
340 item
->next
= pos
->next
;
344 item
->next
->prev
= item
;
348 pool
->item_list
= item
;
351 allocated
+= item
->size_in_dw
;
356 void compute_memory_free(struct compute_memory_pool
* pool
, int64_t id
)
358 struct compute_memory_item
*item
, *next
;
360 COMPUTE_DBG("* compute_memory_free() id + %ld \n", id
);
362 for (item
= pool
->item_list
; item
; item
= next
) {
365 if (item
->id
== id
) {
367 item
->prev
->next
= item
->next
;
370 pool
->item_list
= item
->next
;
374 item
->next
->prev
= item
->prev
;
383 fprintf(stderr
, "Internal error, invalid id %"PRIi64
" "
384 "for compute_memory_free\n", id
);
386 assert(0 && "error");
390 * Creates pending allocations
392 struct compute_memory_item
* compute_memory_alloc(
393 struct compute_memory_pool
* pool
,
396 struct compute_memory_item
*new_item
;
398 COMPUTE_DBG("* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
399 size_in_dw
, 4 * size_in_dw
);
401 new_item
= (struct compute_memory_item
*)
402 CALLOC(sizeof(struct compute_memory_item
), 1);
403 new_item
->size_in_dw
= size_in_dw
;
404 new_item
->start_in_dw
= -1; /* mark pending */
405 new_item
->id
= pool
->next_id
++;
406 new_item
->pool
= pool
;
408 struct compute_memory_item
*last_item
;
410 if (pool
->item_list
) {
411 for (last_item
= pool
->item_list
; last_item
->next
;
412 last_item
= last_item
->next
);
414 last_item
->next
= new_item
;
415 new_item
->prev
= last_item
;
418 pool
->item_list
= new_item
;
421 COMPUTE_DBG(" + Adding item %p id = %u size = %u (%u bytes)\n",
422 new_item
, new_item
->id
, new_item
->size_in_dw
,
423 new_item
->size_in_dw
* 4);
428 * Transfer data host<->device, offset and size is in bytes
430 void compute_memory_transfer(
431 struct compute_memory_pool
* pool
,
432 struct pipe_context
* pipe
,
434 struct compute_memory_item
* chunk
,
439 int64_t aligned_size
= pool
->size_in_dw
;
440 struct pipe_resource
* gart
= (struct pipe_resource
*)pool
->bo
;
441 int64_t internal_offset
= chunk
->start_in_dw
*4 + offset_in_chunk
;
443 struct pipe_transfer
*xfer
;
448 COMPUTE_DBG("* compute_memory_transfer() device_to_host = %d, "
449 "offset_in_chunk = %d, size = %d\n", device_to_host
,
450 offset_in_chunk
, size
);
454 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_READ
,
455 &(struct pipe_box
) { .width
= aligned_size
,
456 .height
= 1, .depth
= 1 });
458 map
= pipe
->transfer_map(pipe
, xfer
);
460 memcpy(data
, map
+ internal_offset
, size
);
461 pipe
->transfer_unmap(pipe
, xfer
);
462 pipe
->transfer_destroy(pipe
, xfer
);
464 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_WRITE
,
465 &(struct pipe_box
) { .width
= aligned_size
,
466 .height
= 1, .depth
= 1 });
468 map
= pipe
->transfer_map(pipe
, xfer
);
470 memcpy(map
+ internal_offset
, data
, size
);
471 pipe
->transfer_unmap(pipe
, xfer
);
472 pipe
->transfer_destroy(pipe
, xfer
);
477 * Transfer data between chunk<->data, it is for VRAM<->GART transfers
479 void compute_memory_transfer_direct(
480 struct compute_memory_pool
* pool
,
482 struct compute_memory_item
* chunk
,
483 struct r600_resource
* data
,