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_math.h"
34 #include "util/u_memory.h"
35 #include "util/u_inlines.h"
36 #include "util/u_framebuffer.h"
37 #include "r600_shader.h"
38 #include "r600_pipe.h"
39 #include "r600_formats.h"
40 #include "compute_memory_pool.h"
41 #include "evergreen_compute.h"
42 #include "evergreen_compute_internal.h"
45 #define ITEM_ALIGNMENT 1024
49 struct compute_memory_pool
* compute_memory_pool_new(
50 struct r600_screen
* rscreen
)
52 struct compute_memory_pool
* pool
= (struct compute_memory_pool
*)
53 CALLOC(sizeof(struct compute_memory_pool
), 1);
57 COMPUTE_DBG(rscreen
, "* compute_memory_pool_new()\n");
59 pool
->screen
= rscreen
;
63 static void compute_memory_pool_init(struct compute_memory_pool
* pool
,
64 unsigned initial_size_in_dw
)
67 COMPUTE_DBG(pool
->screen
, "* compute_memory_pool_init() initial_size_in_dw = %ld\n",
70 pool
->shadow
= (uint32_t*)CALLOC(initial_size_in_dw
, 4);
71 if (pool
->shadow
== NULL
)
75 pool
->size_in_dw
= initial_size_in_dw
;
76 pool
->bo
= (struct r600_resource
*)r600_compute_buffer_alloc_vram(pool
->screen
,
77 pool
->size_in_dw
* 4);
81 * Frees all stuff in the pool and the pool struct itself too
83 void compute_memory_pool_delete(struct compute_memory_pool
* pool
)
85 COMPUTE_DBG(pool
->screen
, "* compute_memory_pool_delete()\n");
88 pool
->screen
->b
.b
.resource_destroy((struct pipe_screen
*)
89 pool
->screen
, (struct pipe_resource
*)pool
->bo
);
95 * Searches for an empty space in the pool, return with the pointer to the
96 * allocatable space in the pool, returns -1 on failure.
98 int64_t compute_memory_prealloc_chunk(
99 struct compute_memory_pool
* pool
,
102 struct compute_memory_item
*item
;
106 assert(size_in_dw
<= pool
->size_in_dw
);
108 COMPUTE_DBG(pool
->screen
, "* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
111 for (item
= pool
->item_list
; item
; item
= item
->next
) {
112 if (item
->start_in_dw
> -1) {
113 if (item
->start_in_dw
-last_end
> size_in_dw
) {
117 last_end
= item
->start_in_dw
+ align(item
->size_in_dw
, ITEM_ALIGNMENT
);
121 if (pool
->size_in_dw
- last_end
< size_in_dw
) {
129 * Search for the chunk where we can link our new chunk after it.
131 struct compute_memory_item
* compute_memory_postalloc_chunk(
132 struct compute_memory_pool
* pool
,
135 struct compute_memory_item
* item
;
137 COMPUTE_DBG(pool
->screen
, "* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
140 /* Check if we can insert it in the front of the list */
141 if (pool
->item_list
&& pool
->item_list
->start_in_dw
> start_in_dw
) {
145 for (item
= pool
->item_list
; item
; item
= item
->next
) {
147 if (item
->start_in_dw
< start_in_dw
148 && item
->next
->start_in_dw
> start_in_dw
) {
154 assert(item
->start_in_dw
< start_in_dw
);
159 assert(0 && "unreachable");
164 * Reallocates pool, conserves data.
165 * @returns -1 if it fails, 0 otherwise
167 int compute_memory_grow_pool(struct compute_memory_pool
* pool
,
168 struct pipe_context
* pipe
, int new_size_in_dw
)
170 COMPUTE_DBG(pool
->screen
, "* compute_memory_grow_pool() "
171 "new_size_in_dw = %d (%d bytes)\n",
172 new_size_in_dw
, new_size_in_dw
* 4);
174 assert(new_size_in_dw
>= pool
->size_in_dw
);
177 compute_memory_pool_init(pool
, MAX2(new_size_in_dw
, 1024 * 16));
178 if (pool
->shadow
== NULL
)
181 new_size_in_dw
= align(new_size_in_dw
, ITEM_ALIGNMENT
);
183 COMPUTE_DBG(pool
->screen
, " Aligned size = %d (%d bytes)\n",
184 new_size_in_dw
, new_size_in_dw
* 4);
186 compute_memory_shadow(pool
, pipe
, 1);
187 pool
->shadow
= realloc(pool
->shadow
, new_size_in_dw
*4);
188 if (pool
->shadow
== NULL
)
191 pool
->size_in_dw
= new_size_in_dw
;
192 pool
->screen
->b
.b
.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);
205 * Copy pool from device to host, or host to device.
207 void compute_memory_shadow(struct compute_memory_pool
* pool
,
208 struct pipe_context
* pipe
, int device_to_host
)
210 struct compute_memory_item chunk
;
212 COMPUTE_DBG(pool
->screen
, "* compute_memory_shadow() device_to_host = %d\n",
216 chunk
.start_in_dw
= 0;
217 chunk
.size_in_dw
= pool
->size_in_dw
;
218 chunk
.prev
= chunk
.next
= NULL
;
219 compute_memory_transfer(pool
, pipe
, device_to_host
, &chunk
,
220 pool
->shadow
, 0, pool
->size_in_dw
*4);
224 * Allocates pending allocations in the pool
225 * @returns -1 if it fails, 0 otherwise
227 int compute_memory_finalize_pending(struct compute_memory_pool
* pool
,
228 struct pipe_context
* pipe
)
230 struct compute_memory_item
*pending_list
= NULL
, *end_p
= NULL
;
231 struct compute_memory_item
*item
, *next
;
233 int64_t allocated
= 0;
234 int64_t unallocated
= 0;
236 int64_t start_in_dw
= 0;
240 COMPUTE_DBG(pool
->screen
, "* compute_memory_finalize_pending()\n");
242 for (item
= pool
->item_list
; item
; item
= item
->next
) {
243 COMPUTE_DBG(pool
->screen
, " + list: offset = %i id = %i size = %i "
244 "(%i bytes)\n",item
->start_in_dw
, item
->id
,
245 item
->size_in_dw
, item
->size_in_dw
* 4);
248 /* Search through the list of memory items in the pool */
249 for (item
= pool
->item_list
; item
; item
= next
) {
252 /* Check if the item is pending. */
253 if (item
->start_in_dw
== -1) {
254 /* It is pending, so add it to the pending_list... */
262 /* ... and then remove it from the item list. */
264 item
->prev
->next
= next
;
267 pool
->item_list
= next
;
271 next
->prev
= item
->prev
;
274 /* This sequence makes the item be at the end of the list */
279 /* Update the amount of space we will need to allocate. */
280 unallocated
+= item
->size_in_dw
+1024;
283 /* The item is not pending, so update the amount of space
284 * that has already been allocated. */
285 allocated
+= item
->size_in_dw
;
289 /* If we require more space than the size of the pool, then grow the
292 * XXX: I'm pretty sure this won't work. Imagine this scenario:
300 * Allocated size = 150
301 * Pending Item D Size = 200
303 * In this case, there are 300 units of free space in the pool, but
304 * they aren't contiguous, so it will be impossible to allocate Item D.
306 if (pool
->size_in_dw
< allocated
+unallocated
) {
307 err
= compute_memory_grow_pool(pool
, pipe
, allocated
+unallocated
);
312 /* Loop through all the pending items, allocate space for them and
313 * add them back to the item_list. */
314 for (item
= pending_list
; item
; item
= next
) {
317 /* Search for free space in the pool for this item. */
318 while ((start_in_dw
=compute_memory_prealloc_chunk(pool
,
319 item
->size_in_dw
)) == -1) {
320 int64_t need
= item
->size_in_dw
+2048 -
321 (pool
->size_in_dw
- allocated
);
324 need
= pool
->size_in_dw
/ 10;
327 need
= align(need
, ITEM_ALIGNMENT
);
329 err
= compute_memory_grow_pool(pool
,
331 pool
->size_in_dw
+ need
);
336 COMPUTE_DBG(pool
->screen
, " + Found space for Item %p id = %u "
337 "start_in_dw = %u (%u bytes) size_in_dw = %u (%u bytes)\n",
338 item
, item
->id
, start_in_dw
, start_in_dw
* 4,
339 item
->size_in_dw
, item
->size_in_dw
* 4);
341 item
->start_in_dw
= start_in_dw
;
345 if (pool
->item_list
) {
346 struct compute_memory_item
*pos
;
348 pos
= compute_memory_postalloc_chunk(pool
, start_in_dw
);
351 item
->next
= pos
->next
;
354 item
->next
->prev
= item
;
357 /* Add item to the front of the list */
358 item
->next
= pool
->item_list
;
359 item
->prev
= pool
->item_list
->prev
;
360 pool
->item_list
->prev
= item
;
361 pool
->item_list
= item
;
365 pool
->item_list
= item
;
368 allocated
+= item
->size_in_dw
;
375 void compute_memory_free(struct compute_memory_pool
* pool
, int64_t id
)
377 struct compute_memory_item
*item
, *next
;
379 COMPUTE_DBG(pool
->screen
, "* compute_memory_free() id + %ld \n", id
);
381 for (item
= pool
->item_list
; item
; item
= next
) {
384 if (item
->id
== id
) {
386 item
->prev
->next
= item
->next
;
389 pool
->item_list
= item
->next
;
393 item
->next
->prev
= item
->prev
;
402 fprintf(stderr
, "Internal error, invalid id %"PRIi64
" "
403 "for compute_memory_free\n", id
);
405 assert(0 && "error");
409 * Creates pending allocations
411 struct compute_memory_item
* compute_memory_alloc(
412 struct compute_memory_pool
* pool
,
415 struct compute_memory_item
*new_item
= NULL
, *last_item
= NULL
;
417 COMPUTE_DBG(pool
->screen
, "* compute_memory_alloc() size_in_dw = %ld (%ld bytes)\n",
418 size_in_dw
, 4 * size_in_dw
);
420 new_item
= (struct compute_memory_item
*)
421 CALLOC(sizeof(struct compute_memory_item
), 1);
422 if (new_item
== NULL
)
425 new_item
->size_in_dw
= size_in_dw
;
426 new_item
->start_in_dw
= -1; /* mark pending */
427 new_item
->id
= pool
->next_id
++;
428 new_item
->pool
= pool
;
430 if (pool
->item_list
) {
431 for (last_item
= pool
->item_list
; last_item
->next
;
432 last_item
= last_item
->next
);
434 last_item
->next
= new_item
;
435 new_item
->prev
= last_item
;
438 pool
->item_list
= new_item
;
441 COMPUTE_DBG(pool
->screen
, " + Adding item %p id = %u size = %u (%u bytes)\n",
442 new_item
, new_item
->id
, new_item
->size_in_dw
,
443 new_item
->size_in_dw
* 4);
448 * Transfer data host<->device, offset and size is in bytes
450 void compute_memory_transfer(
451 struct compute_memory_pool
* pool
,
452 struct pipe_context
* pipe
,
454 struct compute_memory_item
* chunk
,
459 int64_t aligned_size
= pool
->size_in_dw
;
460 struct pipe_resource
* gart
= (struct pipe_resource
*)pool
->bo
;
461 int64_t internal_offset
= chunk
->start_in_dw
*4 + offset_in_chunk
;
463 struct pipe_transfer
*xfer
;
468 COMPUTE_DBG(pool
->screen
, "* compute_memory_transfer() device_to_host = %d, "
469 "offset_in_chunk = %d, size = %d\n", device_to_host
,
470 offset_in_chunk
, size
);
472 if (device_to_host
) {
473 map
= pipe
->transfer_map(pipe
, gart
, 0, PIPE_TRANSFER_READ
,
474 &(struct pipe_box
) { .width
= aligned_size
* 4,
475 .height
= 1, .depth
= 1 }, &xfer
);
478 memcpy(data
, map
+ internal_offset
, size
);
479 pipe
->transfer_unmap(pipe
, xfer
);
481 map
= pipe
->transfer_map(pipe
, gart
, 0, PIPE_TRANSFER_WRITE
,
482 &(struct pipe_box
) { .width
= aligned_size
* 4,
483 .height
= 1, .depth
= 1 }, &xfer
);
486 memcpy(map
+ internal_offset
, data
, size
);
487 pipe
->transfer_unmap(pipe
, xfer
);
492 * Transfer data between chunk<->data, it is for VRAM<->GART transfers
494 void compute_memory_transfer_direct(
495 struct compute_memory_pool
* pool
,
497 struct compute_memory_item
* chunk
,
498 struct r600_resource
* data
,