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"
44 static struct r600_resource_texture
* create_pool_texture(struct r600_screen
* screen
,
48 struct pipe_resource templ
;
49 struct r600_resource_texture
* tex
;
51 if (size_in_dw
== 0) {
54 memset(&templ
, 0, sizeof(templ
));
55 templ
.target
= PIPE_TEXTURE_1D
;
56 templ
.format
= PIPE_FORMAT_R32_UINT
;
57 templ
.bind
= PIPE_BIND_CUSTOM
;
58 templ
.usage
= PIPE_USAGE_IMMUTABLE
;
60 templ
.width0
= size_in_dw
;
65 tex
= (struct r600_resource_texture
*)r600_texture_create(
66 &screen
->screen
, &templ
);
67 /* XXX: Propagate this error */
68 assert(tex
&& "Out of memory");
76 struct compute_memory_pool
* compute_memory_pool_new(
77 struct r600_screen
* rscreen
)
79 struct compute_memory_pool
* pool
= (struct compute_memory_pool
*)
80 CALLOC(sizeof(struct compute_memory_pool
), 1);
82 COMPUTE_DBG("* compute_memory_pool_new()\n");
84 pool
->screen
= rscreen
;
88 static void compute_memory_pool_init(struct compute_memory_pool
* pool
,
89 unsigned initial_size_in_dw
)
92 COMPUTE_DBG("* compute_memory_pool_init() initial_size_in_dw = %ld\n",
95 /* XXX: pool->shadow is used when the buffer needs to be resized, but
96 * resizing does not work at the moment.
97 * pool->shadow = (uint32_t*)CALLOC(4, pool->size_in_dw);
100 pool
->size_in_dw
= initial_size_in_dw
;
101 pool
->bo
= (struct r600_resource
*)create_pool_texture(pool
->screen
,
106 * Frees all stuff in the pool and the pool struct itself too
108 void compute_memory_pool_delete(struct compute_memory_pool
* pool
)
110 COMPUTE_DBG("* compute_memory_pool_delete()\n");
113 pool
->screen
->screen
.resource_destroy((struct pipe_screen
*)
114 pool
->screen
, (struct pipe_resource
*)pool
->bo
);
120 * Searches for an empty space in the pool, return with the pointer to the
121 * allocatable space in the pool, returns -1 on failure.
123 int64_t compute_memory_prealloc_chunk(
124 struct compute_memory_pool
* pool
,
127 assert(size_in_dw
<= pool
->size_in_dw
);
129 struct compute_memory_item
*item
;
133 COMPUTE_DBG("* compute_memory_prealloc_chunk() size_in_dw = %ld\n",
136 for (item
= pool
->item_list
; item
; item
= item
->next
) {
137 if (item
->start_in_dw
> -1) {
138 if (item
->start_in_dw
-last_end
> size_in_dw
) {
142 last_end
= item
->start_in_dw
+ item
->size_in_dw
;
143 last_end
+= (1024 - last_end
% 1024);
147 if (pool
->size_in_dw
- last_end
< size_in_dw
) {
155 * Search for the chunk where we can link our new chunk after it.
157 struct compute_memory_item
* compute_memory_postalloc_chunk(
158 struct compute_memory_pool
* pool
,
161 struct compute_memory_item
* item
;
163 COMPUTE_DBG("* compute_memory_postalloc_chunck() start_in_dw = %ld\n",
166 for (item
= pool
->item_list
; item
; item
= item
->next
) {
168 if (item
->start_in_dw
< start_in_dw
169 && item
->next
->start_in_dw
> start_in_dw
) {
175 assert(item
->start_in_dw
< start_in_dw
);
180 assert(0 && "unreachable");
185 * Reallocates pool, conserves data
187 void compute_memory_grow_pool(struct compute_memory_pool
* pool
,
188 struct pipe_context
* pipe
, int new_size_in_dw
)
190 COMPUTE_DBG("* compute_memory_grow_pool() new_size_in_dw = %d\n",
193 assert(new_size_in_dw
>= pool
->size_in_dw
);
195 assert(!pool
->bo
&& "Growing the global memory pool is not yet "
196 "supported. You will see this message if you are trying to"
197 "use more than 64 kb of memory");
200 compute_memory_pool_init(pool
, 1024 * 16);
202 /* XXX: Growing memory pools does not work at the moment. I think
203 * it is because we are using fragment shaders to copy data to
204 * the new texture and some of the compute registers are being
205 * included in the 3D command stream. */
206 fprintf(stderr
, "Warning: growing the global memory pool to"
207 "more than 64 kb is not yet supported\n");
208 new_size_in_dw
+= 1024 - (new_size_in_dw
% 1024);
210 COMPUTE_DBG(" Aligned size = %d\n", new_size_in_dw
);
212 compute_memory_shadow(pool
, pipe
, 1);
213 pool
->shadow
= (uint32_t*)realloc(pool
->shadow
, new_size_in_dw
*4);
214 pool
->size_in_dw
= new_size_in_dw
;
215 pool
->screen
->screen
.resource_destroy(
216 (struct pipe_screen
*)pool
->screen
,
217 (struct pipe_resource
*)pool
->bo
);
218 pool
->bo
= (struct r600_resource
*)create_pool_texture(
221 compute_memory_shadow(pool
, pipe
, 0);
226 * Copy pool from device to host, or host to device.
228 void compute_memory_shadow(struct compute_memory_pool
* pool
,
229 struct pipe_context
* pipe
, int device_to_host
)
231 struct compute_memory_item chunk
;
233 COMPUTE_DBG("* compute_memory_shadow() device_to_host = %d\n",
237 chunk
.start_in_dw
= 0;
238 chunk
.size_in_dw
= pool
->size_in_dw
;
239 chunk
.prev
= chunk
.next
= NULL
;
240 compute_memory_transfer(pool
, pipe
, device_to_host
, &chunk
,
241 pool
->shadow
, 0, pool
->size_in_dw
*4);
245 * Allocates pending allocations in the pool
247 void compute_memory_finalize_pending(struct compute_memory_pool
* pool
,
248 struct pipe_context
* pipe
)
250 struct compute_memory_item
*pending_list
= NULL
, *end_p
= NULL
;
251 struct compute_memory_item
*item
, *next
;
253 int64_t allocated
= 0;
254 int64_t unallocated
= 0;
256 COMPUTE_DBG("* compute_memory_finalize_pending()\n");
258 for (item
= pool
->item_list
; item
; item
= item
->next
) {
259 COMPUTE_DBG("list: %i %p\n", item
->start_in_dw
, item
->next
);
262 for (item
= pool
->item_list
; item
; item
= next
) {
266 if (item
->start_in_dw
== -1) {
275 item
->prev
->next
= next
;
278 pool
->item_list
= next
;
282 next
->prev
= item
->prev
;
289 unallocated
+= item
->size_in_dw
+1024;
292 allocated
+= item
->size_in_dw
;
296 if (pool
->size_in_dw
< allocated
+unallocated
) {
297 compute_memory_grow_pool(pool
, pipe
, allocated
+unallocated
);
300 for (item
= pending_list
; item
; item
= next
) {
305 while ((start_in_dw
=compute_memory_prealloc_chunk(pool
,
306 item
->size_in_dw
)) == -1) {
307 int64_t need
= item
->size_in_dw
+2048 -
308 (pool
->size_in_dw
- allocated
);
310 need
+= 1024 - (need
% 1024);
313 compute_memory_grow_pool(pool
,
315 pool
->size_in_dw
+ need
);
318 need
= pool
->size_in_dw
/ 10;
319 need
+= 1024 - (need
% 1024);
320 compute_memory_grow_pool(pool
,
322 pool
->size_in_dw
+ need
);
326 item
->start_in_dw
= start_in_dw
;
330 if (pool
->item_list
) {
331 struct compute_memory_item
*pos
;
333 pos
= compute_memory_postalloc_chunk(pool
, start_in_dw
);
335 item
->next
= pos
->next
;
339 item
->next
->prev
= item
;
343 pool
->item_list
= item
;
346 allocated
+= item
->size_in_dw
;
351 void compute_memory_free(struct compute_memory_pool
* pool
, int64_t id
)
353 struct compute_memory_item
*item
, *next
;
355 COMPUTE_DBG("* compute_memory_free() id + %ld \n", id
);
357 for (item
= pool
->item_list
; item
; item
= next
) {
360 if (item
->id
== id
) {
362 item
->prev
->next
= item
->next
;
365 pool
->item_list
= item
->next
;
369 item
->next
->prev
= item
->prev
;
378 fprintf(stderr
, "Internal error, invalid id %ld "
379 "for compute_memory_free\n", id
);
381 assert(0 && "error");
385 * Creates pending allocations
387 struct compute_memory_item
* compute_memory_alloc(
388 struct compute_memory_pool
* pool
,
391 struct compute_memory_item
*new_item
;
393 COMPUTE_DBG("* compute_memory_alloc() size_in_dw = %ld\n", size_in_dw
);
395 new_item
= (struct compute_memory_item
*)
396 CALLOC(sizeof(struct compute_memory_item
), 1);
397 new_item
->size_in_dw
= size_in_dw
;
398 new_item
->start_in_dw
= -1; /* mark pending */
399 new_item
->id
= pool
->next_id
++;
400 new_item
->pool
= pool
;
402 struct compute_memory_item
*last_item
;
404 if (pool
->item_list
) {
405 for (last_item
= pool
->item_list
; last_item
->next
;
406 last_item
= last_item
->next
);
408 last_item
->next
= new_item
;
409 new_item
->prev
= last_item
;
412 pool
->item_list
= new_item
;
419 * Transfer data host<->device, offset and size is in bytes
421 void compute_memory_transfer(
422 struct compute_memory_pool
* pool
,
423 struct pipe_context
* pipe
,
425 struct compute_memory_item
* chunk
,
430 int64_t aligned_size
= pool
->size_in_dw
;
431 struct pipe_resource
* gart
= (struct pipe_resource
*)pool
->bo
;
432 int64_t internal_offset
= chunk
->start_in_dw
*4 + offset_in_chunk
;
434 struct pipe_transfer
*xfer
;
439 COMPUTE_DBG("* compute_memory_transfer() device_to_host = %d, "
440 "offset_in_chunk = %d, size = %d\n", device_to_host
,
441 offset_in_chunk
, size
);
445 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_READ
,
446 &(struct pipe_box
) { .width
= aligned_size
,
447 .height
= 1, .depth
= 1 });
449 map
= pipe
->transfer_map(pipe
, xfer
);
451 memcpy(data
, map
+ internal_offset
, size
);
452 pipe
->transfer_unmap(pipe
, xfer
);
453 pipe
->transfer_destroy(pipe
, xfer
);
455 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_WRITE
,
456 &(struct pipe_box
) { .width
= aligned_size
,
457 .height
= 1, .depth
= 1 });
459 map
= pipe
->transfer_map(pipe
, xfer
);
461 memcpy(map
+ internal_offset
, data
, size
);
462 pipe
->transfer_unmap(pipe
, xfer
);
463 pipe
->transfer_destroy(pipe
, xfer
);
468 * Transfer data between chunk<->data, it is for VRAM<->GART transfers
470 void compute_memory_transfer_direct(
471 struct compute_memory_pool
* pool
,
473 struct compute_memory_item
* chunk
,
474 struct r600_resource
* data
,