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 int64_t initial_size_in_dw
,
78 struct r600_screen
* rscreen
)
80 struct compute_memory_pool
* pool
= (struct compute_memory_pool
*)
81 CALLOC(sizeof(struct compute_memory_pool
), 1);
84 pool
->size_in_dw
= initial_size_in_dw
;
85 pool
->screen
= rscreen
;
86 pool
->bo
= (struct r600_resource
*)create_pool_texture(pool
->screen
,
88 pool
->shadow
= (uint32_t*)CALLOC(4, pool
->size_in_dw
);
94 * Frees all stuff in the pool and the pool struct itself too
96 void compute_memory_pool_delete(struct compute_memory_pool
* pool
)
100 pool
->screen
->screen
.resource_destroy((struct pipe_screen
*)
101 pool
->screen
, (struct pipe_resource
*)pool
->bo
);
107 * Searches for an empty space in the pool, return with the pointer to the
108 * allocatable space in the pool, returns -1 on failure.
110 int64_t compute_memory_prealloc_chunk(
111 struct compute_memory_pool
* pool
,
114 assert(size_in_dw
<= pool
->size_in_dw
);
116 struct compute_memory_item
*item
;
120 for (item
= pool
->item_list
; item
; item
= item
->next
) {
121 if (item
->start_in_dw
> -1) {
122 if (item
->start_in_dw
-last_end
> size_in_dw
) {
126 last_end
= item
->start_in_dw
+ item
->size_in_dw
;
127 last_end
+= (1024 - last_end
% 1024);
131 if (pool
->size_in_dw
- last_end
< size_in_dw
) {
139 * Search for the chunk where we can link our new chunk after it.
141 struct compute_memory_item
* compute_memory_postalloc_chunk(
142 struct compute_memory_pool
* pool
,
145 struct compute_memory_item
* item
;
147 for (item
= pool
->item_list
; item
; item
= item
->next
) {
149 if (item
->start_in_dw
< start_in_dw
150 && item
->next
->start_in_dw
> start_in_dw
) {
156 assert(item
->start_in_dw
< start_in_dw
);
161 assert(0 && "unreachable");
166 * Reallocates pool, conserves data
168 void compute_memory_grow_pool(struct compute_memory_pool
* pool
,
169 struct pipe_context
* pipe
, int new_size_in_dw
)
171 assert(new_size_in_dw
>= pool
->size_in_dw
);
173 new_size_in_dw
+= 1024 - (new_size_in_dw
% 1024);
176 compute_memory_shadow(pool
, pipe
, 1);
178 pool
->shadow
= (uint32_t*)realloc(pool
->shadow
, new_size_in_dw
*4);
179 pool
->size_in_dw
= new_size_in_dw
;
181 pool
->screen
->screen
.resource_destroy(
182 (struct pipe_screen
*)pool
->screen
,
183 (struct pipe_resource
*)pool
->bo
);
184 pool
->bo
= (struct r600_resource
*)create_pool_texture(
187 compute_memory_shadow(pool
, pipe
, 0);
189 pool
->bo
= (struct r600_resource
*)create_pool_texture(
196 * Copy pool from device to host, or host to device.
198 void compute_memory_shadow(struct compute_memory_pool
* pool
,
199 struct pipe_context
* pipe
, int device_to_host
)
201 struct compute_memory_item chunk
;
204 chunk
.start_in_dw
= 0;
205 chunk
.size_in_dw
= pool
->size_in_dw
;
206 chunk
.prev
= chunk
.next
= NULL
;
207 compute_memory_transfer(pool
, pipe
, device_to_host
, &chunk
,
208 pool
->shadow
, 0, pool
->size_in_dw
*4);
212 * Allocates pending allocations in the pool
214 void compute_memory_finalize_pending(struct compute_memory_pool
* pool
,
215 struct pipe_context
* pipe
)
217 struct compute_memory_item
*pending_list
= NULL
, *end_p
= NULL
;
218 struct compute_memory_item
*item
, *next
;
220 int64_t allocated
= 0;
221 int64_t unallocated
= 0;
223 for (item
= pool
->item_list
; item
; item
= item
->next
) {
224 COMPUTE_DBG("list: %i %p\n", item
->start_in_dw
, item
->next
);
227 for (item
= pool
->item_list
; item
; item
= next
) {
231 if (item
->start_in_dw
== -1) {
240 item
->prev
->next
= next
;
243 pool
->item_list
= next
;
247 next
->prev
= item
->prev
;
254 unallocated
+= item
->size_in_dw
+1024;
257 allocated
+= item
->size_in_dw
;
261 if (pool
->size_in_dw
< allocated
+unallocated
) {
262 compute_memory_grow_pool(pool
, pipe
, allocated
+unallocated
);
265 for (item
= pending_list
; item
; item
= next
) {
270 while ((start_in_dw
=compute_memory_prealloc_chunk(pool
,
271 item
->size_in_dw
)) == -1) {
272 int64_t need
= item
->size_in_dw
+2048 -
273 (pool
->size_in_dw
- allocated
);
275 need
+= 1024 - (need
% 1024);
278 compute_memory_grow_pool(pool
,
280 pool
->size_in_dw
+ need
);
283 need
= pool
->size_in_dw
/ 10;
284 need
+= 1024 - (need
% 1024);
285 compute_memory_grow_pool(pool
,
287 pool
->size_in_dw
+ need
);
291 item
->start_in_dw
= start_in_dw
;
295 if (pool
->item_list
) {
296 struct compute_memory_item
*pos
;
298 pos
= compute_memory_postalloc_chunk(pool
, start_in_dw
);
300 item
->next
= pos
->next
;
304 item
->next
->prev
= item
;
308 pool
->item_list
= item
;
311 allocated
+= item
->size_in_dw
;
316 void compute_memory_free(struct compute_memory_pool
* pool
, int64_t id
)
318 struct compute_memory_item
*item
, *next
;
320 for (item
= pool
->item_list
; item
; item
= next
) {
323 if (item
->id
== id
) {
325 item
->prev
->next
= item
->next
;
328 pool
->item_list
= item
->next
;
332 item
->next
->prev
= item
->prev
;
341 fprintf(stderr
, "Internal error, invalid id %ld "
342 "for compute_memory_free\n", id
);
344 assert(0 && "error");
348 * Creates pending allocations
350 struct compute_memory_item
* compute_memory_alloc(
351 struct compute_memory_pool
* pool
,
354 struct compute_memory_item
*new_item
;
356 COMPUTE_DBG("Alloc: %i\n", size_in_dw
);
358 new_item
= (struct compute_memory_item
*)
359 CALLOC(sizeof(struct compute_memory_item
), 1);
360 new_item
->size_in_dw
= size_in_dw
;
361 new_item
->start_in_dw
= -1; /* mark pending */
362 new_item
->id
= pool
->next_id
++;
363 new_item
->pool
= pool
;
365 struct compute_memory_item
*last_item
;
367 if (pool
->item_list
) {
368 for (last_item
= pool
->item_list
; last_item
->next
;
369 last_item
= last_item
->next
);
371 last_item
->next
= new_item
;
372 new_item
->prev
= last_item
;
375 pool
->item_list
= new_item
;
382 * Transfer data host<->device, offset and size is in bytes
384 void compute_memory_transfer(
385 struct compute_memory_pool
* pool
,
386 struct pipe_context
* pipe
,
388 struct compute_memory_item
* chunk
,
393 int64_t aligned_size
= pool
->size_in_dw
;
394 struct pipe_resource
* gart
= (struct pipe_resource
*)pool
->bo
;
395 int64_t internal_offset
= chunk
->start_in_dw
*4 + offset_in_chunk
;
397 struct pipe_transfer
*xfer
;
404 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_READ
,
405 &(struct pipe_box
) { .width
= aligned_size
,
406 .height
= 1, .depth
= 1 });
408 map
= pipe
->transfer_map(pipe
, xfer
);
410 memcpy(data
, map
+ internal_offset
, size
);
411 pipe
->transfer_unmap(pipe
, xfer
);
412 pipe
->transfer_destroy(pipe
, xfer
);
414 xfer
= pipe
->get_transfer(pipe
, gart
, 0, PIPE_TRANSFER_WRITE
,
415 &(struct pipe_box
) { .width
= aligned_size
,
416 .height
= 1, .depth
= 1 });
418 map
= pipe
->transfer_map(pipe
, xfer
);
420 memcpy(map
+ internal_offset
, data
, size
);
421 pipe
->transfer_unmap(pipe
, xfer
);
422 pipe
->transfer_destroy(pipe
, xfer
);
427 * Transfer data between chunk<->data, it is for VRAM<->GART transfers
429 void compute_memory_transfer_direct(
430 struct compute_memory_pool
* pool
,
432 struct compute_memory_item
* chunk
,
433 struct r600_resource
* data
,