src/gallium/drivers/r600/compute_memory_pool.c

   1 /*
   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:
   8  *
   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
  11  * Software.
  12  *
  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.
  20  *
  21  * Authors:
  22  *      Adam Rak <adam.rak@streamnovation.com>
  23  */
  24
  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"
  36 #include "r600.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"
  43
  44 static struct r600_resource_texture * create_pool_texture(struct r600_screen * screen,
  45                 unsigned size_in_dw)
  46 {
  47
  48         struct pipe_resource templ;
  49         struct r600_resource_texture * tex;
  50
  51         if (size_in_dw == 0) {
  52                 return NULL;
  53         }
  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;
  59         templ.flags = 0;
  60         templ.width0 = size_in_dw;
  61         templ.height0 = 1;
  62         templ.depth0 = 1;
  63         templ.array_size = 1;
  64
  65         tex = (struct r600_resource_texture *)r600_texture_create(
  66                                                 &screen->screen, &templ);
  67         /* XXX: Propagate this error */
  68         assert(tex && "Out of memory");
  69         tex->is_rat = 1;
  70         return tex;
  71 }
  72
  73 /**
  74  * Creates a new pool
  75  */
  76 struct compute_memory_pool* compute_memory_pool_new(
  77         int64_t initial_size_in_dw,
  78         struct r600_screen * rscreen)
  79 {
  80         struct compute_memory_pool* pool = (struct compute_memory_pool*)
  81                                 CALLOC(sizeof(struct compute_memory_pool), 1);
  82
  83         pool->next_id = 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,
  87                                                         pool->size_in_dw);
  88         pool->shadow = (uint32_t*)CALLOC(4, pool->size_in_dw);
  89
  90         return pool;
  91 }
  92
  93 /**
  94  * Frees all stuff in the pool and the pool struct itself too
  95  */
  96 void compute_memory_pool_delete(struct compute_memory_pool* pool)
  97 {
  98         free(pool->shadow);
  99         if (pool->bo) {
 100                 pool->screen->screen.resource_destroy((struct pipe_screen *)
 101                         pool->screen, (struct pipe_resource *)pool->bo);
 102         }
 103         free(pool);
 104 }
 105
 106 /**
 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.
 109  */
 110 int64_t compute_memory_prealloc_chunk(
 111         struct compute_memory_pool* pool,
 112         int64_t size_in_dw)
 113 {
 114         assert(size_in_dw <= pool->size_in_dw);
 115
 116         struct compute_memory_item *item;
 117
 118         int last_end = 0;
 119
 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) {
 123                                 return last_end;
 124                         }
 125
 126                         last_end = item->start_in_dw + item->size_in_dw;
 127                         last_end += (1024 - last_end % 1024);
 128                 }
 129         }
 130
 131         if (pool->size_in_dw - last_end < size_in_dw) {
 132                 return -1;
 133         }
 134
 135         return last_end;
 136 }
 137
 138 /**
 139  *  Search for the chunk where we can link our new chunk after it.
 140  */
 141 struct compute_memory_item* compute_memory_postalloc_chunk(
 142         struct compute_memory_pool* pool,
 143         int64_t start_in_dw)
 144 {
 145         struct compute_memory_item* item;
 146
 147         for (item = pool->item_list; item; item = item->next) {
 148                 if (item->next) {
 149                         if (item->start_in_dw < start_in_dw
 150                                 && item->next->start_in_dw > start_in_dw) {
 151                                 return item;
 152                         }
 153                 }
 154                 else {
 155                         /* end of chain */
 156                         assert(item->start_in_dw < start_in_dw);
 157                         return item;
 158                 }
 159         }
 160
 161         assert(0 && "unreachable");
 162         return NULL;
 163 }
 164
 165 /**
 166  * Reallocates pool, conserves data
 167  */
 168 void compute_memory_grow_pool(struct compute_memory_pool* pool,
 169         struct pipe_context * pipe, int new_size_in_dw)
 170 {
 171         assert(new_size_in_dw >= pool->size_in_dw);
 172
 173         new_size_in_dw += 1024 - (new_size_in_dw % 1024);
 174
 175         if (pool->bo) {
 176                 compute_memory_shadow(pool, pipe, 1);
 177         }
 178         pool->shadow = (uint32_t*)realloc(pool->shadow, new_size_in_dw*4);
 179         pool->size_in_dw = new_size_in_dw;
 180         if (pool->bo) {
 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(
 185                                                         pool->screen,
 186                                                         pool->size_in_dw);
 187                 compute_memory_shadow(pool, pipe, 0);
 188         } else {
 189                 pool->bo = (struct r600_resource*)create_pool_texture(
 190                                                         pool->screen,
 191                                                         pool->size_in_dw);
 192         }
 193 }
 194
 195 /**
 196  * Copy pool from device to host, or host to device.
 197  */
 198 void compute_memory_shadow(struct compute_memory_pool* pool,
 199         struct pipe_context * pipe, int device_to_host)
 200 {
 201         struct compute_memory_item chunk;
 202
 203         chunk.id = 0;
 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);
 209 }
 210
 211 /**
 212  * Allocates pending allocations in the pool
 213  */
 214 void compute_memory_finalize_pending(struct compute_memory_pool* pool,
 215         struct pipe_context * pipe)
 216 {
 217         struct compute_memory_item *pending_list = NULL, *end_p = NULL;
 218         struct compute_memory_item *item, *next;
 219
 220         int64_t allocated = 0;
 221         int64_t unallocated = 0;
 222
 223         for (item = pool->item_list; item; item = item->next) {
 224                 COMPUTE_DBG("list: %i %p\n", item->start_in_dw, item->next);
 225         }
 226
 227         for (item = pool->item_list; item; item = next) {
 228                 next = item->next;
 229
 230
 231                 if (item->start_in_dw == -1) {
 232                         if (end_p) {
 233                                 end_p->next = item;
 234                         }
 235                         else {
 236                                 pending_list = item;
 237                         }
 238
 239                         if (item->prev) {
 240                                 item->prev->next = next;
 241                         }
 242                         else {
 243                                 pool->item_list = next;
 244                         }
 245
 246                         if (next) {
 247                                 next->prev = item->prev;
 248                         }
 249
 250                         item->prev = end_p;
 251                         item->next = NULL;
 252                         end_p = item;
 253
 254                         unallocated += item->size_in_dw+1024;
 255                 }
 256                 else {
 257                         allocated += item->size_in_dw;
 258                 }
 259         }
 260
 261         if (pool->size_in_dw < allocated+unallocated) {
 262                 compute_memory_grow_pool(pool, pipe, allocated+unallocated);
 263         }
 264
 265         for (item = pending_list; item; item = next) {
 266                 next = item->next;
 267
 268                 int64_t start_in_dw;
 269
 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);
 274
 275                         need += 1024 - (need % 1024);
 276
 277                         if (need > 0) {
 278                                 compute_memory_grow_pool(pool,
 279                                                 pipe,
 280                                                 pool->size_in_dw + need);
 281                         }
 282                         else {
 283                                 need = pool->size_in_dw / 10;
 284                                 need += 1024 - (need % 1024);
 285                                 compute_memory_grow_pool(pool,
 286                                                 pipe,
 287                                                 pool->size_in_dw + need);
 288                         }
 289                 }
 290
 291                 item->start_in_dw = start_in_dw;
 292                 item->next = NULL;
 293                 item->prev = NULL;
 294
 295                 if (pool->item_list) {
 296                         struct compute_memory_item *pos;
 297
 298                         pos = compute_memory_postalloc_chunk(pool, start_in_dw);
 299                         item->prev = pos;
 300                         item->next = pos->next;
 301                         pos->next = item;
 302
 303                         if (item->next) {
 304                                 item->next->prev = item;
 305                         }
 306                 }
 307                 else {
 308                         pool->item_list = item;
 309                 }
 310
 311                 allocated += item->size_in_dw;
 312         }
 313 }
 314
 315
 316 void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 317 {
 318         struct compute_memory_item *item, *next;
 319
 320         for (item = pool->item_list; item; item = next) {
 321                 next = item->next;
 322
 323                 if (item->id == id) {
 324                         if (item->prev) {
 325                                 item->prev->next = item->next;
 326                         }
 327                         else {
 328                                 pool->item_list = item->next;
 329                         }
 330
 331                         if (item->next) {
 332                                 item->next->prev = item->prev;
 333                         }
 334
 335                         free(item);
 336
 337                         return;
 338                 }
 339         }
 340
 341         fprintf(stderr, "Internal error, invalid id %ld "
 342                 "for compute_memory_free\n", id);
 343
 344         assert(0 && "error");
 345 }
 346
 347 /**
 348  * Creates pending allocations
 349  */
 350 struct compute_memory_item* compute_memory_alloc(
 351         struct compute_memory_pool* pool,
 352         int64_t size_in_dw)
 353 {
 354         struct compute_memory_item *new_item;
 355
 356         COMPUTE_DBG("Alloc: %i\n", size_in_dw);
 357
 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;
 364
 365         struct compute_memory_item *last_item;
 366
 367         if (pool->item_list) {
 368                 for (last_item = pool->item_list; last_item->next;
 369                                                 last_item = last_item->next);
 370
 371                 last_item->next = new_item;
 372                 new_item->prev = last_item;
 373         }
 374         else {
 375                 pool->item_list = new_item;
 376         }
 377
 378         return new_item;
 379 }
 380
 381 /**
 382  * Transfer data host<->device, offset and size is in bytes
 383  */
 384 void compute_memory_transfer(
 385         struct compute_memory_pool* pool,
 386         struct pipe_context * pipe,
 387         int device_to_host,
 388         struct compute_memory_item* chunk,
 389         void* data,
 390         int offset_in_chunk,
 391         int size)
 392 {
 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;
 396
 397         struct pipe_transfer *xfer;
 398         uint32_t *map;
 399
 400         assert(gart);
 401
 402         if (device_to_host)
 403         {
 404                 xfer = pipe->get_transfer(pipe, gart, 0, PIPE_TRANSFER_READ,
 405                         &(struct pipe_box) { .width = aligned_size,
 406                         .height = 1, .depth = 1 });
 407                 assert(xfer);
 408                 map = pipe->transfer_map(pipe, xfer);
 409                 assert(map);
 410                 memcpy(data, map + internal_offset, size);
 411                 pipe->transfer_unmap(pipe, xfer);
 412                 pipe->transfer_destroy(pipe, xfer);
 413         } else {
 414                 xfer = pipe->get_transfer(pipe, gart, 0, PIPE_TRANSFER_WRITE,
 415                         &(struct pipe_box) { .width = aligned_size,
 416                         .height = 1, .depth = 1 });
 417                 assert(xfer);
 418                 map = pipe->transfer_map(pipe, xfer);
 419                 assert(map);
 420                 memcpy(map + internal_offset, data, size);
 421                 pipe->transfer_unmap(pipe, xfer);
 422                 pipe->transfer_destroy(pipe, xfer);
 423         }
 424 }
 425
 426 /**
 427  * Transfer data between chunk<->data, it is for VRAM<->GART transfers
 428  */
 429 void compute_memory_transfer_direct(
 430         struct compute_memory_pool* pool,
 431         int chunk_to_data,
 432         struct compute_memory_item* chunk,
 433         struct r600_resource* data,
 434         int offset_in_chunk,
 435         int offset_in_data,
 436         int size)
 437 {
 438         ///TODO: DMA
 439 }