1 /* 2 * Copyright 2013 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * on the rights to use, copy, modify, merge, publish, distribute, sub 8 * license, and/or sell copies of the Software, and to permit persons to whom 9 * the Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, 19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 21 * USE OR OTHER DEALINGS IN THE SOFTWARE. 22 * 23 * Authors: 24 * Marek Olk 25 */ 26 27 #include "r600_cs.h" 28 #include "util/u_memory.h" 29 #include "util/u_upload_mgr.h" 30 #include <inttypes.h> 31 #include <stdio.h> 32 33 bool r600_rings_is_buffer_referenced(struct r600_common_context *ctx, 34 struct pb_buffer *buf, 35 enum radeon_bo_usage usage) 36 { 37 if (ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, buf, usage)) { 38 return true; 39 } 40 if (radeon_emitted(ctx->dma.cs, 0) && 41 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs, buf, usage)) { 42 return true; 43 } 44 return false; 45 } 46 47 void *r600_buffer_map_sync_with_rings(struct r600_common_context *ctx, 48 struct r600_resource *resource, 49 unsigned usage) 50 { 51 enum radeon_bo_usage rusage = RADEON_USAGE_READWRITE; 52 bool busy = false; 53 54 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) { 55 return ctx->ws->buffer_map(resource->buf, NULL, usage); 56 } 57 58 if (!(usage & PIPE_TRANSFER_WRITE)) { 59 /* have to wait for the last write */ 60 rusage = RADEON_USAGE_WRITE; 61 } 62 63 if (radeon_emitted(ctx->gfx.cs, ctx->initial_gfx_cs_size) && 64 ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, 65 resource->buf, rusage)) { 66 if (usage & PIPE_TRANSFER_DONTBLOCK) { 67 ctx->gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL); 68 return NULL; 69 } else { 70 ctx->gfx.flush(ctx, 0, NULL); 71 busy = true; 72 } 73 } 74 if (radeon_emitted(ctx->dma.cs, 0) && 75 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs, 76 resource->buf, rusage)) { 77 if (usage & PIPE_TRANSFER_DONTBLOCK) { 78 ctx->dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL); 79 return NULL; 80 } else { 81 ctx->dma.flush(ctx, 0, NULL); 82 busy = true; 83 } 84 } 85 86 if (busy || !ctx->ws->buffer_wait(resource->buf, 0, rusage)) { 87 if (usage & PIPE_TRANSFER_DONTBLOCK) { 88 return NULL; 89 } else { 90 /* We will be wait for the GPU. Wait for any offloaded 91 * CS flush to complete to avoid busy-waiting in the winsys. */ 92 ctx->ws->cs_sync_flush(ctx->gfx.cs); 93 if (ctx->dma.cs) 94 ctx->ws->cs_sync_flush(ctx->dma.cs); 95 } 96 } 97 98 /* Setting the CS to NULL will prevent doing checks we have done already. */ 99 return ctx->ws->buffer_map(resource->buf, NULL, usage); 100 } 101 102 void r600_init_resource_fields(struct r600_common_screen *rscreen, 103 struct r600_resource *res, 104 uint64_t size, unsigned alignment) 105 { 106 struct r600_texture *rtex = (struct r600_texture*)res; 107 108 res->bo_size = size; 109 res->bo_alignment = alignment; 110 res->flags = 0; 111 112 switch (res->b.b.usage) { 113 case PIPE_USAGE_STREAM: 114 res->flags = RADEON_FLAG_GTT_WC; 115 /* fall through */ 116 case PIPE_USAGE_STAGING: 117 /* Transfers are likely to occur more often with these 118 * resources. */ 119 res->domains = RADEON_DOMAIN_GTT; 120 break; 121 case PIPE_USAGE_DYNAMIC: 122 /* Older kernels didn't always flush the HDP cache before 123 * CS execution 124 */ 125 if (rscreen->info.drm_major == 2 && 126 rscreen->info.drm_minor < 40) { 127 res->domains = RADEON_DOMAIN_GTT; 128 res->flags |= RADEON_FLAG_GTT_WC; 129 break; 130 } 131 res->flags |= RADEON_FLAG_CPU_ACCESS; 132 /* fall through */ 133 case PIPE_USAGE_DEFAULT: 134 case PIPE_USAGE_IMMUTABLE: 135 default: 136 /* Not listing GTT here improves performance in some 137 * apps. */ 138 res->domains = RADEON_DOMAIN_VRAM; 139 res->flags |= RADEON_FLAG_GTT_WC; 140 break; 141 } 142 143 if (res->b.b.target == PIPE_BUFFER && 144 res->b.b.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT | 145 PIPE_RESOURCE_FLAG_MAP_COHERENT)) { 146 /* Use GTT for all persistent mappings with older 147 * kernels, because they didn't always flush the HDP 148 * cache before CS execution. 149 * 150 * Write-combined CPU mappings are fine, the kernel 151 * ensures all CPU writes finish before the GPU 152 * executes a command stream. 153 */ 154 if (rscreen->info.drm_major == 2 && 155 rscreen->info.drm_minor < 40) 156 res->domains = RADEON_DOMAIN_GTT; 157 else if (res->domains & RADEON_DOMAIN_VRAM) 158 res->flags |= RADEON_FLAG_CPU_ACCESS; 159 } 160 161 /* Tiled textures are unmappable. Always put them in VRAM. */ 162 if (res->b.b.target != PIPE_BUFFER && 163 !rtex->surface.is_linear) { 164 res->domains = RADEON_DOMAIN_VRAM; 165 res->flags &= ~RADEON_FLAG_CPU_ACCESS; 166 res->flags |= RADEON_FLAG_NO_CPU_ACCESS | 167 RADEON_FLAG_GTT_WC; 168 } 169 170 /* If VRAM is just stolen system memory, allow both VRAM and 171 * GTT, whichever has free space. If a buffer is evicted from 172 * VRAM to GTT, it will stay there. 173 */ 174 if (!rscreen->info.has_dedicated_vram && 175 res->domains == RADEON_DOMAIN_VRAM) 176 res->domains = RADEON_DOMAIN_VRAM_GTT; 177 178 if (rscreen->debug_flags & DBG_NO_WC) 179 res->flags &= ~RADEON_FLAG_GTT_WC; 180 181 /* Set expected VRAM and GART usage for the buffer. */ 182 res->vram_usage = 0; 183 res->gart_usage = 0; 184 185 if (res->domains & RADEON_DOMAIN_VRAM) 186 res->vram_usage = size; 187 else if (res->domains & RADEON_DOMAIN_GTT) 188 res->gart_usage = size; 189 } 190 191 bool r600_alloc_resource(struct r600_common_screen *rscreen, 192 struct r600_resource *res) 193 { 194 struct pb_buffer *old_buf, *new_buf; 195 196 /* Allocate a new resource. */ 197 new_buf = rscreen->ws->buffer_create(rscreen->ws, res->bo_size, 198 res->bo_alignment, 199 res->domains, res->flags); 200 if (!new_buf) { 201 return false; 202 } 203 204 /* Replace the pointer such that if res->buf wasn't NULL, it won't be 205 * NULL. This should prevent crashes with multiple contexts using 206 * the same buffer where one of the contexts invalidates it while 207 * the others are using it. */ 208 old_buf = res->buf; 209 res->buf = new_buf; /* should be atomic */ 210 211 if (rscreen->info.has_virtual_memory) 212 res->gpu_address = rscreen->ws->buffer_get_virtual_address(res->buf); 213 else 214 res->gpu_address = 0; 215 216 pb_reference(&old_buf, NULL); 217 218 util_range_set_empty(&res->valid_buffer_range); 219 res->TC_L2_dirty = false; 220 221 /* Print debug information. */ 222 if (rscreen->debug_flags & DBG_VM && res->b.b.target == PIPE_BUFFER) { 223 fprintf(stderr, "VM start=0x%"PRIX64" end=0x%"PRIX64" | Buffer %"PRIu64" bytes\n", 224 res->gpu_address, res->gpu_address + res->buf->size, 225 res->buf->size); 226 } 227 return true; 228 } 229 230 static void r600_buffer_destroy(struct pipe_screen *screen, 231 struct pipe_resource *buf) 232 { 233 struct r600_resource *rbuffer = r600_resource(buf); 234 235 util_range_destroy(&rbuffer->valid_buffer_range); 236 pb_reference(&rbuffer->buf, NULL); 237 FREE(rbuffer); 238 } 239 240 static bool 241 r600_invalidate_buffer(struct r600_common_context *rctx, 242 struct r600_resource *rbuffer) 243 { 244 /* Shared buffers can't be reallocated. */ 245 if (rbuffer->is_shared) 246 return false; 247 248 /* In AMD_pinned_memory, the user pointer association only gets 249 * broken when the buffer is explicitly re-allocated. 250 */ 251 if (rctx->ws->buffer_is_user_ptr(rbuffer->buf)) 252 return false; 253 254 /* Check if mapping this buffer would cause waiting for the GPU. */ 255 if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) || 256 !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) { 257 rctx->invalidate_buffer(&rctx->b, &rbuffer->b.b); 258 } else { 259 util_range_set_empty(&rbuffer->valid_buffer_range); 260 } 261 262 return true; 263 } 264 265 void r600_invalidate_resource(struct pipe_context *ctx, 266 struct pipe_resource *resource) 267 { 268 struct r600_common_context *rctx = (struct r600_common_context*)ctx; 269 struct r600_resource *rbuffer = r600_resource(resource); 270 271 /* We currently only do anyting here for buffers */ 272 if (resource->target == PIPE_BUFFER) 273 (void)r600_invalidate_buffer(rctx, rbuffer); 274 } 275 276 static void *r600_buffer_get_transfer(struct pipe_context *ctx, 277 struct pipe_resource *resource, 278 unsigned level, 279 unsigned usage, 280 const struct pipe_box *box, 281 struct pipe_transfer **ptransfer, 282 void *data, struct r600_resource *staging, 283 unsigned offset) 284 { 285 struct r600_common_context *rctx = (struct r600_common_context*)ctx; 286 struct r600_transfer *transfer = slab_alloc(&rctx->pool_transfers); 287 288 transfer->transfer.resource = resource; 289 transfer->transfer.level = level; 290 transfer->transfer.usage = usage; 291 transfer->transfer.box = *box; 292 transfer->transfer.stride = 0; 293 transfer->transfer.layer_stride = 0; 294 transfer->offset = offset; 295 transfer->staging = staging; 296 *ptransfer = &transfer->transfer; 297 return data; 298 } 299 300 static bool r600_can_dma_copy_buffer(struct r600_common_context *rctx, 301 unsigned dstx, unsigned srcx, unsigned size) 302 { 303 bool dword_aligned = !(dstx % 4) && !(srcx % 4) && !(size % 4); 304 305 return rctx->screen->has_cp_dma || 306 (dword_aligned && (rctx->dma.cs || 307 rctx->screen->has_streamout)); 308 309 } 310 311 static void *r600_buffer_transfer_map(struct pipe_context *ctx, 312 struct pipe_resource *resource, 313 unsigned level, 314 unsigned usage, 315 const struct pipe_box *box, 316 struct pipe_transfer **ptransfer) 317 { 318 struct r600_common_context *rctx = (struct r600_common_context*)ctx; 319 struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen; 320 struct r600_resource *rbuffer = r600_resource(resource); 321 uint8_t *data; 322 323 assert(box->x + box->width <= resource->width0); 324 325 /* See if the buffer range being mapped has never been initialized, 326 * in which case it can be mapped unsynchronized. */ 327 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) && 328 usage & PIPE_TRANSFER_WRITE && 329 !rbuffer->is_shared && 330 !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) { 331 usage |= PIPE_TRANSFER_UNSYNCHRONIZED; 332 } 333 334 /* If discarding the entire range, discard the whole resource instead. */ 335 if (usage & PIPE_TRANSFER_DISCARD_RANGE && 336 box->x == 0 && box->width == resource->width0) { 337 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE; 338 } 339 340 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE && 341 !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) { 342 assert(usage & PIPE_TRANSFER_WRITE); 343 344 if (r600_invalidate_buffer(rctx, rbuffer)) { 345 /* At this point, the buffer is always idle. */ 346 usage |= PIPE_TRANSFER_UNSYNCHRONIZED; 347 } else { 348 /* Fall back to a temporary buffer. */ 349 usage |= PIPE_TRANSFER_DISCARD_RANGE; 350 } 351 } 352 353 if ((usage & PIPE_TRANSFER_DISCARD_RANGE) && 354 !(usage & (PIPE_TRANSFER_UNSYNCHRONIZED | 355 PIPE_TRANSFER_PERSISTENT)) && 356 !(rscreen->debug_flags & DBG_NO_DISCARD_RANGE) && 357 r600_can_dma_copy_buffer(rctx, box->x, 0, box->width)) { 358 assert(usage & PIPE_TRANSFER_WRITE); 359 360 /* Check if mapping this buffer would cause waiting for the GPU. */ 361 if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) || 362 !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) { 363 /* Do a wait-free write-only transfer using a temporary buffer. */ 364 unsigned offset; 365 struct r600_resource *staging = NULL; 366 367 u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT), 368 256, &offset, (struct pipe_resource**)&staging, (void**)&data); 369 370 if (staging) { 371 data += box->x % R600_MAP_BUFFER_ALIGNMENT; 372 return r600_buffer_get_transfer(ctx, resource, level, usage, box, 373 ptransfer, data, staging, offset); 374 } 375 } else { 376 /* At this point, the buffer is always idle (we checked it above). */ 377 usage |= PIPE_TRANSFER_UNSYNCHRONIZED; 378 } 379 } 380 /* Use a staging buffer in cached GTT for reads. */ 381 else if ((usage & PIPE_TRANSFER_READ) && 382 !(usage & PIPE_TRANSFER_PERSISTENT) && 383 (rbuffer->domains & RADEON_DOMAIN_VRAM || 384 rbuffer->flags & RADEON_FLAG_GTT_WC) && 385 r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) { 386 struct r600_resource *staging; 387 388 staging = (struct r600_resource*) pipe_buffer_create( 389 ctx->screen, 0, PIPE_USAGE_STAGING, 390 box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT)); 391 if (staging) { 392 /* Copy the VRAM buffer to the staging buffer. */ 393 rctx->dma_copy(ctx, &staging->b.b, 0, 394 box->x % R600_MAP_BUFFER_ALIGNMENT, 395 0, 0, resource, 0, box); 396 397 data = r600_buffer_map_sync_with_rings(rctx, staging, 398 usage & ~PIPE_TRANSFER_UNSYNCHRONIZED); 399 if (!data) { 400 r600_resource_reference(&staging, NULL); 401 return NULL; 402 } 403 data += box->x % R600_MAP_BUFFER_ALIGNMENT; 404 405 return r600_buffer_get_transfer(ctx, resource, level, usage, box, 406 ptransfer, data, staging, 0); 407 } 408 } 409 410 data = r600_buffer_map_sync_with_rings(rctx, rbuffer, usage); 411 if (!data) { 412 return NULL; 413 } 414 data += box->x; 415 416 return r600_buffer_get_transfer(ctx, resource, level, usage, box, 417 ptransfer, data, NULL, 0); 418 } 419 420 static void r600_buffer_do_flush_region(struct pipe_context *ctx, 421 struct pipe_transfer *transfer, 422 const struct pipe_box *box) 423 { 424 struct r600_transfer *rtransfer = (struct r600_transfer*)transfer; 425 struct r600_resource *rbuffer = r600_resource(transfer->resource); 426 427 if (rtransfer->staging) { 428 struct pipe_resource *dst, *src; 429 unsigned soffset; 430 struct pipe_box dma_box; 431 432 dst = transfer->resource; 433 src = &rtransfer->staging->b.b; 434 soffset = rtransfer->offset + box->x % R600_MAP_BUFFER_ALIGNMENT; 435 436 u_box_1d(soffset, box->width, &dma_box); 437 438 /* Copy the staging buffer into the original one. */ 439 ctx->resource_copy_region(ctx, dst, 0, box->x, 0, 0, src, 0, &dma_box); 440 } 441 442 util_range_add(&rbuffer->valid_buffer_range, box->x, 443 box->x + box->width); 444 } 445 446 static void r600_buffer_flush_region(struct pipe_context *ctx, 447 struct pipe_transfer *transfer, 448 const struct pipe_box *rel_box) 449 { 450 if (transfer->usage & (PIPE_TRANSFER_WRITE | 451 PIPE_TRANSFER_FLUSH_EXPLICIT)) { 452 struct pipe_box box; 453 454 u_box_1d(transfer->box.x + rel_box->x, rel_box->width, &box); 455 r600_buffer_do_flush_region(ctx, transfer, &box); 456 } 457 } 458 459 static void r600_buffer_transfer_unmap(struct pipe_context *ctx, 460 struct pipe_transfer *transfer) 461 { 462 struct r600_common_context *rctx = (struct r600_common_context*)ctx; 463 struct r600_transfer *rtransfer = (struct r600_transfer*)transfer; 464 465 if (transfer->usage & PIPE_TRANSFER_WRITE && 466 !(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) 467 r600_buffer_do_flush_region(ctx, transfer, &transfer->box); 468 469 if (rtransfer->staging) 470 r600_resource_reference(&rtransfer->staging, NULL); 471 472 slab_free(&rctx->pool_transfers, transfer); 473 } 474 475 void r600_buffer_subdata(struct pipe_context *ctx, 476 struct pipe_resource *buffer, 477 unsigned usage, unsigned offset, 478 unsigned size, const void *data) 479 { 480 struct pipe_transfer *transfer = NULL; 481 struct pipe_box box; 482 uint8_t *map = NULL; 483 484 u_box_1d(offset, size, &box); 485 map = r600_buffer_transfer_map(ctx, buffer, 0, 486 PIPE_TRANSFER_WRITE | 487 PIPE_TRANSFER_DISCARD_RANGE | 488 usage, 489 &box, &transfer); 490 if (!map) 491 return; 492 493 memcpy(map, data, size); 494 r600_buffer_transfer_unmap(ctx, transfer); 495 } 496 497 static const struct u_resource_vtbl r600_buffer_vtbl = 498 { 499 NULL, /* get_handle */ 500 r600_buffer_destroy, /* resource_destroy */ 501 r600_buffer_transfer_map, /* transfer_map */ 502 r600_buffer_flush_region, /* transfer_flush_region */ 503 r600_buffer_transfer_unmap, /* transfer_unmap */ 504 }; 505 506 static struct r600_resource * 507 r600_alloc_buffer_struct(struct pipe_screen *screen, 508 const struct pipe_resource *templ) 509 { 510 struct r600_resource *rbuffer; 511 512 rbuffer = MALLOC_STRUCT(r600_resource); 513 514 rbuffer->b.b = *templ; 515 rbuffer->b.b.next = NULL; 516 pipe_reference_init(&rbuffer->b.b.reference, 1); 517 rbuffer->b.b.screen = screen; 518 rbuffer->b.vtbl = &r600_buffer_vtbl; 519 rbuffer->buf = NULL; 520 rbuffer->bind_history = 0; 521 rbuffer->TC_L2_dirty = false; 522 rbuffer->is_shared = false; 523 util_range_init(&rbuffer->valid_buffer_range); 524 return rbuffer; 525 } 526 527 struct pipe_resource *r600_buffer_create(struct pipe_screen *screen, 528 const struct pipe_resource *templ, 529 unsigned alignment) 530 { 531 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen; 532 struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ); 533 534 r600_init_resource_fields(rscreen, rbuffer, templ->width0, alignment); 535 536 if (templ->bind & PIPE_BIND_SHARED) 537 rbuffer->flags |= RADEON_FLAG_HANDLE; 538 539 if (!r600_alloc_resource(rscreen, rbuffer)) { 540 FREE(rbuffer); 541 return NULL; 542 } 543 return &rbuffer->b.b; 544 } 545 546 struct pipe_resource *r600_aligned_buffer_create(struct pipe_screen *screen, 547 unsigned bind, 548 unsigned usage, 549 unsigned size, 550 unsigned alignment) 551 { 552 struct pipe_resource buffer; 553 554 memset(&buffer, 0, sizeof buffer); 555 buffer.target = PIPE_BUFFER; 556 buffer.format = PIPE_FORMAT_R8_UNORM; 557 buffer.bind = bind; 558 buffer.usage = usage; 559 buffer.flags = 0; 560 buffer.width0 = size; 561 buffer.height0 = 1; 562 buffer.depth0 = 1; 563 buffer.array_size = 1; 564 return r600_buffer_create(screen, &buffer, alignment); 565 } 566 567 struct pipe_resource * 568 r600_buffer_from_user_memory(struct pipe_screen *screen, 569 const struct pipe_resource *templ, 570 void *user_memory) 571 { 572 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen; 573 struct radeon_winsys *ws = rscreen->ws; 574 struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ); 575 576 rbuffer->domains = RADEON_DOMAIN_GTT; 577 util_range_add(&rbuffer->valid_buffer_range, 0, templ->width0); 578 579 /* Convert a user pointer to a buffer. */ 580 rbuffer->buf = ws->buffer_from_ptr(ws, user_memory, templ->width0); 581 if (!rbuffer->buf) { 582 FREE(rbuffer); 583 return NULL; 584 } 585 586 if (rscreen->info.has_virtual_memory) 587 rbuffer->gpu_address = 588 ws->buffer_get_virtual_address(rbuffer->buf); 589 else 590 rbuffer->gpu_address = 0; 591 592 return &rbuffer->b.b; 593 } 594