u_vbuf.c revision 634066946e265aa5d2628ddca16c6556d2091a66
1/************************************************************************** 2 * 3 * Copyright 2011 Marek Olšák <maraeo@gmail.com> 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21 * IN NO EVENT SHALL AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR 22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 28#include "util/u_vbuf.h" 29 30#include "util/u_dump.h" 31#include "util/u_format.h" 32#include "util/u_inlines.h" 33#include "util/u_memory.h" 34#include "util/u_upload_mgr.h" 35#include "translate/translate.h" 36#include "translate/translate_cache.h" 37#include "cso_cache/cso_cache.h" 38#include "cso_cache/cso_hash.h" 39 40struct u_vbuf_elements { 41 unsigned count; 42 struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS]; 43 44 unsigned src_format_size[PIPE_MAX_ATTRIBS]; 45 46 /* If (velem[i].src_format != native_format[i]), the vertex buffer 47 * referenced by the vertex element cannot be used for rendering and 48 * its vertex data must be translated to native_format[i]. */ 49 enum pipe_format native_format[PIPE_MAX_ATTRIBS]; 50 unsigned native_format_size[PIPE_MAX_ATTRIBS]; 51 52 /* This might mean two things: 53 * - src_format != native_format, as discussed above. 54 * - src_offset % 4 != 0 (if the caps don't allow such an offset). */ 55 uint32_t incompatible_elem_mask; /* each bit describes a corresp. attrib */ 56 /* Which buffer has at least one vertex element referencing it 57 * incompatible. */ 58 uint32_t incompatible_vb_mask_any; 59 /* Which buffer has all vertex elements referencing it incompatible. */ 60 uint32_t incompatible_vb_mask_all; 61 /* Which buffer has at least one vertex element referencing it 62 * compatible. */ 63 uint32_t compatible_vb_mask_any; 64 /* Which buffer has all vertex elements referencing it compatible. */ 65 uint32_t compatible_vb_mask_all; 66 67 /* Which buffer has at least one vertex element referencing it 68 * non-instanced. */ 69 uint32_t noninstance_vb_mask_any; 70 71 void *driver_cso; 72}; 73 74enum { 75 VB_VERTEX = 0, 76 VB_INSTANCE = 1, 77 VB_CONST = 2, 78 VB_NUM = 3 79}; 80 81struct u_vbuf { 82 struct u_vbuf_caps caps; 83 84 struct pipe_context *pipe; 85 struct translate_cache *translate_cache; 86 struct cso_cache *cso_cache; 87 struct u_upload_mgr *uploader; 88 89 /* This is what was set in set_vertex_buffers. 90 * May contain user buffers. */ 91 struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS]; 92 unsigned nr_vertex_buffers; 93 94 /* Saved vertex buffers. */ 95 struct pipe_vertex_buffer vertex_buffer_saved[PIPE_MAX_ATTRIBS]; 96 unsigned nr_vertex_buffers_saved; 97 98 /* Vertex buffers for the driver. 99 * There are no user buffers. */ 100 struct pipe_vertex_buffer real_vertex_buffer[PIPE_MAX_ATTRIBS]; 101 int nr_real_vertex_buffers; 102 boolean vertex_buffers_dirty; 103 104 /* The index buffer. */ 105 struct pipe_index_buffer index_buffer; 106 107 /* Vertex elements. */ 108 struct u_vbuf_elements *ve, *ve_saved; 109 110 /* Vertex elements used for the translate fallback. */ 111 struct pipe_vertex_element fallback_velems[PIPE_MAX_ATTRIBS]; 112 /* If non-NULL, this is a vertex element state used for the translate 113 * fallback and therefore used for rendering too. */ 114 boolean using_translate; 115 /* The vertex buffer slot index where translated vertices have been 116 * stored in. */ 117 unsigned fallback_vbs[VB_NUM]; 118 119 /* Which buffer is a user buffer. */ 120 uint32_t user_vb_mask; /* each bit describes a corresp. buffer */ 121 /* Which buffer is incompatible (unaligned). */ 122 uint32_t incompatible_vb_mask; /* each bit describes a corresp. buffer */ 123 /* Which buffer has a non-zero stride. */ 124 uint32_t nonzero_stride_vb_mask; /* each bit describes a corresp. buffer */ 125}; 126 127static void * 128u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count, 129 const struct pipe_vertex_element *attribs); 130static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso); 131 132 133void u_vbuf_get_caps(struct pipe_screen *screen, struct u_vbuf_caps *caps) 134{ 135 caps->format_fixed32 = 136 screen->is_format_supported(screen, PIPE_FORMAT_R32_FIXED, PIPE_BUFFER, 137 0, PIPE_BIND_VERTEX_BUFFER); 138 139 caps->format_float16 = 140 screen->is_format_supported(screen, PIPE_FORMAT_R16_FLOAT, PIPE_BUFFER, 141 0, PIPE_BIND_VERTEX_BUFFER); 142 143 caps->format_float64 = 144 screen->is_format_supported(screen, PIPE_FORMAT_R64_FLOAT, PIPE_BUFFER, 145 0, PIPE_BIND_VERTEX_BUFFER); 146 147 caps->format_norm32 = 148 screen->is_format_supported(screen, PIPE_FORMAT_R32_UNORM, PIPE_BUFFER, 149 0, PIPE_BIND_VERTEX_BUFFER) && 150 screen->is_format_supported(screen, PIPE_FORMAT_R32_SNORM, PIPE_BUFFER, 151 0, PIPE_BIND_VERTEX_BUFFER); 152 153 caps->format_scaled32 = 154 screen->is_format_supported(screen, PIPE_FORMAT_R32_USCALED, PIPE_BUFFER, 155 0, PIPE_BIND_VERTEX_BUFFER) && 156 screen->is_format_supported(screen, PIPE_FORMAT_R32_SSCALED, PIPE_BUFFER, 157 0, PIPE_BIND_VERTEX_BUFFER); 158 159 caps->buffer_offset_unaligned = 160 !screen->get_param(screen, 161 PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY); 162 163 caps->buffer_stride_unaligned = 164 !screen->get_param(screen, 165 PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY); 166 167 caps->velem_src_offset_unaligned = 168 !screen->get_param(screen, 169 PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY); 170 171 caps->user_vertex_buffers = 172 screen->get_param(screen, PIPE_CAP_USER_VERTEX_BUFFERS); 173} 174 175struct u_vbuf * 176u_vbuf_create(struct pipe_context *pipe, 177 struct u_vbuf_caps *caps) 178{ 179 struct u_vbuf *mgr = CALLOC_STRUCT(u_vbuf); 180 181 mgr->caps = *caps; 182 mgr->pipe = pipe; 183 mgr->cso_cache = cso_cache_create(); 184 mgr->translate_cache = translate_cache_create(); 185 memset(mgr->fallback_vbs, ~0, sizeof(mgr->fallback_vbs)); 186 187 mgr->uploader = u_upload_create(pipe, 1024 * 1024, 4, 188 PIPE_BIND_VERTEX_BUFFER); 189 190 return mgr; 191} 192 193/* u_vbuf uses its own caching for vertex elements, because it needs to keep 194 * its own preprocessed state per vertex element CSO. */ 195static struct u_vbuf_elements * 196u_vbuf_set_vertex_elements_internal(struct u_vbuf *mgr, unsigned count, 197 const struct pipe_vertex_element *states) 198{ 199 struct pipe_context *pipe = mgr->pipe; 200 unsigned key_size, hash_key; 201 struct cso_hash_iter iter; 202 struct u_vbuf_elements *ve; 203 struct cso_velems_state velems_state; 204 205 /* need to include the count into the stored state data too. */ 206 key_size = sizeof(struct pipe_vertex_element) * count + sizeof(unsigned); 207 velems_state.count = count; 208 memcpy(velems_state.velems, states, 209 sizeof(struct pipe_vertex_element) * count); 210 hash_key = cso_construct_key((void*)&velems_state, key_size); 211 iter = cso_find_state_template(mgr->cso_cache, hash_key, CSO_VELEMENTS, 212 (void*)&velems_state, key_size); 213 214 if (cso_hash_iter_is_null(iter)) { 215 struct cso_velements *cso = MALLOC_STRUCT(cso_velements); 216 memcpy(&cso->state, &velems_state, key_size); 217 cso->data = u_vbuf_create_vertex_elements(mgr, count, states); 218 cso->delete_state = (cso_state_callback)u_vbuf_delete_vertex_elements; 219 cso->context = (void*)mgr; 220 221 iter = cso_insert_state(mgr->cso_cache, hash_key, CSO_VELEMENTS, cso); 222 ve = cso->data; 223 } else { 224 ve = ((struct cso_velements *)cso_hash_iter_data(iter))->data; 225 } 226 227 assert(ve); 228 pipe->bind_vertex_elements_state(pipe, ve->driver_cso); 229 return ve; 230} 231 232void u_vbuf_set_vertex_elements(struct u_vbuf *mgr, unsigned count, 233 const struct pipe_vertex_element *states) 234{ 235 mgr->ve = u_vbuf_set_vertex_elements_internal(mgr, count, states); 236} 237 238void u_vbuf_destroy(struct u_vbuf *mgr) 239{ 240 unsigned i; 241 242 for (i = 0; i < mgr->nr_vertex_buffers; i++) { 243 pipe_resource_reference(&mgr->vertex_buffer[i].buffer, NULL); 244 } 245 for (i = 0; i < mgr->nr_real_vertex_buffers; i++) { 246 pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL); 247 } 248 249 translate_cache_destroy(mgr->translate_cache); 250 u_upload_destroy(mgr->uploader); 251 cso_cache_delete(mgr->cso_cache); 252 FREE(mgr); 253} 254 255static void 256u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key, 257 unsigned vb_mask, unsigned out_vb, 258 int start_vertex, unsigned num_vertices, 259 int start_index, unsigned num_indices, int min_index, 260 boolean unroll_indices) 261{ 262 struct translate *tr; 263 struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0}; 264 struct pipe_resource *out_buffer = NULL; 265 uint8_t *out_map; 266 unsigned i, out_offset; 267 268 /* Get a translate object. */ 269 tr = translate_cache_find(mgr->translate_cache, key); 270 271 /* Map buffers we want to translate. */ 272 for (i = 0; i < mgr->nr_vertex_buffers; i++) { 273 if (vb_mask & (1 << i)) { 274 struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[i]; 275 unsigned offset = vb->buffer_offset + vb->stride * start_vertex; 276 uint8_t *map; 277 278 if (vb->buffer->user_ptr) { 279 map = vb->buffer->user_ptr + offset; 280 } else { 281 unsigned size = vb->stride ? num_vertices * vb->stride 282 : sizeof(double)*4; 283 284 if (offset+size > vb->buffer->width0) { 285 size = vb->buffer->width0 - offset; 286 } 287 288 map = pipe_buffer_map_range(mgr->pipe, vb->buffer, offset, size, 289 PIPE_TRANSFER_READ, &vb_transfer[i]); 290 } 291 292 /* Subtract min_index so that indexing with the index buffer works. */ 293 if (unroll_indices) { 294 map -= vb->stride * min_index; 295 } 296 297 tr->set_buffer(tr, i, map, vb->stride, ~0); 298 } 299 } 300 301 /* Translate. */ 302 if (unroll_indices) { 303 struct pipe_index_buffer *ib = &mgr->index_buffer; 304 struct pipe_transfer *transfer = NULL; 305 unsigned offset = ib->offset + start_index * ib->index_size; 306 uint8_t *map; 307 308 assert(ib->buffer && ib->index_size); 309 310 if (ib->buffer->user_ptr) { 311 map = ib->buffer->user_ptr + offset; 312 } else { 313 map = pipe_buffer_map_range(mgr->pipe, ib->buffer, offset, 314 num_indices * ib->index_size, 315 PIPE_TRANSFER_READ, &transfer); 316 } 317 318 /* Create and map the output buffer. */ 319 u_upload_alloc(mgr->uploader, 0, 320 key->output_stride * num_indices, 321 &out_offset, &out_buffer, 322 (void**)&out_map); 323 324 switch (ib->index_size) { 325 case 4: 326 tr->run_elts(tr, (unsigned*)map, num_indices, 0, out_map); 327 break; 328 case 2: 329 tr->run_elts16(tr, (uint16_t*)map, num_indices, 0, out_map); 330 break; 331 case 1: 332 tr->run_elts8(tr, map, num_indices, 0, out_map); 333 break; 334 } 335 336 if (transfer) { 337 pipe_buffer_unmap(mgr->pipe, transfer); 338 } 339 } else { 340 /* Create and map the output buffer. */ 341 u_upload_alloc(mgr->uploader, 342 key->output_stride * start_vertex, 343 key->output_stride * num_vertices, 344 &out_offset, &out_buffer, 345 (void**)&out_map); 346 347 out_offset -= key->output_stride * start_vertex; 348 349 tr->run(tr, 0, num_vertices, 0, out_map); 350 } 351 352 /* Unmap all buffers. */ 353 for (i = 0; i < mgr->nr_vertex_buffers; i++) { 354 if (vb_transfer[i]) { 355 pipe_buffer_unmap(mgr->pipe, vb_transfer[i]); 356 } 357 } 358 359 /* Setup the new vertex buffer. */ 360 mgr->real_vertex_buffer[out_vb].buffer_offset = out_offset; 361 mgr->real_vertex_buffer[out_vb].stride = key->output_stride; 362 363 /* Move the buffer reference. */ 364 pipe_resource_reference( 365 &mgr->real_vertex_buffer[out_vb].buffer, NULL); 366 mgr->real_vertex_buffer[out_vb].buffer = out_buffer; 367} 368 369static boolean 370u_vbuf_translate_find_free_vb_slots(struct u_vbuf *mgr, 371 unsigned mask[VB_NUM]) 372{ 373 unsigned type; 374 unsigned fallback_vbs[VB_NUM]; 375 /* Set the bit for each buffer which is incompatible, or isn't set. */ 376 uint32_t unused_vb_mask = 377 mgr->ve->incompatible_vb_mask_all | mgr->incompatible_vb_mask | 378 ~((1 << mgr->nr_vertex_buffers) - 1); 379 380 memset(fallback_vbs, ~0, sizeof(fallback_vbs)); 381 382 /* Find free slots for each type if needed. */ 383 for (type = 0; type < VB_NUM; type++) { 384 if (mask[type]) { 385 uint32_t index; 386 387 if (!unused_vb_mask) { 388 /* fail, reset the number to its original value */ 389 mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers; 390 return FALSE; 391 } 392 393 index = ffs(unused_vb_mask) - 1; 394 fallback_vbs[type] = index; 395 if (index >= mgr->nr_real_vertex_buffers) { 396 mgr->nr_real_vertex_buffers = index + 1; 397 } 398 /*printf("found slot=%i for type=%i\n", index, type);*/ 399 } 400 } 401 402 memcpy(mgr->fallback_vbs, fallback_vbs, sizeof(fallback_vbs)); 403 return TRUE; 404} 405 406static boolean 407u_vbuf_translate_begin(struct u_vbuf *mgr, 408 int start_vertex, unsigned num_vertices, 409 int start_instance, unsigned num_instances, 410 int start_index, unsigned num_indices, int min_index, 411 boolean unroll_indices) 412{ 413 unsigned mask[VB_NUM] = {0}; 414 struct translate_key key[VB_NUM]; 415 unsigned elem_index[VB_NUM][PIPE_MAX_ATTRIBS]; /* ... into key.elements */ 416 unsigned i, type; 417 418 int start[VB_NUM] = { 419 start_vertex, /* VERTEX */ 420 start_instance, /* INSTANCE */ 421 0 /* CONST */ 422 }; 423 424 unsigned num[VB_NUM] = { 425 num_vertices, /* VERTEX */ 426 num_instances, /* INSTANCE */ 427 1 /* CONST */ 428 }; 429 430 memset(key, 0, sizeof(key)); 431 memset(elem_index, ~0, sizeof(elem_index)); 432 433 /* See if there are vertex attribs of each type to translate and 434 * which ones. */ 435 for (i = 0; i < mgr->ve->count; i++) { 436 unsigned vb_index = mgr->ve->ve[i].vertex_buffer_index; 437 438 if (!mgr->vertex_buffer[vb_index].stride) { 439 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 440 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 441 continue; 442 } 443 mask[VB_CONST] |= 1 << vb_index; 444 } else if (mgr->ve->ve[i].instance_divisor) { 445 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 446 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 447 continue; 448 } 449 mask[VB_INSTANCE] |= 1 << vb_index; 450 } else { 451 if (!unroll_indices && 452 !(mgr->ve->incompatible_elem_mask & (1 << i)) && 453 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 454 continue; 455 } 456 mask[VB_VERTEX] |= 1 << vb_index; 457 } 458 } 459 460 assert(mask[VB_VERTEX] || mask[VB_INSTANCE] || mask[VB_CONST]); 461 462 /* Find free vertex buffer slots. */ 463 if (!u_vbuf_translate_find_free_vb_slots(mgr, mask)) { 464 return FALSE; 465 } 466 467 /* Initialize the translate keys. */ 468 for (i = 0; i < mgr->ve->count; i++) { 469 struct translate_key *k; 470 struct translate_element *te; 471 unsigned bit, vb_index = mgr->ve->ve[i].vertex_buffer_index; 472 bit = 1 << vb_index; 473 474 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 475 !(mgr->incompatible_vb_mask & (1 << vb_index)) && 476 (!unroll_indices || !(mask[VB_VERTEX] & bit))) { 477 continue; 478 } 479 480 /* Set type to what we will translate. 481 * Whether vertex, instance, or constant attribs. */ 482 for (type = 0; type < VB_NUM; type++) { 483 if (mask[type] & bit) { 484 break; 485 } 486 } 487 assert(type < VB_NUM); 488 assert(translate_is_output_format_supported(mgr->ve->native_format[i])); 489 /*printf("velem=%i type=%i\n", i, type);*/ 490 491 /* Add the vertex element. */ 492 k = &key[type]; 493 elem_index[type][i] = k->nr_elements; 494 495 te = &k->element[k->nr_elements]; 496 te->type = TRANSLATE_ELEMENT_NORMAL; 497 te->instance_divisor = 0; 498 te->input_buffer = vb_index; 499 te->input_format = mgr->ve->ve[i].src_format; 500 te->input_offset = mgr->ve->ve[i].src_offset; 501 te->output_format = mgr->ve->native_format[i]; 502 te->output_offset = k->output_stride; 503 504 k->output_stride += mgr->ve->native_format_size[i]; 505 k->nr_elements++; 506 } 507 508 /* Translate buffers. */ 509 for (type = 0; type < VB_NUM; type++) { 510 if (key[type].nr_elements) { 511 u_vbuf_translate_buffers(mgr, &key[type], mask[type], 512 mgr->fallback_vbs[type], 513 start[type], num[type], 514 start_index, num_indices, min_index, 515 unroll_indices && type == VB_VERTEX); 516 517 /* Fixup the stride for constant attribs. */ 518 if (type == VB_CONST) { 519 mgr->real_vertex_buffer[mgr->fallback_vbs[VB_CONST]].stride = 0; 520 } 521 } 522 } 523 524 /* Setup new vertex elements. */ 525 for (i = 0; i < mgr->ve->count; i++) { 526 for (type = 0; type < VB_NUM; type++) { 527 if (elem_index[type][i] < key[type].nr_elements) { 528 struct translate_element *te = &key[type].element[elem_index[type][i]]; 529 mgr->fallback_velems[i].instance_divisor = mgr->ve->ve[i].instance_divisor; 530 mgr->fallback_velems[i].src_format = te->output_format; 531 mgr->fallback_velems[i].src_offset = te->output_offset; 532 mgr->fallback_velems[i].vertex_buffer_index = mgr->fallback_vbs[type]; 533 534 /* elem_index[type][i] can only be set for one type. */ 535 assert(type > VB_INSTANCE || elem_index[type+1][i] == ~0); 536 assert(type > VB_VERTEX || elem_index[type+2][i] == ~0); 537 break; 538 } 539 } 540 /* No translating, just copy the original vertex element over. */ 541 if (type == VB_NUM) { 542 memcpy(&mgr->fallback_velems[i], &mgr->ve->ve[i], 543 sizeof(struct pipe_vertex_element)); 544 } 545 } 546 547 u_vbuf_set_vertex_elements_internal(mgr, mgr->ve->count, 548 mgr->fallback_velems); 549 mgr->using_translate = TRUE; 550 return TRUE; 551} 552 553static void u_vbuf_translate_end(struct u_vbuf *mgr) 554{ 555 unsigned i; 556 557 /* Restore vertex elements. */ 558 mgr->pipe->bind_vertex_elements_state(mgr->pipe, mgr->ve->driver_cso); 559 mgr->using_translate = FALSE; 560 561 /* Unreference the now-unused VBOs. */ 562 for (i = 0; i < VB_NUM; i++) { 563 unsigned vb = mgr->fallback_vbs[i]; 564 if (vb != ~0) { 565 pipe_resource_reference(&mgr->real_vertex_buffer[vb].buffer, NULL); 566 mgr->fallback_vbs[i] = ~0; 567 } 568 } 569 mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers; 570} 571 572#define FORMAT_REPLACE(what, withwhat) \ 573 case PIPE_FORMAT_##what: format = PIPE_FORMAT_##withwhat; break 574 575static void * 576u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count, 577 const struct pipe_vertex_element *attribs) 578{ 579 struct pipe_context *pipe = mgr->pipe; 580 unsigned i; 581 struct pipe_vertex_element driver_attribs[PIPE_MAX_ATTRIBS]; 582 struct u_vbuf_elements *ve = CALLOC_STRUCT(u_vbuf_elements); 583 uint32_t used_buffers = 0; 584 585 ve->count = count; 586 587 memcpy(ve->ve, attribs, sizeof(struct pipe_vertex_element) * count); 588 memcpy(driver_attribs, attribs, sizeof(struct pipe_vertex_element) * count); 589 590 /* Set the best native format in case the original format is not 591 * supported. */ 592 for (i = 0; i < count; i++) { 593 enum pipe_format format = ve->ve[i].src_format; 594 595 ve->src_format_size[i] = util_format_get_blocksize(format); 596 597 used_buffers |= 1 << ve->ve[i].vertex_buffer_index; 598 599 if (!ve->ve[i].instance_divisor) { 600 ve->noninstance_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 601 } 602 603 /* Choose a native format. 604 * For now we don't care about the alignment, that's going to 605 * be sorted out later. */ 606 if (!mgr->caps.format_fixed32) { 607 switch (format) { 608 FORMAT_REPLACE(R32_FIXED, R32_FLOAT); 609 FORMAT_REPLACE(R32G32_FIXED, R32G32_FLOAT); 610 FORMAT_REPLACE(R32G32B32_FIXED, R32G32B32_FLOAT); 611 FORMAT_REPLACE(R32G32B32A32_FIXED, R32G32B32A32_FLOAT); 612 default:; 613 } 614 } 615 if (!mgr->caps.format_float16) { 616 switch (format) { 617 FORMAT_REPLACE(R16_FLOAT, R32_FLOAT); 618 FORMAT_REPLACE(R16G16_FLOAT, R32G32_FLOAT); 619 FORMAT_REPLACE(R16G16B16_FLOAT, R32G32B32_FLOAT); 620 FORMAT_REPLACE(R16G16B16A16_FLOAT, R32G32B32A32_FLOAT); 621 default:; 622 } 623 } 624 if (!mgr->caps.format_float64) { 625 switch (format) { 626 FORMAT_REPLACE(R64_FLOAT, R32_FLOAT); 627 FORMAT_REPLACE(R64G64_FLOAT, R32G32_FLOAT); 628 FORMAT_REPLACE(R64G64B64_FLOAT, R32G32B32_FLOAT); 629 FORMAT_REPLACE(R64G64B64A64_FLOAT, R32G32B32A32_FLOAT); 630 default:; 631 } 632 } 633 if (!mgr->caps.format_norm32) { 634 switch (format) { 635 FORMAT_REPLACE(R32_UNORM, R32_FLOAT); 636 FORMAT_REPLACE(R32G32_UNORM, R32G32_FLOAT); 637 FORMAT_REPLACE(R32G32B32_UNORM, R32G32B32_FLOAT); 638 FORMAT_REPLACE(R32G32B32A32_UNORM, R32G32B32A32_FLOAT); 639 FORMAT_REPLACE(R32_SNORM, R32_FLOAT); 640 FORMAT_REPLACE(R32G32_SNORM, R32G32_FLOAT); 641 FORMAT_REPLACE(R32G32B32_SNORM, R32G32B32_FLOAT); 642 FORMAT_REPLACE(R32G32B32A32_SNORM, R32G32B32A32_FLOAT); 643 default:; 644 } 645 } 646 if (!mgr->caps.format_scaled32) { 647 switch (format) { 648 FORMAT_REPLACE(R32_USCALED, R32_FLOAT); 649 FORMAT_REPLACE(R32G32_USCALED, R32G32_FLOAT); 650 FORMAT_REPLACE(R32G32B32_USCALED, R32G32B32_FLOAT); 651 FORMAT_REPLACE(R32G32B32A32_USCALED,R32G32B32A32_FLOAT); 652 FORMAT_REPLACE(R32_SSCALED, R32_FLOAT); 653 FORMAT_REPLACE(R32G32_SSCALED, R32G32_FLOAT); 654 FORMAT_REPLACE(R32G32B32_SSCALED, R32G32B32_FLOAT); 655 FORMAT_REPLACE(R32G32B32A32_SSCALED,R32G32B32A32_FLOAT); 656 default:; 657 } 658 } 659 660 driver_attribs[i].src_format = format; 661 ve->native_format[i] = format; 662 ve->native_format_size[i] = 663 util_format_get_blocksize(ve->native_format[i]); 664 665 if (ve->ve[i].src_format != format || 666 (!mgr->caps.velem_src_offset_unaligned && 667 ve->ve[i].src_offset % 4 != 0)) { 668 ve->incompatible_elem_mask |= 1 << i; 669 ve->incompatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 670 } else { 671 ve->compatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 672 } 673 } 674 675 ve->compatible_vb_mask_all = ~ve->incompatible_vb_mask_any & used_buffers; 676 ve->incompatible_vb_mask_all = ~ve->compatible_vb_mask_any & used_buffers; 677 678 /* Align the formats to the size of DWORD if needed. */ 679 if (!mgr->caps.velem_src_offset_unaligned) { 680 for (i = 0; i < count; i++) { 681 ve->native_format_size[i] = align(ve->native_format_size[i], 4); 682 } 683 } 684 685 ve->driver_cso = 686 pipe->create_vertex_elements_state(pipe, count, driver_attribs); 687 return ve; 688} 689 690static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso) 691{ 692 struct pipe_context *pipe = mgr->pipe; 693 struct u_vbuf_elements *ve = cso; 694 695 pipe->delete_vertex_elements_state(pipe, ve->driver_cso); 696 FREE(ve); 697} 698 699void u_vbuf_set_vertex_buffers(struct u_vbuf *mgr, unsigned count, 700 const struct pipe_vertex_buffer *bufs) 701{ 702 unsigned i; 703 704 mgr->user_vb_mask = 0; 705 mgr->incompatible_vb_mask = 0; 706 mgr->nonzero_stride_vb_mask = 0; 707 708 for (i = 0; i < count; i++) { 709 const struct pipe_vertex_buffer *vb = &bufs[i]; 710 struct pipe_vertex_buffer *orig_vb = &mgr->vertex_buffer[i]; 711 struct pipe_vertex_buffer *real_vb = &mgr->real_vertex_buffer[i]; 712 713 pipe_resource_reference(&orig_vb->buffer, vb->buffer); 714 715 real_vb->buffer_offset = orig_vb->buffer_offset = vb->buffer_offset; 716 real_vb->stride = orig_vb->stride = vb->stride; 717 718 if (vb->stride) { 719 mgr->nonzero_stride_vb_mask |= 1 << i; 720 } 721 722 if (!vb->buffer) { 723 pipe_resource_reference(&real_vb->buffer, NULL); 724 continue; 725 } 726 727 if ((!mgr->caps.buffer_offset_unaligned && vb->buffer_offset % 4 != 0) || 728 (!mgr->caps.buffer_stride_unaligned && vb->stride % 4 != 0)) { 729 mgr->incompatible_vb_mask |= 1 << i; 730 pipe_resource_reference(&real_vb->buffer, NULL); 731 continue; 732 } 733 734 if (!mgr->caps.user_vertex_buffers && vb->buffer->user_ptr) { 735 mgr->user_vb_mask |= 1 << i; 736 pipe_resource_reference(&real_vb->buffer, NULL); 737 continue; 738 } 739 740 pipe_resource_reference(&real_vb->buffer, vb->buffer); 741 } 742 743 for (i = count; i < mgr->nr_vertex_buffers; i++) { 744 pipe_resource_reference(&mgr->vertex_buffer[i].buffer, NULL); 745 } 746 for (i = count; i < mgr->nr_real_vertex_buffers; i++) { 747 pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL); 748 } 749 750 mgr->nr_vertex_buffers = count; 751 mgr->nr_real_vertex_buffers = count; 752 mgr->vertex_buffers_dirty = TRUE; 753} 754 755void u_vbuf_set_index_buffer(struct u_vbuf *mgr, 756 const struct pipe_index_buffer *ib) 757{ 758 struct pipe_context *pipe = mgr->pipe; 759 760 if (ib && ib->buffer) { 761 assert(ib->offset % ib->index_size == 0); 762 pipe_resource_reference(&mgr->index_buffer.buffer, ib->buffer); 763 mgr->index_buffer.offset = ib->offset; 764 mgr->index_buffer.index_size = ib->index_size; 765 } else { 766 pipe_resource_reference(&mgr->index_buffer.buffer, NULL); 767 } 768 769 pipe->set_index_buffer(pipe, ib); 770} 771 772static void 773u_vbuf_upload_buffers(struct u_vbuf *mgr, 774 int start_vertex, unsigned num_vertices, 775 int start_instance, unsigned num_instances) 776{ 777 unsigned i; 778 unsigned nr_velems = mgr->ve->count; 779 unsigned nr_vbufs = mgr->nr_vertex_buffers; 780 struct pipe_vertex_element *velems = 781 mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve; 782 unsigned start_offset[PIPE_MAX_ATTRIBS]; 783 unsigned end_offset[PIPE_MAX_ATTRIBS] = {0}; 784 785 /* Determine how much data needs to be uploaded. */ 786 for (i = 0; i < nr_velems; i++) { 787 struct pipe_vertex_element *velem = &velems[i]; 788 unsigned index = velem->vertex_buffer_index; 789 struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index]; 790 unsigned instance_div, first, size; 791 792 /* Skip the buffers generated by translate. */ 793 if (index == mgr->fallback_vbs[VB_VERTEX] || 794 index == mgr->fallback_vbs[VB_INSTANCE] || 795 index == mgr->fallback_vbs[VB_CONST]) { 796 continue; 797 } 798 799 assert(vb->buffer); 800 801 if (!vb->buffer->user_ptr) { 802 continue; 803 } 804 805 instance_div = velem->instance_divisor; 806 first = vb->buffer_offset + velem->src_offset; 807 808 if (!vb->stride) { 809 /* Constant attrib. */ 810 size = mgr->ve->src_format_size[i]; 811 } else if (instance_div) { 812 /* Per-instance attrib. */ 813 unsigned count = (num_instances + instance_div - 1) / instance_div; 814 first += vb->stride * start_instance; 815 size = vb->stride * (count - 1) + mgr->ve->src_format_size[i]; 816 } else { 817 /* Per-vertex attrib. */ 818 first += vb->stride * start_vertex; 819 size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i]; 820 } 821 822 /* Update offsets. */ 823 if (!end_offset[index]) { 824 start_offset[index] = first; 825 end_offset[index] = first + size; 826 } else { 827 if (first < start_offset[index]) 828 start_offset[index] = first; 829 if (first + size > end_offset[index]) 830 end_offset[index] = first + size; 831 } 832 } 833 834 /* Upload buffers. */ 835 for (i = 0; i < nr_vbufs; i++) { 836 unsigned start, end = end_offset[i]; 837 struct pipe_vertex_buffer *real_vb; 838 uint8_t *ptr; 839 840 if (!end) { 841 continue; 842 } 843 844 start = start_offset[i]; 845 assert(start < end); 846 847 real_vb = &mgr->real_vertex_buffer[i]; 848 ptr = mgr->vertex_buffer[i].buffer->user_ptr; 849 850 u_upload_data(mgr->uploader, start, end - start, ptr + start, 851 &real_vb->buffer_offset, &real_vb->buffer); 852 853 real_vb->buffer_offset -= start; 854 } 855} 856 857static boolean u_vbuf_need_minmax_index(struct u_vbuf *mgr) 858{ 859 /* See if there are any per-vertex attribs which will be uploaded or 860 * translated. Use bitmasks to get the info instead of looping over vertex 861 * elements. */ 862 return ((mgr->user_vb_mask | mgr->incompatible_vb_mask | 863 mgr->ve->incompatible_vb_mask_any) & 864 mgr->ve->noninstance_vb_mask_any & mgr->nonzero_stride_vb_mask) != 0; 865} 866 867static boolean u_vbuf_mapping_vertex_buffer_blocks(struct u_vbuf *mgr) 868{ 869 /* Return true if there are hw buffers which don't need to be translated. 870 * 871 * We could query whether each buffer is busy, but that would 872 * be way more costly than this. */ 873 return (~mgr->user_vb_mask & ~mgr->incompatible_vb_mask & 874 mgr->ve->compatible_vb_mask_all & mgr->ve->noninstance_vb_mask_any & 875 mgr->nonzero_stride_vb_mask) != 0; 876} 877 878static void u_vbuf_get_minmax_index(struct pipe_context *pipe, 879 struct pipe_index_buffer *ib, 880 const struct pipe_draw_info *info, 881 int *out_min_index, 882 int *out_max_index) 883{ 884 struct pipe_transfer *transfer = NULL; 885 const void *indices; 886 unsigned i; 887 unsigned restart_index = info->restart_index; 888 889 if (ib->buffer->user_ptr) { 890 indices = ib->buffer->user_ptr + 891 ib->offset + info->start * ib->index_size; 892 } else { 893 indices = pipe_buffer_map_range(pipe, ib->buffer, 894 ib->offset + info->start * ib->index_size, 895 info->count * ib->index_size, 896 PIPE_TRANSFER_READ, &transfer); 897 } 898 899 switch (ib->index_size) { 900 case 4: { 901 const unsigned *ui_indices = (const unsigned*)indices; 902 unsigned max_ui = 0; 903 unsigned min_ui = ~0U; 904 if (info->primitive_restart) { 905 for (i = 0; i < info->count; i++) { 906 if (ui_indices[i] != restart_index) { 907 if (ui_indices[i] > max_ui) max_ui = ui_indices[i]; 908 if (ui_indices[i] < min_ui) min_ui = ui_indices[i]; 909 } 910 } 911 } 912 else { 913 for (i = 0; i < info->count; i++) { 914 if (ui_indices[i] > max_ui) max_ui = ui_indices[i]; 915 if (ui_indices[i] < min_ui) min_ui = ui_indices[i]; 916 } 917 } 918 *out_min_index = min_ui; 919 *out_max_index = max_ui; 920 break; 921 } 922 case 2: { 923 const unsigned short *us_indices = (const unsigned short*)indices; 924 unsigned max_us = 0; 925 unsigned min_us = ~0U; 926 if (info->primitive_restart) { 927 for (i = 0; i < info->count; i++) { 928 if (us_indices[i] != restart_index) { 929 if (us_indices[i] > max_us) max_us = us_indices[i]; 930 if (us_indices[i] < min_us) min_us = us_indices[i]; 931 } 932 } 933 } 934 else { 935 for (i = 0; i < info->count; i++) { 936 if (us_indices[i] > max_us) max_us = us_indices[i]; 937 if (us_indices[i] < min_us) min_us = us_indices[i]; 938 } 939 } 940 *out_min_index = min_us; 941 *out_max_index = max_us; 942 break; 943 } 944 case 1: { 945 const unsigned char *ub_indices = (const unsigned char*)indices; 946 unsigned max_ub = 0; 947 unsigned min_ub = ~0U; 948 if (info->primitive_restart) { 949 for (i = 0; i < info->count; i++) { 950 if (ub_indices[i] != restart_index) { 951 if (ub_indices[i] > max_ub) max_ub = ub_indices[i]; 952 if (ub_indices[i] < min_ub) min_ub = ub_indices[i]; 953 } 954 } 955 } 956 else { 957 for (i = 0; i < info->count; i++) { 958 if (ub_indices[i] > max_ub) max_ub = ub_indices[i]; 959 if (ub_indices[i] < min_ub) min_ub = ub_indices[i]; 960 } 961 } 962 *out_min_index = min_ub; 963 *out_max_index = max_ub; 964 break; 965 } 966 default: 967 assert(0); 968 *out_min_index = 0; 969 *out_max_index = 0; 970 } 971 972 if (transfer) { 973 pipe_buffer_unmap(pipe, transfer); 974 } 975} 976 977void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info) 978{ 979 struct pipe_context *pipe = mgr->pipe; 980 int start_vertex, min_index; 981 unsigned num_vertices; 982 boolean unroll_indices = FALSE; 983 984 /* Normal draw. No fallback and no user buffers. */ 985 if (!mgr->incompatible_vb_mask && 986 !mgr->ve->incompatible_elem_mask && 987 !mgr->user_vb_mask) { 988 /* Set vertex buffers if needed. */ 989 if (mgr->vertex_buffers_dirty) { 990 pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers, 991 mgr->real_vertex_buffer); 992 mgr->vertex_buffers_dirty = FALSE; 993 } 994 995 pipe->draw_vbo(pipe, info); 996 return; 997 } 998 999 if (info->indexed) { 1000 /* See if anything needs to be done for per-vertex attribs. */ 1001 if (u_vbuf_need_minmax_index(mgr)) { 1002 int max_index; 1003 1004 if (info->max_index != ~0) { 1005 min_index = info->min_index; 1006 max_index = info->max_index; 1007 } else { 1008 u_vbuf_get_minmax_index(mgr->pipe, &mgr->index_buffer, info, 1009 &min_index, &max_index); 1010 } 1011 1012 assert(min_index <= max_index); 1013 1014 start_vertex = min_index + info->index_bias; 1015 num_vertices = max_index + 1 - min_index; 1016 1017 /* Primitive restart doesn't work when unrolling indices. 1018 * We would have to break this drawing operation into several ones. */ 1019 /* Use some heuristic to see if unrolling indices improves 1020 * performance. */ 1021 if (!info->primitive_restart && 1022 num_vertices > info->count*2 && 1023 num_vertices-info->count > 32 && 1024 !u_vbuf_mapping_vertex_buffer_blocks(mgr)) { 1025 /*printf("num_vertices=%i count=%i\n", num_vertices, info->count);*/ 1026 unroll_indices = TRUE; 1027 } 1028 } else { 1029 /* Nothing to do for per-vertex attribs. */ 1030 start_vertex = 0; 1031 num_vertices = 0; 1032 min_index = 0; 1033 } 1034 } else { 1035 start_vertex = info->start; 1036 num_vertices = info->count; 1037 min_index = 0; 1038 } 1039 1040 /* Translate vertices with non-native layouts or formats. */ 1041 if (unroll_indices || 1042 mgr->incompatible_vb_mask || 1043 mgr->ve->incompatible_elem_mask) { 1044 /* XXX check the return value */ 1045 u_vbuf_translate_begin(mgr, start_vertex, num_vertices, 1046 info->start_instance, info->instance_count, 1047 info->start, info->count, min_index, 1048 unroll_indices); 1049 } 1050 1051 /* Upload user buffers. */ 1052 if (mgr->user_vb_mask) { 1053 u_vbuf_upload_buffers(mgr, start_vertex, num_vertices, 1054 info->start_instance, info->instance_count); 1055 } 1056 1057 /* 1058 if (unroll_indices) { 1059 printf("unrolling indices: start_vertex = %i, num_vertices = %i\n", 1060 start_vertex, num_vertices); 1061 util_dump_draw_info(stdout, info); 1062 printf("\n"); 1063 } 1064 1065 unsigned i; 1066 for (i = 0; i < mgr->nr_vertex_buffers; i++) { 1067 printf("input %i: ", i); 1068 util_dump_vertex_buffer(stdout, mgr->vertex_buffer+i); 1069 printf("\n"); 1070 } 1071 for (i = 0; i < mgr->nr_real_vertex_buffers; i++) { 1072 printf("real %i: ", i); 1073 util_dump_vertex_buffer(stdout, mgr->real_vertex_buffer+i); 1074 printf("\n"); 1075 } 1076 */ 1077 1078 u_upload_unmap(mgr->uploader); 1079 pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers, 1080 mgr->real_vertex_buffer); 1081 1082 if (unlikely(unroll_indices)) { 1083 struct pipe_draw_info new_info = *info; 1084 new_info.indexed = FALSE; 1085 new_info.index_bias = 0; 1086 new_info.min_index = 0; 1087 new_info.max_index = info->count - 1; 1088 new_info.start = 0; 1089 1090 pipe->draw_vbo(pipe, &new_info); 1091 } else { 1092 pipe->draw_vbo(pipe, info); 1093 } 1094 1095 if (mgr->using_translate) { 1096 u_vbuf_translate_end(mgr); 1097 } 1098 mgr->vertex_buffers_dirty = TRUE; 1099} 1100 1101void u_vbuf_save_vertex_elements(struct u_vbuf *mgr) 1102{ 1103 assert(!mgr->ve_saved); 1104 mgr->ve_saved = mgr->ve; 1105} 1106 1107void u_vbuf_restore_vertex_elements(struct u_vbuf *mgr) 1108{ 1109 if (mgr->ve != mgr->ve_saved) { 1110 struct pipe_context *pipe = mgr->pipe; 1111 1112 mgr->ve = mgr->ve_saved; 1113 pipe->bind_vertex_elements_state(pipe, 1114 mgr->ve ? mgr->ve->driver_cso : NULL); 1115 } 1116 mgr->ve_saved = NULL; 1117} 1118 1119void u_vbuf_save_vertex_buffers(struct u_vbuf *mgr) 1120{ 1121 util_copy_vertex_buffers(mgr->vertex_buffer_saved, 1122 &mgr->nr_vertex_buffers_saved, 1123 mgr->vertex_buffer, 1124 mgr->nr_vertex_buffers); 1125} 1126 1127void u_vbuf_restore_vertex_buffers(struct u_vbuf *mgr) 1128{ 1129 unsigned i; 1130 1131 u_vbuf_set_vertex_buffers(mgr, mgr->nr_vertex_buffers_saved, 1132 mgr->vertex_buffer_saved); 1133 for (i = 0; i < mgr->nr_vertex_buffers_saved; i++) { 1134 pipe_resource_reference(&mgr->vertex_buffer_saved[i].buffer, NULL); 1135 } 1136 mgr->nr_vertex_buffers_saved = 0; 1137} 1138