u_vbuf.c revision b5e5e61439bda7a3cf1f909b48467371ea53d9d7
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 i, type; 374 unsigned nr = mgr->ve->count; 375 boolean used_vb[PIPE_MAX_ATTRIBS] = {0}; 376 unsigned fallback_vbs[VB_NUM]; 377 378 memset(fallback_vbs, ~0, sizeof(fallback_vbs)); 379 380 /* Mark used vertex buffers as... used. */ 381 for (i = 0; i < nr; i++) { 382 if (!(mgr->ve->incompatible_elem_mask & (1 << i))) { 383 unsigned index = mgr->ve->ve[i].vertex_buffer_index; 384 385 if (!(mgr->incompatible_vb_mask & (1 << index))) { 386 used_vb[index] = TRUE; 387 } 388 } 389 } 390 391 /* Find free slots for each type if needed. */ 392 i = 0; 393 for (type = 0; type < VB_NUM; type++) { 394 if (mask[type]) { 395 for (; i < PIPE_MAX_ATTRIBS; i++) { 396 if (!used_vb[i]) { 397 /*printf("found slot=%i for type=%i\n", i, type);*/ 398 fallback_vbs[type] = i; 399 i++; 400 if (i > mgr->nr_real_vertex_buffers) { 401 mgr->nr_real_vertex_buffers = i; 402 } 403 break; 404 } 405 } 406 if (i == PIPE_MAX_ATTRIBS) { 407 /* fail, reset the number to its original value */ 408 mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers; 409 return FALSE; 410 } 411 } 412 } 413 414 memcpy(mgr->fallback_vbs, fallback_vbs, sizeof(fallback_vbs)); 415 return TRUE; 416} 417 418static boolean 419u_vbuf_translate_begin(struct u_vbuf *mgr, 420 int start_vertex, unsigned num_vertices, 421 int start_instance, unsigned num_instances, 422 int start_index, unsigned num_indices, int min_index, 423 boolean unroll_indices) 424{ 425 unsigned mask[VB_NUM] = {0}; 426 struct translate_key key[VB_NUM]; 427 unsigned elem_index[VB_NUM][PIPE_MAX_ATTRIBS]; /* ... into key.elements */ 428 unsigned i, type; 429 430 int start[VB_NUM] = { 431 start_vertex, /* VERTEX */ 432 start_instance, /* INSTANCE */ 433 0 /* CONST */ 434 }; 435 436 unsigned num[VB_NUM] = { 437 num_vertices, /* VERTEX */ 438 num_instances, /* INSTANCE */ 439 1 /* CONST */ 440 }; 441 442 memset(key, 0, sizeof(key)); 443 memset(elem_index, ~0, sizeof(elem_index)); 444 445 /* See if there are vertex attribs of each type to translate and 446 * which ones. */ 447 for (i = 0; i < mgr->ve->count; i++) { 448 unsigned vb_index = mgr->ve->ve[i].vertex_buffer_index; 449 450 if (!mgr->vertex_buffer[vb_index].stride) { 451 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 452 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 453 continue; 454 } 455 mask[VB_CONST] |= 1 << vb_index; 456 } else if (mgr->ve->ve[i].instance_divisor) { 457 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 458 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 459 continue; 460 } 461 mask[VB_INSTANCE] |= 1 << vb_index; 462 } else { 463 if (!unroll_indices && 464 !(mgr->ve->incompatible_elem_mask & (1 << i)) && 465 !(mgr->incompatible_vb_mask & (1 << vb_index))) { 466 continue; 467 } 468 mask[VB_VERTEX] |= 1 << vb_index; 469 } 470 } 471 472 assert(mask[VB_VERTEX] || mask[VB_INSTANCE] || mask[VB_CONST]); 473 474 /* Find free vertex buffer slots. */ 475 if (!u_vbuf_translate_find_free_vb_slots(mgr, mask)) { 476 return FALSE; 477 } 478 479 /* Initialize the translate keys. */ 480 for (i = 0; i < mgr->ve->count; i++) { 481 struct translate_key *k; 482 struct translate_element *te; 483 unsigned bit, vb_index = mgr->ve->ve[i].vertex_buffer_index; 484 bit = 1 << vb_index; 485 486 if (!(mgr->ve->incompatible_elem_mask & (1 << i)) && 487 !(mgr->incompatible_vb_mask & (1 << vb_index)) && 488 (!unroll_indices || !(mask[VB_VERTEX] & bit))) { 489 continue; 490 } 491 492 /* Set type to what we will translate. 493 * Whether vertex, instance, or constant attribs. */ 494 for (type = 0; type < VB_NUM; type++) { 495 if (mask[type] & bit) { 496 break; 497 } 498 } 499 assert(type < VB_NUM); 500 assert(translate_is_output_format_supported(mgr->ve->native_format[i])); 501 /*printf("velem=%i type=%i\n", i, type);*/ 502 503 /* Add the vertex element. */ 504 k = &key[type]; 505 elem_index[type][i] = k->nr_elements; 506 507 te = &k->element[k->nr_elements]; 508 te->type = TRANSLATE_ELEMENT_NORMAL; 509 te->instance_divisor = 0; 510 te->input_buffer = vb_index; 511 te->input_format = mgr->ve->ve[i].src_format; 512 te->input_offset = mgr->ve->ve[i].src_offset; 513 te->output_format = mgr->ve->native_format[i]; 514 te->output_offset = k->output_stride; 515 516 k->output_stride += mgr->ve->native_format_size[i]; 517 k->nr_elements++; 518 } 519 520 /* Translate buffers. */ 521 for (type = 0; type < VB_NUM; type++) { 522 if (key[type].nr_elements) { 523 u_vbuf_translate_buffers(mgr, &key[type], mask[type], 524 mgr->fallback_vbs[type], 525 start[type], num[type], 526 start_index, num_indices, min_index, 527 unroll_indices && type == VB_VERTEX); 528 529 /* Fixup the stride for constant attribs. */ 530 if (type == VB_CONST) { 531 mgr->real_vertex_buffer[mgr->fallback_vbs[VB_CONST]].stride = 0; 532 } 533 } 534 } 535 536 /* Setup new vertex elements. */ 537 for (i = 0; i < mgr->ve->count; i++) { 538 for (type = 0; type < VB_NUM; type++) { 539 if (elem_index[type][i] < key[type].nr_elements) { 540 struct translate_element *te = &key[type].element[elem_index[type][i]]; 541 mgr->fallback_velems[i].instance_divisor = mgr->ve->ve[i].instance_divisor; 542 mgr->fallback_velems[i].src_format = te->output_format; 543 mgr->fallback_velems[i].src_offset = te->output_offset; 544 mgr->fallback_velems[i].vertex_buffer_index = mgr->fallback_vbs[type]; 545 546 /* elem_index[type][i] can only be set for one type. */ 547 assert(type > VB_INSTANCE || elem_index[type+1][i] == ~0); 548 assert(type > VB_VERTEX || elem_index[type+2][i] == ~0); 549 break; 550 } 551 } 552 /* No translating, just copy the original vertex element over. */ 553 if (type == VB_NUM) { 554 memcpy(&mgr->fallback_velems[i], &mgr->ve->ve[i], 555 sizeof(struct pipe_vertex_element)); 556 } 557 } 558 559 u_vbuf_set_vertex_elements_internal(mgr, mgr->ve->count, 560 mgr->fallback_velems); 561 mgr->using_translate = TRUE; 562 return TRUE; 563} 564 565static void u_vbuf_translate_end(struct u_vbuf *mgr) 566{ 567 unsigned i; 568 569 /* Restore vertex elements. */ 570 mgr->pipe->bind_vertex_elements_state(mgr->pipe, mgr->ve->driver_cso); 571 mgr->using_translate = FALSE; 572 573 /* Unreference the now-unused VBOs. */ 574 for (i = 0; i < VB_NUM; i++) { 575 unsigned vb = mgr->fallback_vbs[i]; 576 if (vb != ~0) { 577 pipe_resource_reference(&mgr->real_vertex_buffer[vb].buffer, NULL); 578 mgr->fallback_vbs[i] = ~0; 579 } 580 } 581 mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers; 582} 583 584#define FORMAT_REPLACE(what, withwhat) \ 585 case PIPE_FORMAT_##what: format = PIPE_FORMAT_##withwhat; break 586 587static void * 588u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count, 589 const struct pipe_vertex_element *attribs) 590{ 591 struct pipe_context *pipe = mgr->pipe; 592 unsigned i; 593 struct pipe_vertex_element driver_attribs[PIPE_MAX_ATTRIBS]; 594 struct u_vbuf_elements *ve = CALLOC_STRUCT(u_vbuf_elements); 595 uint32_t used_buffers = 0; 596 597 ve->count = count; 598 599 memcpy(ve->ve, attribs, sizeof(struct pipe_vertex_element) * count); 600 memcpy(driver_attribs, attribs, sizeof(struct pipe_vertex_element) * count); 601 602 /* Set the best native format in case the original format is not 603 * supported. */ 604 for (i = 0; i < count; i++) { 605 enum pipe_format format = ve->ve[i].src_format; 606 607 ve->src_format_size[i] = util_format_get_blocksize(format); 608 609 used_buffers |= 1 << ve->ve[i].vertex_buffer_index; 610 611 if (!ve->ve[i].instance_divisor) { 612 ve->noninstance_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 613 } 614 615 /* Choose a native format. 616 * For now we don't care about the alignment, that's going to 617 * be sorted out later. */ 618 if (!mgr->caps.format_fixed32) { 619 switch (format) { 620 FORMAT_REPLACE(R32_FIXED, R32_FLOAT); 621 FORMAT_REPLACE(R32G32_FIXED, R32G32_FLOAT); 622 FORMAT_REPLACE(R32G32B32_FIXED, R32G32B32_FLOAT); 623 FORMAT_REPLACE(R32G32B32A32_FIXED, R32G32B32A32_FLOAT); 624 default:; 625 } 626 } 627 if (!mgr->caps.format_float16) { 628 switch (format) { 629 FORMAT_REPLACE(R16_FLOAT, R32_FLOAT); 630 FORMAT_REPLACE(R16G16_FLOAT, R32G32_FLOAT); 631 FORMAT_REPLACE(R16G16B16_FLOAT, R32G32B32_FLOAT); 632 FORMAT_REPLACE(R16G16B16A16_FLOAT, R32G32B32A32_FLOAT); 633 default:; 634 } 635 } 636 if (!mgr->caps.format_float64) { 637 switch (format) { 638 FORMAT_REPLACE(R64_FLOAT, R32_FLOAT); 639 FORMAT_REPLACE(R64G64_FLOAT, R32G32_FLOAT); 640 FORMAT_REPLACE(R64G64B64_FLOAT, R32G32B32_FLOAT); 641 FORMAT_REPLACE(R64G64B64A64_FLOAT, R32G32B32A32_FLOAT); 642 default:; 643 } 644 } 645 if (!mgr->caps.format_norm32) { 646 switch (format) { 647 FORMAT_REPLACE(R32_UNORM, R32_FLOAT); 648 FORMAT_REPLACE(R32G32_UNORM, R32G32_FLOAT); 649 FORMAT_REPLACE(R32G32B32_UNORM, R32G32B32_FLOAT); 650 FORMAT_REPLACE(R32G32B32A32_UNORM, R32G32B32A32_FLOAT); 651 FORMAT_REPLACE(R32_SNORM, R32_FLOAT); 652 FORMAT_REPLACE(R32G32_SNORM, R32G32_FLOAT); 653 FORMAT_REPLACE(R32G32B32_SNORM, R32G32B32_FLOAT); 654 FORMAT_REPLACE(R32G32B32A32_SNORM, R32G32B32A32_FLOAT); 655 default:; 656 } 657 } 658 if (!mgr->caps.format_scaled32) { 659 switch (format) { 660 FORMAT_REPLACE(R32_USCALED, R32_FLOAT); 661 FORMAT_REPLACE(R32G32_USCALED, R32G32_FLOAT); 662 FORMAT_REPLACE(R32G32B32_USCALED, R32G32B32_FLOAT); 663 FORMAT_REPLACE(R32G32B32A32_USCALED,R32G32B32A32_FLOAT); 664 FORMAT_REPLACE(R32_SSCALED, R32_FLOAT); 665 FORMAT_REPLACE(R32G32_SSCALED, R32G32_FLOAT); 666 FORMAT_REPLACE(R32G32B32_SSCALED, R32G32B32_FLOAT); 667 FORMAT_REPLACE(R32G32B32A32_SSCALED,R32G32B32A32_FLOAT); 668 default:; 669 } 670 } 671 672 driver_attribs[i].src_format = format; 673 ve->native_format[i] = format; 674 ve->native_format_size[i] = 675 util_format_get_blocksize(ve->native_format[i]); 676 677 if (ve->ve[i].src_format != format || 678 (!mgr->caps.velem_src_offset_unaligned && 679 ve->ve[i].src_offset % 4 != 0)) { 680 ve->incompatible_elem_mask |= 1 << i; 681 ve->incompatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 682 } else { 683 ve->compatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index; 684 } 685 } 686 687 ve->compatible_vb_mask_all = ~ve->incompatible_vb_mask_any & used_buffers; 688 ve->incompatible_vb_mask_all = ~ve->compatible_vb_mask_any & used_buffers; 689 690 /* Align the formats to the size of DWORD if needed. */ 691 if (!mgr->caps.velem_src_offset_unaligned) { 692 for (i = 0; i < count; i++) { 693 ve->native_format_size[i] = align(ve->native_format_size[i], 4); 694 } 695 } 696 697 ve->driver_cso = 698 pipe->create_vertex_elements_state(pipe, count, driver_attribs); 699 return ve; 700} 701 702static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso) 703{ 704 struct pipe_context *pipe = mgr->pipe; 705 struct u_vbuf_elements *ve = cso; 706 707 pipe->delete_vertex_elements_state(pipe, ve->driver_cso); 708 FREE(ve); 709} 710 711void u_vbuf_set_vertex_buffers(struct u_vbuf *mgr, unsigned count, 712 const struct pipe_vertex_buffer *bufs) 713{ 714 unsigned i; 715 716 mgr->user_vb_mask = 0; 717 mgr->incompatible_vb_mask = 0; 718 mgr->nonzero_stride_vb_mask = 0; 719 720 for (i = 0; i < count; i++) { 721 const struct pipe_vertex_buffer *vb = &bufs[i]; 722 struct pipe_vertex_buffer *orig_vb = &mgr->vertex_buffer[i]; 723 struct pipe_vertex_buffer *real_vb = &mgr->real_vertex_buffer[i]; 724 725 pipe_resource_reference(&orig_vb->buffer, vb->buffer); 726 727 real_vb->buffer_offset = orig_vb->buffer_offset = vb->buffer_offset; 728 real_vb->stride = orig_vb->stride = vb->stride; 729 730 if (vb->stride) { 731 mgr->nonzero_stride_vb_mask |= 1 << i; 732 } 733 734 if (!vb->buffer) { 735 pipe_resource_reference(&real_vb->buffer, NULL); 736 continue; 737 } 738 739 if ((!mgr->caps.buffer_offset_unaligned && vb->buffer_offset % 4 != 0) || 740 (!mgr->caps.buffer_stride_unaligned && vb->stride % 4 != 0)) { 741 mgr->incompatible_vb_mask |= 1 << i; 742 pipe_resource_reference(&real_vb->buffer, NULL); 743 continue; 744 } 745 746 if (vb->buffer->user_ptr) { 747 mgr->user_vb_mask |= 1 << i; 748 pipe_resource_reference(&real_vb->buffer, NULL); 749 continue; 750 } 751 752 pipe_resource_reference(&real_vb->buffer, vb->buffer); 753 } 754 755 for (i = count; i < mgr->nr_vertex_buffers; i++) { 756 pipe_resource_reference(&mgr->vertex_buffer[i].buffer, NULL); 757 } 758 for (i = count; i < mgr->nr_real_vertex_buffers; i++) { 759 pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL); 760 } 761 762 mgr->nr_vertex_buffers = count; 763 mgr->nr_real_vertex_buffers = count; 764 mgr->vertex_buffers_dirty = TRUE; 765} 766 767void u_vbuf_set_index_buffer(struct u_vbuf *mgr, 768 const struct pipe_index_buffer *ib) 769{ 770 struct pipe_context *pipe = mgr->pipe; 771 772 if (ib && ib->buffer) { 773 assert(ib->offset % ib->index_size == 0); 774 pipe_resource_reference(&mgr->index_buffer.buffer, ib->buffer); 775 mgr->index_buffer.offset = ib->offset; 776 mgr->index_buffer.index_size = ib->index_size; 777 } else { 778 pipe_resource_reference(&mgr->index_buffer.buffer, NULL); 779 } 780 781 pipe->set_index_buffer(pipe, ib); 782} 783 784static void 785u_vbuf_upload_buffers(struct u_vbuf *mgr, 786 int start_vertex, unsigned num_vertices, 787 int start_instance, unsigned num_instances) 788{ 789 unsigned i; 790 unsigned nr_velems = mgr->ve->count; 791 unsigned nr_vbufs = mgr->nr_vertex_buffers; 792 struct pipe_vertex_element *velems = 793 mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve; 794 unsigned start_offset[PIPE_MAX_ATTRIBS]; 795 unsigned end_offset[PIPE_MAX_ATTRIBS] = {0}; 796 797 /* Determine how much data needs to be uploaded. */ 798 for (i = 0; i < nr_velems; i++) { 799 struct pipe_vertex_element *velem = &velems[i]; 800 unsigned index = velem->vertex_buffer_index; 801 struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index]; 802 unsigned instance_div, first, size; 803 804 /* Skip the buffers generated by translate. */ 805 if (index == mgr->fallback_vbs[VB_VERTEX] || 806 index == mgr->fallback_vbs[VB_INSTANCE] || 807 index == mgr->fallback_vbs[VB_CONST]) { 808 continue; 809 } 810 811 assert(vb->buffer); 812 813 if (!vb->buffer->user_ptr) { 814 continue; 815 } 816 817 instance_div = velem->instance_divisor; 818 first = vb->buffer_offset + velem->src_offset; 819 820 if (!vb->stride) { 821 /* Constant attrib. */ 822 size = mgr->ve->src_format_size[i]; 823 } else if (instance_div) { 824 /* Per-instance attrib. */ 825 unsigned count = (num_instances + instance_div - 1) / instance_div; 826 first += vb->stride * start_instance; 827 size = vb->stride * (count - 1) + mgr->ve->src_format_size[i]; 828 } else { 829 /* Per-vertex attrib. */ 830 first += vb->stride * start_vertex; 831 size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i]; 832 } 833 834 /* Update offsets. */ 835 if (!end_offset[index]) { 836 start_offset[index] = first; 837 end_offset[index] = first + size; 838 } else { 839 if (first < start_offset[index]) 840 start_offset[index] = first; 841 if (first + size > end_offset[index]) 842 end_offset[index] = first + size; 843 } 844 } 845 846 /* Upload buffers. */ 847 for (i = 0; i < nr_vbufs; i++) { 848 unsigned start, end = end_offset[i]; 849 struct pipe_vertex_buffer *real_vb; 850 uint8_t *ptr; 851 852 if (!end) { 853 continue; 854 } 855 856 start = start_offset[i]; 857 assert(start < end); 858 859 real_vb = &mgr->real_vertex_buffer[i]; 860 ptr = mgr->vertex_buffer[i].buffer->user_ptr; 861 862 u_upload_data(mgr->uploader, start, end - start, ptr + start, 863 &real_vb->buffer_offset, &real_vb->buffer); 864 865 real_vb->buffer_offset -= start; 866 } 867} 868 869static boolean u_vbuf_need_minmax_index(struct u_vbuf *mgr) 870{ 871 /* See if there are any per-vertex attribs which will be uploaded or 872 * translated. Use bitmasks to get the info instead of looping over vertex 873 * elements. */ 874 return ((mgr->user_vb_mask | mgr->incompatible_vb_mask | 875 mgr->ve->incompatible_vb_mask_any) & 876 mgr->ve->noninstance_vb_mask_any & mgr->nonzero_stride_vb_mask) != 0; 877} 878 879static boolean u_vbuf_mapping_vertex_buffer_blocks(struct u_vbuf *mgr) 880{ 881 /* Return true if there are hw buffers which don't need to be translated. 882 * 883 * We could query whether each buffer is busy, but that would 884 * be way more costly than this. */ 885 return (~mgr->user_vb_mask & ~mgr->incompatible_vb_mask & 886 mgr->ve->compatible_vb_mask_all & mgr->ve->noninstance_vb_mask_any & 887 mgr->nonzero_stride_vb_mask) != 0; 888} 889 890static void u_vbuf_get_minmax_index(struct pipe_context *pipe, 891 struct pipe_index_buffer *ib, 892 const struct pipe_draw_info *info, 893 int *out_min_index, 894 int *out_max_index) 895{ 896 struct pipe_transfer *transfer = NULL; 897 const void *indices; 898 unsigned i; 899 unsigned restart_index = info->restart_index; 900 901 if (ib->buffer->user_ptr) { 902 indices = ib->buffer->user_ptr + 903 ib->offset + info->start * ib->index_size; 904 } else { 905 indices = pipe_buffer_map_range(pipe, ib->buffer, 906 ib->offset + info->start * ib->index_size, 907 info->count * ib->index_size, 908 PIPE_TRANSFER_READ, &transfer); 909 } 910 911 switch (ib->index_size) { 912 case 4: { 913 const unsigned *ui_indices = (const unsigned*)indices; 914 unsigned max_ui = 0; 915 unsigned min_ui = ~0U; 916 if (info->primitive_restart) { 917 for (i = 0; i < info->count; i++) { 918 if (ui_indices[i] != restart_index) { 919 if (ui_indices[i] > max_ui) max_ui = ui_indices[i]; 920 if (ui_indices[i] < min_ui) min_ui = ui_indices[i]; 921 } 922 } 923 } 924 else { 925 for (i = 0; i < info->count; i++) { 926 if (ui_indices[i] > max_ui) max_ui = ui_indices[i]; 927 if (ui_indices[i] < min_ui) min_ui = ui_indices[i]; 928 } 929 } 930 *out_min_index = min_ui; 931 *out_max_index = max_ui; 932 break; 933 } 934 case 2: { 935 const unsigned short *us_indices = (const unsigned short*)indices; 936 unsigned max_us = 0; 937 unsigned min_us = ~0U; 938 if (info->primitive_restart) { 939 for (i = 0; i < info->count; i++) { 940 if (us_indices[i] != restart_index) { 941 if (us_indices[i] > max_us) max_us = us_indices[i]; 942 if (us_indices[i] < min_us) min_us = us_indices[i]; 943 } 944 } 945 } 946 else { 947 for (i = 0; i < info->count; i++) { 948 if (us_indices[i] > max_us) max_us = us_indices[i]; 949 if (us_indices[i] < min_us) min_us = us_indices[i]; 950 } 951 } 952 *out_min_index = min_us; 953 *out_max_index = max_us; 954 break; 955 } 956 case 1: { 957 const unsigned char *ub_indices = (const unsigned char*)indices; 958 unsigned max_ub = 0; 959 unsigned min_ub = ~0U; 960 if (info->primitive_restart) { 961 for (i = 0; i < info->count; i++) { 962 if (ub_indices[i] != restart_index) { 963 if (ub_indices[i] > max_ub) max_ub = ub_indices[i]; 964 if (ub_indices[i] < min_ub) min_ub = ub_indices[i]; 965 } 966 } 967 } 968 else { 969 for (i = 0; i < info->count; i++) { 970 if (ub_indices[i] > max_ub) max_ub = ub_indices[i]; 971 if (ub_indices[i] < min_ub) min_ub = ub_indices[i]; 972 } 973 } 974 *out_min_index = min_ub; 975 *out_max_index = max_ub; 976 break; 977 } 978 default: 979 assert(0); 980 *out_min_index = 0; 981 *out_max_index = 0; 982 } 983 984 if (transfer) { 985 pipe_buffer_unmap(pipe, transfer); 986 } 987} 988 989void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info) 990{ 991 struct pipe_context *pipe = mgr->pipe; 992 int start_vertex, min_index; 993 unsigned num_vertices; 994 boolean unroll_indices = FALSE; 995 996 /* Normal draw. No fallback and no user buffers. */ 997 if (!mgr->incompatible_vb_mask && 998 !mgr->ve->incompatible_elem_mask && 999 !mgr->user_vb_mask) { 1000 /* Set vertex buffers if needed. */ 1001 if (mgr->vertex_buffers_dirty) { 1002 pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers, 1003 mgr->real_vertex_buffer); 1004 mgr->vertex_buffers_dirty = FALSE; 1005 } 1006 1007 pipe->draw_vbo(pipe, info); 1008 return; 1009 } 1010 1011 if (info->indexed) { 1012 /* See if anything needs to be done for per-vertex attribs. */ 1013 if (u_vbuf_need_minmax_index(mgr)) { 1014 int max_index; 1015 1016 if (info->max_index != ~0) { 1017 min_index = info->min_index; 1018 max_index = info->max_index; 1019 } else { 1020 u_vbuf_get_minmax_index(mgr->pipe, &mgr->index_buffer, info, 1021 &min_index, &max_index); 1022 } 1023 1024 assert(min_index <= max_index); 1025 1026 start_vertex = min_index + info->index_bias; 1027 num_vertices = max_index + 1 - min_index; 1028 1029 /* Primitive restart doesn't work when unrolling indices. 1030 * We would have to break this drawing operation into several ones. */ 1031 /* Use some heuristic to see if unrolling indices improves 1032 * performance. */ 1033 if (!info->primitive_restart && 1034 num_vertices > info->count*2 && 1035 num_vertices-info->count > 32 && 1036 !u_vbuf_mapping_vertex_buffer_blocks(mgr)) { 1037 /*printf("num_vertices=%i count=%i\n", num_vertices, info->count);*/ 1038 unroll_indices = TRUE; 1039 } 1040 } else { 1041 /* Nothing to do for per-vertex attribs. */ 1042 start_vertex = 0; 1043 num_vertices = 0; 1044 min_index = 0; 1045 } 1046 } else { 1047 start_vertex = info->start; 1048 num_vertices = info->count; 1049 min_index = 0; 1050 } 1051 1052 /* Translate vertices with non-native layouts or formats. */ 1053 if (unroll_indices || 1054 mgr->incompatible_vb_mask || 1055 mgr->ve->incompatible_elem_mask) { 1056 /* XXX check the return value */ 1057 u_vbuf_translate_begin(mgr, start_vertex, num_vertices, 1058 info->start_instance, info->instance_count, 1059 info->start, info->count, min_index, 1060 unroll_indices); 1061 } 1062 1063 /* Upload user buffers. */ 1064 if (mgr->user_vb_mask) { 1065 u_vbuf_upload_buffers(mgr, start_vertex, num_vertices, 1066 info->start_instance, info->instance_count); 1067 } 1068 1069 /* 1070 if (unroll_indices) { 1071 printf("unrolling indices: start_vertex = %i, num_vertices = %i\n", 1072 start_vertex, num_vertices); 1073 util_dump_draw_info(stdout, info); 1074 printf("\n"); 1075 } 1076 1077 unsigned i; 1078 for (i = 0; i < mgr->nr_vertex_buffers; i++) { 1079 printf("input %i: ", i); 1080 util_dump_vertex_buffer(stdout, mgr->vertex_buffer+i); 1081 printf("\n"); 1082 } 1083 for (i = 0; i < mgr->nr_real_vertex_buffers; i++) { 1084 printf("real %i: ", i); 1085 util_dump_vertex_buffer(stdout, mgr->real_vertex_buffer+i); 1086 printf("\n"); 1087 } 1088 */ 1089 1090 u_upload_unmap(mgr->uploader); 1091 pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers, 1092 mgr->real_vertex_buffer); 1093 1094 if (unlikely(unroll_indices)) { 1095 struct pipe_draw_info new_info = *info; 1096 new_info.indexed = FALSE; 1097 new_info.index_bias = 0; 1098 new_info.min_index = 0; 1099 new_info.max_index = info->count - 1; 1100 new_info.start = 0; 1101 1102 pipe->draw_vbo(pipe, &new_info); 1103 } else { 1104 pipe->draw_vbo(pipe, info); 1105 } 1106 1107 if (mgr->using_translate) { 1108 u_vbuf_translate_end(mgr); 1109 } 1110 mgr->vertex_buffers_dirty = TRUE; 1111} 1112 1113void u_vbuf_save_vertex_elements(struct u_vbuf *mgr) 1114{ 1115 assert(!mgr->ve_saved); 1116 mgr->ve_saved = mgr->ve; 1117} 1118 1119void u_vbuf_restore_vertex_elements(struct u_vbuf *mgr) 1120{ 1121 if (mgr->ve != mgr->ve_saved) { 1122 struct pipe_context *pipe = mgr->pipe; 1123 1124 mgr->ve = mgr->ve_saved; 1125 pipe->bind_vertex_elements_state(pipe, 1126 mgr->ve ? mgr->ve->driver_cso : NULL); 1127 } 1128 mgr->ve_saved = NULL; 1129} 1130 1131void u_vbuf_save_vertex_buffers(struct u_vbuf *mgr) 1132{ 1133 util_copy_vertex_buffers(mgr->vertex_buffer_saved, 1134 &mgr->nr_vertex_buffers_saved, 1135 mgr->vertex_buffer, 1136 mgr->nr_vertex_buffers); 1137} 1138 1139void u_vbuf_restore_vertex_buffers(struct u_vbuf *mgr) 1140{ 1141 unsigned i; 1142 1143 u_vbuf_set_vertex_buffers(mgr, mgr->nr_vertex_buffers_saved, 1144 mgr->vertex_buffer_saved); 1145 for (i = 0; i < mgr->nr_vertex_buffers_saved; i++) { 1146 pipe_resource_reference(&mgr->vertex_buffer_saved[i].buffer, NULL); 1147 } 1148 mgr->nr_vertex_buffers_saved = 0; 1149} 1150