1 /* 2 * Copyright 2008 Intel Corporation 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 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * 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 NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 * 23 * Authors: 24 * Eric Anholt <eric (at) anholt.net> 25 * Kenneth Graunke <kenneth (at) whitecape.org> 26 */ 27 28 /** @file gen6_queryobj.c 29 * 30 * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query, 31 * GL_EXT_transform_feedback, and friends) on platforms that support 32 * hardware contexts (Gen6+). 33 */ 34 #include "main/imports.h" 35 36 #include "brw_context.h" 37 #include "brw_defines.h" 38 #include "brw_state.h" 39 #include "intel_batchbuffer.h" 40 #include "intel_buffer_objects.h" 41 42 static inline void 43 set_query_availability(struct brw_context *brw, struct brw_query_object *query, 44 bool available) 45 { 46 /* For platforms that support ARB_query_buffer_object, we write the 47 * query availability for "pipelined" queries. 48 * 49 * Most counter snapshots are written by the command streamer, by 50 * doing a CS stall and then MI_STORE_REGISTER_MEM. For these 51 * counters, the CS stall guarantees that the results will be 52 * available when subsequent CS commands run. So we don't need to 53 * do any additional tracking. 54 * 55 * Other counters (occlusion queries and timestamp) are written by 56 * PIPE_CONTROL, without a CS stall. This means that we can't be 57 * sure whether the writes have landed yet or not. Performing a 58 * PIPE_CONTROL with an immediate write will synchronize with 59 * those earlier writes, so we write 1 when the value has landed. 60 */ 61 if (brw->ctx.Extensions.ARB_query_buffer_object && 62 brw_is_query_pipelined(query)) { 63 unsigned flags = PIPE_CONTROL_WRITE_IMMEDIATE; 64 65 if (available) { 66 /* Order available *after* the query results. */ 67 flags |= PIPE_CONTROL_FLUSH_ENABLE; 68 } else { 69 /* Make it unavailable *before* any pipelined reads. */ 70 flags |= PIPE_CONTROL_CS_STALL; 71 } 72 73 brw_emit_pipe_control_write(brw, flags, 74 query->bo, 2 * sizeof(uint64_t), 75 available); 76 } 77 } 78 79 static void 80 write_primitives_generated(struct brw_context *brw, 81 struct brw_bo *query_bo, int stream, int idx) 82 { 83 const struct gen_device_info *devinfo = &brw->screen->devinfo; 84 85 brw_emit_mi_flush(brw); 86 87 if (devinfo->gen >= 7 && stream > 0) { 88 brw_store_register_mem64(brw, query_bo, 89 GEN7_SO_PRIM_STORAGE_NEEDED(stream), 90 idx * sizeof(uint64_t)); 91 } else { 92 brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT, 93 idx * sizeof(uint64_t)); 94 } 95 } 96 97 static void 98 write_xfb_primitives_written(struct brw_context *brw, 99 struct brw_bo *bo, int stream, int idx) 100 { 101 const struct gen_device_info *devinfo = &brw->screen->devinfo; 102 103 brw_emit_mi_flush(brw); 104 105 if (devinfo->gen >= 7) { 106 brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream), 107 idx * sizeof(uint64_t)); 108 } else { 109 brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN, 110 idx * sizeof(uint64_t)); 111 } 112 } 113 114 static void 115 write_xfb_overflow_streams(struct gl_context *ctx, 116 struct brw_bo *bo, int stream, int count, 117 int idx) 118 { 119 struct brw_context *brw = brw_context(ctx); 120 const struct gen_device_info *devinfo = &brw->screen->devinfo; 121 122 brw_emit_mi_flush(brw); 123 124 for (int i = 0; i < count; i++) { 125 int w_idx = 4 * i + idx; 126 int g_idx = 4 * i + idx + 2; 127 128 if (devinfo->gen >= 7) { 129 brw_store_register_mem64(brw, bo, 130 GEN7_SO_NUM_PRIMS_WRITTEN(stream + i), 131 g_idx * sizeof(uint64_t)); 132 brw_store_register_mem64(brw, bo, 133 GEN7_SO_PRIM_STORAGE_NEEDED(stream + i), 134 w_idx * sizeof(uint64_t)); 135 } else { 136 brw_store_register_mem64(brw, bo, 137 GEN6_SO_NUM_PRIMS_WRITTEN, 138 g_idx * sizeof(uint64_t)); 139 brw_store_register_mem64(brw, bo, 140 GEN6_SO_PRIM_STORAGE_NEEDED, 141 w_idx * sizeof(uint64_t)); 142 } 143 } 144 } 145 146 static bool 147 check_xfb_overflow_streams(uint64_t *results, int count) 148 { 149 bool overflow = false; 150 151 for (int i = 0; i < count; i++) { 152 uint64_t *result_i = &results[4 * i]; 153 154 if ((result_i[3] - result_i[2]) != (result_i[1] - result_i[0])) { 155 overflow = true; 156 break; 157 } 158 } 159 160 return overflow; 161 } 162 163 static inline int 164 pipeline_target_to_index(int target) 165 { 166 if (target == GL_GEOMETRY_SHADER_INVOCATIONS) 167 return MAX_PIPELINE_STATISTICS - 1; 168 else 169 return target - GL_VERTICES_SUBMITTED_ARB; 170 } 171 172 static void 173 emit_pipeline_stat(struct brw_context *brw, struct brw_bo *bo, 174 int stream, int target, int idx) 175 { 176 const struct gen_device_info *devinfo = &brw->screen->devinfo; 177 178 /* One source of confusion is the tessellation shader statistics. The 179 * hardware has no statistics specific to the TE unit. Ideally we could have 180 * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS 181 * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB. 182 * Unfortunately we don't have HS primitives, we only have HS invocations. 183 */ 184 185 /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple 186 * lookup table 187 */ 188 static const uint32_t target_to_register[] = { 189 IA_VERTICES_COUNT, /* VERTICES_SUBMITTED */ 190 IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */ 191 VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */ 192 HS_INVOCATION_COUNT, /* TESS_CONTROL_SHADER_PATCHES */ 193 DS_INVOCATION_COUNT, /* TESS_EVALUATION_SHADER_INVOCATIONS */ 194 GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */ 195 PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */ 196 CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */ 197 CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */ 198 CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */ 199 GS_INVOCATION_COUNT /* This one is special... */ 200 }; 201 STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS); 202 uint32_t reg = target_to_register[pipeline_target_to_index(target)]; 203 /* Gen6 GS code counts full primitives, that is, it won't count individual 204 * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that. 205 */ 206 if (devinfo->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB) 207 reg = CL_INVOCATION_COUNT; 208 assert(reg != 0); 209 210 /* Emit a flush to make sure various parts of the pipeline are complete and 211 * we get an accurate value 212 */ 213 brw_emit_mi_flush(brw); 214 215 brw_store_register_mem64(brw, bo, reg, idx * sizeof(uint64_t)); 216 } 217 218 219 /** 220 * Wait on the query object's BO and calculate the final result. 221 */ 222 static void 223 gen6_queryobj_get_results(struct gl_context *ctx, 224 struct brw_query_object *query) 225 { 226 struct brw_context *brw = brw_context(ctx); 227 const struct gen_device_info *devinfo = &brw->screen->devinfo; 228 229 if (query->bo == NULL) 230 return; 231 232 uint64_t *results = brw_bo_map(brw, query->bo, MAP_READ); 233 switch (query->Base.Target) { 234 case GL_TIME_ELAPSED: 235 /* The query BO contains the starting and ending timestamps. 236 * Subtract the two and convert to nanoseconds. 237 */ 238 query->Base.Result = brw_raw_timestamp_delta(brw, results[0], results[1]); 239 query->Base.Result = brw_timebase_scale(brw, query->Base.Result); 240 break; 241 242 case GL_TIMESTAMP: 243 /* The query BO contains a single timestamp value in results[0]. */ 244 query->Base.Result = brw_timebase_scale(brw, results[0]); 245 246 /* Ensure the scaled timestamp overflows according to 247 * GL_QUERY_COUNTER_BITS 248 */ 249 query->Base.Result &= (1ull << ctx->Const.QueryCounterBits.Timestamp) - 1; 250 break; 251 252 case GL_SAMPLES_PASSED_ARB: 253 /* We need to use += rather than = here since some BLT-based operations 254 * may have added additional samples to our occlusion query value. 255 */ 256 query->Base.Result += results[1] - results[0]; 257 break; 258 259 case GL_ANY_SAMPLES_PASSED: 260 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: 261 if (results[0] != results[1]) 262 query->Base.Result = true; 263 break; 264 265 case GL_PRIMITIVES_GENERATED: 266 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN: 267 case GL_VERTICES_SUBMITTED_ARB: 268 case GL_PRIMITIVES_SUBMITTED_ARB: 269 case GL_VERTEX_SHADER_INVOCATIONS_ARB: 270 case GL_GEOMETRY_SHADER_INVOCATIONS: 271 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB: 272 case GL_CLIPPING_INPUT_PRIMITIVES_ARB: 273 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB: 274 case GL_COMPUTE_SHADER_INVOCATIONS_ARB: 275 case GL_TESS_CONTROL_SHADER_PATCHES_ARB: 276 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB: 277 query->Base.Result = results[1] - results[0]; 278 break; 279 280 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB: 281 query->Base.Result = check_xfb_overflow_streams(results, 1); 282 break; 283 284 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB: 285 query->Base.Result = check_xfb_overflow_streams(results, MAX_VERTEX_STREAMS); 286 break; 287 288 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB: 289 query->Base.Result = (results[1] - results[0]); 290 /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround: 291 * "Invocation counter is 4 times actual. WA: SW to divide HW reported 292 * PS Invocations value by 4." 293 * 294 * Prior to Haswell, invocation count was counted by the WM, and it 295 * buggily counted invocations in units of subspans (2x2 unit). To get the 296 * correct value, the CS multiplied this by 4. With HSW the logic moved, 297 * and correctly emitted the number of pixel shader invocations, but, 298 * whomever forgot to undo the multiply by 4. 299 */ 300 if (devinfo->gen == 8 || devinfo->is_haswell) 301 query->Base.Result /= 4; 302 break; 303 304 default: 305 unreachable("Unrecognized query target in brw_queryobj_get_results()"); 306 } 307 brw_bo_unmap(query->bo); 308 309 /* Now that we've processed the data stored in the query's buffer object, 310 * we can release it. 311 */ 312 brw_bo_unreference(query->bo); 313 query->bo = NULL; 314 315 query->Base.Ready = true; 316 } 317 318 /** 319 * Driver hook for glBeginQuery(). 320 * 321 * Initializes driver structures and emits any GPU commands required to begin 322 * recording data for the query. 323 */ 324 static void 325 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q) 326 { 327 struct brw_context *brw = brw_context(ctx); 328 struct brw_query_object *query = (struct brw_query_object *)q; 329 330 /* Since we're starting a new query, we need to throw away old results. */ 331 brw_bo_unreference(query->bo); 332 query->bo = brw_bo_alloc(brw->bufmgr, "query results", 4096, 4096); 333 334 /* For ARB_query_buffer_object: The result is not available */ 335 set_query_availability(brw, query, false); 336 337 switch (query->Base.Target) { 338 case GL_TIME_ELAPSED: 339 /* For timestamp queries, we record the starting time right away so that 340 * we measure the full time between BeginQuery and EndQuery. There's 341 * some debate about whether this is the right thing to do. Our decision 342 * is based on the following text from the ARB_timer_query extension: 343 * 344 * "(5) Should the extension measure total time elapsed between the full 345 * completion of the BeginQuery and EndQuery commands, or just time 346 * spent in the graphics library? 347 * 348 * RESOLVED: This extension will measure the total time elapsed 349 * between the full completion of these commands. Future extensions 350 * may implement a query to determine time elapsed at different stages 351 * of the graphics pipeline." 352 * 353 * We write a starting timestamp now (at index 0). At EndQuery() time, 354 * we'll write a second timestamp (at index 1), and subtract the two to 355 * obtain the time elapsed. Notably, this includes time elapsed while 356 * the system was doing other work, such as running other applications. 357 */ 358 brw_write_timestamp(brw, query->bo, 0); 359 break; 360 361 case GL_ANY_SAMPLES_PASSED: 362 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: 363 case GL_SAMPLES_PASSED_ARB: 364 brw_write_depth_count(brw, query->bo, 0); 365 break; 366 367 case GL_PRIMITIVES_GENERATED: 368 write_primitives_generated(brw, query->bo, query->Base.Stream, 0); 369 if (query->Base.Stream == 0) 370 ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD; 371 break; 372 373 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN: 374 write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0); 375 break; 376 377 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB: 378 write_xfb_overflow_streams(ctx, query->bo, query->Base.Stream, 1, 0); 379 break; 380 381 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB: 382 write_xfb_overflow_streams(ctx, query->bo, 0, MAX_VERTEX_STREAMS, 0); 383 break; 384 385 case GL_VERTICES_SUBMITTED_ARB: 386 case GL_PRIMITIVES_SUBMITTED_ARB: 387 case GL_VERTEX_SHADER_INVOCATIONS_ARB: 388 case GL_GEOMETRY_SHADER_INVOCATIONS: 389 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB: 390 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB: 391 case GL_CLIPPING_INPUT_PRIMITIVES_ARB: 392 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB: 393 case GL_COMPUTE_SHADER_INVOCATIONS_ARB: 394 case GL_TESS_CONTROL_SHADER_PATCHES_ARB: 395 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB: 396 emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0); 397 break; 398 399 default: 400 unreachable("Unrecognized query target in brw_begin_query()"); 401 } 402 } 403 404 /** 405 * Driver hook for glEndQuery(). 406 * 407 * Emits GPU commands to record a final query value, ending any data capturing. 408 * However, the final result isn't necessarily available until the GPU processes 409 * those commands. brw_queryobj_get_results() processes the captured data to 410 * produce the final result. 411 */ 412 static void 413 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q) 414 { 415 struct brw_context *brw = brw_context(ctx); 416 struct brw_query_object *query = (struct brw_query_object *)q; 417 418 switch (query->Base.Target) { 419 case GL_TIME_ELAPSED: 420 brw_write_timestamp(brw, query->bo, 1); 421 break; 422 423 case GL_ANY_SAMPLES_PASSED: 424 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE: 425 case GL_SAMPLES_PASSED_ARB: 426 brw_write_depth_count(brw, query->bo, 1); 427 break; 428 429 case GL_PRIMITIVES_GENERATED: 430 write_primitives_generated(brw, query->bo, query->Base.Stream, 1); 431 if (query->Base.Stream == 0) 432 ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD; 433 break; 434 435 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN: 436 write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1); 437 break; 438 439 case GL_TRANSFORM_FEEDBACK_STREAM_OVERFLOW_ARB: 440 write_xfb_overflow_streams(ctx, query->bo, query->Base.Stream, 1, 1); 441 break; 442 443 case GL_TRANSFORM_FEEDBACK_OVERFLOW_ARB: 444 write_xfb_overflow_streams(ctx, query->bo, 0, MAX_VERTEX_STREAMS, 1); 445 break; 446 447 /* calculate overflow here */ 448 case GL_VERTICES_SUBMITTED_ARB: 449 case GL_PRIMITIVES_SUBMITTED_ARB: 450 case GL_VERTEX_SHADER_INVOCATIONS_ARB: 451 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB: 452 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB: 453 case GL_COMPUTE_SHADER_INVOCATIONS_ARB: 454 case GL_CLIPPING_INPUT_PRIMITIVES_ARB: 455 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB: 456 case GL_GEOMETRY_SHADER_INVOCATIONS: 457 case GL_TESS_CONTROL_SHADER_PATCHES_ARB: 458 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB: 459 emit_pipeline_stat(brw, query->bo, 460 query->Base.Stream, query->Base.Target, 1); 461 break; 462 463 default: 464 unreachable("Unrecognized query target in brw_end_query()"); 465 } 466 467 /* The current batch contains the commands to handle EndQuery(), 468 * but they won't actually execute until it is flushed. 469 */ 470 query->flushed = false; 471 472 /* For ARB_query_buffer_object: The result is now available */ 473 set_query_availability(brw, query, true); 474 } 475 476 /** 477 * Flush the batch if it still references the query object BO. 478 */ 479 static void 480 flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query) 481 { 482 /* If the batch doesn't reference the BO, it must have been flushed 483 * (for example, due to being full). Record that it's been flushed. 484 */ 485 query->flushed = query->flushed || 486 !brw_batch_references(&brw->batch, query->bo); 487 488 if (!query->flushed) 489 intel_batchbuffer_flush(brw); 490 } 491 492 /** 493 * The WaitQuery() driver hook. 494 * 495 * Wait for a query result to become available and return it. This is the 496 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname. 497 */ 498 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q) 499 { 500 struct brw_context *brw = brw_context(ctx); 501 struct brw_query_object *query = (struct brw_query_object *)q; 502 503 /* If the application has requested the query result, but this batch is 504 * still contributing to it, flush it now to finish that work so the 505 * result will become available (eventually). 506 */ 507 flush_batch_if_needed(brw, query); 508 509 gen6_queryobj_get_results(ctx, query); 510 } 511 512 /** 513 * The CheckQuery() driver hook. 514 * 515 * Checks whether a query result is ready yet. If not, flushes. 516 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname. 517 */ 518 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q) 519 { 520 struct brw_context *brw = brw_context(ctx); 521 struct brw_query_object *query = (struct brw_query_object *)q; 522 523 /* If query->bo is NULL, we've already gathered the results - this is a 524 * redundant CheckQuery call. Ignore it. 525 */ 526 if (query->bo == NULL) 527 return; 528 529 /* From the GL_ARB_occlusion_query spec: 530 * 531 * "Instead of allowing for an infinite loop, performing a 532 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is 533 * not ready yet on the first time it is queried. This ensures that 534 * the async query will return true in finite time. 535 */ 536 flush_batch_if_needed(brw, query); 537 538 if (!brw_bo_busy(query->bo)) { 539 gen6_queryobj_get_results(ctx, query); 540 } 541 } 542 543 static void 544 gen6_query_counter(struct gl_context *ctx, struct gl_query_object *q) 545 { 546 struct brw_context *brw = brw_context(ctx); 547 struct brw_query_object *query = (struct brw_query_object *)q; 548 brw_query_counter(ctx, q); 549 set_query_availability(brw, query, true); 550 } 551 552 /* Initialize Gen6+-specific query object functions. */ 553 void gen6_init_queryobj_functions(struct dd_function_table *functions) 554 { 555 functions->BeginQuery = gen6_begin_query; 556 functions->EndQuery = gen6_end_query; 557 functions->CheckQuery = gen6_check_query; 558 functions->WaitQuery = gen6_wait_query; 559 functions->QueryCounter = gen6_query_counter; 560 } 561