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@anholt.net>
25 * Kenneth Graunke <kenneth@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
set_query_availability(struct brw_context * brw,struct brw_query_object * query,bool available)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 brw_emit_pipe_control_write(brw,
64 PIPE_CONTROL_WRITE_IMMEDIATE,
65 query->bo, 2 * sizeof(uint64_t),
66 available, 0);
67 }
68 }
69
70 static void
write_primitives_generated(struct brw_context * brw,drm_intel_bo * query_bo,int stream,int idx)71 write_primitives_generated(struct brw_context *brw,
72 drm_intel_bo *query_bo, int stream, int idx)
73 {
74 brw_emit_mi_flush(brw);
75
76 if (brw->gen >= 7 && stream > 0) {
77 brw_store_register_mem64(brw, query_bo,
78 GEN7_SO_PRIM_STORAGE_NEEDED(stream),
79 idx * sizeof(uint64_t));
80 } else {
81 brw_store_register_mem64(brw, query_bo, CL_INVOCATION_COUNT,
82 idx * sizeof(uint64_t));
83 }
84 }
85
86 static void
write_xfb_primitives_written(struct brw_context * brw,drm_intel_bo * bo,int stream,int idx)87 write_xfb_primitives_written(struct brw_context *brw,
88 drm_intel_bo *bo, int stream, int idx)
89 {
90 brw_emit_mi_flush(brw);
91
92 if (brw->gen >= 7) {
93 brw_store_register_mem64(brw, bo, GEN7_SO_NUM_PRIMS_WRITTEN(stream),
94 idx * sizeof(uint64_t));
95 } else {
96 brw_store_register_mem64(brw, bo, GEN6_SO_NUM_PRIMS_WRITTEN,
97 idx * sizeof(uint64_t));
98 }
99 }
100
101 static inline int
pipeline_target_to_index(int target)102 pipeline_target_to_index(int target)
103 {
104 if (target == GL_GEOMETRY_SHADER_INVOCATIONS)
105 return MAX_PIPELINE_STATISTICS - 1;
106 else
107 return target - GL_VERTICES_SUBMITTED_ARB;
108 }
109
110 static void
emit_pipeline_stat(struct brw_context * brw,drm_intel_bo * bo,int stream,int target,int idx)111 emit_pipeline_stat(struct brw_context *brw, drm_intel_bo *bo,
112 int stream, int target, int idx)
113 {
114 /* One source of confusion is the tessellation shader statistics. The
115 * hardware has no statistics specific to the TE unit. Ideally we could have
116 * the HS primitives for TESS_CONTROL_SHADER_PATCHES_ARB, and the DS
117 * invocations as the register for TESS_CONTROL_SHADER_PATCHES_ARB.
118 * Unfortunately we don't have HS primitives, we only have HS invocations.
119 */
120
121 /* Everything except GEOMETRY_SHADER_INVOCATIONS can be kept in a simple
122 * lookup table
123 */
124 static const uint32_t target_to_register[] = {
125 IA_VERTICES_COUNT, /* VERTICES_SUBMITTED */
126 IA_PRIMITIVES_COUNT, /* PRIMITIVES_SUBMITTED */
127 VS_INVOCATION_COUNT, /* VERTEX_SHADER_INVOCATIONS */
128 HS_INVOCATION_COUNT, /* TESS_CONTROL_SHADER_PATCHES */
129 DS_INVOCATION_COUNT, /* TESS_EVALUATION_SHADER_INVOCATIONS */
130 GS_PRIMITIVES_COUNT, /* GEOMETRY_SHADER_PRIMITIVES_EMITTED */
131 PS_INVOCATION_COUNT, /* FRAGMENT_SHADER_INVOCATIONS */
132 CS_INVOCATION_COUNT, /* COMPUTE_SHADER_INVOCATIONS */
133 CL_INVOCATION_COUNT, /* CLIPPING_INPUT_PRIMITIVES */
134 CL_PRIMITIVES_COUNT, /* CLIPPING_OUTPUT_PRIMITIVES */
135 GS_INVOCATION_COUNT /* This one is special... */
136 };
137 STATIC_ASSERT(ARRAY_SIZE(target_to_register) == MAX_PIPELINE_STATISTICS);
138 uint32_t reg = target_to_register[pipeline_target_to_index(target)];
139 /* Gen6 GS code counts full primitives, that is, it won't count individual
140 * triangles in a triangle strip. Use CL_INVOCATION_COUNT for that.
141 */
142 if (brw->gen == 6 && target == GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB)
143 reg = CL_INVOCATION_COUNT;
144 assert(reg != 0);
145
146 /* Emit a flush to make sure various parts of the pipeline are complete and
147 * we get an accurate value
148 */
149 brw_emit_mi_flush(brw);
150
151 brw_store_register_mem64(brw, bo, reg, idx * sizeof(uint64_t));
152 }
153
154
155 /**
156 * Wait on the query object's BO and calculate the final result.
157 */
158 static void
gen6_queryobj_get_results(struct gl_context * ctx,struct brw_query_object * query)159 gen6_queryobj_get_results(struct gl_context *ctx,
160 struct brw_query_object *query)
161 {
162 struct brw_context *brw = brw_context(ctx);
163
164 if (query->bo == NULL)
165 return;
166
167 brw_bo_map(brw, query->bo, false, "query object");
168 uint64_t *results = query->bo->virtual;
169 switch (query->Base.Target) {
170 case GL_TIME_ELAPSED:
171 /* The query BO contains the starting and ending timestamps.
172 * Subtract the two and convert to nanoseconds.
173 */
174 query->Base.Result += 80 * (results[1] - results[0]);
175 break;
176
177 case GL_TIMESTAMP:
178 /* Our timer is a clock that increments every 80ns (regardless of
179 * other clock scaling in the system). The timestamp register we can
180 * read for glGetTimestamp() masks out the top 32 bits, so we do that
181 * here too to let the two counters be compared against each other.
182 *
183 * If we just multiplied that 32 bits of data by 80, it would roll
184 * over at a non-power-of-two, so an application couldn't use
185 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
186 * report 36 bits and truncate at that (rolling over 5 times as often
187 * as the HW counter), and when the 32-bit counter rolls over, it
188 * happens to also be at a rollover in the reported value from near
189 * (1<<36) to 0.
190 *
191 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
192 * rolls over every ~69 seconds.
193 *
194 * The query BO contains a single timestamp value in results[0].
195 */
196 query->Base.Result = 80 * (results[0] & 0xffffffff);
197 query->Base.Result &= (1ull << 36) - 1;
198 break;
199
200 case GL_SAMPLES_PASSED_ARB:
201 /* We need to use += rather than = here since some BLT-based operations
202 * may have added additional samples to our occlusion query value.
203 */
204 query->Base.Result += results[1] - results[0];
205 break;
206
207 case GL_ANY_SAMPLES_PASSED:
208 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
209 if (results[0] != results[1])
210 query->Base.Result = true;
211 break;
212
213 case GL_PRIMITIVES_GENERATED:
214 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
215 case GL_VERTICES_SUBMITTED_ARB:
216 case GL_PRIMITIVES_SUBMITTED_ARB:
217 case GL_VERTEX_SHADER_INVOCATIONS_ARB:
218 case GL_GEOMETRY_SHADER_INVOCATIONS:
219 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
220 case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
221 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
222 case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
223 case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
224 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
225 query->Base.Result = results[1] - results[0];
226 break;
227
228 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
229 query->Base.Result = (results[1] - results[0]);
230 /* Implement the "WaDividePSInvocationCountBy4:HSW,BDW" workaround:
231 * "Invocation counter is 4 times actual. WA: SW to divide HW reported
232 * PS Invocations value by 4."
233 *
234 * Prior to Haswell, invocation count was counted by the WM, and it
235 * buggily counted invocations in units of subspans (2x2 unit). To get the
236 * correct value, the CS multiplied this by 4. With HSW the logic moved,
237 * and correctly emitted the number of pixel shader invocations, but,
238 * whomever forgot to undo the multiply by 4.
239 */
240 if (brw->gen == 8 || brw->is_haswell)
241 query->Base.Result /= 4;
242 break;
243
244 default:
245 unreachable("Unrecognized query target in brw_queryobj_get_results()");
246 }
247 drm_intel_bo_unmap(query->bo);
248
249 /* Now that we've processed the data stored in the query's buffer object,
250 * we can release it.
251 */
252 drm_intel_bo_unreference(query->bo);
253 query->bo = NULL;
254
255 query->Base.Ready = true;
256 }
257
258 /**
259 * Driver hook for glBeginQuery().
260 *
261 * Initializes driver structures and emits any GPU commands required to begin
262 * recording data for the query.
263 */
264 static void
gen6_begin_query(struct gl_context * ctx,struct gl_query_object * q)265 gen6_begin_query(struct gl_context *ctx, struct gl_query_object *q)
266 {
267 struct brw_context *brw = brw_context(ctx);
268 struct brw_query_object *query = (struct brw_query_object *)q;
269
270 /* Since we're starting a new query, we need to throw away old results. */
271 drm_intel_bo_unreference(query->bo);
272 query->bo = drm_intel_bo_alloc(brw->bufmgr, "query results", 4096, 4096);
273
274 /* For ARB_query_buffer_object: The result is not available */
275 set_query_availability(brw, query, false);
276
277 switch (query->Base.Target) {
278 case GL_TIME_ELAPSED:
279 /* For timestamp queries, we record the starting time right away so that
280 * we measure the full time between BeginQuery and EndQuery. There's
281 * some debate about whether this is the right thing to do. Our decision
282 * is based on the following text from the ARB_timer_query extension:
283 *
284 * "(5) Should the extension measure total time elapsed between the full
285 * completion of the BeginQuery and EndQuery commands, or just time
286 * spent in the graphics library?
287 *
288 * RESOLVED: This extension will measure the total time elapsed
289 * between the full completion of these commands. Future extensions
290 * may implement a query to determine time elapsed at different stages
291 * of the graphics pipeline."
292 *
293 * We write a starting timestamp now (at index 0). At EndQuery() time,
294 * we'll write a second timestamp (at index 1), and subtract the two to
295 * obtain the time elapsed. Notably, this includes time elapsed while
296 * the system was doing other work, such as running other applications.
297 */
298 brw_write_timestamp(brw, query->bo, 0);
299 break;
300
301 case GL_ANY_SAMPLES_PASSED:
302 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
303 case GL_SAMPLES_PASSED_ARB:
304 brw_write_depth_count(brw, query->bo, 0);
305 break;
306
307 case GL_PRIMITIVES_GENERATED:
308 write_primitives_generated(brw, query->bo, query->Base.Stream, 0);
309 if (query->Base.Stream == 0)
310 ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
311 break;
312
313 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
314 write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 0);
315 break;
316
317 case GL_VERTICES_SUBMITTED_ARB:
318 case GL_PRIMITIVES_SUBMITTED_ARB:
319 case GL_VERTEX_SHADER_INVOCATIONS_ARB:
320 case GL_GEOMETRY_SHADER_INVOCATIONS:
321 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
322 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
323 case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
324 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
325 case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
326 case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
327 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
328 emit_pipeline_stat(brw, query->bo, query->Base.Stream, query->Base.Target, 0);
329 break;
330
331 default:
332 unreachable("Unrecognized query target in brw_begin_query()");
333 }
334 }
335
336 /**
337 * Driver hook for glEndQuery().
338 *
339 * Emits GPU commands to record a final query value, ending any data capturing.
340 * However, the final result isn't necessarily available until the GPU processes
341 * those commands. brw_queryobj_get_results() processes the captured data to
342 * produce the final result.
343 */
344 static void
gen6_end_query(struct gl_context * ctx,struct gl_query_object * q)345 gen6_end_query(struct gl_context *ctx, struct gl_query_object *q)
346 {
347 struct brw_context *brw = brw_context(ctx);
348 struct brw_query_object *query = (struct brw_query_object *)q;
349
350 switch (query->Base.Target) {
351 case GL_TIME_ELAPSED:
352 brw_write_timestamp(brw, query->bo, 1);
353 break;
354
355 case GL_ANY_SAMPLES_PASSED:
356 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
357 case GL_SAMPLES_PASSED_ARB:
358 brw_write_depth_count(brw, query->bo, 1);
359 break;
360
361 case GL_PRIMITIVES_GENERATED:
362 write_primitives_generated(brw, query->bo, query->Base.Stream, 1);
363 if (query->Base.Stream == 0)
364 ctx->NewDriverState |= BRW_NEW_RASTERIZER_DISCARD;
365 break;
366
367 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
368 write_xfb_primitives_written(brw, query->bo, query->Base.Stream, 1);
369 break;
370
371 case GL_VERTICES_SUBMITTED_ARB:
372 case GL_PRIMITIVES_SUBMITTED_ARB:
373 case GL_VERTEX_SHADER_INVOCATIONS_ARB:
374 case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
375 case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
376 case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
377 case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
378 case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
379 case GL_GEOMETRY_SHADER_INVOCATIONS:
380 case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
381 case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
382 emit_pipeline_stat(brw, query->bo,
383 query->Base.Stream, query->Base.Target, 1);
384 break;
385
386 default:
387 unreachable("Unrecognized query target in brw_end_query()");
388 }
389
390 /* The current batch contains the commands to handle EndQuery(),
391 * but they won't actually execute until it is flushed.
392 */
393 query->flushed = false;
394
395 /* For ARB_query_buffer_object: The result is now available */
396 set_query_availability(brw, query, true);
397 }
398
399 /**
400 * Flush the batch if it still references the query object BO.
401 */
402 static void
flush_batch_if_needed(struct brw_context * brw,struct brw_query_object * query)403 flush_batch_if_needed(struct brw_context *brw, struct brw_query_object *query)
404 {
405 /* If the batch doesn't reference the BO, it must have been flushed
406 * (for example, due to being full). Record that it's been flushed.
407 */
408 query->flushed = query->flushed ||
409 !drm_intel_bo_references(brw->batch.bo, query->bo);
410
411 if (!query->flushed)
412 intel_batchbuffer_flush(brw);
413 }
414
415 /**
416 * The WaitQuery() driver hook.
417 *
418 * Wait for a query result to become available and return it. This is the
419 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
420 */
gen6_wait_query(struct gl_context * ctx,struct gl_query_object * q)421 static void gen6_wait_query(struct gl_context *ctx, struct gl_query_object *q)
422 {
423 struct brw_context *brw = brw_context(ctx);
424 struct brw_query_object *query = (struct brw_query_object *)q;
425
426 /* If the application has requested the query result, but this batch is
427 * still contributing to it, flush it now to finish that work so the
428 * result will become available (eventually).
429 */
430 flush_batch_if_needed(brw, query);
431
432 gen6_queryobj_get_results(ctx, query);
433 }
434
435 /**
436 * The CheckQuery() driver hook.
437 *
438 * Checks whether a query result is ready yet. If not, flushes.
439 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
440 */
gen6_check_query(struct gl_context * ctx,struct gl_query_object * q)441 static void gen6_check_query(struct gl_context *ctx, struct gl_query_object *q)
442 {
443 struct brw_context *brw = brw_context(ctx);
444 struct brw_query_object *query = (struct brw_query_object *)q;
445
446 /* If query->bo is NULL, we've already gathered the results - this is a
447 * redundant CheckQuery call. Ignore it.
448 */
449 if (query->bo == NULL)
450 return;
451
452 /* From the GL_ARB_occlusion_query spec:
453 *
454 * "Instead of allowing for an infinite loop, performing a
455 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
456 * not ready yet on the first time it is queried. This ensures that
457 * the async query will return true in finite time.
458 */
459 flush_batch_if_needed(brw, query);
460
461 if (!drm_intel_bo_busy(query->bo)) {
462 gen6_queryobj_get_results(ctx, query);
463 }
464 }
465
466 static void
gen6_query_counter(struct gl_context * ctx,struct gl_query_object * q)467 gen6_query_counter(struct gl_context *ctx, struct gl_query_object *q)
468 {
469 struct brw_context *brw = brw_context(ctx);
470 struct brw_query_object *query = (struct brw_query_object *)q;
471 brw_query_counter(ctx, q);
472 set_query_availability(brw, query, true);
473 }
474
475 /* Initialize Gen6+-specific query object functions. */
gen6_init_queryobj_functions(struct dd_function_table * functions)476 void gen6_init_queryobj_functions(struct dd_function_table *functions)
477 {
478 functions->BeginQuery = gen6_begin_query;
479 functions->EndQuery = gen6_end_query;
480 functions->CheckQuery = gen6_check_query;
481 functions->WaitQuery = gen6_wait_query;
482 functions->QueryCounter = gen6_query_counter;
483 }
484