2 * Copyright © 2008 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
28 /** @file brw_queryobj.c
30 * Support for query objects (GL_ARB_occlusion_query, GL_ARB_timer_query,
31 * GL_EXT_transform_feedback, and friends).
33 * The hardware provides a PIPE_CONTROL command that can report the number of
34 * fragments that passed the depth test, or the hardware timer. They are
35 * appropriately synced with the stage of the pipeline for our extensions'
38 * To avoid getting samples from another context's rendering in our results,
39 * we capture the counts at the start and end of every batchbuffer while the
40 * query is active, and sum up the differences. (We should do so for
41 * GL_TIME_ELAPSED as well, but don't).
43 #include "main/imports.h"
45 #include "brw_context.h"
46 #include "brw_defines.h"
47 #include "brw_state.h"
48 #include "intel_batchbuffer.h"
49 #include "intel_reg.h"
52 write_timestamp(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
54 if (intel->gen >= 6) {
55 /* Emit workaround flushes: */
56 if (intel->gen == 6) {
57 /* The timestamp write below is a non-zero post-sync op, which on
58 * Gen6 necessitates a CS stall. CS stalls need stall at scoreboard
59 * set. See the comments for intel_emit_post_sync_nonzero_flush().
62 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2));
63 OUT_BATCH(PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD);
70 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
71 OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP);
73 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
74 PIPE_CONTROL_GLOBAL_GTT_WRITE |
75 idx * sizeof(uint64_t));
81 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
82 PIPE_CONTROL_WRITE_TIMESTAMP);
84 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
85 PIPE_CONTROL_GLOBAL_GTT_WRITE |
86 idx * sizeof(uint64_t));
94 write_depth_count(struct intel_context *intel, drm_intel_bo *query_bo, int idx)
96 if (intel->gen >= 6) {
97 /* Emit Sandybridge workaround flush: */
99 intel_emit_post_sync_nonzero_flush(intel);
102 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (5 - 2));
103 OUT_BATCH(PIPE_CONTROL_DEPTH_STALL |
104 PIPE_CONTROL_WRITE_DEPTH_COUNT);
106 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
107 PIPE_CONTROL_GLOBAL_GTT_WRITE |
108 (idx * sizeof(uint64_t)));
114 OUT_BATCH(_3DSTATE_PIPE_CONTROL | (4 - 2) |
115 PIPE_CONTROL_DEPTH_STALL |
116 PIPE_CONTROL_WRITE_DEPTH_COUNT);
117 /* This object could be mapped cacheable, but we don't have an exposed
118 * mechanism to support that. Since it's going uncached, tell GEM that
119 * we're writing to it. The usual clflush should be all that's required
120 * to pick up the results.
123 I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
124 PIPE_CONTROL_GLOBAL_GTT_WRITE |
125 (idx * sizeof(uint64_t)));
132 /** Waits on the query object's BO and totals the results for this query */
134 brw_queryobj_get_results(struct gl_context *ctx,
135 struct brw_query_object *query)
137 struct intel_context *intel = intel_context(ctx);
142 if (query->bo == NULL)
145 if (unlikely(INTEL_DEBUG & DEBUG_PERF)) {
146 if (drm_intel_bo_busy(query->bo)) {
147 perf_debug("Stalling on the GPU waiting for a query object.\n");
151 drm_intel_bo_map(query->bo, false);
152 results = query->bo->virtual;
153 switch (query->Base.Target) {
154 case GL_TIME_ELAPSED_EXT:
156 query->Base.Result += 80 * (results[1] - results[0]);
158 query->Base.Result += 1000 * ((results[1] >> 32) - (results[0] >> 32));
162 if (intel->gen >= 6) {
163 /* Our timer is a clock that increments every 80ns (regardless of
164 * other clock scaling in the system). The timestamp register we can
165 * read for glGetTimestamp() masks out the top 32 bits, so we do that
166 * here too to let the two counters be compared against each other.
168 * If we just multiplied that 32 bits of data by 80, it would roll
169 * over at a non-power-of-two, so an application couldn't use
170 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
171 * report 36 bits and truncate at that (rolling over 5 times as often
172 * as the HW counter), and when the 32-bit counter rolls over, it
173 * happens to also be at a rollover in the reported value from near
176 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
177 * rolls over every ~69 seconds.
179 query->Base.Result = 80 * (results[1] & 0xffffffff);
180 query->Base.Result &= (1ull << 36) - 1;
182 query->Base.Result = 1000 * (results[1] >> 32);
187 case GL_SAMPLES_PASSED_ARB:
188 /* Map and count the pixels from the current query BO */
189 for (i = query->first_index; i <= query->last_index; i++) {
190 query->Base.Result += results[i * 2 + 1] - results[i * 2];
194 case GL_ANY_SAMPLES_PASSED:
195 /* Set true if any of the sub-queries passed. */
196 for (i = query->first_index; i <= query->last_index; i++) {
197 if (results[i * 2 + 1] != results[i * 2]) {
198 query->Base.Result = GL_TRUE;
204 case GL_PRIMITIVES_GENERATED:
205 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
206 /* We don't actually query the hardware for this value, so query->bo
207 * should always be NULL and execution should never reach here.
209 assert(!"Unreachable");
213 assert(!"Unrecognized query target in brw_queryobj_get_results()");
216 drm_intel_bo_unmap(query->bo);
218 drm_intel_bo_unreference(query->bo);
222 static struct gl_query_object *
223 brw_new_query_object(struct gl_context *ctx, GLuint id)
225 struct brw_query_object *query;
227 query = calloc(1, sizeof(struct brw_query_object));
230 query->Base.Result = 0;
231 query->Base.Active = false;
232 query->Base.Ready = true;
238 brw_delete_query(struct gl_context *ctx, struct gl_query_object *q)
240 struct brw_query_object *query = (struct brw_query_object *)q;
242 drm_intel_bo_unreference(query->bo);
247 brw_begin_query(struct gl_context *ctx, struct gl_query_object *q)
249 struct brw_context *brw = brw_context(ctx);
250 struct intel_context *intel = intel_context(ctx);
251 struct brw_query_object *query = (struct brw_query_object *)q;
253 switch (query->Base.Target) {
254 case GL_TIME_ELAPSED_EXT:
255 drm_intel_bo_unreference(query->bo);
256 query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query", 4096, 4096);
257 write_timestamp(intel, query->bo, 0);
260 case GL_ANY_SAMPLES_PASSED:
261 case GL_SAMPLES_PASSED_ARB:
262 /* Reset our driver's tracking of query state. */
263 drm_intel_bo_unreference(query->bo);
265 query->first_index = -1;
266 query->last_index = -1;
268 brw->query.obj = query;
272 case GL_PRIMITIVES_GENERATED:
273 /* We don't actually query the hardware for this value; we keep track of
274 * it a software counter. So just reset the counter.
276 brw->sol.primitives_generated = 0;
277 brw->sol.counting_primitives_generated = true;
280 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
281 /* We don't actually query the hardware for this value; we keep track of
282 * it a software counter. So just reset the counter.
284 brw->sol.primitives_written = 0;
285 brw->sol.counting_primitives_written = true;
289 assert(!"Unrecognized query target in brw_begin_query()");
295 * Begin the ARB_occlusion_query query on a query object.
298 brw_end_query(struct gl_context *ctx, struct gl_query_object *q)
300 struct brw_context *brw = brw_context(ctx);
301 struct intel_context *intel = intel_context(ctx);
302 struct brw_query_object *query = (struct brw_query_object *)q;
304 switch (query->Base.Target) {
306 drm_intel_bo_unreference(query->bo);
307 query->bo = drm_intel_bo_alloc(intel->bufmgr, "timer query",
311 case GL_TIME_ELAPSED_EXT:
312 write_timestamp(intel, query->bo, 1);
313 intel_batchbuffer_flush(intel);
316 case GL_ANY_SAMPLES_PASSED:
317 case GL_SAMPLES_PASSED_ARB:
318 /* Flush the batchbuffer in case it has writes to our query BO.
319 * Have later queries write to a new query BO so that further rendering
320 * doesn't delay the collection of our results.
323 brw_emit_query_end(brw);
324 intel_batchbuffer_flush(intel);
326 drm_intel_bo_unreference(brw->query.bo);
327 brw->query.bo = NULL;
330 brw->query.obj = NULL;
335 case GL_PRIMITIVES_GENERATED:
336 /* We don't actually query the hardware for this value; we keep track of
337 * it in a software counter. So just read the counter and store it in
340 query->Base.Result = brw->sol.primitives_generated;
341 brw->sol.counting_primitives_generated = false;
343 /* And set brw->query.obj to NULL so that this query won't try to wait
344 * for any rendering to complete.
349 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
350 /* We don't actually query the hardware for this value; we keep track of
351 * it in a software counter. So just read the counter and store it in
354 query->Base.Result = brw->sol.primitives_written;
355 brw->sol.counting_primitives_written = false;
357 /* And set brw->query.obj to NULL so that this query won't try to wait
358 * for any rendering to complete.
364 assert(!"Unrecognized query target in brw_end_query()");
369 static void brw_wait_query(struct gl_context *ctx, struct gl_query_object *q)
371 struct brw_query_object *query = (struct brw_query_object *)q;
373 brw_queryobj_get_results(ctx, query);
374 query->Base.Ready = true;
377 static void brw_check_query(struct gl_context *ctx, struct gl_query_object *q)
379 struct brw_query_object *query = (struct brw_query_object *)q;
381 if (query->bo == NULL || !drm_intel_bo_busy(query->bo)) {
382 brw_queryobj_get_results(ctx, query);
383 query->Base.Ready = true;
387 /** Called to set up the query BO and account for its aperture space */
389 brw_prepare_query_begin(struct brw_context *brw)
391 struct intel_context *intel = &brw->intel;
393 /* Skip if we're not doing any queries. */
397 /* Get a new query BO if we're going to need it. */
398 if (brw->query.bo == NULL ||
399 brw->query.index * 2 + 1 >= 4096 / sizeof(uint64_t)) {
400 drm_intel_bo_unreference(brw->query.bo);
401 brw->query.bo = NULL;
403 brw->query.bo = drm_intel_bo_alloc(intel->bufmgr, "query", 4096, 1);
405 /* clear target buffer */
406 drm_intel_bo_map(brw->query.bo, true);
407 memset((char *)brw->query.bo->virtual, 0, 4096);
408 drm_intel_bo_unmap(brw->query.bo);
410 brw->query.index = 0;
414 /** Called just before primitive drawing to get a beginning PS_DEPTH_COUNT. */
416 brw_emit_query_begin(struct brw_context *brw)
418 struct intel_context *intel = &brw->intel;
419 struct gl_context *ctx = &intel->ctx;
420 struct brw_query_object *query = brw->query.obj;
422 /* Skip if we're not doing any queries, or we've emitted the start. */
423 if (!query || brw->query.active)
426 write_depth_count(intel, brw->query.bo, brw->query.index * 2);
428 if (query->bo != brw->query.bo) {
429 if (query->bo != NULL)
430 brw_queryobj_get_results(ctx, query);
431 drm_intel_bo_reference(brw->query.bo);
432 query->bo = brw->query.bo;
433 query->first_index = brw->query.index;
435 query->last_index = brw->query.index;
436 brw->query.active = true;
439 /** Called at batchbuffer flush to get an ending PS_DEPTH_COUNT */
441 brw_emit_query_end(struct brw_context *brw)
443 struct intel_context *intel = &brw->intel;
445 if (!brw->query.active)
448 write_depth_count(intel, brw->query.bo, brw->query.index * 2 + 1);
450 brw->query.active = false;
455 brw_get_timestamp(struct gl_context *ctx)
457 struct intel_context *intel = intel_context(ctx);
460 drm_intel_reg_read(intel->bufmgr, TIMESTAMP, &result);
462 /* See logic in brw_queryobj_get_results() */
463 result = result >> 32;
465 result &= (1ull << 36) - 1;
470 void brw_init_queryobj_functions(struct dd_function_table *functions)
472 functions->NewQueryObject = brw_new_query_object;
473 functions->DeleteQuery = brw_delete_query;
474 functions->BeginQuery = brw_begin_query;
475 functions->EndQuery = brw_end_query;
476 functions->CheckQuery = brw_check_query;
477 functions->WaitQuery = brw_wait_query;
478 functions->GetTimestamp = brw_get_timestamp;