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>
25 * Kenneth Graunke <kenneth@whitecape.org>
28 /** @file gen6_queryobj.c
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+).
34 #include "main/imports.h"
36 #include "brw_context.h"
37 #include "brw_defines.h"
38 #include "brw_state.h"
39 #include "intel_batchbuffer.h"
40 #include "intel_reg.h"
43 * Emit PIPE_CONTROLs to write the current GPU timestamp into a buffer.
46 write_timestamp(struct brw_context
*brw
, drm_intel_bo
*query_bo
, int idx
)
48 struct intel_context
*intel
= &brw
->intel
;
49 /* Emit workaround flushes: */
50 if (intel
->gen
== 6) {
51 /* The timestamp write below is a non-zero post-sync op, which on
52 * Gen6 necessitates a CS stall. CS stalls need stall at scoreboard
53 * set. See the comments for intel_emit_post_sync_nonzero_flush().
56 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (4 - 2));
57 OUT_BATCH(PIPE_CONTROL_CS_STALL
| PIPE_CONTROL_STALL_AT_SCOREBOARD
);
64 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (5 - 2));
65 OUT_BATCH(PIPE_CONTROL_WRITE_TIMESTAMP
);
67 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
68 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
69 idx
* sizeof(uint64_t));
76 * Emit PIPE_CONTROLs to write the PS_DEPTH_COUNT register into a buffer.
79 write_depth_count(struct brw_context
*brw
, drm_intel_bo
*query_bo
, int idx
)
81 struct intel_context
*intel
= &brw
->intel
;
82 /* Emit Sandybridge workaround flush: */
84 intel_emit_post_sync_nonzero_flush(brw
);
87 OUT_BATCH(_3DSTATE_PIPE_CONTROL
| (5 - 2));
88 OUT_BATCH(PIPE_CONTROL_DEPTH_STALL
|
89 PIPE_CONTROL_WRITE_DEPTH_COUNT
);
91 I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
92 PIPE_CONTROL_GLOBAL_GTT_WRITE
|
93 (idx
* sizeof(uint64_t)));
100 * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
102 * Only TIMESTAMP and PS_DEPTH_COUNT have special PIPE_CONTROL support; other
103 * counters have to be read via the generic MI_STORE_REGISTER_MEM. This
104 * function also performs a pipeline flush for proper synchronization.
107 write_reg(struct brw_context
*brw
,
108 drm_intel_bo
*query_bo
, uint32_t reg
, int idx
)
110 struct intel_context
*intel
= &brw
->intel
;
111 assert(intel
->gen
>= 6);
113 intel_batchbuffer_emit_mi_flush(brw
);
115 /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
116 * read a full 64-bit register, we need to do two of them.
119 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
121 OUT_RELOC(query_bo
, I915_GEM_DOMAIN_RENDER
, I915_GEM_DOMAIN_RENDER
,
122 idx
* sizeof(uint64_t));
126 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
127 OUT_BATCH(reg
+ sizeof(uint32_t));
128 OUT_RELOC(query_bo
, I915_GEM_DOMAIN_RENDER
, I915_GEM_DOMAIN_RENDER
,
129 sizeof(uint32_t) + idx
* sizeof(uint64_t));
134 write_primitives_generated(struct brw_context
*brw
,
135 drm_intel_bo
*query_bo
, int idx
)
137 write_reg(brw
, query_bo
, CL_INVOCATION_COUNT
, idx
);
141 write_xfb_primitives_written(struct brw_context
*brw
,
142 drm_intel_bo
*query_bo
, int idx
)
144 struct intel_context
*intel
= &brw
->intel
;
145 if (intel
->gen
>= 7) {
146 write_reg(brw
, query_bo
, SO_NUM_PRIMS_WRITTEN0_IVB
, idx
);
148 write_reg(brw
, query_bo
, SO_NUM_PRIMS_WRITTEN
, idx
);
153 * Wait on the query object's BO and calculate the final result.
156 gen6_queryobj_get_results(struct gl_context
*ctx
,
157 struct brw_query_object
*query
)
159 struct brw_context
*brw
= brw_context(ctx
);
161 if (query
->bo
== NULL
)
164 /* If the application has requested the query result, but this batch is
165 * still contributing to it, flush it now so the results will be present
168 if (drm_intel_bo_references(brw
->batch
.bo
, query
->bo
))
169 intel_batchbuffer_flush(brw
);
171 if (unlikely(brw
->perf_debug
)) {
172 if (drm_intel_bo_busy(query
->bo
)) {
173 perf_debug("Stalling on the GPU waiting for a query object.\n");
177 drm_intel_bo_map(query
->bo
, false);
178 uint64_t *results
= query
->bo
->virtual;
179 switch (query
->Base
.Target
) {
180 case GL_TIME_ELAPSED
:
181 /* The query BO contains the starting and ending timestamps.
182 * Subtract the two and convert to nanoseconds.
184 query
->Base
.Result
+= 80 * (results
[1] - results
[0]);
188 /* Our timer is a clock that increments every 80ns (regardless of
189 * other clock scaling in the system). The timestamp register we can
190 * read for glGetTimestamp() masks out the top 32 bits, so we do that
191 * here too to let the two counters be compared against each other.
193 * If we just multiplied that 32 bits of data by 80, it would roll
194 * over at a non-power-of-two, so an application couldn't use
195 * GL_QUERY_COUNTER_BITS to handle rollover correctly. Instead, we
196 * report 36 bits and truncate at that (rolling over 5 times as often
197 * as the HW counter), and when the 32-bit counter rolls over, it
198 * happens to also be at a rollover in the reported value from near
201 * The low 32 bits rolls over in ~343 seconds. Our 36-bit result
202 * rolls over every ~69 seconds.
204 * The query BO contains a single timestamp value in results[0].
206 query
->Base
.Result
= 80 * (results
[0] & 0xffffffff);
207 query
->Base
.Result
&= (1ull << 36) - 1;
210 case GL_SAMPLES_PASSED_ARB
:
211 /* We need to use += rather than = here since some BLT-based operations
212 * may have added additional samples to our occlusion query value.
214 query
->Base
.Result
+= results
[1] - results
[0];
217 case GL_ANY_SAMPLES_PASSED
:
218 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
219 if (results
[0] != results
[1])
220 query
->Base
.Result
= true;
223 case GL_PRIMITIVES_GENERATED
:
224 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
225 query
->Base
.Result
= results
[1] - results
[0];
229 assert(!"Unrecognized query target in brw_queryobj_get_results()");
232 drm_intel_bo_unmap(query
->bo
);
234 /* Now that we've processed the data stored in the query's buffer object,
237 drm_intel_bo_unreference(query
->bo
);
242 * Driver hook for glBeginQuery().
244 * Initializes driver structures and emits any GPU commands required to begin
245 * recording data for the query.
248 gen6_begin_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
250 struct brw_context
*brw
= brw_context(ctx
);
251 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
253 /* Since we're starting a new query, we need to throw away old results. */
254 drm_intel_bo_unreference(query
->bo
);
255 query
->bo
= drm_intel_bo_alloc(brw
->bufmgr
, "query results", 4096, 4096);
257 switch (query
->Base
.Target
) {
258 case GL_TIME_ELAPSED
:
259 /* For timestamp queries, we record the starting time right away so that
260 * we measure the full time between BeginQuery and EndQuery. There's
261 * some debate about whether this is the right thing to do. Our decision
262 * is based on the following text from the ARB_timer_query extension:
264 * "(5) Should the extension measure total time elapsed between the full
265 * completion of the BeginQuery and EndQuery commands, or just time
266 * spent in the graphics library?
268 * RESOLVED: This extension will measure the total time elapsed
269 * between the full completion of these commands. Future extensions
270 * may implement a query to determine time elapsed at different stages
271 * of the graphics pipeline."
273 * We write a starting timestamp now (at index 0). At EndQuery() time,
274 * we'll write a second timestamp (at index 1), and subtract the two to
275 * obtain the time elapsed. Notably, this includes time elapsed while
276 * the system was doing other work, such as running other applications.
278 write_timestamp(brw
, query
->bo
, 0);
281 case GL_ANY_SAMPLES_PASSED
:
282 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
283 case GL_SAMPLES_PASSED_ARB
:
284 write_depth_count(brw
, query
->bo
, 0);
287 case GL_PRIMITIVES_GENERATED
:
288 write_primitives_generated(brw
, query
->bo
, 0);
291 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
292 write_xfb_primitives_written(brw
, query
->bo
, 0);
296 assert(!"Unrecognized query target in brw_begin_query()");
302 * Driver hook for glEndQuery().
304 * Emits GPU commands to record a final query value, ending any data capturing.
305 * However, the final result isn't necessarily available until the GPU processes
306 * those commands. brw_queryobj_get_results() processes the captured data to
307 * produce the final result.
310 gen6_end_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
312 struct brw_context
*brw
= brw_context(ctx
);
313 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
315 switch (query
->Base
.Target
) {
316 case GL_TIME_ELAPSED
:
317 write_timestamp(brw
, query
->bo
, 1);
320 case GL_ANY_SAMPLES_PASSED
:
321 case GL_ANY_SAMPLES_PASSED_CONSERVATIVE
:
322 case GL_SAMPLES_PASSED_ARB
:
323 write_depth_count(brw
, query
->bo
, 1);
326 case GL_PRIMITIVES_GENERATED
:
327 write_primitives_generated(brw
, query
->bo
, 1);
330 case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
:
331 write_xfb_primitives_written(brw
, query
->bo
, 1);
335 assert(!"Unrecognized query target in brw_end_query()");
341 * The WaitQuery() driver hook.
343 * Wait for a query result to become available and return it. This is the
344 * backing for glGetQueryObjectiv() with the GL_QUERY_RESULT pname.
346 static void gen6_wait_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
348 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
350 gen6_queryobj_get_results(ctx
, query
);
351 query
->Base
.Ready
= true;
355 * The CheckQuery() driver hook.
357 * Checks whether a query result is ready yet. If not, flushes.
358 * This is the backing for glGetQueryObjectiv()'s QUERY_RESULT_AVAILABLE pname.
360 static void gen6_check_query(struct gl_context
*ctx
, struct gl_query_object
*q
)
362 struct brw_context
*brw
= brw_context(ctx
);
363 struct brw_query_object
*query
= (struct brw_query_object
*)q
;
365 /* From the GL_ARB_occlusion_query spec:
367 * "Instead of allowing for an infinite loop, performing a
368 * QUERY_RESULT_AVAILABLE_ARB will perform a flush if the result is
369 * not ready yet on the first time it is queried. This ensures that
370 * the async query will return true in finite time.
372 if (query
->bo
&& drm_intel_bo_references(brw
->batch
.bo
, query
->bo
))
373 intel_batchbuffer_flush(brw
);
375 if (query
->bo
== NULL
|| !drm_intel_bo_busy(query
->bo
)) {
376 gen6_queryobj_get_results(ctx
, query
);
377 query
->Base
.Ready
= true;
381 /* Initialize Gen6+-specific query object functions. */
382 void gen6_init_queryobj_functions(struct dd_function_table
*functions
)
384 functions
->BeginQuery
= gen6_begin_query
;
385 functions
->EndQuery
= gen6_end_query
;
386 functions
->CheckQuery
= gen6_check_query
;
387 functions
->WaitQuery
= gen6_wait_query
;