2 * Copyright 2006 VMware, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
21 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
22 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
23 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 #include "intel_batchbuffer.h"
27 #include "intel_buffer_objects.h"
28 #include "intel_reg.h"
29 #include "intel_bufmgr.h"
30 #include "intel_buffers.h"
31 #include "intel_fbo.h"
32 #include "brw_context.h"
33 #include "brw_defines.h"
34 #include "brw_state.h"
40 intel_batchbuffer_reset(struct brw_context
*brw
);
43 intel_batchbuffer_init(struct brw_context
*brw
)
45 intel_batchbuffer_reset(brw
);
48 brw
->batch
.cpu_map
= malloc(BATCH_SZ
);
49 brw
->batch
.map
= brw
->batch
.cpu_map
;
50 brw
->batch
.map_next
= brw
->batch
.cpu_map
;
55 intel_batchbuffer_reset(struct brw_context
*brw
)
57 if (brw
->batch
.last_bo
!= NULL
) {
58 drm_intel_bo_unreference(brw
->batch
.last_bo
);
59 brw
->batch
.last_bo
= NULL
;
61 brw
->batch
.last_bo
= brw
->batch
.bo
;
63 brw_render_cache_set_clear(brw
);
65 brw
->batch
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "batchbuffer",
68 drm_intel_bo_map(brw
->batch
.bo
, true);
69 brw
->batch
.map
= brw
->batch
.bo
->virtual;
71 brw
->batch
.map_next
= brw
->batch
.map
;
73 brw
->batch
.reserved_space
= BATCH_RESERVED
;
74 brw
->batch
.state_batch_offset
= brw
->batch
.bo
->size
;
75 brw
->batch
.needs_sol_reset
= false;
77 /* We don't know what ring the new batch will be sent to until we see the
78 * first BEGIN_BATCH or BEGIN_BATCH_BLT. Mark it as unknown.
80 brw
->batch
.ring
= UNKNOWN_RING
;
84 intel_batchbuffer_save_state(struct brw_context
*brw
)
86 brw
->batch
.saved
.map_next
= brw
->batch
.map_next
;
87 brw
->batch
.saved
.reloc_count
=
88 drm_intel_gem_bo_get_reloc_count(brw
->batch
.bo
);
92 intel_batchbuffer_reset_to_saved(struct brw_context
*brw
)
94 drm_intel_gem_bo_clear_relocs(brw
->batch
.bo
, brw
->batch
.saved
.reloc_count
);
96 brw
->batch
.map_next
= brw
->batch
.saved
.map_next
;
97 if (USED_BATCH(brw
->batch
) == 0)
98 brw
->batch
.ring
= UNKNOWN_RING
;
102 intel_batchbuffer_free(struct brw_context
*brw
)
104 free(brw
->batch
.cpu_map
);
105 drm_intel_bo_unreference(brw
->batch
.last_bo
);
106 drm_intel_bo_unreference(brw
->batch
.bo
);
110 intel_batchbuffer_require_space(struct brw_context
*brw
, GLuint sz
,
111 enum brw_gpu_ring ring
)
113 /* If we're switching rings, implicitly flush the batch. */
114 if (unlikely(ring
!= brw
->batch
.ring
) && brw
->batch
.ring
!= UNKNOWN_RING
&&
116 intel_batchbuffer_flush(brw
);
120 assert(sz
< BATCH_SZ
- BATCH_RESERVED
);
122 if (intel_batchbuffer_space(brw
) < sz
)
123 intel_batchbuffer_flush(brw
);
125 enum brw_gpu_ring prev_ring
= brw
->batch
.ring
;
126 /* The intel_batchbuffer_flush() calls above might have changed
127 * brw->batch.ring to UNKNOWN_RING, so we need to set it here at the end.
129 brw
->batch
.ring
= ring
;
131 if (unlikely(prev_ring
== UNKNOWN_RING
&& ring
== RENDER_RING
))
132 intel_batchbuffer_emit_render_ring_prelude(brw
);
136 do_batch_dump(struct brw_context
*brw
)
138 struct drm_intel_decode
*decode
;
139 struct intel_batchbuffer
*batch
= &brw
->batch
;
142 decode
= drm_intel_decode_context_alloc(brw
->intelScreen
->deviceID
);
146 ret
= drm_intel_bo_map(batch
->bo
, false);
148 drm_intel_decode_set_batch_pointer(decode
,
154 "WARNING: failed to map batchbuffer (%s), "
155 "dumping uploaded data instead.\n", strerror(ret
));
157 drm_intel_decode_set_batch_pointer(decode
,
163 drm_intel_decode_set_output_file(decode
, stderr
);
164 drm_intel_decode(decode
);
166 drm_intel_decode_context_free(decode
);
169 drm_intel_bo_unmap(batch
->bo
);
171 brw_debug_batch(brw
);
176 intel_batchbuffer_emit_render_ring_prelude(struct brw_context
*brw
)
178 /* We may need to enable and snapshot OA counters. */
179 brw_perf_monitor_new_batch(brw
);
183 * Called when starting a new batch buffer.
186 brw_new_batch(struct brw_context
*brw
)
188 /* Create a new batchbuffer and reset the associated state: */
189 drm_intel_gem_bo_clear_relocs(brw
->batch
.bo
, 0);
190 intel_batchbuffer_reset(brw
);
192 /* If the kernel supports hardware contexts, then most hardware state is
193 * preserved between batches; we only need to re-emit state that is required
194 * to be in every batch. Otherwise we need to re-emit all the state that
195 * would otherwise be stored in the context (which for all intents and
196 * purposes means everything).
198 if (brw
->hw_ctx
== NULL
)
199 brw
->ctx
.NewDriverState
|= BRW_NEW_CONTEXT
;
201 brw
->ctx
.NewDriverState
|= BRW_NEW_BATCH
;
203 brw
->state_batch_count
= 0;
207 /* We need to periodically reap the shader time results, because rollover
208 * happens every few seconds. We also want to see results every once in a
209 * while, because many programs won't cleanly destroy our context, so the
210 * end-of-run printout may not happen.
212 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
213 brw_collect_and_report_shader_time(brw
);
215 if (INTEL_DEBUG
& DEBUG_PERFMON
)
216 brw_dump_perf_monitors(brw
);
220 * Called from intel_batchbuffer_flush before emitting MI_BATCHBUFFER_END and
223 * This function can emit state (say, to preserve registers that aren't saved
224 * between batches). All of this state MUST fit in the reserved space at the
225 * end of the batchbuffer. If you add more GPU state, increase the reserved
226 * space by updating the BATCH_RESERVED macro.
229 brw_finish_batch(struct brw_context
*brw
)
231 /* Capture the closing pipeline statistics register values necessary to
232 * support query objects (in the non-hardware context world).
234 brw_emit_query_end(brw
);
236 if (brw
->batch
.ring
== RENDER_RING
) {
237 /* Work around L3 state leaks into contexts set MI_RESTORE_INHIBIT which
238 * assume that the L3 cache is configured according to the hardware
242 gen7_restore_default_l3_config(brw
);
244 /* We may also need to snapshot and disable OA counters. */
245 brw_perf_monitor_finish_batch(brw
);
247 if (brw
->is_haswell
) {
248 /* From the Haswell PRM, Volume 2b, Command Reference: Instructions,
249 * 3DSTATE_CC_STATE_POINTERS > "Note":
251 * "SW must program 3DSTATE_CC_STATE_POINTERS command at the end of every
252 * 3D batch buffer followed by a PIPE_CONTROL with RC flush and CS stall."
254 * From the example in the docs, it seems to expect a regular pipe control
255 * flush here as well. We may have done it already, but meh.
257 * See also WaAvoidRCZCounterRollover.
259 brw_emit_mi_flush(brw
);
261 OUT_BATCH(_3DSTATE_CC_STATE_POINTERS
<< 16 | (2 - 2));
262 OUT_BATCH(brw
->cc
.state_offset
| 1);
264 brw_emit_pipe_control_flush(brw
, PIPE_CONTROL_RENDER_TARGET_FLUSH
|
265 PIPE_CONTROL_CS_STALL
);
269 /* Mark that the current program cache BO has been used by the GPU.
270 * It will be reallocated if we need to put new programs in for the
273 brw
->cache
.bo_used_by_gpu
= true;
277 throttle(struct brw_context
*brw
)
279 /* Wait for the swapbuffers before the one we just emitted, so we
280 * don't get too many swaps outstanding for apps that are GPU-heavy
283 * We're using intelDRI2Flush (called from the loader before
284 * swapbuffer) and glFlush (for front buffer rendering) as the
285 * indicator that a frame is done and then throttle when we get
286 * here as we prepare to render the next frame. At this point for
287 * round trips for swap/copy and getting new buffers are done and
288 * we'll spend less time waiting on the GPU.
290 * Unfortunately, we don't have a handle to the batch containing
291 * the swap, and getting our hands on that doesn't seem worth it,
292 * so we just use the first batch we emitted after the last swap.
294 if (brw
->need_swap_throttle
&& brw
->throttle_batch
[0]) {
295 if (brw
->throttle_batch
[1]) {
296 if (!brw
->disable_throttling
)
297 drm_intel_bo_wait_rendering(brw
->throttle_batch
[1]);
298 drm_intel_bo_unreference(brw
->throttle_batch
[1]);
300 brw
->throttle_batch
[1] = brw
->throttle_batch
[0];
301 brw
->throttle_batch
[0] = NULL
;
302 brw
->need_swap_throttle
= false;
303 /* Throttling here is more precise than the throttle ioctl, so skip it */
304 brw
->need_flush_throttle
= false;
307 if (brw
->need_flush_throttle
) {
308 __DRIscreen
*psp
= brw
->intelScreen
->driScrnPriv
;
309 drmCommandNone(psp
->fd
, DRM_I915_GEM_THROTTLE
);
310 brw
->need_flush_throttle
= false;
314 /* Drop when RS headers get pulled to libdrm */
315 #ifndef I915_EXEC_RESOURCE_STREAMER
316 #define I915_EXEC_RESOURCE_STREAMER (1<<15)
319 /* TODO: Push this whole function into bufmgr.
322 do_flush_locked(struct brw_context
*brw
)
324 struct intel_batchbuffer
*batch
= &brw
->batch
;
328 drm_intel_bo_unmap(batch
->bo
);
330 ret
= drm_intel_bo_subdata(batch
->bo
, 0, 4 * USED_BATCH(*batch
), batch
->map
);
331 if (ret
== 0 && batch
->state_batch_offset
!= batch
->bo
->size
) {
332 ret
= drm_intel_bo_subdata(batch
->bo
,
333 batch
->state_batch_offset
,
334 batch
->bo
->size
- batch
->state_batch_offset
,
335 (char *)batch
->map
+ batch
->state_batch_offset
);
339 if (!brw
->intelScreen
->no_hw
) {
342 if (brw
->gen
>= 6 && batch
->ring
== BLT_RING
) {
343 flags
= I915_EXEC_BLT
;
345 flags
= I915_EXEC_RENDER
|
346 (brw
->use_resource_streamer
? I915_EXEC_RESOURCE_STREAMER
: 0);
348 if (batch
->needs_sol_reset
)
349 flags
|= I915_EXEC_GEN7_SOL_RESET
;
352 if (unlikely(INTEL_DEBUG
& DEBUG_AUB
))
353 brw_annotate_aub(brw
);
355 if (brw
->hw_ctx
== NULL
|| batch
->ring
!= RENDER_RING
) {
356 ret
= drm_intel_bo_mrb_exec(batch
->bo
, 4 * USED_BATCH(*batch
),
359 ret
= drm_intel_gem_bo_context_exec(batch
->bo
, brw
->hw_ctx
,
360 4 * USED_BATCH(*batch
), flags
);
367 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
))
371 fprintf(stderr
, "intel_do_flush_locked failed: %s\n", strerror(-ret
));
379 _intel_batchbuffer_flush(struct brw_context
*brw
,
380 const char *file
, int line
)
384 if (USED_BATCH(brw
->batch
) == 0)
387 if (brw
->throttle_batch
[0] == NULL
) {
388 brw
->throttle_batch
[0] = brw
->batch
.bo
;
389 drm_intel_bo_reference(brw
->throttle_batch
[0]);
392 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
)) {
393 int bytes_for_commands
= 4 * USED_BATCH(brw
->batch
);
394 int bytes_for_state
= brw
->batch
.bo
->size
- brw
->batch
.state_batch_offset
;
395 int total_bytes
= bytes_for_commands
+ bytes_for_state
;
396 fprintf(stderr
, "%s:%d: Batchbuffer flush with %4db (pkt) + "
397 "%4db (state) = %4db (%0.1f%%)\n", file
, line
,
398 bytes_for_commands
, bytes_for_state
,
400 100.0f
* total_bytes
/ BATCH_SZ
);
403 brw
->batch
.reserved_space
= 0;
405 brw_finish_batch(brw
);
407 /* Mark the end of the buffer. */
408 intel_batchbuffer_emit_dword(brw
, MI_BATCH_BUFFER_END
);
409 if (USED_BATCH(brw
->batch
) & 1) {
410 /* Round batchbuffer usage to 2 DWORDs. */
411 intel_batchbuffer_emit_dword(brw
, MI_NOOP
);
414 intel_upload_finish(brw
);
416 /* Check that we didn't just wrap our batchbuffer at a bad time. */
417 assert(!brw
->no_batch_wrap
);
419 ret
= do_flush_locked(brw
);
421 if (unlikely(INTEL_DEBUG
& DEBUG_SYNC
)) {
422 fprintf(stderr
, "waiting for idle\n");
423 drm_intel_bo_wait_rendering(brw
->batch
.bo
);
426 if (brw
->use_resource_streamer
)
427 gen7_reset_hw_bt_pool_offsets(brw
);
429 /* Start a new batch buffer. */
436 /* This is the only way buffers get added to the validate list.
439 intel_batchbuffer_reloc(struct brw_context
*brw
,
440 drm_intel_bo
*buffer
, uint32_t offset
,
441 uint32_t read_domains
, uint32_t write_domain
,
446 ret
= drm_intel_bo_emit_reloc(brw
->batch
.bo
, offset
,
448 read_domains
, write_domain
);
452 /* Using the old buffer offset, write in what the right data would be, in
453 * case the buffer doesn't move and we can short-circuit the relocation
454 * processing in the kernel
456 return buffer
->offset64
+ delta
;
460 intel_batchbuffer_reloc64(struct brw_context
*brw
,
461 drm_intel_bo
*buffer
, uint32_t offset
,
462 uint32_t read_domains
, uint32_t write_domain
,
465 int ret
= drm_intel_bo_emit_reloc(brw
->batch
.bo
, offset
,
467 read_domains
, write_domain
);
471 /* Using the old buffer offset, write in what the right data would be, in
472 * case the buffer doesn't move and we can short-circuit the relocation
473 * processing in the kernel
475 return buffer
->offset64
+ delta
;
480 intel_batchbuffer_data(struct brw_context
*brw
,
481 const void *data
, GLuint bytes
, enum brw_gpu_ring ring
)
483 assert((bytes
& 3) == 0);
484 intel_batchbuffer_require_space(brw
, bytes
, ring
);
485 memcpy(brw
->batch
.map_next
, data
, bytes
);
486 brw
->batch
.map_next
+= bytes
>> 2;
490 load_sized_register_mem(struct brw_context
*brw
,
493 uint32_t read_domains
, uint32_t write_domain
,
499 /* MI_LOAD_REGISTER_MEM only exists on Gen7+. */
500 assert(brw
->gen
>= 7);
503 BEGIN_BATCH(4 * size
);
504 for (i
= 0; i
< size
; i
++) {
505 OUT_BATCH(GEN7_MI_LOAD_REGISTER_MEM
| (4 - 2));
506 OUT_BATCH(reg
+ i
* 4);
507 OUT_RELOC64(bo
, read_domains
, write_domain
, offset
+ i
* 4);
511 BEGIN_BATCH(3 * size
);
512 for (i
= 0; i
< size
; i
++) {
513 OUT_BATCH(GEN7_MI_LOAD_REGISTER_MEM
| (3 - 2));
514 OUT_BATCH(reg
+ i
* 4);
515 OUT_RELOC(bo
, read_domains
, write_domain
, offset
+ i
* 4);
522 brw_load_register_mem(struct brw_context
*brw
,
525 uint32_t read_domains
, uint32_t write_domain
,
528 load_sized_register_mem(brw
, reg
, bo
, read_domains
, write_domain
, offset
, 1);
532 brw_load_register_mem64(struct brw_context
*brw
,
535 uint32_t read_domains
, uint32_t write_domain
,
538 load_sized_register_mem(brw
, reg
, bo
, read_domains
, write_domain
, offset
, 2);
542 * Write an arbitrary 32-bit register to a buffer via MI_STORE_REGISTER_MEM.
545 brw_store_register_mem32(struct brw_context
*brw
,
546 drm_intel_bo
*bo
, uint32_t reg
, uint32_t offset
)
548 assert(brw
->gen
>= 6);
552 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
554 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
559 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
561 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
568 * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
571 brw_store_register_mem64(struct brw_context
*brw
,
572 drm_intel_bo
*bo
, uint32_t reg
, uint32_t offset
)
574 assert(brw
->gen
>= 6);
576 /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
577 * read a full 64-bit register, we need to do two of them.
581 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
583 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
585 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
586 OUT_BATCH(reg
+ sizeof(uint32_t));
587 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
588 offset
+ sizeof(uint32_t));
592 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
594 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
596 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
597 OUT_BATCH(reg
+ sizeof(uint32_t));
598 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
599 offset
+ sizeof(uint32_t));
605 * Write a 32-bit register using immediate data.
608 brw_load_register_imm32(struct brw_context
*brw
, uint32_t reg
, uint32_t imm
)
610 assert(brw
->gen
>= 6);
613 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (3 - 2));
620 * Write a 64-bit register using immediate data.
623 brw_load_register_imm64(struct brw_context
*brw
, uint32_t reg
, uint64_t imm
)
625 assert(brw
->gen
>= 6);
628 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (5 - 2));
630 OUT_BATCH(imm
& 0xffffffff);
632 OUT_BATCH(imm
>> 32);
637 * Copies a 32-bit register.
640 brw_load_register_reg(struct brw_context
*brw
, uint32_t src
, uint32_t dest
)
642 assert(brw
->gen
>= 8 || brw
->is_haswell
);
645 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
652 * Copies a 64-bit register.
655 brw_load_register_reg64(struct brw_context
*brw
, uint32_t src
, uint32_t dest
)
657 assert(brw
->gen
>= 8 || brw
->is_haswell
);
660 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
663 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
664 OUT_BATCH(src
+ sizeof(uint32_t));
665 OUT_BATCH(dest
+ sizeof(uint32_t));
670 * Write 32-bits of immediate data to a GPU memory buffer.
673 brw_store_data_imm32(struct brw_context
*brw
, drm_intel_bo
*bo
,
674 uint32_t offset
, uint32_t imm
)
676 assert(brw
->gen
>= 6);
679 OUT_BATCH(MI_STORE_DATA_IMM
| (4 - 2));
681 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
684 OUT_BATCH(0); /* MBZ */
685 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
693 * Write 64-bits of immediate data to a GPU memory buffer.
696 brw_store_data_imm64(struct brw_context
*brw
, drm_intel_bo
*bo
,
697 uint32_t offset
, uint64_t imm
)
699 assert(brw
->gen
>= 6);
702 OUT_BATCH(MI_STORE_DATA_IMM
| (5 - 2));
704 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
707 OUT_BATCH(0); /* MBZ */
708 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
711 OUT_BATCH(imm
& 0xffffffffu
);
712 OUT_BATCH(imm
>> 32);