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_bufmgr.h"
29 #include "intel_buffers.h"
30 #include "intel_fbo.h"
31 #include "brw_context.h"
32 #include "brw_defines.h"
33 #include "brw_state.h"
39 intel_batchbuffer_reset(struct brw_context
*brw
);
42 intel_batchbuffer_init(struct brw_context
*brw
)
44 intel_batchbuffer_reset(brw
);
47 brw
->batch
.cpu_map
= malloc(BATCH_SZ
);
48 brw
->batch
.map
= brw
->batch
.cpu_map
;
49 brw
->batch
.map_next
= brw
->batch
.cpu_map
;
54 intel_batchbuffer_reset(struct brw_context
*brw
)
56 if (brw
->batch
.last_bo
!= NULL
) {
57 drm_intel_bo_unreference(brw
->batch
.last_bo
);
58 brw
->batch
.last_bo
= NULL
;
60 brw
->batch
.last_bo
= brw
->batch
.bo
;
62 brw_render_cache_set_clear(brw
);
64 brw
->batch
.bo
= drm_intel_bo_alloc(brw
->bufmgr
, "batchbuffer",
67 drm_intel_bo_map(brw
->batch
.bo
, true);
68 brw
->batch
.map
= brw
->batch
.bo
->virtual;
70 brw
->batch
.map_next
= brw
->batch
.map
;
72 brw
->batch
.reserved_space
= BATCH_RESERVED
;
73 brw
->batch
.state_batch_offset
= brw
->batch
.bo
->size
;
74 brw
->batch
.needs_sol_reset
= false;
75 brw
->batch
.state_base_address_emitted
= 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
->screen
->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 /* Un-used currently */
182 * Called when starting a new batch buffer.
185 brw_new_batch(struct brw_context
*brw
)
187 /* Create a new batchbuffer and reset the associated state: */
188 drm_intel_gem_bo_clear_relocs(brw
->batch
.bo
, 0);
189 intel_batchbuffer_reset(brw
);
191 /* If the kernel supports hardware contexts, then most hardware state is
192 * preserved between batches; we only need to re-emit state that is required
193 * to be in every batch. Otherwise we need to re-emit all the state that
194 * would otherwise be stored in the context (which for all intents and
195 * purposes means everything).
197 if (brw
->hw_ctx
== NULL
)
198 brw
->ctx
.NewDriverState
|= BRW_NEW_CONTEXT
;
200 brw
->ctx
.NewDriverState
|= BRW_NEW_BATCH
;
202 brw
->state_batch_count
= 0;
206 /* We need to periodically reap the shader time results, because rollover
207 * happens every few seconds. We also want to see results every once in a
208 * while, because many programs won't cleanly destroy our context, so the
209 * end-of-run printout may not happen.
211 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
212 brw_collect_and_report_shader_time(brw
);
216 * Called from intel_batchbuffer_flush before emitting MI_BATCHBUFFER_END and
219 * This function can emit state (say, to preserve registers that aren't saved
220 * between batches). All of this state MUST fit in the reserved space at the
221 * end of the batchbuffer. If you add more GPU state, increase the reserved
222 * space by updating the BATCH_RESERVED macro.
225 brw_finish_batch(struct brw_context
*brw
)
227 /* Capture the closing pipeline statistics register values necessary to
228 * support query objects (in the non-hardware context world).
230 brw_emit_query_end(brw
);
232 if (brw
->batch
.ring
== RENDER_RING
) {
233 /* Work around L3 state leaks into contexts set MI_RESTORE_INHIBIT which
234 * assume that the L3 cache is configured according to the hardware
238 gen7_restore_default_l3_config(brw
);
240 if (brw
->is_haswell
) {
241 /* From the Haswell PRM, Volume 2b, Command Reference: Instructions,
242 * 3DSTATE_CC_STATE_POINTERS > "Note":
244 * "SW must program 3DSTATE_CC_STATE_POINTERS command at the end of every
245 * 3D batch buffer followed by a PIPE_CONTROL with RC flush and CS stall."
247 * From the example in the docs, it seems to expect a regular pipe control
248 * flush here as well. We may have done it already, but meh.
250 * See also WaAvoidRCZCounterRollover.
252 brw_emit_mi_flush(brw
);
254 OUT_BATCH(_3DSTATE_CC_STATE_POINTERS
<< 16 | (2 - 2));
255 OUT_BATCH(brw
->cc
.state_offset
| 1);
257 brw_emit_pipe_control_flush(brw
, PIPE_CONTROL_RENDER_TARGET_FLUSH
|
258 PIPE_CONTROL_CS_STALL
);
262 /* Mark that the current program cache BO has been used by the GPU.
263 * It will be reallocated if we need to put new programs in for the
266 brw
->cache
.bo_used_by_gpu
= true;
270 throttle(struct brw_context
*brw
)
272 /* Wait for the swapbuffers before the one we just emitted, so we
273 * don't get too many swaps outstanding for apps that are GPU-heavy
276 * We're using intelDRI2Flush (called from the loader before
277 * swapbuffer) and glFlush (for front buffer rendering) as the
278 * indicator that a frame is done and then throttle when we get
279 * here as we prepare to render the next frame. At this point for
280 * round trips for swap/copy and getting new buffers are done and
281 * we'll spend less time waiting on the GPU.
283 * Unfortunately, we don't have a handle to the batch containing
284 * the swap, and getting our hands on that doesn't seem worth it,
285 * so we just use the first batch we emitted after the last swap.
287 if (brw
->need_swap_throttle
&& brw
->throttle_batch
[0]) {
288 if (brw
->throttle_batch
[1]) {
289 if (!brw
->disable_throttling
)
290 drm_intel_bo_wait_rendering(brw
->throttle_batch
[1]);
291 drm_intel_bo_unreference(brw
->throttle_batch
[1]);
293 brw
->throttle_batch
[1] = brw
->throttle_batch
[0];
294 brw
->throttle_batch
[0] = NULL
;
295 brw
->need_swap_throttle
= false;
296 /* Throttling here is more precise than the throttle ioctl, so skip it */
297 brw
->need_flush_throttle
= false;
300 if (brw
->need_flush_throttle
) {
301 __DRIscreen
*dri_screen
= brw
->screen
->driScrnPriv
;
302 drmCommandNone(dri_screen
->fd
, DRM_I915_GEM_THROTTLE
);
303 brw
->need_flush_throttle
= false;
307 /* Drop when RS headers get pulled to libdrm */
308 #ifndef I915_EXEC_RESOURCE_STREAMER
309 #define I915_EXEC_RESOURCE_STREAMER (1<<15)
312 /* TODO: Push this whole function into bufmgr.
315 do_flush_locked(struct brw_context
*brw
)
317 struct intel_batchbuffer
*batch
= &brw
->batch
;
321 drm_intel_bo_unmap(batch
->bo
);
323 ret
= drm_intel_bo_subdata(batch
->bo
, 0, 4 * USED_BATCH(*batch
), batch
->map
);
324 if (ret
== 0 && batch
->state_batch_offset
!= batch
->bo
->size
) {
325 ret
= drm_intel_bo_subdata(batch
->bo
,
326 batch
->state_batch_offset
,
327 batch
->bo
->size
- batch
->state_batch_offset
,
328 (char *)batch
->map
+ batch
->state_batch_offset
);
332 if (!brw
->screen
->no_hw
) {
335 if (brw
->gen
>= 6 && batch
->ring
== BLT_RING
) {
336 flags
= I915_EXEC_BLT
;
338 flags
= I915_EXEC_RENDER
|
339 (brw
->use_resource_streamer
? I915_EXEC_RESOURCE_STREAMER
: 0);
341 if (batch
->needs_sol_reset
)
342 flags
|= I915_EXEC_GEN7_SOL_RESET
;
345 if (unlikely(INTEL_DEBUG
& DEBUG_AUB
))
346 brw_annotate_aub(brw
);
348 if (brw
->hw_ctx
== NULL
|| batch
->ring
!= RENDER_RING
) {
349 ret
= drm_intel_bo_mrb_exec(batch
->bo
, 4 * USED_BATCH(*batch
),
352 ret
= drm_intel_gem_bo_context_exec(batch
->bo
, brw
->hw_ctx
,
353 4 * USED_BATCH(*batch
), flags
);
360 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
))
363 if (brw
->ctx
.Const
.ResetStrategy
== GL_LOSE_CONTEXT_ON_RESET_ARB
)
364 brw_check_for_reset(brw
);
367 fprintf(stderr
, "intel_do_flush_locked failed: %s\n", strerror(-ret
));
375 _intel_batchbuffer_flush(struct brw_context
*brw
,
376 const char *file
, int line
)
380 if (USED_BATCH(brw
->batch
) == 0)
383 if (brw
->throttle_batch
[0] == NULL
) {
384 brw
->throttle_batch
[0] = brw
->batch
.bo
;
385 drm_intel_bo_reference(brw
->throttle_batch
[0]);
388 if (unlikely(INTEL_DEBUG
& DEBUG_BATCH
)) {
389 int bytes_for_commands
= 4 * USED_BATCH(brw
->batch
);
390 int bytes_for_state
= brw
->batch
.bo
->size
- brw
->batch
.state_batch_offset
;
391 int total_bytes
= bytes_for_commands
+ bytes_for_state
;
392 fprintf(stderr
, "%s:%d: Batchbuffer flush with %4db (pkt) + "
393 "%4db (state) = %4db (%0.1f%%)\n", file
, line
,
394 bytes_for_commands
, bytes_for_state
,
396 100.0f
* total_bytes
/ BATCH_SZ
);
399 brw
->batch
.reserved_space
= 0;
401 brw_finish_batch(brw
);
403 /* Mark the end of the buffer. */
404 intel_batchbuffer_emit_dword(brw
, MI_BATCH_BUFFER_END
);
405 if (USED_BATCH(brw
->batch
) & 1) {
406 /* Round batchbuffer usage to 2 DWORDs. */
407 intel_batchbuffer_emit_dword(brw
, MI_NOOP
);
410 intel_upload_finish(brw
);
412 /* Check that we didn't just wrap our batchbuffer at a bad time. */
413 assert(!brw
->no_batch_wrap
);
415 ret
= do_flush_locked(brw
);
417 if (unlikely(INTEL_DEBUG
& DEBUG_SYNC
)) {
418 fprintf(stderr
, "waiting for idle\n");
419 drm_intel_bo_wait_rendering(brw
->batch
.bo
);
422 if (brw
->use_resource_streamer
)
423 gen7_reset_hw_bt_pool_offsets(brw
);
425 /* Start a new batch buffer. */
432 /* This is the only way buffers get added to the validate list.
435 intel_batchbuffer_reloc(struct intel_batchbuffer
*batch
,
436 drm_intel_bo
*buffer
, uint32_t offset
,
437 uint32_t read_domains
, uint32_t write_domain
,
442 ret
= drm_intel_bo_emit_reloc(batch
->bo
, offset
,
444 read_domains
, write_domain
);
448 /* Using the old buffer offset, write in what the right data would be, in
449 * case the buffer doesn't move and we can short-circuit the relocation
450 * processing in the kernel
452 return buffer
->offset64
+ delta
;
456 intel_batchbuffer_reloc64(struct intel_batchbuffer
*batch
,
457 drm_intel_bo
*buffer
, uint32_t offset
,
458 uint32_t read_domains
, uint32_t write_domain
,
461 int ret
= drm_intel_bo_emit_reloc(batch
->bo
, offset
,
463 read_domains
, write_domain
);
467 /* Using the old buffer offset, write in what the right data would be, in
468 * case the buffer doesn't move and we can short-circuit the relocation
469 * processing in the kernel
471 return buffer
->offset64
+ delta
;
476 intel_batchbuffer_data(struct brw_context
*brw
,
477 const void *data
, GLuint bytes
, enum brw_gpu_ring ring
)
479 assert((bytes
& 3) == 0);
480 intel_batchbuffer_require_space(brw
, bytes
, ring
);
481 memcpy(brw
->batch
.map_next
, data
, bytes
);
482 brw
->batch
.map_next
+= bytes
>> 2;
486 load_sized_register_mem(struct brw_context
*brw
,
489 uint32_t read_domains
, uint32_t write_domain
,
495 /* MI_LOAD_REGISTER_MEM only exists on Gen7+. */
496 assert(brw
->gen
>= 7);
499 BEGIN_BATCH(4 * size
);
500 for (i
= 0; i
< size
; i
++) {
501 OUT_BATCH(GEN7_MI_LOAD_REGISTER_MEM
| (4 - 2));
502 OUT_BATCH(reg
+ i
* 4);
503 OUT_RELOC64(bo
, read_domains
, write_domain
, offset
+ i
* 4);
507 BEGIN_BATCH(3 * size
);
508 for (i
= 0; i
< size
; i
++) {
509 OUT_BATCH(GEN7_MI_LOAD_REGISTER_MEM
| (3 - 2));
510 OUT_BATCH(reg
+ i
* 4);
511 OUT_RELOC(bo
, read_domains
, write_domain
, offset
+ i
* 4);
518 brw_load_register_mem(struct brw_context
*brw
,
521 uint32_t read_domains
, uint32_t write_domain
,
524 load_sized_register_mem(brw
, reg
, bo
, read_domains
, write_domain
, offset
, 1);
528 brw_load_register_mem64(struct brw_context
*brw
,
531 uint32_t read_domains
, uint32_t write_domain
,
534 load_sized_register_mem(brw
, reg
, bo
, read_domains
, write_domain
, offset
, 2);
538 * Write an arbitrary 32-bit register to a buffer via MI_STORE_REGISTER_MEM.
541 brw_store_register_mem32(struct brw_context
*brw
,
542 drm_intel_bo
*bo
, uint32_t reg
, uint32_t offset
)
544 assert(brw
->gen
>= 6);
548 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
550 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
555 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
557 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
564 * Write an arbitrary 64-bit register to a buffer via MI_STORE_REGISTER_MEM.
567 brw_store_register_mem64(struct brw_context
*brw
,
568 drm_intel_bo
*bo
, uint32_t reg
, uint32_t offset
)
570 assert(brw
->gen
>= 6);
572 /* MI_STORE_REGISTER_MEM only stores a single 32-bit value, so to
573 * read a full 64-bit register, we need to do two of them.
577 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
579 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
581 OUT_BATCH(MI_STORE_REGISTER_MEM
| (4 - 2));
582 OUT_BATCH(reg
+ sizeof(uint32_t));
583 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
584 offset
+ sizeof(uint32_t));
588 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
590 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
592 OUT_BATCH(MI_STORE_REGISTER_MEM
| (3 - 2));
593 OUT_BATCH(reg
+ sizeof(uint32_t));
594 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
595 offset
+ sizeof(uint32_t));
601 * Write a 32-bit register using immediate data.
604 brw_load_register_imm32(struct brw_context
*brw
, uint32_t reg
, uint32_t imm
)
606 assert(brw
->gen
>= 6);
609 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (3 - 2));
616 * Write a 64-bit register using immediate data.
619 brw_load_register_imm64(struct brw_context
*brw
, uint32_t reg
, uint64_t imm
)
621 assert(brw
->gen
>= 6);
624 OUT_BATCH(MI_LOAD_REGISTER_IMM
| (5 - 2));
626 OUT_BATCH(imm
& 0xffffffff);
628 OUT_BATCH(imm
>> 32);
633 * Copies a 32-bit register.
636 brw_load_register_reg(struct brw_context
*brw
, uint32_t src
, uint32_t dest
)
638 assert(brw
->gen
>= 8 || brw
->is_haswell
);
641 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
648 * Copies a 64-bit register.
651 brw_load_register_reg64(struct brw_context
*brw
, uint32_t src
, uint32_t dest
)
653 assert(brw
->gen
>= 8 || brw
->is_haswell
);
656 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
659 OUT_BATCH(MI_LOAD_REGISTER_REG
| (3 - 2));
660 OUT_BATCH(src
+ sizeof(uint32_t));
661 OUT_BATCH(dest
+ sizeof(uint32_t));
666 * Write 32-bits of immediate data to a GPU memory buffer.
669 brw_store_data_imm32(struct brw_context
*brw
, drm_intel_bo
*bo
,
670 uint32_t offset
, uint32_t imm
)
672 assert(brw
->gen
>= 6);
675 OUT_BATCH(MI_STORE_DATA_IMM
| (4 - 2));
677 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
680 OUT_BATCH(0); /* MBZ */
681 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
689 * Write 64-bits of immediate data to a GPU memory buffer.
692 brw_store_data_imm64(struct brw_context
*brw
, drm_intel_bo
*bo
,
693 uint32_t offset
, uint64_t imm
)
695 assert(brw
->gen
>= 6);
698 OUT_BATCH(MI_STORE_DATA_IMM
| (5 - 2));
700 OUT_RELOC64(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
703 OUT_BATCH(0); /* MBZ */
704 OUT_RELOC(bo
, I915_GEM_DOMAIN_INSTRUCTION
, I915_GEM_DOMAIN_INSTRUCTION
,
707 OUT_BATCH(imm
& 0xffffffffu
);
708 OUT_BATCH(imm
>> 32);