2 * Copyright 2013 Advanced Micro Devices, Inc.
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #include "r600_pipe_common.h"
26 #include "util/u_memory.h"
27 #include "util/u_upload_mgr.h"
28 #include "radeon/radeon_video.h"
37 * \param event EVENT_TYPE_*
38 * \param event_flags Optional cache flush flags (TC)
39 * \param data_sel 1 = fence, 3 = timestamp
41 * \param va GPU address
42 * \param old_value Previous fence value (for a bug workaround)
43 * \param new_value Fence value to write for this event.
45 void si_gfx_write_event_eop(struct r600_common_context
*ctx
,
46 unsigned event
, unsigned event_flags
,
48 struct r600_resource
*buf
, uint64_t va
,
49 uint32_t new_fence
, unsigned query_type
)
51 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
52 unsigned op
= EVENT_TYPE(event
) |
55 unsigned sel
= EOP_DATA_SEL(data_sel
);
57 /* Wait for write confirmation before writing data, but don't send
59 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
60 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
62 if (ctx
->chip_class
>= GFX9
) {
63 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
64 * counters) must immediately precede every timestamp event to
65 * prevent a GPU hang on GFX9.
67 * Occlusion queries don't need to do it here, because they
68 * always do ZPASS_DONE before the timestamp.
70 if (ctx
->chip_class
== GFX9
&&
71 query_type
!= PIPE_QUERY_OCCLUSION_COUNTER
&&
72 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE
&&
73 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
) {
74 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
76 assert(16 * ctx
->screen
->info
.num_render_backends
<=
78 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
79 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
80 radeon_emit(cs
, scratch
->gpu_address
);
81 radeon_emit(cs
, scratch
->gpu_address
>> 32);
83 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
84 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
87 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, 6, 0));
90 radeon_emit(cs
, va
); /* address lo */
91 radeon_emit(cs
, va
>> 32); /* address hi */
92 radeon_emit(cs
, new_fence
); /* immediate data lo */
93 radeon_emit(cs
, 0); /* immediate data hi */
94 radeon_emit(cs
, 0); /* unused */
96 if (ctx
->chip_class
== CIK
||
97 ctx
->chip_class
== VI
) {
98 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
99 uint64_t va
= scratch
->gpu_address
;
101 /* Two EOP events are required to make all engines go idle
102 * (and optional cache flushes executed) before the timestamp
105 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
108 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
109 radeon_emit(cs
, 0); /* immediate data */
110 radeon_emit(cs
, 0); /* unused */
112 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
113 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
116 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
119 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
120 radeon_emit(cs
, new_fence
); /* immediate data */
121 radeon_emit(cs
, 0); /* unused */
125 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, buf
, RADEON_USAGE_WRITE
,
130 unsigned si_gfx_write_fence_dwords(struct si_screen
*screen
)
134 if (screen
->info
.chip_class
== CIK
||
135 screen
->info
.chip_class
== VI
)
141 void si_gfx_wait_fence(struct r600_common_context
*ctx
,
142 uint64_t va
, uint32_t ref
, uint32_t mask
)
144 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
146 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, 0));
147 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
149 radeon_emit(cs
, va
>> 32);
150 radeon_emit(cs
, ref
); /* reference value */
151 radeon_emit(cs
, mask
); /* mask */
152 radeon_emit(cs
, 4); /* poll interval */
155 static void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
157 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
159 /* NOP waits for idle on Evergreen and later. */
160 if (rctx
->chip_class
>= CIK
)
161 radeon_emit(cs
, 0x00000000); /* NOP */
163 radeon_emit(cs
, 0xf0000000); /* NOP */
166 void si_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
,
167 struct r600_resource
*dst
, struct r600_resource
*src
)
169 uint64_t vram
= ctx
->dma
.cs
->used_vram
;
170 uint64_t gtt
= ctx
->dma
.cs
->used_gart
;
173 vram
+= dst
->vram_usage
;
174 gtt
+= dst
->gart_usage
;
177 vram
+= src
->vram_usage
;
178 gtt
+= src
->gart_usage
;
181 /* Flush the GFX IB if DMA depends on it. */
182 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
184 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
185 RADEON_USAGE_READWRITE
)) ||
187 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
188 RADEON_USAGE_WRITE
))))
189 ctx
->gfx
.flush(ctx
, PIPE_FLUSH_ASYNC
, NULL
);
191 /* Flush if there's not enough space, or if the memory usage per IB
194 * IBs using too little memory are limited by the IB submission overhead.
195 * IBs using too much memory are limited by the kernel/TTM overhead.
196 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
198 * This heuristic makes sure that DMA requests are executed
199 * very soon after the call is made and lowers memory usage.
200 * It improves texture upload performance by keeping the DMA
201 * engine busy while uploads are being submitted.
203 num_dw
++; /* for emit_wait_idle below */
204 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
205 ctx
->dma
.cs
->used_vram
+ ctx
->dma
.cs
->used_gart
> 64 * 1024 * 1024 ||
206 !radeon_cs_memory_below_limit(ctx
->screen
, ctx
->dma
.cs
, vram
, gtt
)) {
207 ctx
->dma
.flush(ctx
, PIPE_FLUSH_ASYNC
, NULL
);
208 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
211 /* Wait for idle if either buffer has been used in the IB before to
212 * prevent read-after-write hazards.
215 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, dst
->buf
,
216 RADEON_USAGE_READWRITE
)) ||
218 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, src
->buf
,
219 RADEON_USAGE_WRITE
)))
220 r600_dma_emit_wait_idle(ctx
);
223 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
225 RADEON_PRIO_SDMA_BUFFER
);
228 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
230 RADEON_PRIO_SDMA_BUFFER
);
233 /* this function is called before all DMA calls, so increment this. */
234 ctx
->num_dma_calls
++;
237 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
238 struct pipe_fence_handle
**fence
)
240 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
241 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
242 struct radeon_saved_cs saved
;
244 (rctx
->screen
->debug_flags
& DBG(CHECK_VM
)) &&
245 rctx
->check_vm_faults
;
247 if (!radeon_emitted(cs
, 0)) {
249 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
254 si_save_cs(rctx
->ws
, cs
, &saved
, true);
256 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
258 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
261 /* Use conservative timeout 800ms, after which we won't wait any
262 * longer and assume the GPU is hung.
264 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
266 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
267 si_clear_saved_cs(&saved
);
272 * Store a linearized copy of all chunks of \p cs together with the buffer
275 void si_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
276 struct radeon_saved_cs
*saved
, bool get_buffer_list
)
281 /* Save the IB chunks. */
282 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
283 saved
->ib
= MALLOC(4 * saved
->num_dw
);
288 for (i
= 0; i
< cs
->num_prev
; ++i
) {
289 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
290 buf
+= cs
->prev
[i
].cdw
;
292 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
294 if (!get_buffer_list
)
297 /* Save the buffer list. */
298 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
299 saved
->bo_list
= CALLOC(saved
->bo_count
,
300 sizeof(saved
->bo_list
[0]));
301 if (!saved
->bo_list
) {
305 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
310 fprintf(stderr
, "%s: out of memory\n", __func__
);
311 memset(saved
, 0, sizeof(*saved
));
314 void si_clear_saved_cs(struct radeon_saved_cs
*saved
)
317 FREE(saved
->bo_list
);
319 memset(saved
, 0, sizeof(*saved
));
322 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
324 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
325 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
326 RADEON_GPU_RESET_COUNTER
);
328 if (rctx
->gpu_reset_counter
== latest
)
329 return PIPE_NO_RESET
;
331 rctx
->gpu_reset_counter
= latest
;
332 return PIPE_UNKNOWN_CONTEXT_RESET
;
335 static void r600_set_device_reset_callback(struct pipe_context
*ctx
,
336 const struct pipe_device_reset_callback
*cb
)
338 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
341 rctx
->device_reset_callback
= *cb
;
343 memset(&rctx
->device_reset_callback
, 0,
344 sizeof(rctx
->device_reset_callback
));
347 bool si_check_device_reset(struct r600_common_context
*rctx
)
349 enum pipe_reset_status status
;
351 if (!rctx
->device_reset_callback
.reset
)
354 if (!rctx
->b
.get_device_reset_status
)
357 status
= rctx
->b
.get_device_reset_status(&rctx
->b
);
358 if (status
== PIPE_NO_RESET
)
361 rctx
->device_reset_callback
.reset(rctx
->device_reset_callback
.data
, status
);
365 static bool r600_resource_commit(struct pipe_context
*pctx
,
366 struct pipe_resource
*resource
,
367 unsigned level
, struct pipe_box
*box
,
370 struct r600_common_context
*ctx
= (struct r600_common_context
*)pctx
;
371 struct r600_resource
*res
= r600_resource(resource
);
374 * Since buffer commitment changes cannot be pipelined, we need to
375 * (a) flush any pending commands that refer to the buffer we're about
377 * (b) wait for threaded submit to finish, including those that were
378 * triggered by some other, earlier operation.
380 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
381 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
,
382 res
->buf
, RADEON_USAGE_READWRITE
)) {
383 ctx
->gfx
.flush(ctx
, PIPE_FLUSH_ASYNC
, NULL
);
385 if (radeon_emitted(ctx
->dma
.cs
, 0) &&
386 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
,
387 res
->buf
, RADEON_USAGE_READWRITE
)) {
388 ctx
->dma
.flush(ctx
, PIPE_FLUSH_ASYNC
, NULL
);
391 ctx
->ws
->cs_sync_flush(ctx
->dma
.cs
);
392 ctx
->ws
->cs_sync_flush(ctx
->gfx
.cs
);
394 assert(resource
->target
== PIPE_BUFFER
);
396 return ctx
->ws
->buffer_commit(res
->buf
, box
->x
, box
->width
, commit
);
399 bool si_common_context_init(struct r600_common_context
*rctx
,
400 struct si_screen
*sscreen
,
401 unsigned context_flags
)
403 slab_create_child(&rctx
->pool_transfers
, &sscreen
->pool_transfers
);
404 slab_create_child(&rctx
->pool_transfers_unsync
, &sscreen
->pool_transfers
);
406 rctx
->screen
= sscreen
;
407 rctx
->ws
= sscreen
->ws
;
408 rctx
->family
= sscreen
->info
.family
;
409 rctx
->chip_class
= sscreen
->info
.chip_class
;
411 rctx
->b
.resource_commit
= r600_resource_commit
;
413 if (sscreen
->info
.drm_major
== 2 && sscreen
->info
.drm_minor
>= 43) {
414 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
415 rctx
->gpu_reset_counter
=
416 rctx
->ws
->query_value(rctx
->ws
,
417 RADEON_GPU_RESET_COUNTER
);
420 rctx
->b
.set_device_reset_callback
= r600_set_device_reset_callback
;
422 si_init_context_texture_functions(rctx
);
423 si_init_query_functions(rctx
);
425 if (rctx
->chip_class
== CIK
||
426 rctx
->chip_class
== VI
||
427 rctx
->chip_class
== GFX9
) {
428 rctx
->eop_bug_scratch
= (struct r600_resource
*)
429 pipe_buffer_create(&sscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
430 16 * sscreen
->info
.num_render_backends
);
431 if (!rctx
->eop_bug_scratch
)
435 rctx
->allocator_zeroed_memory
=
436 u_suballocator_create(&rctx
->b
, sscreen
->info
.gart_page_size
,
437 0, PIPE_USAGE_DEFAULT
, 0, true);
438 if (!rctx
->allocator_zeroed_memory
)
441 rctx
->b
.stream_uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
442 0, PIPE_USAGE_STREAM
,
443 R600_RESOURCE_FLAG_READ_ONLY
);
444 if (!rctx
->b
.stream_uploader
)
447 rctx
->b
.const_uploader
= u_upload_create(&rctx
->b
, 128 * 1024,
448 0, PIPE_USAGE_DEFAULT
,
449 R600_RESOURCE_FLAG_32BIT
|
450 (sscreen
->cpdma_prefetch_writes_memory
?
451 0 : R600_RESOURCE_FLAG_READ_ONLY
));
452 if (!rctx
->b
.const_uploader
)
455 rctx
->cached_gtt_allocator
= u_upload_create(&rctx
->b
, 16 * 1024,
456 0, PIPE_USAGE_STAGING
, 0);
457 if (!rctx
->cached_gtt_allocator
)
460 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
464 if (sscreen
->info
.num_sdma_rings
&& !(sscreen
->debug_flags
& DBG(NO_ASYNC_DMA
))) {
465 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
468 rctx
->dma
.flush
= r600_flush_dma_ring
;
474 void si_common_context_cleanup(struct r600_common_context
*rctx
)
478 /* Release DCC stats. */
479 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
480 assert(!rctx
->dcc_stats
[i
].query_active
);
482 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
483 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
484 rctx
->b
.destroy_query(&rctx
->b
,
485 rctx
->dcc_stats
[i
].ps_stats
[j
]);
487 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
490 if (rctx
->query_result_shader
)
491 rctx
->b
.delete_compute_state(&rctx
->b
, rctx
->query_result_shader
);
494 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
496 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
498 rctx
->ws
->ctx_destroy(rctx
->ctx
);
500 if (rctx
->b
.stream_uploader
)
501 u_upload_destroy(rctx
->b
.stream_uploader
);
502 if (rctx
->b
.const_uploader
)
503 u_upload_destroy(rctx
->b
.const_uploader
);
504 if (rctx
->cached_gtt_allocator
)
505 u_upload_destroy(rctx
->cached_gtt_allocator
);
507 slab_destroy_child(&rctx
->pool_transfers
);
508 slab_destroy_child(&rctx
->pool_transfers_unsync
);
510 if (rctx
->allocator_zeroed_memory
) {
511 u_suballocator_destroy(rctx
->allocator_zeroed_memory
);
513 rctx
->ws
->fence_reference(&rctx
->last_gfx_fence
, NULL
);
514 rctx
->ws
->fence_reference(&rctx
->last_sdma_fence
, NULL
);
515 r600_resource_reference(&rctx
->eop_bug_scratch
, NULL
);
519 void si_screen_clear_buffer(struct si_screen
*sscreen
, struct pipe_resource
*dst
,
520 uint64_t offset
, uint64_t size
, unsigned value
)
522 struct r600_common_context
*rctx
= (struct r600_common_context
*)sscreen
->aux_context
;
524 mtx_lock(&sscreen
->aux_context_lock
);
525 rctx
->dma_clear_buffer(&rctx
->b
, dst
, offset
, size
, value
);
526 sscreen
->aux_context
->flush(sscreen
->aux_context
, NULL
, 0);
527 mtx_unlock(&sscreen
->aux_context_lock
);