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
23 * Authors: Marek Olšák <maraeo@gmail.com>
27 #include "r600_pipe_common.h"
29 #include "tgsi/tgsi_parse.h"
30 #include "util/list.h"
31 #include "util/u_draw_quad.h"
32 #include "util/u_memory.h"
33 #include "util/u_format_s3tc.h"
34 #include "util/u_upload_mgr.h"
35 #include "os/os_time.h"
36 #include "vl/vl_decoder.h"
37 #include "vl/vl_video_buffer.h"
38 #include "radeon/radeon_video.h"
40 #include <sys/utsname.h>
42 #include <llvm-c/TargetMachine.h>
45 struct r600_multi_fence
{
46 struct pipe_reference reference
;
47 struct pipe_fence_handle
*gfx
;
48 struct pipe_fence_handle
*sdma
;
50 /* If the context wasn't flushed at fence creation, this is non-NULL. */
52 struct r600_common_context
*ctx
;
58 * shader binary helpers.
60 void si_radeon_shader_binary_init(struct ac_shader_binary
*b
)
62 memset(b
, 0, sizeof(*b
));
65 void si_radeon_shader_binary_clean(struct ac_shader_binary
*b
)
72 FREE(b
->global_symbol_offsets
);
74 FREE(b
->disasm_string
);
75 FREE(b
->llvm_ir_string
);
85 * \param event EVENT_TYPE_*
86 * \param event_flags Optional cache flush flags (TC)
87 * \param data_sel 1 = fence, 3 = timestamp
89 * \param va GPU address
90 * \param old_value Previous fence value (for a bug workaround)
91 * \param new_value Fence value to write for this event.
93 void si_gfx_write_event_eop(struct r600_common_context
*ctx
,
94 unsigned event
, unsigned event_flags
,
96 struct r600_resource
*buf
, uint64_t va
,
97 uint32_t new_fence
, unsigned query_type
)
99 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
100 unsigned op
= EVENT_TYPE(event
) |
103 unsigned sel
= EOP_DATA_SEL(data_sel
);
105 /* Wait for write confirmation before writing data, but don't send
107 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
108 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
110 if (ctx
->chip_class
>= GFX9
) {
111 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
112 * counters) must immediately precede every timestamp event to
113 * prevent a GPU hang on GFX9.
115 * Occlusion queries don't need to do it here, because they
116 * always do ZPASS_DONE before the timestamp.
118 if (ctx
->chip_class
== GFX9
&&
119 query_type
!= PIPE_QUERY_OCCLUSION_COUNTER
&&
120 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE
&&
121 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
) {
122 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
124 assert(16 * ctx
->screen
->info
.num_render_backends
<=
125 scratch
->b
.b
.width0
);
126 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
127 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
128 radeon_emit(cs
, scratch
->gpu_address
);
129 radeon_emit(cs
, scratch
->gpu_address
>> 32);
131 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
132 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
135 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, 6, 0));
137 radeon_emit(cs
, sel
);
138 radeon_emit(cs
, va
); /* address lo */
139 radeon_emit(cs
, va
>> 32); /* address hi */
140 radeon_emit(cs
, new_fence
); /* immediate data lo */
141 radeon_emit(cs
, 0); /* immediate data hi */
142 radeon_emit(cs
, 0); /* unused */
144 if (ctx
->chip_class
== CIK
||
145 ctx
->chip_class
== VI
) {
146 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
147 uint64_t va
= scratch
->gpu_address
;
149 /* Two EOP events are required to make all engines go idle
150 * (and optional cache flushes executed) before the timestamp
153 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
156 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
157 radeon_emit(cs
, 0); /* immediate data */
158 radeon_emit(cs
, 0); /* unused */
160 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
161 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
164 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
167 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
168 radeon_emit(cs
, new_fence
); /* immediate data */
169 radeon_emit(cs
, 0); /* unused */
173 r600_emit_reloc(ctx
, &ctx
->gfx
, buf
, RADEON_USAGE_WRITE
,
177 unsigned si_gfx_write_fence_dwords(struct r600_common_screen
*screen
)
181 if (screen
->chip_class
== CIK
||
182 screen
->chip_class
== VI
)
185 if (!screen
->info
.has_virtual_memory
)
191 void si_gfx_wait_fence(struct r600_common_context
*ctx
,
192 uint64_t va
, uint32_t ref
, uint32_t mask
)
194 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
196 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, 0));
197 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
199 radeon_emit(cs
, va
>> 32);
200 radeon_emit(cs
, ref
); /* reference value */
201 radeon_emit(cs
, mask
); /* mask */
202 radeon_emit(cs
, 4); /* poll interval */
205 static void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
207 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
209 /* NOP waits for idle on Evergreen and later. */
210 if (rctx
->chip_class
>= CIK
)
211 radeon_emit(cs
, 0x00000000); /* NOP */
213 radeon_emit(cs
, 0xf0000000); /* NOP */
216 void si_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
,
217 struct r600_resource
*dst
, struct r600_resource
*src
)
219 uint64_t vram
= ctx
->dma
.cs
->used_vram
;
220 uint64_t gtt
= ctx
->dma
.cs
->used_gart
;
223 vram
+= dst
->vram_usage
;
224 gtt
+= dst
->gart_usage
;
227 vram
+= src
->vram_usage
;
228 gtt
+= src
->gart_usage
;
231 /* Flush the GFX IB if DMA depends on it. */
232 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
234 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
235 RADEON_USAGE_READWRITE
)) ||
237 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
238 RADEON_USAGE_WRITE
))))
239 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
241 /* Flush if there's not enough space, or if the memory usage per IB
244 * IBs using too little memory are limited by the IB submission overhead.
245 * IBs using too much memory are limited by the kernel/TTM overhead.
246 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
248 * This heuristic makes sure that DMA requests are executed
249 * very soon after the call is made and lowers memory usage.
250 * It improves texture upload performance by keeping the DMA
251 * engine busy while uploads are being submitted.
253 num_dw
++; /* for emit_wait_idle below */
254 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
255 ctx
->dma
.cs
->used_vram
+ ctx
->dma
.cs
->used_gart
> 64 * 1024 * 1024 ||
256 !radeon_cs_memory_below_limit(ctx
->screen
, ctx
->dma
.cs
, vram
, gtt
)) {
257 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
258 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
261 /* Wait for idle if either buffer has been used in the IB before to
262 * prevent read-after-write hazards.
265 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, dst
->buf
,
266 RADEON_USAGE_READWRITE
)) ||
268 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, src
->buf
,
269 RADEON_USAGE_WRITE
)))
270 r600_dma_emit_wait_idle(ctx
);
272 /* If GPUVM is not supported, the CS checker needs 2 entries
273 * in the buffer list per packet, which has to be done manually.
275 if (ctx
->screen
->info
.has_virtual_memory
) {
277 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
279 RADEON_PRIO_SDMA_BUFFER
);
281 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
283 RADEON_PRIO_SDMA_BUFFER
);
286 /* this function is called before all DMA calls, so increment this. */
287 ctx
->num_dma_calls
++;
290 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
294 void si_preflush_suspend_features(struct r600_common_context
*ctx
)
296 /* suspend queries */
297 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
298 si_suspend_queries(ctx
);
300 ctx
->streamout
.suspended
= false;
301 if (ctx
->streamout
.begin_emitted
) {
302 si_emit_streamout_end(ctx
);
303 ctx
->streamout
.suspended
= true;
307 void si_postflush_resume_features(struct r600_common_context
*ctx
)
309 if (ctx
->streamout
.suspended
) {
310 ctx
->streamout
.append_bitmask
= ctx
->streamout
.enabled_mask
;
311 si_streamout_buffers_dirty(ctx
);
315 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
316 si_resume_queries(ctx
);
319 static void r600_add_fence_dependency(struct r600_common_context
*rctx
,
320 struct pipe_fence_handle
*fence
)
322 struct radeon_winsys
*ws
= rctx
->ws
;
325 ws
->cs_add_fence_dependency(rctx
->dma
.cs
, fence
);
326 ws
->cs_add_fence_dependency(rctx
->gfx
.cs
, fence
);
329 static void r600_fence_server_sync(struct pipe_context
*ctx
,
330 struct pipe_fence_handle
*fence
)
332 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
333 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
335 /* Only amdgpu needs to handle fence dependencies (for fence imports).
336 * radeon synchronizes all rings by default and will not implement
339 if (rctx
->screen
->info
.drm_major
== 2)
342 /* Only imported fences need to be handled by fence_server_sync,
343 * because the winsys handles synchronizations automatically for BOs
344 * within the process.
346 * Simply skip unflushed fences here, and the winsys will drop no-op
347 * dependencies (i.e. dependencies within the same ring).
349 if (rfence
->gfx_unflushed
.ctx
)
352 /* All unflushed commands will not start execution before
353 * this fence dependency is signalled.
355 * Should we flush the context to allow more GPU parallelism?
358 r600_add_fence_dependency(rctx
, rfence
->sdma
);
360 r600_add_fence_dependency(rctx
, rfence
->gfx
);
363 static void r600_flush_from_st(struct pipe_context
*ctx
,
364 struct pipe_fence_handle
**fence
,
367 struct pipe_screen
*screen
= ctx
->screen
;
368 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
369 struct radeon_winsys
*ws
= rctx
->ws
;
370 struct pipe_fence_handle
*gfx_fence
= NULL
;
371 struct pipe_fence_handle
*sdma_fence
= NULL
;
372 bool deferred_fence
= false;
373 unsigned rflags
= RADEON_FLUSH_ASYNC
;
375 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
376 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
378 /* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
380 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
382 if (!radeon_emitted(rctx
->gfx
.cs
, rctx
->initial_gfx_cs_size
)) {
384 ws
->fence_reference(&gfx_fence
, rctx
->last_gfx_fence
);
385 if (!(flags
& PIPE_FLUSH_DEFERRED
))
386 ws
->cs_sync_flush(rctx
->gfx
.cs
);
388 /* Instead of flushing, create a deferred fence. Constraints:
389 * - The state tracker must allow a deferred flush.
390 * - The state tracker must request a fence.
391 * Thread safety in fence_finish must be ensured by the state tracker.
393 if (flags
& PIPE_FLUSH_DEFERRED
&& fence
) {
394 gfx_fence
= rctx
->ws
->cs_get_next_fence(rctx
->gfx
.cs
);
395 deferred_fence
= true;
397 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
401 /* Both engines can signal out of order, so we need to keep both fences. */
403 struct r600_multi_fence
*multi_fence
=
404 CALLOC_STRUCT(r600_multi_fence
);
406 ws
->fence_reference(&sdma_fence
, NULL
);
407 ws
->fence_reference(&gfx_fence
, NULL
);
411 multi_fence
->reference
.count
= 1;
412 /* If both fences are NULL, fence_finish will always return true. */
413 multi_fence
->gfx
= gfx_fence
;
414 multi_fence
->sdma
= sdma_fence
;
416 if (deferred_fence
) {
417 multi_fence
->gfx_unflushed
.ctx
= rctx
;
418 multi_fence
->gfx_unflushed
.ib_index
= rctx
->num_gfx_cs_flushes
;
421 screen
->fence_reference(screen
, fence
, NULL
);
422 *fence
= (struct pipe_fence_handle
*)multi_fence
;
425 if (!(flags
& PIPE_FLUSH_DEFERRED
)) {
427 ws
->cs_sync_flush(rctx
->dma
.cs
);
428 ws
->cs_sync_flush(rctx
->gfx
.cs
);
432 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
433 struct pipe_fence_handle
**fence
)
435 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
436 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
437 struct radeon_saved_cs saved
;
439 (rctx
->screen
->debug_flags
& DBG(CHECK_VM
)) &&
440 rctx
->check_vm_faults
;
442 if (!radeon_emitted(cs
, 0)) {
444 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
449 si_save_cs(rctx
->ws
, cs
, &saved
, true);
451 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
453 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
456 /* Use conservative timeout 800ms, after which we won't wait any
457 * longer and assume the GPU is hung.
459 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
461 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
462 si_clear_saved_cs(&saved
);
467 * Store a linearized copy of all chunks of \p cs together with the buffer
470 void si_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
471 struct radeon_saved_cs
*saved
, bool get_buffer_list
)
476 /* Save the IB chunks. */
477 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
478 saved
->ib
= MALLOC(4 * saved
->num_dw
);
483 for (i
= 0; i
< cs
->num_prev
; ++i
) {
484 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
485 buf
+= cs
->prev
[i
].cdw
;
487 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
489 if (!get_buffer_list
)
492 /* Save the buffer list. */
493 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
494 saved
->bo_list
= CALLOC(saved
->bo_count
,
495 sizeof(saved
->bo_list
[0]));
496 if (!saved
->bo_list
) {
500 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
505 fprintf(stderr
, "%s: out of memory\n", __func__
);
506 memset(saved
, 0, sizeof(*saved
));
509 void si_clear_saved_cs(struct radeon_saved_cs
*saved
)
512 FREE(saved
->bo_list
);
514 memset(saved
, 0, sizeof(*saved
));
517 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
519 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
520 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
521 RADEON_GPU_RESET_COUNTER
);
523 if (rctx
->gpu_reset_counter
== latest
)
524 return PIPE_NO_RESET
;
526 rctx
->gpu_reset_counter
= latest
;
527 return PIPE_UNKNOWN_CONTEXT_RESET
;
530 static void r600_set_debug_callback(struct pipe_context
*ctx
,
531 const struct pipe_debug_callback
*cb
)
533 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
538 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
541 static void r600_set_device_reset_callback(struct pipe_context
*ctx
,
542 const struct pipe_device_reset_callback
*cb
)
544 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
547 rctx
->device_reset_callback
= *cb
;
549 memset(&rctx
->device_reset_callback
, 0,
550 sizeof(rctx
->device_reset_callback
));
553 bool si_check_device_reset(struct r600_common_context
*rctx
)
555 enum pipe_reset_status status
;
557 if (!rctx
->device_reset_callback
.reset
)
560 if (!rctx
->b
.get_device_reset_status
)
563 status
= rctx
->b
.get_device_reset_status(&rctx
->b
);
564 if (status
== PIPE_NO_RESET
)
567 rctx
->device_reset_callback
.reset(rctx
->device_reset_callback
.data
, status
);
571 static void r600_dma_clear_buffer_fallback(struct pipe_context
*ctx
,
572 struct pipe_resource
*dst
,
573 uint64_t offset
, uint64_t size
,
576 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
578 rctx
->clear_buffer(ctx
, dst
, offset
, size
, value
, R600_COHERENCY_NONE
);
581 static bool r600_resource_commit(struct pipe_context
*pctx
,
582 struct pipe_resource
*resource
,
583 unsigned level
, struct pipe_box
*box
,
586 struct r600_common_context
*ctx
= (struct r600_common_context
*)pctx
;
587 struct r600_resource
*res
= r600_resource(resource
);
590 * Since buffer commitment changes cannot be pipelined, we need to
591 * (a) flush any pending commands that refer to the buffer we're about
593 * (b) wait for threaded submit to finish, including those that were
594 * triggered by some other, earlier operation.
596 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
597 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
,
598 res
->buf
, RADEON_USAGE_READWRITE
)) {
599 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
601 if (radeon_emitted(ctx
->dma
.cs
, 0) &&
602 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
,
603 res
->buf
, RADEON_USAGE_READWRITE
)) {
604 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
607 ctx
->ws
->cs_sync_flush(ctx
->dma
.cs
);
608 ctx
->ws
->cs_sync_flush(ctx
->gfx
.cs
);
610 assert(resource
->target
== PIPE_BUFFER
);
612 return ctx
->ws
->buffer_commit(res
->buf
, box
->x
, box
->width
, commit
);
615 bool si_common_context_init(struct r600_common_context
*rctx
,
616 struct r600_common_screen
*rscreen
,
617 unsigned context_flags
)
619 slab_create_child(&rctx
->pool_transfers
, &rscreen
->pool_transfers
);
620 slab_create_child(&rctx
->pool_transfers_unsync
, &rscreen
->pool_transfers
);
622 rctx
->screen
= rscreen
;
623 rctx
->ws
= rscreen
->ws
;
624 rctx
->family
= rscreen
->family
;
625 rctx
->chip_class
= rscreen
->chip_class
;
627 rctx
->b
.invalidate_resource
= si_invalidate_resource
;
628 rctx
->b
.resource_commit
= r600_resource_commit
;
629 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
630 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
631 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
632 rctx
->b
.texture_subdata
= u_default_texture_subdata
;
633 rctx
->b
.memory_barrier
= r600_memory_barrier
;
634 rctx
->b
.flush
= r600_flush_from_st
;
635 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
636 rctx
->b
.fence_server_sync
= r600_fence_server_sync
;
637 rctx
->dma_clear_buffer
= r600_dma_clear_buffer_fallback
;
638 rctx
->b
.buffer_subdata
= si_buffer_subdata
;
640 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
641 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
642 rctx
->gpu_reset_counter
=
643 rctx
->ws
->query_value(rctx
->ws
,
644 RADEON_GPU_RESET_COUNTER
);
647 rctx
->b
.set_device_reset_callback
= r600_set_device_reset_callback
;
649 si_init_context_texture_functions(rctx
);
650 si_streamout_init(rctx
);
651 si_init_query_functions(rctx
);
652 si_init_msaa(&rctx
->b
);
654 if (rctx
->chip_class
== CIK
||
655 rctx
->chip_class
== VI
||
656 rctx
->chip_class
== GFX9
) {
657 rctx
->eop_bug_scratch
= (struct r600_resource
*)
658 pipe_buffer_create(&rscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
659 16 * rscreen
->info
.num_render_backends
);
660 if (!rctx
->eop_bug_scratch
)
664 rctx
->allocator_zeroed_memory
=
665 u_suballocator_create(&rctx
->b
, rscreen
->info
.gart_page_size
,
666 0, PIPE_USAGE_DEFAULT
, 0, true);
667 if (!rctx
->allocator_zeroed_memory
)
670 rctx
->b
.stream_uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
671 0, PIPE_USAGE_STREAM
);
672 if (!rctx
->b
.stream_uploader
)
675 rctx
->b
.const_uploader
= u_upload_create(&rctx
->b
, 128 * 1024,
676 0, PIPE_USAGE_DEFAULT
);
677 if (!rctx
->b
.const_uploader
)
680 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
684 if (rscreen
->info
.num_sdma_rings
&& !(rscreen
->debug_flags
& DBG(NO_ASYNC_DMA
))) {
685 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
688 rctx
->dma
.flush
= r600_flush_dma_ring
;
694 void si_common_context_cleanup(struct r600_common_context
*rctx
)
698 /* Release DCC stats. */
699 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
700 assert(!rctx
->dcc_stats
[i
].query_active
);
702 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
703 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
704 rctx
->b
.destroy_query(&rctx
->b
,
705 rctx
->dcc_stats
[i
].ps_stats
[j
]);
707 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
710 if (rctx
->query_result_shader
)
711 rctx
->b
.delete_compute_state(&rctx
->b
, rctx
->query_result_shader
);
714 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
716 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
718 rctx
->ws
->ctx_destroy(rctx
->ctx
);
720 if (rctx
->b
.stream_uploader
)
721 u_upload_destroy(rctx
->b
.stream_uploader
);
722 if (rctx
->b
.const_uploader
)
723 u_upload_destroy(rctx
->b
.const_uploader
);
725 slab_destroy_child(&rctx
->pool_transfers
);
726 slab_destroy_child(&rctx
->pool_transfers_unsync
);
728 if (rctx
->allocator_zeroed_memory
) {
729 u_suballocator_destroy(rctx
->allocator_zeroed_memory
);
731 rctx
->ws
->fence_reference(&rctx
->last_gfx_fence
, NULL
);
732 rctx
->ws
->fence_reference(&rctx
->last_sdma_fence
, NULL
);
733 r600_resource_reference(&rctx
->eop_bug_scratch
, NULL
);
740 static const struct debug_named_value common_debug_options
[] = {
742 { "tex", DBG(TEX
), "Print texture info" },
743 { "nir", DBG(NIR
), "Enable experimental NIR shaders" },
744 { "compute", DBG(COMPUTE
), "Print compute info" },
745 { "vm", DBG(VM
), "Print virtual addresses when creating resources" },
746 { "info", DBG(INFO
), "Print driver information" },
749 { "vs", DBG(VS
), "Print vertex shaders" },
750 { "gs", DBG(GS
), "Print geometry shaders" },
751 { "ps", DBG(PS
), "Print pixel shaders" },
752 { "cs", DBG(CS
), "Print compute shaders" },
753 { "tcs", DBG(TCS
), "Print tessellation control shaders" },
754 { "tes", DBG(TES
), "Print tessellation evaluation shaders" },
755 { "noir", DBG(NO_IR
), "Don't print the LLVM IR"},
756 { "notgsi", DBG(NO_TGSI
), "Don't print the TGSI"},
757 { "noasm", DBG(NO_ASM
), "Don't print disassembled shaders"},
758 { "preoptir", DBG(PREOPT_IR
), "Print the LLVM IR before initial optimizations" },
759 { "checkir", DBG(CHECK_IR
), "Enable additional sanity checks on shader IR" },
760 { "nooptvariant", DBG(NO_OPT_VARIANT
), "Disable compiling optimized shader variants." },
762 { "testdma", DBG(TEST_DMA
), "Invoke SDMA tests and exit." },
763 { "testvmfaultcp", DBG(TEST_VMFAULT_CP
), "Invoke a CP VM fault test and exit." },
764 { "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA
), "Invoke a SDMA VM fault test and exit." },
765 { "testvmfaultshader", DBG(TEST_VMFAULT_SHADER
), "Invoke a shader VM fault test and exit." },
768 { "nodma", DBG(NO_ASYNC_DMA
), "Disable asynchronous DMA" },
769 { "nohyperz", DBG(NO_HYPERZ
), "Disable Hyper-Z" },
770 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
771 { "noinvalrange", DBG(NO_DISCARD_RANGE
), "Disable handling of INVALIDATE_RANGE map flags" },
772 { "no2d", DBG(NO_2D_TILING
), "Disable 2D tiling" },
773 { "notiling", DBG(NO_TILING
), "Disable tiling" },
774 { "switch_on_eop", DBG(SWITCH_ON_EOP
), "Program WD/IA to switch on end-of-packet." },
775 { "forcedma", DBG(FORCE_DMA
), "Use asynchronous DMA for all operations when possible." },
776 { "precompile", DBG(PRECOMPILE
), "Compile one shader variant at shader creation." },
777 { "nowc", DBG(NO_WC
), "Disable GTT write combining" },
778 { "check_vm", DBG(CHECK_VM
), "Check VM faults and dump debug info." },
779 { "nodcc", DBG(NO_DCC
), "Disable DCC." },
780 { "nodccclear", DBG(NO_DCC_CLEAR
), "Disable DCC fast clear." },
781 { "norbplus", DBG(NO_RB_PLUS
), "Disable RB+." },
782 { "sisched", DBG(SI_SCHED
), "Enable LLVM SI Machine Instruction Scheduler." },
783 { "mono", DBG(MONOLITHIC_SHADERS
), "Use old-style monolithic shaders compiled on demand" },
784 { "unsafemath", DBG(UNSAFE_MATH
), "Enable unsafe math shader optimizations" },
785 { "nodccfb", DBG(NO_DCC_FB
), "Disable separate DCC on the main framebuffer" },
786 { "nodpbb", DBG(NO_DPBB
), "Disable DPBB." },
787 { "nodfsm", DBG(NO_DFSM
), "Disable DFSM." },
788 { "dpbb", DBG(DPBB
), "Enable DPBB." },
789 { "dfsm", DBG(DFSM
), "Enable DFSM." },
790 { "nooutoforder", DBG(NO_OUT_OF_ORDER
), "Disable out-of-order rasterization" },
792 DEBUG_NAMED_VALUE_END
/* must be last */
795 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
800 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
805 static const char *r600_get_marketing_name(struct radeon_winsys
*ws
)
807 if (!ws
->get_chip_name
)
809 return ws
->get_chip_name(ws
);
812 static const char *r600_get_family_name(const struct r600_common_screen
*rscreen
)
814 switch (rscreen
->info
.family
) {
815 case CHIP_TAHITI
: return "AMD TAHITI";
816 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
817 case CHIP_VERDE
: return "AMD CAPE VERDE";
818 case CHIP_OLAND
: return "AMD OLAND";
819 case CHIP_HAINAN
: return "AMD HAINAN";
820 case CHIP_BONAIRE
: return "AMD BONAIRE";
821 case CHIP_KAVERI
: return "AMD KAVERI";
822 case CHIP_KABINI
: return "AMD KABINI";
823 case CHIP_HAWAII
: return "AMD HAWAII";
824 case CHIP_MULLINS
: return "AMD MULLINS";
825 case CHIP_TONGA
: return "AMD TONGA";
826 case CHIP_ICELAND
: return "AMD ICELAND";
827 case CHIP_CARRIZO
: return "AMD CARRIZO";
828 case CHIP_FIJI
: return "AMD FIJI";
829 case CHIP_POLARIS10
: return "AMD POLARIS10";
830 case CHIP_POLARIS11
: return "AMD POLARIS11";
831 case CHIP_POLARIS12
: return "AMD POLARIS12";
832 case CHIP_STONEY
: return "AMD STONEY";
833 case CHIP_VEGA10
: return "AMD VEGA10";
834 case CHIP_RAVEN
: return "AMD RAVEN";
835 default: return "AMD unknown";
839 static void r600_disk_cache_create(struct r600_common_screen
*rscreen
)
841 /* Don't use the cache if shader dumping is enabled. */
842 if (rscreen
->debug_flags
& DBG_ALL_SHADERS
)
845 uint32_t mesa_timestamp
;
846 if (disk_cache_get_function_timestamp(r600_disk_cache_create
,
850 uint32_t llvm_timestamp
;
852 if (disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
,
854 res
= asprintf(×tamp_str
, "%u_%u",
855 mesa_timestamp
, llvm_timestamp
);
859 /* These flags affect shader compilation. */
860 uint64_t shader_debug_flags
=
861 rscreen
->debug_flags
&
862 (DBG(FS_CORRECT_DERIVS_AFTER_KILL
) |
866 rscreen
->disk_shader_cache
=
867 disk_cache_create(r600_get_family_name(rscreen
),
875 static struct disk_cache
*r600_get_disk_shader_cache(struct pipe_screen
*pscreen
)
877 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
878 return rscreen
->disk_shader_cache
;
881 static const char* r600_get_name(struct pipe_screen
* pscreen
)
883 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
885 return rscreen
->renderer_string
;
888 static float r600_get_paramf(struct pipe_screen
* pscreen
,
889 enum pipe_capf param
)
892 case PIPE_CAPF_MAX_LINE_WIDTH
:
893 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
894 case PIPE_CAPF_MAX_POINT_WIDTH
:
895 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
897 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
899 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
901 case PIPE_CAPF_GUARD_BAND_LEFT
:
902 case PIPE_CAPF_GUARD_BAND_TOP
:
903 case PIPE_CAPF_GUARD_BAND_RIGHT
:
904 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
910 static int r600_get_video_param(struct pipe_screen
*screen
,
911 enum pipe_video_profile profile
,
912 enum pipe_video_entrypoint entrypoint
,
913 enum pipe_video_cap param
)
916 case PIPE_VIDEO_CAP_SUPPORTED
:
917 return vl_profile_supported(screen
, profile
, entrypoint
);
918 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
920 case PIPE_VIDEO_CAP_MAX_WIDTH
:
921 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
922 return vl_video_buffer_max_size(screen
);
923 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
924 return PIPE_FORMAT_NV12
;
925 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
927 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
929 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
931 case PIPE_VIDEO_CAP_MAX_LEVEL
:
932 return vl_level_supported(screen
, profile
);
938 const char *si_get_llvm_processor_name(enum radeon_family family
)
941 case CHIP_TAHITI
: return "tahiti";
942 case CHIP_PITCAIRN
: return "pitcairn";
943 case CHIP_VERDE
: return "verde";
944 case CHIP_OLAND
: return "oland";
945 case CHIP_HAINAN
: return "hainan";
946 case CHIP_BONAIRE
: return "bonaire";
947 case CHIP_KABINI
: return "kabini";
948 case CHIP_KAVERI
: return "kaveri";
949 case CHIP_HAWAII
: return "hawaii";
952 case CHIP_TONGA
: return "tonga";
953 case CHIP_ICELAND
: return "iceland";
954 case CHIP_CARRIZO
: return "carrizo";
962 case CHIP_POLARIS12
: /* same as polaris11 */
972 static unsigned get_max_threads_per_block(struct r600_common_screen
*screen
,
973 enum pipe_shader_ir ir_type
)
975 if (ir_type
!= PIPE_SHADER_IR_TGSI
)
978 /* Only 16 waves per thread-group on gfx9. */
979 if (screen
->chip_class
>= GFX9
)
982 /* Up to 40 waves per thread-group on GCN < gfx9. Expose a nice
988 static int r600_get_compute_param(struct pipe_screen
*screen
,
989 enum pipe_shader_ir ir_type
,
990 enum pipe_compute_cap param
,
993 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
995 //TODO: select these params by asic
997 case PIPE_COMPUTE_CAP_IR_TARGET
: {
1001 if (HAVE_LLVM
< 0x0400)
1002 triple
= "amdgcn--";
1004 triple
= "amdgcn-mesa-mesa3d";
1006 gpu
= si_get_llvm_processor_name(rscreen
->family
);
1008 sprintf(ret
, "%s-%s", gpu
, triple
);
1010 /* +2 for dash and terminating NIL byte */
1011 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
1013 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
1015 uint64_t *grid_dimension
= ret
;
1016 grid_dimension
[0] = 3;
1018 return 1 * sizeof(uint64_t);
1020 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
1022 uint64_t *grid_size
= ret
;
1023 grid_size
[0] = 65535;
1024 grid_size
[1] = 65535;
1025 grid_size
[2] = 65535;
1027 return 3 * sizeof(uint64_t) ;
1029 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
1031 uint64_t *block_size
= ret
;
1032 unsigned threads_per_block
= get_max_threads_per_block(rscreen
, ir_type
);
1033 block_size
[0] = threads_per_block
;
1034 block_size
[1] = threads_per_block
;
1035 block_size
[2] = threads_per_block
;
1037 return 3 * sizeof(uint64_t);
1039 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
1041 uint64_t *max_threads_per_block
= ret
;
1042 *max_threads_per_block
= get_max_threads_per_block(rscreen
, ir_type
);
1044 return sizeof(uint64_t);
1045 case PIPE_COMPUTE_CAP_ADDRESS_BITS
:
1047 uint32_t *address_bits
= ret
;
1048 address_bits
[0] = 64;
1050 return 1 * sizeof(uint32_t);
1052 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
1054 uint64_t *max_global_size
= ret
;
1055 uint64_t max_mem_alloc_size
;
1057 r600_get_compute_param(screen
, ir_type
,
1058 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
1059 &max_mem_alloc_size
);
1061 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
1062 * 1/4 of the MAX_GLOBAL_SIZE. Since the
1063 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
1064 * make sure we never report more than
1065 * 4 * MAX_MEM_ALLOC_SIZE.
1067 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
1068 MAX2(rscreen
->info
.gart_size
,
1069 rscreen
->info
.vram_size
));
1071 return sizeof(uint64_t);
1073 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
1075 uint64_t *max_local_size
= ret
;
1076 /* Value reported by the closed source driver. */
1077 *max_local_size
= 32768;
1079 return sizeof(uint64_t);
1081 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
1083 uint64_t *max_input_size
= ret
;
1084 /* Value reported by the closed source driver. */
1085 *max_input_size
= 1024;
1087 return sizeof(uint64_t);
1089 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
1091 uint64_t *max_mem_alloc_size
= ret
;
1093 *max_mem_alloc_size
= rscreen
->info
.max_alloc_size
;
1095 return sizeof(uint64_t);
1097 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
1099 uint32_t *max_clock_frequency
= ret
;
1100 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
1102 return sizeof(uint32_t);
1104 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
1106 uint32_t *max_compute_units
= ret
;
1107 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
1109 return sizeof(uint32_t);
1111 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
1113 uint32_t *images_supported
= ret
;
1114 *images_supported
= 0;
1116 return sizeof(uint32_t);
1117 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
1119 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
1121 uint32_t *subgroup_size
= ret
;
1122 *subgroup_size
= 64;
1124 return sizeof(uint32_t);
1125 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK
:
1127 uint64_t *max_variable_threads_per_block
= ret
;
1128 if (ir_type
== PIPE_SHADER_IR_TGSI
)
1129 *max_variable_threads_per_block
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
1131 *max_variable_threads_per_block
= 0;
1133 return sizeof(uint64_t);
1136 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
1140 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
1142 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1144 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
1145 rscreen
->info
.clock_crystal_freq
;
1148 static void r600_fence_reference(struct pipe_screen
*screen
,
1149 struct pipe_fence_handle
**dst
,
1150 struct pipe_fence_handle
*src
)
1152 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
1153 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
1154 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
1156 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
1157 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
1158 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
1164 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
1165 struct pipe_context
*ctx
,
1166 struct pipe_fence_handle
*fence
,
1169 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
1170 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
1171 struct r600_common_context
*rctx
;
1172 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
1174 ctx
= threaded_context_unwrap_sync(ctx
);
1175 rctx
= ctx
? (struct r600_common_context
*)ctx
: NULL
;
1178 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
1181 /* Recompute the timeout after waiting. */
1182 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1183 int64_t time
= os_time_get_nano();
1184 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1191 /* Flush the gfx IB if it hasn't been flushed yet. */
1193 rfence
->gfx_unflushed
.ctx
== rctx
&&
1194 rfence
->gfx_unflushed
.ib_index
== rctx
->num_gfx_cs_flushes
) {
1195 rctx
->gfx
.flush(rctx
, timeout
? 0 : RADEON_FLUSH_ASYNC
, NULL
);
1196 rfence
->gfx_unflushed
.ctx
= NULL
;
1201 /* Recompute the timeout after all that. */
1202 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1203 int64_t time
= os_time_get_nano();
1204 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1208 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
1211 static void r600_query_memory_info(struct pipe_screen
*screen
,
1212 struct pipe_memory_info
*info
)
1214 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1215 struct radeon_winsys
*ws
= rscreen
->ws
;
1216 unsigned vram_usage
, gtt_usage
;
1218 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
1219 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
1221 /* The real TTM memory usage is somewhat random, because:
1223 * 1) TTM delays freeing memory, because it can only free it after
1226 * 2) The memory usage can be really low if big VRAM evictions are
1227 * taking place, but the real usage is well above the size of VRAM.
1229 * Instead, return statistics of this process.
1231 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
1232 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
1234 info
->avail_device_memory
=
1235 vram_usage
<= info
->total_device_memory
?
1236 info
->total_device_memory
- vram_usage
: 0;
1237 info
->avail_staging_memory
=
1238 gtt_usage
<= info
->total_staging_memory
?
1239 info
->total_staging_memory
- gtt_usage
: 0;
1241 info
->device_memory_evicted
=
1242 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
1244 if (rscreen
->info
.drm_major
== 3 && rscreen
->info
.drm_minor
>= 4)
1245 info
->nr_device_memory_evictions
=
1246 ws
->query_value(ws
, RADEON_NUM_EVICTIONS
);
1248 /* Just return the number of evicted 64KB pages. */
1249 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
1252 struct pipe_resource
*si_resource_create_common(struct pipe_screen
*screen
,
1253 const struct pipe_resource
*templ
)
1255 if (templ
->target
== PIPE_BUFFER
) {
1256 return si_buffer_create(screen
, templ
, 256);
1258 return si_texture_create(screen
, templ
);
1262 bool si_common_screen_init(struct r600_common_screen
*rscreen
,
1263 struct radeon_winsys
*ws
)
1265 char family_name
[32] = {}, llvm_string
[32] = {}, kernel_version
[128] = {};
1266 struct utsname uname_data
;
1267 const char *chip_name
;
1269 ws
->query_info(ws
, &rscreen
->info
);
1272 if ((chip_name
= r600_get_marketing_name(ws
)))
1273 snprintf(family_name
, sizeof(family_name
), "%s / ",
1274 r600_get_family_name(rscreen
) + 4);
1276 chip_name
= r600_get_family_name(rscreen
);
1278 if (uname(&uname_data
) == 0)
1279 snprintf(kernel_version
, sizeof(kernel_version
),
1280 " / %s", uname_data
.release
);
1282 if (HAVE_LLVM
> 0) {
1283 snprintf(llvm_string
, sizeof(llvm_string
),
1284 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
1285 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
1288 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
1289 "%s (%sDRM %i.%i.%i%s%s)",
1290 chip_name
, family_name
, rscreen
->info
.drm_major
,
1291 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
1292 kernel_version
, llvm_string
);
1294 rscreen
->b
.get_name
= r600_get_name
;
1295 rscreen
->b
.get_vendor
= r600_get_vendor
;
1296 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
1297 rscreen
->b
.get_disk_shader_cache
= r600_get_disk_shader_cache
;
1298 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
1299 rscreen
->b
.get_paramf
= r600_get_paramf
;
1300 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
1301 rscreen
->b
.fence_finish
= r600_fence_finish
;
1302 rscreen
->b
.fence_reference
= r600_fence_reference
;
1303 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
1304 rscreen
->b
.resource_from_user_memory
= si_buffer_from_user_memory
;
1305 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
1307 if (rscreen
->info
.has_hw_decode
) {
1308 rscreen
->b
.get_video_param
= si_vid_get_video_param
;
1309 rscreen
->b
.is_video_format_supported
= si_vid_is_format_supported
;
1311 rscreen
->b
.get_video_param
= r600_get_video_param
;
1312 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
1315 si_init_screen_texture_functions(rscreen
);
1316 si_init_screen_query_functions(rscreen
);
1318 rscreen
->family
= rscreen
->info
.family
;
1319 rscreen
->chip_class
= rscreen
->info
.chip_class
;
1320 rscreen
->debug_flags
|= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
1321 rscreen
->has_rbplus
= false;
1322 rscreen
->rbplus_allowed
= false;
1324 r600_disk_cache_create(rscreen
);
1326 slab_create_parent(&rscreen
->pool_transfers
, sizeof(struct r600_transfer
), 64);
1328 rscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1329 if (rscreen
->force_aniso
>= 0) {
1330 printf("radeon: Forcing anisotropy filter to %ix\n",
1331 /* round down to a power of two */
1332 1 << util_logbase2(rscreen
->force_aniso
));
1335 (void) mtx_init(&rscreen
->aux_context_lock
, mtx_plain
);
1336 (void) mtx_init(&rscreen
->gpu_load_mutex
, mtx_plain
);
1338 if (rscreen
->debug_flags
& DBG(INFO
)) {
1339 printf("pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
1340 rscreen
->info
.pci_domain
, rscreen
->info
.pci_bus
,
1341 rscreen
->info
.pci_dev
, rscreen
->info
.pci_func
);
1342 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
1343 printf("family = %i (%s)\n", rscreen
->info
.family
,
1344 r600_get_family_name(rscreen
));
1345 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
1346 printf("pte_fragment_size = %u\n", rscreen
->info
.pte_fragment_size
);
1347 printf("gart_page_size = %u\n", rscreen
->info
.gart_page_size
);
1348 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
1349 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
1350 printf("vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_vis_size
, 1024*1024));
1351 printf("max_alloc_size = %i MB\n",
1352 (int)DIV_ROUND_UP(rscreen
->info
.max_alloc_size
, 1024*1024));
1353 printf("min_alloc_size = %u\n", rscreen
->info
.min_alloc_size
);
1354 printf("has_dedicated_vram = %u\n", rscreen
->info
.has_dedicated_vram
);
1355 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
1356 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
1357 printf("has_hw_decode = %u\n", rscreen
->info
.has_hw_decode
);
1358 printf("num_sdma_rings = %i\n", rscreen
->info
.num_sdma_rings
);
1359 printf("num_compute_rings = %u\n", rscreen
->info
.num_compute_rings
);
1360 printf("uvd_fw_version = %u\n", rscreen
->info
.uvd_fw_version
);
1361 printf("vce_fw_version = %u\n", rscreen
->info
.vce_fw_version
);
1362 printf("me_fw_version = %i\n", rscreen
->info
.me_fw_version
);
1363 printf("me_fw_feature = %i\n", rscreen
->info
.me_fw_feature
);
1364 printf("pfp_fw_version = %i\n", rscreen
->info
.pfp_fw_version
);
1365 printf("pfp_fw_feature = %i\n", rscreen
->info
.pfp_fw_feature
);
1366 printf("ce_fw_version = %i\n", rscreen
->info
.ce_fw_version
);
1367 printf("ce_fw_feature = %i\n", rscreen
->info
.ce_fw_feature
);
1368 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
1369 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
1370 printf("tcc_cache_line_size = %u\n", rscreen
->info
.tcc_cache_line_size
);
1371 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
1372 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
1373 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
1374 printf("has_syncobj = %u\n", rscreen
->info
.has_syncobj
);
1376 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
1377 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
1378 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
1379 printf("max_se = %i\n", rscreen
->info
.max_se
);
1380 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
1382 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
1383 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
1384 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
1385 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
1386 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
1387 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
1388 printf("enabled_rb_mask = 0x%x\n", rscreen
->info
.enabled_rb_mask
);
1389 printf("max_alignment = %u\n", (unsigned)rscreen
->info
.max_alignment
);
1394 void si_destroy_common_screen(struct r600_common_screen
*rscreen
)
1396 si_perfcounters_destroy(rscreen
);
1397 si_gpu_load_kill_thread(rscreen
);
1399 mtx_destroy(&rscreen
->gpu_load_mutex
);
1400 mtx_destroy(&rscreen
->aux_context_lock
);
1401 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
1403 slab_destroy_parent(&rscreen
->pool_transfers
);
1405 disk_cache_destroy(rscreen
->disk_shader_cache
);
1406 rscreen
->ws
->destroy(rscreen
->ws
);
1410 bool si_can_dump_shader(struct r600_common_screen
*rscreen
,
1413 return rscreen
->debug_flags
& (1 << processor
);
1416 bool si_extra_shader_checks(struct r600_common_screen
*rscreen
, unsigned processor
)
1418 return (rscreen
->debug_flags
& DBG(CHECK_IR
)) ||
1419 si_can_dump_shader(rscreen
, processor
);
1422 void si_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
1423 uint64_t offset
, uint64_t size
, unsigned value
)
1425 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
1427 mtx_lock(&rscreen
->aux_context_lock
);
1428 rctx
->dma_clear_buffer(&rctx
->b
, dst
, offset
, size
, value
);
1429 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
1430 mtx_unlock(&rscreen
->aux_context_lock
);