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"
39 #include "amd/common/sid.h"
41 #include <sys/utsname.h>
43 #include <llvm-c/TargetMachine.h>
46 struct r600_multi_fence
{
47 struct pipe_reference reference
;
48 struct pipe_fence_handle
*gfx
;
49 struct pipe_fence_handle
*sdma
;
51 /* If the context wasn't flushed at fence creation, this is non-NULL. */
53 struct r600_common_context
*ctx
;
59 * shader binary helpers.
61 void si_radeon_shader_binary_init(struct ac_shader_binary
*b
)
63 memset(b
, 0, sizeof(*b
));
66 void si_radeon_shader_binary_clean(struct ac_shader_binary
*b
)
73 FREE(b
->global_symbol_offsets
);
75 FREE(b
->disasm_string
);
76 FREE(b
->llvm_ir_string
);
86 * \param event EVENT_TYPE_*
87 * \param event_flags Optional cache flush flags (TC)
88 * \param data_sel 1 = fence, 3 = timestamp
90 * \param va GPU address
91 * \param old_value Previous fence value (for a bug workaround)
92 * \param new_value Fence value to write for this event.
94 void si_gfx_write_event_eop(struct r600_common_context
*ctx
,
95 unsigned event
, unsigned event_flags
,
97 struct r600_resource
*buf
, uint64_t va
,
98 uint32_t new_fence
, unsigned query_type
)
100 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
101 unsigned op
= EVENT_TYPE(event
) |
104 unsigned sel
= EOP_DATA_SEL(data_sel
);
106 /* Wait for write confirmation before writing data, but don't send
108 if (data_sel
!= EOP_DATA_SEL_DISCARD
)
109 sel
|= EOP_INT_SEL(EOP_INT_SEL_SEND_DATA_AFTER_WR_CONFIRM
);
111 if (ctx
->chip_class
>= GFX9
) {
112 /* A ZPASS_DONE or PIXEL_STAT_DUMP_EVENT (of the DB occlusion
113 * counters) must immediately precede every timestamp event to
114 * prevent a GPU hang on GFX9.
116 * Occlusion queries don't need to do it here, because they
117 * always do ZPASS_DONE before the timestamp.
119 if (ctx
->chip_class
== GFX9
&&
120 query_type
!= PIPE_QUERY_OCCLUSION_COUNTER
&&
121 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE
&&
122 query_type
!= PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
) {
123 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
125 assert(16 * ctx
->screen
->info
.num_render_backends
<=
126 scratch
->b
.b
.width0
);
127 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE
, 2, 0));
128 radeon_emit(cs
, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE
) | EVENT_INDEX(1));
129 radeon_emit(cs
, scratch
->gpu_address
);
130 radeon_emit(cs
, scratch
->gpu_address
>> 32);
132 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
133 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
136 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, 6, 0));
138 radeon_emit(cs
, sel
);
139 radeon_emit(cs
, va
); /* address lo */
140 radeon_emit(cs
, va
>> 32); /* address hi */
141 radeon_emit(cs
, new_fence
); /* immediate data lo */
142 radeon_emit(cs
, 0); /* immediate data hi */
143 radeon_emit(cs
, 0); /* unused */
145 if (ctx
->chip_class
== CIK
||
146 ctx
->chip_class
== VI
) {
147 struct r600_resource
*scratch
= ctx
->eop_bug_scratch
;
148 uint64_t va
= scratch
->gpu_address
;
150 /* Two EOP events are required to make all engines go idle
151 * (and optional cache flushes executed) before the timestamp
154 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
157 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
158 radeon_emit(cs
, 0); /* immediate data */
159 radeon_emit(cs
, 0); /* unused */
161 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, scratch
,
162 RADEON_USAGE_WRITE
, RADEON_PRIO_QUERY
);
165 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
168 radeon_emit(cs
, ((va
>> 32) & 0xffff) | sel
);
169 radeon_emit(cs
, new_fence
); /* immediate data */
170 radeon_emit(cs
, 0); /* unused */
174 radeon_add_to_buffer_list(ctx
, &ctx
->gfx
, buf
, RADEON_USAGE_WRITE
,
179 unsigned si_gfx_write_fence_dwords(struct r600_common_screen
*screen
)
183 if (screen
->chip_class
== CIK
||
184 screen
->chip_class
== VI
)
187 if (!screen
->info
.has_virtual_memory
)
193 void si_gfx_wait_fence(struct r600_common_context
*ctx
,
194 uint64_t va
, uint32_t ref
, uint32_t mask
)
196 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
198 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, 0));
199 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
201 radeon_emit(cs
, va
>> 32);
202 radeon_emit(cs
, ref
); /* reference value */
203 radeon_emit(cs
, mask
); /* mask */
204 radeon_emit(cs
, 4); /* poll interval */
207 static void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
209 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
211 /* NOP waits for idle on Evergreen and later. */
212 if (rctx
->chip_class
>= CIK
)
213 radeon_emit(cs
, 0x00000000); /* NOP */
215 radeon_emit(cs
, 0xf0000000); /* NOP */
218 void si_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
,
219 struct r600_resource
*dst
, struct r600_resource
*src
)
221 uint64_t vram
= ctx
->dma
.cs
->used_vram
;
222 uint64_t gtt
= ctx
->dma
.cs
->used_gart
;
225 vram
+= dst
->vram_usage
;
226 gtt
+= dst
->gart_usage
;
229 vram
+= src
->vram_usage
;
230 gtt
+= src
->gart_usage
;
233 /* Flush the GFX IB if DMA depends on it. */
234 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
236 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
237 RADEON_USAGE_READWRITE
)) ||
239 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
240 RADEON_USAGE_WRITE
))))
241 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
243 /* Flush if there's not enough space, or if the memory usage per IB
246 * IBs using too little memory are limited by the IB submission overhead.
247 * IBs using too much memory are limited by the kernel/TTM overhead.
248 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
250 * This heuristic makes sure that DMA requests are executed
251 * very soon after the call is made and lowers memory usage.
252 * It improves texture upload performance by keeping the DMA
253 * engine busy while uploads are being submitted.
255 num_dw
++; /* for emit_wait_idle below */
256 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
257 ctx
->dma
.cs
->used_vram
+ ctx
->dma
.cs
->used_gart
> 64 * 1024 * 1024 ||
258 !radeon_cs_memory_below_limit(ctx
->screen
, ctx
->dma
.cs
, vram
, gtt
)) {
259 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
260 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
263 /* Wait for idle if either buffer has been used in the IB before to
264 * prevent read-after-write hazards.
267 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, dst
->buf
,
268 RADEON_USAGE_READWRITE
)) ||
270 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, src
->buf
,
271 RADEON_USAGE_WRITE
)))
272 r600_dma_emit_wait_idle(ctx
);
274 /* If GPUVM is not supported, the CS checker needs 2 entries
275 * in the buffer list per packet, which has to be done manually.
277 if (ctx
->screen
->info
.has_virtual_memory
) {
279 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
281 RADEON_PRIO_SDMA_BUFFER
);
283 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
285 RADEON_PRIO_SDMA_BUFFER
);
288 /* this function is called before all DMA calls, so increment this. */
289 ctx
->num_dma_calls
++;
292 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
296 void si_preflush_suspend_features(struct r600_common_context
*ctx
)
298 /* suspend queries */
299 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
300 si_suspend_queries(ctx
);
303 void si_postflush_resume_features(struct r600_common_context
*ctx
)
306 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
307 si_resume_queries(ctx
);
310 static void r600_add_fence_dependency(struct r600_common_context
*rctx
,
311 struct pipe_fence_handle
*fence
)
313 struct radeon_winsys
*ws
= rctx
->ws
;
316 ws
->cs_add_fence_dependency(rctx
->dma
.cs
, fence
);
317 ws
->cs_add_fence_dependency(rctx
->gfx
.cs
, fence
);
320 static void r600_fence_server_sync(struct pipe_context
*ctx
,
321 struct pipe_fence_handle
*fence
)
323 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
324 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
326 /* Only amdgpu needs to handle fence dependencies (for fence imports).
327 * radeon synchronizes all rings by default and will not implement
330 if (rctx
->screen
->info
.drm_major
== 2)
333 /* Only imported fences need to be handled by fence_server_sync,
334 * because the winsys handles synchronizations automatically for BOs
335 * within the process.
337 * Simply skip unflushed fences here, and the winsys will drop no-op
338 * dependencies (i.e. dependencies within the same ring).
340 if (rfence
->gfx_unflushed
.ctx
)
343 /* All unflushed commands will not start execution before
344 * this fence dependency is signalled.
346 * Should we flush the context to allow more GPU parallelism?
349 r600_add_fence_dependency(rctx
, rfence
->sdma
);
351 r600_add_fence_dependency(rctx
, rfence
->gfx
);
354 static void r600_flush_from_st(struct pipe_context
*ctx
,
355 struct pipe_fence_handle
**fence
,
358 struct pipe_screen
*screen
= ctx
->screen
;
359 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
360 struct radeon_winsys
*ws
= rctx
->ws
;
361 struct pipe_fence_handle
*gfx_fence
= NULL
;
362 struct pipe_fence_handle
*sdma_fence
= NULL
;
363 bool deferred_fence
= false;
364 unsigned rflags
= RADEON_FLUSH_ASYNC
;
366 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
367 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
369 /* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
371 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
373 if (!radeon_emitted(rctx
->gfx
.cs
, rctx
->initial_gfx_cs_size
)) {
375 ws
->fence_reference(&gfx_fence
, rctx
->last_gfx_fence
);
376 if (!(flags
& PIPE_FLUSH_DEFERRED
))
377 ws
->cs_sync_flush(rctx
->gfx
.cs
);
379 /* Instead of flushing, create a deferred fence. Constraints:
380 * - The state tracker must allow a deferred flush.
381 * - The state tracker must request a fence.
382 * Thread safety in fence_finish must be ensured by the state tracker.
384 if (flags
& PIPE_FLUSH_DEFERRED
&& fence
) {
385 gfx_fence
= rctx
->ws
->cs_get_next_fence(rctx
->gfx
.cs
);
386 deferred_fence
= true;
388 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
392 /* Both engines can signal out of order, so we need to keep both fences. */
394 struct r600_multi_fence
*multi_fence
=
395 CALLOC_STRUCT(r600_multi_fence
);
397 ws
->fence_reference(&sdma_fence
, NULL
);
398 ws
->fence_reference(&gfx_fence
, NULL
);
402 multi_fence
->reference
.count
= 1;
403 /* If both fences are NULL, fence_finish will always return true. */
404 multi_fence
->gfx
= gfx_fence
;
405 multi_fence
->sdma
= sdma_fence
;
407 if (deferred_fence
) {
408 multi_fence
->gfx_unflushed
.ctx
= rctx
;
409 multi_fence
->gfx_unflushed
.ib_index
= rctx
->num_gfx_cs_flushes
;
412 screen
->fence_reference(screen
, fence
, NULL
);
413 *fence
= (struct pipe_fence_handle
*)multi_fence
;
416 if (!(flags
& PIPE_FLUSH_DEFERRED
)) {
418 ws
->cs_sync_flush(rctx
->dma
.cs
);
419 ws
->cs_sync_flush(rctx
->gfx
.cs
);
423 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
424 struct pipe_fence_handle
**fence
)
426 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
427 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
428 struct radeon_saved_cs saved
;
430 (rctx
->screen
->debug_flags
& DBG(CHECK_VM
)) &&
431 rctx
->check_vm_faults
;
433 if (!radeon_emitted(cs
, 0)) {
435 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
440 si_save_cs(rctx
->ws
, cs
, &saved
, true);
442 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
444 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
447 /* Use conservative timeout 800ms, after which we won't wait any
448 * longer and assume the GPU is hung.
450 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
452 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
453 si_clear_saved_cs(&saved
);
458 * Store a linearized copy of all chunks of \p cs together with the buffer
461 void si_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
462 struct radeon_saved_cs
*saved
, bool get_buffer_list
)
467 /* Save the IB chunks. */
468 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
469 saved
->ib
= MALLOC(4 * saved
->num_dw
);
474 for (i
= 0; i
< cs
->num_prev
; ++i
) {
475 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
476 buf
+= cs
->prev
[i
].cdw
;
478 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
480 if (!get_buffer_list
)
483 /* Save the buffer list. */
484 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
485 saved
->bo_list
= CALLOC(saved
->bo_count
,
486 sizeof(saved
->bo_list
[0]));
487 if (!saved
->bo_list
) {
491 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
496 fprintf(stderr
, "%s: out of memory\n", __func__
);
497 memset(saved
, 0, sizeof(*saved
));
500 void si_clear_saved_cs(struct radeon_saved_cs
*saved
)
503 FREE(saved
->bo_list
);
505 memset(saved
, 0, sizeof(*saved
));
508 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
510 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
511 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
512 RADEON_GPU_RESET_COUNTER
);
514 if (rctx
->gpu_reset_counter
== latest
)
515 return PIPE_NO_RESET
;
517 rctx
->gpu_reset_counter
= latest
;
518 return PIPE_UNKNOWN_CONTEXT_RESET
;
521 static void r600_set_debug_callback(struct pipe_context
*ctx
,
522 const struct pipe_debug_callback
*cb
)
524 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
529 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
532 static void r600_set_device_reset_callback(struct pipe_context
*ctx
,
533 const struct pipe_device_reset_callback
*cb
)
535 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
538 rctx
->device_reset_callback
= *cb
;
540 memset(&rctx
->device_reset_callback
, 0,
541 sizeof(rctx
->device_reset_callback
));
544 bool si_check_device_reset(struct r600_common_context
*rctx
)
546 enum pipe_reset_status status
;
548 if (!rctx
->device_reset_callback
.reset
)
551 if (!rctx
->b
.get_device_reset_status
)
554 status
= rctx
->b
.get_device_reset_status(&rctx
->b
);
555 if (status
== PIPE_NO_RESET
)
558 rctx
->device_reset_callback
.reset(rctx
->device_reset_callback
.data
, status
);
562 static void r600_dma_clear_buffer_fallback(struct pipe_context
*ctx
,
563 struct pipe_resource
*dst
,
564 uint64_t offset
, uint64_t size
,
567 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
569 rctx
->clear_buffer(ctx
, dst
, offset
, size
, value
, R600_COHERENCY_NONE
);
572 static bool r600_resource_commit(struct pipe_context
*pctx
,
573 struct pipe_resource
*resource
,
574 unsigned level
, struct pipe_box
*box
,
577 struct r600_common_context
*ctx
= (struct r600_common_context
*)pctx
;
578 struct r600_resource
*res
= r600_resource(resource
);
581 * Since buffer commitment changes cannot be pipelined, we need to
582 * (a) flush any pending commands that refer to the buffer we're about
584 * (b) wait for threaded submit to finish, including those that were
585 * triggered by some other, earlier operation.
587 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
588 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
,
589 res
->buf
, RADEON_USAGE_READWRITE
)) {
590 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
592 if (radeon_emitted(ctx
->dma
.cs
, 0) &&
593 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
,
594 res
->buf
, RADEON_USAGE_READWRITE
)) {
595 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
598 ctx
->ws
->cs_sync_flush(ctx
->dma
.cs
);
599 ctx
->ws
->cs_sync_flush(ctx
->gfx
.cs
);
601 assert(resource
->target
== PIPE_BUFFER
);
603 return ctx
->ws
->buffer_commit(res
->buf
, box
->x
, box
->width
, commit
);
606 bool si_common_context_init(struct r600_common_context
*rctx
,
607 struct r600_common_screen
*rscreen
,
608 unsigned context_flags
)
610 slab_create_child(&rctx
->pool_transfers
, &rscreen
->pool_transfers
);
611 slab_create_child(&rctx
->pool_transfers_unsync
, &rscreen
->pool_transfers
);
613 rctx
->screen
= rscreen
;
614 rctx
->ws
= rscreen
->ws
;
615 rctx
->family
= rscreen
->family
;
616 rctx
->chip_class
= rscreen
->chip_class
;
618 rctx
->b
.invalidate_resource
= si_invalidate_resource
;
619 rctx
->b
.resource_commit
= r600_resource_commit
;
620 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
621 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
622 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
623 rctx
->b
.texture_subdata
= u_default_texture_subdata
;
624 rctx
->b
.memory_barrier
= r600_memory_barrier
;
625 rctx
->b
.flush
= r600_flush_from_st
;
626 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
627 rctx
->b
.fence_server_sync
= r600_fence_server_sync
;
628 rctx
->dma_clear_buffer
= r600_dma_clear_buffer_fallback
;
629 rctx
->b
.buffer_subdata
= si_buffer_subdata
;
631 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
632 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
633 rctx
->gpu_reset_counter
=
634 rctx
->ws
->query_value(rctx
->ws
,
635 RADEON_GPU_RESET_COUNTER
);
638 rctx
->b
.set_device_reset_callback
= r600_set_device_reset_callback
;
640 si_init_context_texture_functions(rctx
);
641 si_init_query_functions(rctx
);
643 if (rctx
->chip_class
== CIK
||
644 rctx
->chip_class
== VI
||
645 rctx
->chip_class
== GFX9
) {
646 rctx
->eop_bug_scratch
= (struct r600_resource
*)
647 pipe_buffer_create(&rscreen
->b
, 0, PIPE_USAGE_DEFAULT
,
648 16 * rscreen
->info
.num_render_backends
);
649 if (!rctx
->eop_bug_scratch
)
653 rctx
->allocator_zeroed_memory
=
654 u_suballocator_create(&rctx
->b
, rscreen
->info
.gart_page_size
,
655 0, PIPE_USAGE_DEFAULT
, 0, true);
656 if (!rctx
->allocator_zeroed_memory
)
659 rctx
->b
.stream_uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
660 0, PIPE_USAGE_STREAM
);
661 if (!rctx
->b
.stream_uploader
)
664 rctx
->b
.const_uploader
= u_upload_create(&rctx
->b
, 128 * 1024,
665 0, PIPE_USAGE_DEFAULT
);
666 if (!rctx
->b
.const_uploader
)
669 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
673 if (rscreen
->info
.num_sdma_rings
&& !(rscreen
->debug_flags
& DBG(NO_ASYNC_DMA
))) {
674 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
677 rctx
->dma
.flush
= r600_flush_dma_ring
;
683 void si_common_context_cleanup(struct r600_common_context
*rctx
)
687 /* Release DCC stats. */
688 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
689 assert(!rctx
->dcc_stats
[i
].query_active
);
691 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
692 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
693 rctx
->b
.destroy_query(&rctx
->b
,
694 rctx
->dcc_stats
[i
].ps_stats
[j
]);
696 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
699 if (rctx
->query_result_shader
)
700 rctx
->b
.delete_compute_state(&rctx
->b
, rctx
->query_result_shader
);
703 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
705 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
707 rctx
->ws
->ctx_destroy(rctx
->ctx
);
709 if (rctx
->b
.stream_uploader
)
710 u_upload_destroy(rctx
->b
.stream_uploader
);
711 if (rctx
->b
.const_uploader
)
712 u_upload_destroy(rctx
->b
.const_uploader
);
714 slab_destroy_child(&rctx
->pool_transfers
);
715 slab_destroy_child(&rctx
->pool_transfers_unsync
);
717 if (rctx
->allocator_zeroed_memory
) {
718 u_suballocator_destroy(rctx
->allocator_zeroed_memory
);
720 rctx
->ws
->fence_reference(&rctx
->last_gfx_fence
, NULL
);
721 rctx
->ws
->fence_reference(&rctx
->last_sdma_fence
, NULL
);
722 r600_resource_reference(&rctx
->eop_bug_scratch
, NULL
);
729 static const struct debug_named_value common_debug_options
[] = {
731 { "tex", DBG(TEX
), "Print texture info" },
732 { "nir", DBG(NIR
), "Enable experimental NIR shaders" },
733 { "compute", DBG(COMPUTE
), "Print compute info" },
734 { "vm", DBG(VM
), "Print virtual addresses when creating resources" },
735 { "info", DBG(INFO
), "Print driver information" },
738 { "vs", DBG(VS
), "Print vertex shaders" },
739 { "gs", DBG(GS
), "Print geometry shaders" },
740 { "ps", DBG(PS
), "Print pixel shaders" },
741 { "cs", DBG(CS
), "Print compute shaders" },
742 { "tcs", DBG(TCS
), "Print tessellation control shaders" },
743 { "tes", DBG(TES
), "Print tessellation evaluation shaders" },
744 { "noir", DBG(NO_IR
), "Don't print the LLVM IR"},
745 { "notgsi", DBG(NO_TGSI
), "Don't print the TGSI"},
746 { "noasm", DBG(NO_ASM
), "Don't print disassembled shaders"},
747 { "preoptir", DBG(PREOPT_IR
), "Print the LLVM IR before initial optimizations" },
748 { "checkir", DBG(CHECK_IR
), "Enable additional sanity checks on shader IR" },
749 { "nooptvariant", DBG(NO_OPT_VARIANT
), "Disable compiling optimized shader variants." },
751 { "testdma", DBG(TEST_DMA
), "Invoke SDMA tests and exit." },
752 { "testvmfaultcp", DBG(TEST_VMFAULT_CP
), "Invoke a CP VM fault test and exit." },
753 { "testvmfaultsdma", DBG(TEST_VMFAULT_SDMA
), "Invoke a SDMA VM fault test and exit." },
754 { "testvmfaultshader", DBG(TEST_VMFAULT_SHADER
), "Invoke a shader VM fault test and exit." },
757 { "nodma", DBG(NO_ASYNC_DMA
), "Disable asynchronous DMA" },
758 { "nohyperz", DBG(NO_HYPERZ
), "Disable Hyper-Z" },
759 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
760 { "noinvalrange", DBG(NO_DISCARD_RANGE
), "Disable handling of INVALIDATE_RANGE map flags" },
761 { "no2d", DBG(NO_2D_TILING
), "Disable 2D tiling" },
762 { "notiling", DBG(NO_TILING
), "Disable tiling" },
763 { "switch_on_eop", DBG(SWITCH_ON_EOP
), "Program WD/IA to switch on end-of-packet." },
764 { "forcedma", DBG(FORCE_DMA
), "Use asynchronous DMA for all operations when possible." },
765 { "precompile", DBG(PRECOMPILE
), "Compile one shader variant at shader creation." },
766 { "nowc", DBG(NO_WC
), "Disable GTT write combining" },
767 { "check_vm", DBG(CHECK_VM
), "Check VM faults and dump debug info." },
768 { "nodcc", DBG(NO_DCC
), "Disable DCC." },
769 { "nodccclear", DBG(NO_DCC_CLEAR
), "Disable DCC fast clear." },
770 { "norbplus", DBG(NO_RB_PLUS
), "Disable RB+." },
771 { "sisched", DBG(SI_SCHED
), "Enable LLVM SI Machine Instruction Scheduler." },
772 { "mono", DBG(MONOLITHIC_SHADERS
), "Use old-style monolithic shaders compiled on demand" },
773 { "unsafemath", DBG(UNSAFE_MATH
), "Enable unsafe math shader optimizations" },
774 { "nodccfb", DBG(NO_DCC_FB
), "Disable separate DCC on the main framebuffer" },
775 { "nodpbb", DBG(NO_DPBB
), "Disable DPBB." },
776 { "nodfsm", DBG(NO_DFSM
), "Disable DFSM." },
777 { "dpbb", DBG(DPBB
), "Enable DPBB." },
778 { "dfsm", DBG(DFSM
), "Enable DFSM." },
779 { "nooutoforder", DBG(NO_OUT_OF_ORDER
), "Disable out-of-order rasterization" },
781 DEBUG_NAMED_VALUE_END
/* must be last */
784 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
789 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
794 static const char *r600_get_marketing_name(struct radeon_winsys
*ws
)
796 if (!ws
->get_chip_name
)
798 return ws
->get_chip_name(ws
);
801 static const char *r600_get_family_name(const struct r600_common_screen
*rscreen
)
803 switch (rscreen
->info
.family
) {
804 case CHIP_TAHITI
: return "AMD TAHITI";
805 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
806 case CHIP_VERDE
: return "AMD CAPE VERDE";
807 case CHIP_OLAND
: return "AMD OLAND";
808 case CHIP_HAINAN
: return "AMD HAINAN";
809 case CHIP_BONAIRE
: return "AMD BONAIRE";
810 case CHIP_KAVERI
: return "AMD KAVERI";
811 case CHIP_KABINI
: return "AMD KABINI";
812 case CHIP_HAWAII
: return "AMD HAWAII";
813 case CHIP_MULLINS
: return "AMD MULLINS";
814 case CHIP_TONGA
: return "AMD TONGA";
815 case CHIP_ICELAND
: return "AMD ICELAND";
816 case CHIP_CARRIZO
: return "AMD CARRIZO";
817 case CHIP_FIJI
: return "AMD FIJI";
818 case CHIP_POLARIS10
: return "AMD POLARIS10";
819 case CHIP_POLARIS11
: return "AMD POLARIS11";
820 case CHIP_POLARIS12
: return "AMD POLARIS12";
821 case CHIP_STONEY
: return "AMD STONEY";
822 case CHIP_VEGA10
: return "AMD VEGA10";
823 case CHIP_RAVEN
: return "AMD RAVEN";
824 default: return "AMD unknown";
828 static void r600_disk_cache_create(struct r600_common_screen
*rscreen
)
830 /* Don't use the cache if shader dumping is enabled. */
831 if (rscreen
->debug_flags
& DBG_ALL_SHADERS
)
834 uint32_t mesa_timestamp
;
835 if (disk_cache_get_function_timestamp(r600_disk_cache_create
,
839 uint32_t llvm_timestamp
;
841 if (disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
,
843 res
= asprintf(×tamp_str
, "%u_%u",
844 mesa_timestamp
, llvm_timestamp
);
848 /* These flags affect shader compilation. */
849 uint64_t shader_debug_flags
=
850 rscreen
->debug_flags
&
851 (DBG(FS_CORRECT_DERIVS_AFTER_KILL
) |
855 rscreen
->disk_shader_cache
=
856 disk_cache_create(r600_get_family_name(rscreen
),
864 static struct disk_cache
*r600_get_disk_shader_cache(struct pipe_screen
*pscreen
)
866 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
867 return rscreen
->disk_shader_cache
;
870 static const char* r600_get_name(struct pipe_screen
* pscreen
)
872 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
874 return rscreen
->renderer_string
;
877 static float r600_get_paramf(struct pipe_screen
* pscreen
,
878 enum pipe_capf param
)
881 case PIPE_CAPF_MAX_LINE_WIDTH
:
882 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
883 case PIPE_CAPF_MAX_POINT_WIDTH
:
884 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
886 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
888 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
890 case PIPE_CAPF_GUARD_BAND_LEFT
:
891 case PIPE_CAPF_GUARD_BAND_TOP
:
892 case PIPE_CAPF_GUARD_BAND_RIGHT
:
893 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
899 static int r600_get_video_param(struct pipe_screen
*screen
,
900 enum pipe_video_profile profile
,
901 enum pipe_video_entrypoint entrypoint
,
902 enum pipe_video_cap param
)
905 case PIPE_VIDEO_CAP_SUPPORTED
:
906 return vl_profile_supported(screen
, profile
, entrypoint
);
907 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
909 case PIPE_VIDEO_CAP_MAX_WIDTH
:
910 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
911 return vl_video_buffer_max_size(screen
);
912 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
913 return PIPE_FORMAT_NV12
;
914 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
916 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
918 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
920 case PIPE_VIDEO_CAP_MAX_LEVEL
:
921 return vl_level_supported(screen
, profile
);
927 const char *si_get_llvm_processor_name(enum radeon_family family
)
930 case CHIP_TAHITI
: return "tahiti";
931 case CHIP_PITCAIRN
: return "pitcairn";
932 case CHIP_VERDE
: return "verde";
933 case CHIP_OLAND
: return "oland";
934 case CHIP_HAINAN
: return "hainan";
935 case CHIP_BONAIRE
: return "bonaire";
936 case CHIP_KABINI
: return "kabini";
937 case CHIP_KAVERI
: return "kaveri";
938 case CHIP_HAWAII
: return "hawaii";
941 case CHIP_TONGA
: return "tonga";
942 case CHIP_ICELAND
: return "iceland";
943 case CHIP_CARRIZO
: return "carrizo";
951 case CHIP_POLARIS12
: /* same as polaris11 */
961 static unsigned get_max_threads_per_block(struct r600_common_screen
*screen
,
962 enum pipe_shader_ir ir_type
)
964 if (ir_type
!= PIPE_SHADER_IR_TGSI
)
967 /* Only 16 waves per thread-group on gfx9. */
968 if (screen
->chip_class
>= GFX9
)
971 /* Up to 40 waves per thread-group on GCN < gfx9. Expose a nice
977 static int r600_get_compute_param(struct pipe_screen
*screen
,
978 enum pipe_shader_ir ir_type
,
979 enum pipe_compute_cap param
,
982 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
984 //TODO: select these params by asic
986 case PIPE_COMPUTE_CAP_IR_TARGET
: {
990 if (HAVE_LLVM
< 0x0400)
993 triple
= "amdgcn-mesa-mesa3d";
995 gpu
= si_get_llvm_processor_name(rscreen
->family
);
997 sprintf(ret
, "%s-%s", gpu
, triple
);
999 /* +2 for dash and terminating NIL byte */
1000 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
1002 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
1004 uint64_t *grid_dimension
= ret
;
1005 grid_dimension
[0] = 3;
1007 return 1 * sizeof(uint64_t);
1009 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
1011 uint64_t *grid_size
= ret
;
1012 grid_size
[0] = 65535;
1013 grid_size
[1] = 65535;
1014 grid_size
[2] = 65535;
1016 return 3 * sizeof(uint64_t) ;
1018 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
1020 uint64_t *block_size
= ret
;
1021 unsigned threads_per_block
= get_max_threads_per_block(rscreen
, ir_type
);
1022 block_size
[0] = threads_per_block
;
1023 block_size
[1] = threads_per_block
;
1024 block_size
[2] = threads_per_block
;
1026 return 3 * sizeof(uint64_t);
1028 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
1030 uint64_t *max_threads_per_block
= ret
;
1031 *max_threads_per_block
= get_max_threads_per_block(rscreen
, ir_type
);
1033 return sizeof(uint64_t);
1034 case PIPE_COMPUTE_CAP_ADDRESS_BITS
:
1036 uint32_t *address_bits
= ret
;
1037 address_bits
[0] = 64;
1039 return 1 * sizeof(uint32_t);
1041 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
1043 uint64_t *max_global_size
= ret
;
1044 uint64_t max_mem_alloc_size
;
1046 r600_get_compute_param(screen
, ir_type
,
1047 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
1048 &max_mem_alloc_size
);
1050 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
1051 * 1/4 of the MAX_GLOBAL_SIZE. Since the
1052 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
1053 * make sure we never report more than
1054 * 4 * MAX_MEM_ALLOC_SIZE.
1056 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
1057 MAX2(rscreen
->info
.gart_size
,
1058 rscreen
->info
.vram_size
));
1060 return sizeof(uint64_t);
1062 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
1064 uint64_t *max_local_size
= ret
;
1065 /* Value reported by the closed source driver. */
1066 *max_local_size
= 32768;
1068 return sizeof(uint64_t);
1070 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
1072 uint64_t *max_input_size
= ret
;
1073 /* Value reported by the closed source driver. */
1074 *max_input_size
= 1024;
1076 return sizeof(uint64_t);
1078 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
1080 uint64_t *max_mem_alloc_size
= ret
;
1082 *max_mem_alloc_size
= rscreen
->info
.max_alloc_size
;
1084 return sizeof(uint64_t);
1086 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
1088 uint32_t *max_clock_frequency
= ret
;
1089 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
1091 return sizeof(uint32_t);
1093 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
1095 uint32_t *max_compute_units
= ret
;
1096 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
1098 return sizeof(uint32_t);
1100 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
1102 uint32_t *images_supported
= ret
;
1103 *images_supported
= 0;
1105 return sizeof(uint32_t);
1106 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
1108 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
1110 uint32_t *subgroup_size
= ret
;
1111 *subgroup_size
= 64;
1113 return sizeof(uint32_t);
1114 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK
:
1116 uint64_t *max_variable_threads_per_block
= ret
;
1117 if (ir_type
== PIPE_SHADER_IR_TGSI
)
1118 *max_variable_threads_per_block
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
1120 *max_variable_threads_per_block
= 0;
1122 return sizeof(uint64_t);
1125 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
1129 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
1131 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1133 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
1134 rscreen
->info
.clock_crystal_freq
;
1137 static void r600_fence_reference(struct pipe_screen
*screen
,
1138 struct pipe_fence_handle
**dst
,
1139 struct pipe_fence_handle
*src
)
1141 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
1142 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
1143 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
1145 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
1146 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
1147 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
1153 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
1154 struct pipe_context
*ctx
,
1155 struct pipe_fence_handle
*fence
,
1158 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
1159 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
1160 struct r600_common_context
*rctx
;
1161 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
1163 ctx
= threaded_context_unwrap_sync(ctx
);
1164 rctx
= ctx
? (struct r600_common_context
*)ctx
: NULL
;
1167 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
1170 /* Recompute the timeout after waiting. */
1171 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1172 int64_t time
= os_time_get_nano();
1173 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1180 /* Flush the gfx IB if it hasn't been flushed yet. */
1182 rfence
->gfx_unflushed
.ctx
== rctx
&&
1183 rfence
->gfx_unflushed
.ib_index
== rctx
->num_gfx_cs_flushes
) {
1184 rctx
->gfx
.flush(rctx
, timeout
? 0 : RADEON_FLUSH_ASYNC
, NULL
);
1185 rfence
->gfx_unflushed
.ctx
= NULL
;
1190 /* Recompute the timeout after all that. */
1191 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1192 int64_t time
= os_time_get_nano();
1193 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1197 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
1200 static void r600_query_memory_info(struct pipe_screen
*screen
,
1201 struct pipe_memory_info
*info
)
1203 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1204 struct radeon_winsys
*ws
= rscreen
->ws
;
1205 unsigned vram_usage
, gtt_usage
;
1207 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
1208 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
1210 /* The real TTM memory usage is somewhat random, because:
1212 * 1) TTM delays freeing memory, because it can only free it after
1215 * 2) The memory usage can be really low if big VRAM evictions are
1216 * taking place, but the real usage is well above the size of VRAM.
1218 * Instead, return statistics of this process.
1220 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
1221 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
1223 info
->avail_device_memory
=
1224 vram_usage
<= info
->total_device_memory
?
1225 info
->total_device_memory
- vram_usage
: 0;
1226 info
->avail_staging_memory
=
1227 gtt_usage
<= info
->total_staging_memory
?
1228 info
->total_staging_memory
- gtt_usage
: 0;
1230 info
->device_memory_evicted
=
1231 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
1233 if (rscreen
->info
.drm_major
== 3 && rscreen
->info
.drm_minor
>= 4)
1234 info
->nr_device_memory_evictions
=
1235 ws
->query_value(ws
, RADEON_NUM_EVICTIONS
);
1237 /* Just return the number of evicted 64KB pages. */
1238 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
1241 struct pipe_resource
*si_resource_create_common(struct pipe_screen
*screen
,
1242 const struct pipe_resource
*templ
)
1244 if (templ
->target
== PIPE_BUFFER
) {
1245 return si_buffer_create(screen
, templ
, 256);
1247 return si_texture_create(screen
, templ
);
1251 bool si_common_screen_init(struct r600_common_screen
*rscreen
,
1252 struct radeon_winsys
*ws
)
1254 char family_name
[32] = {}, llvm_string
[32] = {}, kernel_version
[128] = {};
1255 struct utsname uname_data
;
1256 const char *chip_name
;
1258 ws
->query_info(ws
, &rscreen
->info
);
1261 if ((chip_name
= r600_get_marketing_name(ws
)))
1262 snprintf(family_name
, sizeof(family_name
), "%s / ",
1263 r600_get_family_name(rscreen
) + 4);
1265 chip_name
= r600_get_family_name(rscreen
);
1267 if (uname(&uname_data
) == 0)
1268 snprintf(kernel_version
, sizeof(kernel_version
),
1269 " / %s", uname_data
.release
);
1271 if (HAVE_LLVM
> 0) {
1272 snprintf(llvm_string
, sizeof(llvm_string
),
1273 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
1274 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
1277 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
1278 "%s (%sDRM %i.%i.%i%s%s)",
1279 chip_name
, family_name
, rscreen
->info
.drm_major
,
1280 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
1281 kernel_version
, llvm_string
);
1283 rscreen
->b
.get_name
= r600_get_name
;
1284 rscreen
->b
.get_vendor
= r600_get_vendor
;
1285 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
1286 rscreen
->b
.get_disk_shader_cache
= r600_get_disk_shader_cache
;
1287 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
1288 rscreen
->b
.get_paramf
= r600_get_paramf
;
1289 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
1290 rscreen
->b
.fence_finish
= r600_fence_finish
;
1291 rscreen
->b
.fence_reference
= r600_fence_reference
;
1292 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
1293 rscreen
->b
.resource_from_user_memory
= si_buffer_from_user_memory
;
1294 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
1296 if (rscreen
->info
.has_hw_decode
) {
1297 rscreen
->b
.get_video_param
= si_vid_get_video_param
;
1298 rscreen
->b
.is_video_format_supported
= si_vid_is_format_supported
;
1300 rscreen
->b
.get_video_param
= r600_get_video_param
;
1301 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
1304 si_init_screen_texture_functions(rscreen
);
1305 si_init_screen_query_functions(rscreen
);
1307 rscreen
->family
= rscreen
->info
.family
;
1308 rscreen
->chip_class
= rscreen
->info
.chip_class
;
1309 rscreen
->debug_flags
|= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
1310 rscreen
->has_rbplus
= false;
1311 rscreen
->rbplus_allowed
= false;
1313 r600_disk_cache_create(rscreen
);
1315 slab_create_parent(&rscreen
->pool_transfers
, sizeof(struct r600_transfer
), 64);
1317 rscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1318 if (rscreen
->force_aniso
>= 0) {
1319 printf("radeon: Forcing anisotropy filter to %ix\n",
1320 /* round down to a power of two */
1321 1 << util_logbase2(rscreen
->force_aniso
));
1324 (void) mtx_init(&rscreen
->aux_context_lock
, mtx_plain
);
1325 (void) mtx_init(&rscreen
->gpu_load_mutex
, mtx_plain
);
1327 if (rscreen
->debug_flags
& DBG(INFO
)) {
1328 printf("pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
1329 rscreen
->info
.pci_domain
, rscreen
->info
.pci_bus
,
1330 rscreen
->info
.pci_dev
, rscreen
->info
.pci_func
);
1331 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
1332 printf("family = %i (%s)\n", rscreen
->info
.family
,
1333 r600_get_family_name(rscreen
));
1334 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
1335 printf("pte_fragment_size = %u\n", rscreen
->info
.pte_fragment_size
);
1336 printf("gart_page_size = %u\n", rscreen
->info
.gart_page_size
);
1337 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
1338 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
1339 printf("vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_vis_size
, 1024*1024));
1340 printf("max_alloc_size = %i MB\n",
1341 (int)DIV_ROUND_UP(rscreen
->info
.max_alloc_size
, 1024*1024));
1342 printf("min_alloc_size = %u\n", rscreen
->info
.min_alloc_size
);
1343 printf("has_dedicated_vram = %u\n", rscreen
->info
.has_dedicated_vram
);
1344 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
1345 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
1346 printf("has_hw_decode = %u\n", rscreen
->info
.has_hw_decode
);
1347 printf("num_sdma_rings = %i\n", rscreen
->info
.num_sdma_rings
);
1348 printf("num_compute_rings = %u\n", rscreen
->info
.num_compute_rings
);
1349 printf("uvd_fw_version = %u\n", rscreen
->info
.uvd_fw_version
);
1350 printf("vce_fw_version = %u\n", rscreen
->info
.vce_fw_version
);
1351 printf("me_fw_version = %i\n", rscreen
->info
.me_fw_version
);
1352 printf("me_fw_feature = %i\n", rscreen
->info
.me_fw_feature
);
1353 printf("pfp_fw_version = %i\n", rscreen
->info
.pfp_fw_version
);
1354 printf("pfp_fw_feature = %i\n", rscreen
->info
.pfp_fw_feature
);
1355 printf("ce_fw_version = %i\n", rscreen
->info
.ce_fw_version
);
1356 printf("ce_fw_feature = %i\n", rscreen
->info
.ce_fw_feature
);
1357 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
1358 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
1359 printf("tcc_cache_line_size = %u\n", rscreen
->info
.tcc_cache_line_size
);
1360 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
1361 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
1362 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
1363 printf("has_syncobj = %u\n", rscreen
->info
.has_syncobj
);
1364 printf("has_sync_file = %u\n", rscreen
->info
.has_sync_file
);
1366 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
1367 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
1368 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
1369 printf("max_se = %i\n", rscreen
->info
.max_se
);
1370 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
1372 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
1373 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
1374 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
1375 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
1376 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
1377 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
1378 printf("enabled_rb_mask = 0x%x\n", rscreen
->info
.enabled_rb_mask
);
1379 printf("max_alignment = %u\n", (unsigned)rscreen
->info
.max_alignment
);
1384 void si_destroy_common_screen(struct r600_common_screen
*rscreen
)
1386 si_perfcounters_destroy(rscreen
);
1387 si_gpu_load_kill_thread(rscreen
);
1389 mtx_destroy(&rscreen
->gpu_load_mutex
);
1390 mtx_destroy(&rscreen
->aux_context_lock
);
1391 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
1393 slab_destroy_parent(&rscreen
->pool_transfers
);
1395 disk_cache_destroy(rscreen
->disk_shader_cache
);
1396 rscreen
->ws
->destroy(rscreen
->ws
);
1400 bool si_can_dump_shader(struct r600_common_screen
*rscreen
,
1403 return rscreen
->debug_flags
& (1 << processor
);
1406 bool si_extra_shader_checks(struct r600_common_screen
*rscreen
, unsigned processor
)
1408 return (rscreen
->debug_flags
& DBG(CHECK_IR
)) ||
1409 si_can_dump_shader(rscreen
, processor
);
1412 void si_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
1413 uint64_t offset
, uint64_t size
, unsigned value
)
1415 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
1417 mtx_lock(&rscreen
->aux_context_lock
);
1418 rctx
->dma_clear_buffer(&rctx
->b
, dst
, offset
, size
, value
);
1419 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
1420 mtx_unlock(&rscreen
->aux_context_lock
);