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>
47 #include <llvm-c/TargetMachine.h>
50 #ifndef MESA_LLVM_VERSION_PATCH
51 #define MESA_LLVM_VERSION_PATCH 0
54 struct r600_multi_fence
{
55 struct pipe_reference reference
;
56 struct pipe_fence_handle
*gfx
;
57 struct pipe_fence_handle
*sdma
;
59 /* If the context wasn't flushed at fence creation, this is non-NULL. */
61 struct r600_common_context
*ctx
;
67 * shader binary helpers.
69 void radeon_shader_binary_init(struct ac_shader_binary
*b
)
71 memset(b
, 0, sizeof(*b
));
74 void radeon_shader_binary_clean(struct ac_shader_binary
*b
)
81 FREE(b
->global_symbol_offsets
);
83 FREE(b
->disasm_string
);
84 FREE(b
->llvm_ir_string
);
94 * \param event EVENT_TYPE_*
95 * \param event_flags Optional cache flush flags (TC)
96 * \param data_sel 1 = fence, 3 = timestamp
98 * \param va GPU address
99 * \param old_value Previous fence value (for a bug workaround)
100 * \param new_value Fence value to write for this event.
102 void r600_gfx_write_event_eop(struct r600_common_context
*ctx
,
103 unsigned event
, unsigned event_flags
,
105 struct r600_resource
*buf
, uint64_t va
,
106 uint32_t old_fence
, uint32_t new_fence
)
108 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
109 unsigned op
= EVENT_TYPE(event
) |
113 if (ctx
->chip_class
>= GFX9
) {
114 radeon_emit(cs
, PKT3(PKT3_RELEASE_MEM
, 6, 0));
116 radeon_emit(cs
, EOP_DATA_SEL(data_sel
));
117 radeon_emit(cs
, va
); /* address lo */
118 radeon_emit(cs
, va
>> 32); /* address hi */
119 radeon_emit(cs
, new_fence
); /* immediate data lo */
120 radeon_emit(cs
, 0); /* immediate data hi */
121 radeon_emit(cs
, 0); /* unused */
123 if (ctx
->chip_class
== CIK
||
124 ctx
->chip_class
== VI
) {
125 /* Two EOP events are required to make all engines go idle
126 * (and optional cache flushes executed) before the timestamp
129 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
132 radeon_emit(cs
, ((va
>> 32) & 0xffff) | EOP_DATA_SEL(data_sel
));
133 radeon_emit(cs
, old_fence
); /* immediate data */
134 radeon_emit(cs
, 0); /* unused */
137 radeon_emit(cs
, PKT3(PKT3_EVENT_WRITE_EOP
, 4, 0));
140 radeon_emit(cs
, ((va
>> 32) & 0xffff) | EOP_DATA_SEL(data_sel
));
141 radeon_emit(cs
, new_fence
); /* immediate data */
142 radeon_emit(cs
, 0); /* unused */
146 r600_emit_reloc(ctx
, &ctx
->gfx
, buf
, RADEON_USAGE_WRITE
,
150 unsigned r600_gfx_write_fence_dwords(struct r600_common_screen
*screen
)
154 if (screen
->chip_class
== CIK
||
155 screen
->chip_class
== VI
)
158 if (!screen
->info
.has_virtual_memory
)
164 void r600_gfx_wait_fence(struct r600_common_context
*ctx
,
165 uint64_t va
, uint32_t ref
, uint32_t mask
)
167 struct radeon_winsys_cs
*cs
= ctx
->gfx
.cs
;
169 radeon_emit(cs
, PKT3(PKT3_WAIT_REG_MEM
, 5, 0));
170 radeon_emit(cs
, WAIT_REG_MEM_EQUAL
| WAIT_REG_MEM_MEM_SPACE(1));
172 radeon_emit(cs
, va
>> 32);
173 radeon_emit(cs
, ref
); /* reference value */
174 radeon_emit(cs
, mask
); /* mask */
175 radeon_emit(cs
, 4); /* poll interval */
178 void r600_draw_rectangle(struct blitter_context
*blitter
,
179 int x1
, int y1
, int x2
, int y2
, float depth
,
180 enum blitter_attrib_type type
,
181 const union pipe_color_union
*attrib
)
183 struct r600_common_context
*rctx
=
184 (struct r600_common_context
*)util_blitter_get_pipe(blitter
);
185 struct pipe_viewport_state viewport
;
186 struct pipe_resource
*buf
= NULL
;
190 if (type
== UTIL_BLITTER_ATTRIB_TEXCOORD
) {
191 util_blitter_draw_rectangle(blitter
, x1
, y1
, x2
, y2
, depth
, type
, attrib
);
195 /* Some operations (like color resolve on r6xx) don't work
196 * with the conventional primitive types.
197 * One that works is PT_RECTLIST, which we use here. */
200 viewport
.scale
[0] = 1.0f
;
201 viewport
.scale
[1] = 1.0f
;
202 viewport
.scale
[2] = 1.0f
;
203 viewport
.translate
[0] = 0.0f
;
204 viewport
.translate
[1] = 0.0f
;
205 viewport
.translate
[2] = 0.0f
;
206 rctx
->b
.set_viewport_states(&rctx
->b
, 0, 1, &viewport
);
208 /* Upload vertices. The hw rectangle has only 3 vertices,
209 * I guess the 4th one is derived from the first 3.
210 * The vertex specification should match u_blitter's vertex element state. */
211 u_upload_alloc(rctx
->b
.stream_uploader
, 0, sizeof(float) * 24,
212 rctx
->screen
->info
.tcc_cache_line_size
,
213 &offset
, &buf
, (void**)&vb
);
233 memcpy(vb
+4, attrib
->f
, sizeof(float)*4);
234 memcpy(vb
+12, attrib
->f
, sizeof(float)*4);
235 memcpy(vb
+20, attrib
->f
, sizeof(float)*4);
239 util_draw_vertex_buffer(&rctx
->b
, NULL
, buf
, blitter
->vb_slot
, offset
,
240 R600_PRIM_RECTANGLE_LIST
, 3, 2);
241 pipe_resource_reference(&buf
, NULL
);
244 static void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
246 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
248 /* NOP waits for idle on Evergreen and later. */
249 if (rctx
->chip_class
>= CIK
)
250 radeon_emit(cs
, 0x00000000); /* NOP */
251 else if (rctx
->chip_class
>= EVERGREEN
)
252 radeon_emit(cs
, 0xf0000000); /* NOP */
254 /* TODO: R600-R700 should use the FENCE packet.
255 * CS checker support is required. */
259 void r600_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
,
260 struct r600_resource
*dst
, struct r600_resource
*src
)
262 uint64_t vram
= ctx
->dma
.cs
->used_vram
;
263 uint64_t gtt
= ctx
->dma
.cs
->used_gart
;
266 vram
+= dst
->vram_usage
;
267 gtt
+= dst
->gart_usage
;
270 vram
+= src
->vram_usage
;
271 gtt
+= src
->gart_usage
;
274 /* Flush the GFX IB if DMA depends on it. */
275 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
277 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
278 RADEON_USAGE_READWRITE
)) ||
280 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
281 RADEON_USAGE_WRITE
))))
282 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
284 /* Flush if there's not enough space, or if the memory usage per IB
287 * IBs using too little memory are limited by the IB submission overhead.
288 * IBs using too much memory are limited by the kernel/TTM overhead.
289 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
291 * This heuristic makes sure that DMA requests are executed
292 * very soon after the call is made and lowers memory usage.
293 * It improves texture upload performance by keeping the DMA
294 * engine busy while uploads are being submitted.
296 num_dw
++; /* for emit_wait_idle below */
297 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
298 ctx
->dma
.cs
->used_vram
+ ctx
->dma
.cs
->used_gart
> 64 * 1024 * 1024 ||
299 !radeon_cs_memory_below_limit(ctx
->screen
, ctx
->dma
.cs
, vram
, gtt
)) {
300 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
301 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
304 /* Wait for idle if either buffer has been used in the IB before to
305 * prevent read-after-write hazards.
308 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, dst
->buf
,
309 RADEON_USAGE_READWRITE
)) ||
311 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
, src
->buf
,
312 RADEON_USAGE_WRITE
)))
313 r600_dma_emit_wait_idle(ctx
);
315 /* If GPUVM is not supported, the CS checker needs 2 entries
316 * in the buffer list per packet, which has to be done manually.
318 if (ctx
->screen
->info
.has_virtual_memory
) {
320 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
322 RADEON_PRIO_SDMA_BUFFER
);
324 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
326 RADEON_PRIO_SDMA_BUFFER
);
329 /* this function is called before all DMA calls, so increment this. */
330 ctx
->num_dma_calls
++;
333 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
337 void r600_preflush_suspend_features(struct r600_common_context
*ctx
)
339 /* suspend queries */
340 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
341 r600_suspend_queries(ctx
);
343 ctx
->streamout
.suspended
= false;
344 if (ctx
->streamout
.begin_emitted
) {
345 r600_emit_streamout_end(ctx
);
346 ctx
->streamout
.suspended
= true;
350 void r600_postflush_resume_features(struct r600_common_context
*ctx
)
352 if (ctx
->streamout
.suspended
) {
353 ctx
->streamout
.append_bitmask
= ctx
->streamout
.enabled_mask
;
354 r600_streamout_buffers_dirty(ctx
);
358 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
359 r600_resume_queries(ctx
);
362 static void r600_flush_from_st(struct pipe_context
*ctx
,
363 struct pipe_fence_handle
**fence
,
366 struct pipe_screen
*screen
= ctx
->screen
;
367 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
368 struct radeon_winsys
*ws
= rctx
->ws
;
369 struct pipe_fence_handle
*gfx_fence
= NULL
;
370 struct pipe_fence_handle
*sdma_fence
= NULL
;
371 bool deferred_fence
= false;
372 unsigned rflags
= RADEON_FLUSH_ASYNC
;
374 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
375 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
377 /* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
379 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
381 if (!radeon_emitted(rctx
->gfx
.cs
, rctx
->initial_gfx_cs_size
)) {
383 ws
->fence_reference(&gfx_fence
, rctx
->last_gfx_fence
);
384 if (!(flags
& PIPE_FLUSH_DEFERRED
))
385 ws
->cs_sync_flush(rctx
->gfx
.cs
);
387 /* Instead of flushing, create a deferred fence. Constraints:
388 * - The state tracker must allow a deferred flush.
389 * - The state tracker must request a fence.
390 * Thread safety in fence_finish must be ensured by the state tracker.
392 if (flags
& PIPE_FLUSH_DEFERRED
&& fence
) {
393 gfx_fence
= rctx
->ws
->cs_get_next_fence(rctx
->gfx
.cs
);
394 deferred_fence
= true;
396 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
400 /* Both engines can signal out of order, so we need to keep both fences. */
402 struct r600_multi_fence
*multi_fence
=
403 CALLOC_STRUCT(r600_multi_fence
);
407 multi_fence
->reference
.count
= 1;
408 /* If both fences are NULL, fence_finish will always return true. */
409 multi_fence
->gfx
= gfx_fence
;
410 multi_fence
->sdma
= sdma_fence
;
412 if (deferred_fence
) {
413 multi_fence
->gfx_unflushed
.ctx
= rctx
;
414 multi_fence
->gfx_unflushed
.ib_index
= rctx
->num_gfx_cs_flushes
;
417 screen
->fence_reference(screen
, fence
, NULL
);
418 *fence
= (struct pipe_fence_handle
*)multi_fence
;
421 if (!(flags
& PIPE_FLUSH_DEFERRED
)) {
423 ws
->cs_sync_flush(rctx
->dma
.cs
);
424 ws
->cs_sync_flush(rctx
->gfx
.cs
);
428 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
429 struct pipe_fence_handle
**fence
)
431 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
432 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
433 struct radeon_saved_cs saved
;
435 (rctx
->screen
->debug_flags
& DBG_CHECK_VM
) &&
436 rctx
->check_vm_faults
;
438 if (!radeon_emitted(cs
, 0)) {
440 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
445 radeon_save_cs(rctx
->ws
, cs
, &saved
);
447 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
449 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
452 /* Use conservative timeout 800ms, after which we won't wait any
453 * longer and assume the GPU is hung.
455 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
457 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
458 radeon_clear_saved_cs(&saved
);
463 * Store a linearized copy of all chunks of \p cs together with the buffer
466 void radeon_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
467 struct radeon_saved_cs
*saved
)
472 /* Save the IB chunks. */
473 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
474 saved
->ib
= MALLOC(4 * saved
->num_dw
);
479 for (i
= 0; i
< cs
->num_prev
; ++i
) {
480 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
481 buf
+= cs
->prev
[i
].cdw
;
483 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
485 /* Save the buffer list. */
486 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
487 saved
->bo_list
= CALLOC(saved
->bo_count
,
488 sizeof(saved
->bo_list
[0]));
489 if (!saved
->bo_list
) {
493 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
498 fprintf(stderr
, "%s: out of memory\n", __func__
);
499 memset(saved
, 0, sizeof(*saved
));
502 void radeon_clear_saved_cs(struct radeon_saved_cs
*saved
)
505 FREE(saved
->bo_list
);
507 memset(saved
, 0, sizeof(*saved
));
510 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
512 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
513 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
514 RADEON_GPU_RESET_COUNTER
);
516 if (rctx
->gpu_reset_counter
== latest
)
517 return PIPE_NO_RESET
;
519 rctx
->gpu_reset_counter
= latest
;
520 return PIPE_UNKNOWN_CONTEXT_RESET
;
523 static void r600_set_debug_callback(struct pipe_context
*ctx
,
524 const struct pipe_debug_callback
*cb
)
526 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
531 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
534 static void r600_set_device_reset_callback(struct pipe_context
*ctx
,
535 const struct pipe_device_reset_callback
*cb
)
537 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
540 rctx
->device_reset_callback
= *cb
;
542 memset(&rctx
->device_reset_callback
, 0,
543 sizeof(rctx
->device_reset_callback
));
546 bool r600_check_device_reset(struct r600_common_context
*rctx
)
548 enum pipe_reset_status status
;
550 if (!rctx
->device_reset_callback
.reset
)
553 if (!rctx
->b
.get_device_reset_status
)
556 status
= rctx
->b
.get_device_reset_status(&rctx
->b
);
557 if (status
== PIPE_NO_RESET
)
560 rctx
->device_reset_callback
.reset(rctx
->device_reset_callback
.data
, status
);
564 static void r600_dma_clear_buffer_fallback(struct pipe_context
*ctx
,
565 struct pipe_resource
*dst
,
566 uint64_t offset
, uint64_t size
,
569 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
571 rctx
->clear_buffer(ctx
, dst
, offset
, size
, value
, R600_COHERENCY_NONE
);
574 static bool r600_resource_commit(struct pipe_context
*pctx
,
575 struct pipe_resource
*resource
,
576 unsigned level
, struct pipe_box
*box
,
579 struct r600_common_context
*ctx
= (struct r600_common_context
*)pctx
;
580 struct r600_resource
*res
= r600_resource(resource
);
583 * Since buffer commitment changes cannot be pipelined, we need to
584 * (a) flush any pending commands that refer to the buffer we're about
586 * (b) wait for threaded submit to finish, including those that were
587 * triggered by some other, earlier operation.
589 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
590 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
,
591 res
->buf
, RADEON_USAGE_READWRITE
)) {
592 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
594 if (radeon_emitted(ctx
->dma
.cs
, 0) &&
595 ctx
->ws
->cs_is_buffer_referenced(ctx
->dma
.cs
,
596 res
->buf
, RADEON_USAGE_READWRITE
)) {
597 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
600 ctx
->ws
->cs_sync_flush(ctx
->dma
.cs
);
601 ctx
->ws
->cs_sync_flush(ctx
->gfx
.cs
);
603 assert(resource
->target
== PIPE_BUFFER
);
605 return ctx
->ws
->buffer_commit(res
->buf
, box
->x
, box
->width
, commit
);
608 bool r600_common_context_init(struct r600_common_context
*rctx
,
609 struct r600_common_screen
*rscreen
,
610 unsigned context_flags
)
612 slab_create_child(&rctx
->pool_transfers
, &rscreen
->pool_transfers
);
614 rctx
->screen
= rscreen
;
615 rctx
->ws
= rscreen
->ws
;
616 rctx
->family
= rscreen
->family
;
617 rctx
->chip_class
= rscreen
->chip_class
;
619 rctx
->b
.invalidate_resource
= r600_invalidate_resource
;
620 rctx
->b
.resource_commit
= r600_resource_commit
;
621 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
622 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
623 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
624 rctx
->b
.texture_subdata
= u_default_texture_subdata
;
625 rctx
->b
.memory_barrier
= r600_memory_barrier
;
626 rctx
->b
.flush
= r600_flush_from_st
;
627 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
628 rctx
->dma_clear_buffer
= r600_dma_clear_buffer_fallback
;
630 /* evergreen_compute.c has a special codepath for global buffers.
631 * Everything else can use the direct path.
633 if ((rscreen
->chip_class
== EVERGREEN
|| rscreen
->chip_class
== CAYMAN
) &&
634 (context_flags
& PIPE_CONTEXT_COMPUTE_ONLY
))
635 rctx
->b
.buffer_subdata
= u_default_buffer_subdata
;
637 rctx
->b
.buffer_subdata
= r600_buffer_subdata
;
639 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
640 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
641 rctx
->gpu_reset_counter
=
642 rctx
->ws
->query_value(rctx
->ws
,
643 RADEON_GPU_RESET_COUNTER
);
646 rctx
->b
.set_device_reset_callback
= r600_set_device_reset_callback
;
648 r600_init_context_texture_functions(rctx
);
649 r600_init_viewport_functions(rctx
);
650 r600_streamout_init(rctx
);
651 r600_query_init(rctx
);
652 cayman_init_msaa(&rctx
->b
);
654 rctx
->allocator_zeroed_memory
=
655 u_suballocator_create(&rctx
->b
, rscreen
->info
.gart_page_size
,
656 0, PIPE_USAGE_DEFAULT
, 0, true);
657 if (!rctx
->allocator_zeroed_memory
)
660 rctx
->b
.stream_uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
661 0, PIPE_USAGE_STREAM
);
662 if (!rctx
->b
.stream_uploader
)
665 rctx
->b
.const_uploader
= u_upload_create(&rctx
->b
, 128 * 1024,
666 0, PIPE_USAGE_DEFAULT
);
667 if (!rctx
->b
.const_uploader
)
670 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
674 if (rscreen
->info
.has_sdma
&& !(rscreen
->debug_flags
& DBG_NO_ASYNC_DMA
)) {
675 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
678 rctx
->dma
.flush
= r600_flush_dma_ring
;
684 void r600_common_context_cleanup(struct r600_common_context
*rctx
)
688 /* Release DCC stats. */
689 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
690 assert(!rctx
->dcc_stats
[i
].query_active
);
692 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
693 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
694 rctx
->b
.destroy_query(&rctx
->b
,
695 rctx
->dcc_stats
[i
].ps_stats
[j
]);
697 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
700 if (rctx
->query_result_shader
)
701 rctx
->b
.delete_compute_state(&rctx
->b
, rctx
->query_result_shader
);
704 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
706 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
708 rctx
->ws
->ctx_destroy(rctx
->ctx
);
710 if (rctx
->b
.stream_uploader
)
711 u_upload_destroy(rctx
->b
.stream_uploader
);
712 if (rctx
->b
.const_uploader
)
713 u_upload_destroy(rctx
->b
.const_uploader
);
715 slab_destroy_child(&rctx
->pool_transfers
);
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
);
728 static const struct debug_named_value common_debug_options
[] = {
730 { "tex", DBG_TEX
, "Print texture info" },
731 { "compute", DBG_COMPUTE
, "Print compute info" },
732 { "vm", DBG_VM
, "Print virtual addresses when creating resources" },
733 { "info", DBG_INFO
, "Print driver information" },
736 { "fs", DBG_FS
, "Print fetch shaders" },
737 { "vs", DBG_VS
, "Print vertex shaders" },
738 { "gs", DBG_GS
, "Print geometry shaders" },
739 { "ps", DBG_PS
, "Print pixel shaders" },
740 { "cs", DBG_CS
, "Print compute shaders" },
741 { "tcs", DBG_TCS
, "Print tessellation control shaders" },
742 { "tes", DBG_TES
, "Print tessellation evaluation shaders" },
743 { "noir", DBG_NO_IR
, "Don't print the LLVM IR"},
744 { "notgsi", DBG_NO_TGSI
, "Don't print the TGSI"},
745 { "noasm", DBG_NO_ASM
, "Don't print disassembled shaders"},
746 { "preoptir", DBG_PREOPT_IR
, "Print the LLVM IR before initial optimizations" },
747 { "checkir", DBG_CHECK_IR
, "Enable additional sanity checks on shader IR" },
748 { "nooptvariant", DBG_NO_OPT_VARIANT
, "Disable compiling optimized shader variants." },
750 { "testdma", DBG_TEST_DMA
, "Invoke SDMA tests and exit." },
751 { "testvmfaultcp", DBG_TEST_VMFAULT_CP
, "Invoke a CP VM fault test and exit." },
752 { "testvmfaultsdma", DBG_TEST_VMFAULT_SDMA
, "Invoke a SDMA VM fault test and exit." },
753 { "testvmfaultshader", DBG_TEST_VMFAULT_SHADER
, "Invoke a shader VM fault test and exit." },
756 { "nodma", DBG_NO_ASYNC_DMA
, "Disable asynchronous DMA" },
757 { "nohyperz", DBG_NO_HYPERZ
, "Disable Hyper-Z" },
758 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
759 { "noinvalrange", DBG_NO_DISCARD_RANGE
, "Disable handling of INVALIDATE_RANGE map flags" },
760 { "no2d", DBG_NO_2D_TILING
, "Disable 2D tiling" },
761 { "notiling", DBG_NO_TILING
, "Disable tiling" },
762 { "switch_on_eop", DBG_SWITCH_ON_EOP
, "Program WD/IA to switch on end-of-packet." },
763 { "forcedma", DBG_FORCE_DMA
, "Use asynchronous DMA for all operations when possible." },
764 { "precompile", DBG_PRECOMPILE
, "Compile one shader variant at shader creation." },
765 { "nowc", DBG_NO_WC
, "Disable GTT write combining" },
766 { "check_vm", DBG_CHECK_VM
, "Check VM faults and dump debug info." },
767 { "nodcc", DBG_NO_DCC
, "Disable DCC." },
768 { "nodccclear", DBG_NO_DCC_CLEAR
, "Disable DCC fast clear." },
769 { "norbplus", DBG_NO_RB_PLUS
, "Disable RB+." },
770 { "sisched", DBG_SI_SCHED
, "Enable LLVM SI Machine Instruction Scheduler." },
771 { "mono", DBG_MONOLITHIC_SHADERS
, "Use old-style monolithic shaders compiled on demand" },
772 { "noce", DBG_NO_CE
, "Disable the constant engine"},
773 { "unsafemath", DBG_UNSAFE_MATH
, "Enable unsafe math shader optimizations" },
774 { "nodccfb", DBG_NO_DCC_FB
, "Disable separate DCC on the main framebuffer" },
776 DEBUG_NAMED_VALUE_END
/* must be last */
779 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
784 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
789 static const char* r600_get_chip_name(struct r600_common_screen
*rscreen
)
791 switch (rscreen
->info
.family
) {
792 case CHIP_R600
: return "AMD R600";
793 case CHIP_RV610
: return "AMD RV610";
794 case CHIP_RV630
: return "AMD RV630";
795 case CHIP_RV670
: return "AMD RV670";
796 case CHIP_RV620
: return "AMD RV620";
797 case CHIP_RV635
: return "AMD RV635";
798 case CHIP_RS780
: return "AMD RS780";
799 case CHIP_RS880
: return "AMD RS880";
800 case CHIP_RV770
: return "AMD RV770";
801 case CHIP_RV730
: return "AMD RV730";
802 case CHIP_RV710
: return "AMD RV710";
803 case CHIP_RV740
: return "AMD RV740";
804 case CHIP_CEDAR
: return "AMD CEDAR";
805 case CHIP_REDWOOD
: return "AMD REDWOOD";
806 case CHIP_JUNIPER
: return "AMD JUNIPER";
807 case CHIP_CYPRESS
: return "AMD CYPRESS";
808 case CHIP_HEMLOCK
: return "AMD HEMLOCK";
809 case CHIP_PALM
: return "AMD PALM";
810 case CHIP_SUMO
: return "AMD SUMO";
811 case CHIP_SUMO2
: return "AMD SUMO2";
812 case CHIP_BARTS
: return "AMD BARTS";
813 case CHIP_TURKS
: return "AMD TURKS";
814 case CHIP_CAICOS
: return "AMD CAICOS";
815 case CHIP_CAYMAN
: return "AMD CAYMAN";
816 case CHIP_ARUBA
: return "AMD ARUBA";
817 case CHIP_TAHITI
: return "AMD TAHITI";
818 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
819 case CHIP_VERDE
: return "AMD CAPE VERDE";
820 case CHIP_OLAND
: return "AMD OLAND";
821 case CHIP_HAINAN
: return "AMD HAINAN";
822 case CHIP_BONAIRE
: return "AMD BONAIRE";
823 case CHIP_KAVERI
: return "AMD KAVERI";
824 case CHIP_KABINI
: return "AMD KABINI";
825 case CHIP_HAWAII
: return "AMD HAWAII";
826 case CHIP_MULLINS
: return "AMD MULLINS";
827 case CHIP_TONGA
: return "AMD TONGA";
828 case CHIP_ICELAND
: return "AMD ICELAND";
829 case CHIP_CARRIZO
: return "AMD CARRIZO";
830 case CHIP_FIJI
: return "AMD FIJI";
831 case CHIP_POLARIS10
: return "AMD POLARIS10";
832 case CHIP_POLARIS11
: return "AMD POLARIS11";
833 case CHIP_POLARIS12
: return "AMD POLARIS12";
834 case CHIP_STONEY
: return "AMD STONEY";
835 case CHIP_VEGA10
: return "AMD VEGA10";
836 default: return "AMD unknown";
840 static void r600_disk_cache_create(struct r600_common_screen
*rscreen
)
842 /* Don't use the cache if shader dumping is enabled. */
843 if (rscreen
->debug_flags
&
844 (DBG_FS
| DBG_VS
| DBG_TCS
| DBG_TES
| DBG_GS
| DBG_PS
| DBG_CS
))
847 uint32_t mesa_timestamp
;
848 if (disk_cache_get_function_timestamp(r600_disk_cache_create
,
852 if (rscreen
->chip_class
< SI
) {
853 res
= asprintf(×tamp_str
, "%u",mesa_timestamp
);
857 uint32_t llvm_timestamp
;
858 if (disk_cache_get_function_timestamp(LLVMInitializeAMDGPUTargetInfo
,
860 res
= asprintf(×tamp_str
, "%u_%u",
861 mesa_timestamp
, llvm_timestamp
);
866 rscreen
->disk_shader_cache
=
867 disk_cache_create(r600_get_chip_name(rscreen
),
874 static struct disk_cache
*r600_get_disk_shader_cache(struct pipe_screen
*pscreen
)
876 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
877 return rscreen
->disk_shader_cache
;
880 static const char* r600_get_name(struct pipe_screen
* pscreen
)
882 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
884 return rscreen
->renderer_string
;
887 static float r600_get_paramf(struct pipe_screen
* pscreen
,
888 enum pipe_capf param
)
890 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
893 case PIPE_CAPF_MAX_LINE_WIDTH
:
894 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
895 case PIPE_CAPF_MAX_POINT_WIDTH
:
896 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
897 if (rscreen
->family
>= CHIP_CEDAR
)
901 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
903 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
905 case PIPE_CAPF_GUARD_BAND_LEFT
:
906 case PIPE_CAPF_GUARD_BAND_TOP
:
907 case PIPE_CAPF_GUARD_BAND_RIGHT
:
908 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
914 static int r600_get_video_param(struct pipe_screen
*screen
,
915 enum pipe_video_profile profile
,
916 enum pipe_video_entrypoint entrypoint
,
917 enum pipe_video_cap param
)
920 case PIPE_VIDEO_CAP_SUPPORTED
:
921 return vl_profile_supported(screen
, profile
, entrypoint
);
922 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
924 case PIPE_VIDEO_CAP_MAX_WIDTH
:
925 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
926 return vl_video_buffer_max_size(screen
);
927 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
928 return PIPE_FORMAT_NV12
;
929 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
931 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
933 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
935 case PIPE_VIDEO_CAP_MAX_LEVEL
:
936 return vl_level_supported(screen
, profile
);
942 const char *r600_get_llvm_processor_name(enum radeon_family family
)
985 case CHIP_TAHITI
: return "tahiti";
986 case CHIP_PITCAIRN
: return "pitcairn";
987 case CHIP_VERDE
: return "verde";
988 case CHIP_OLAND
: return "oland";
989 case CHIP_HAINAN
: return "hainan";
990 case CHIP_BONAIRE
: return "bonaire";
991 case CHIP_KABINI
: return "kabini";
992 case CHIP_KAVERI
: return "kaveri";
993 case CHIP_HAWAII
: return "hawaii";
996 case CHIP_TONGA
: return "tonga";
997 case CHIP_ICELAND
: return "iceland";
998 case CHIP_CARRIZO
: return "carrizo";
1003 case CHIP_POLARIS10
:
1005 case CHIP_POLARIS11
:
1006 case CHIP_POLARIS12
: /* same as polaris11 */
1015 static int r600_get_compute_param(struct pipe_screen
*screen
,
1016 enum pipe_shader_ir ir_type
,
1017 enum pipe_compute_cap param
,
1020 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1022 //TODO: select these params by asic
1024 case PIPE_COMPUTE_CAP_IR_TARGET
: {
1027 if (rscreen
->family
<= CHIP_ARUBA
) {
1030 if (HAVE_LLVM
< 0x0400) {
1031 triple
= "amdgcn--";
1033 triple
= "amdgcn-mesa-mesa3d";
1036 switch(rscreen
->family
) {
1037 /* Clang < 3.6 is missing Hainan in its list of
1038 * GPUs, so we need to use the name of a similar GPU.
1041 gpu
= r600_get_llvm_processor_name(rscreen
->family
);
1045 sprintf(ret
, "%s-%s", gpu
, triple
);
1047 /* +2 for dash and terminating NIL byte */
1048 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
1050 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
1052 uint64_t *grid_dimension
= ret
;
1053 grid_dimension
[0] = 3;
1055 return 1 * sizeof(uint64_t);
1057 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
1059 uint64_t *grid_size
= ret
;
1060 grid_size
[0] = 65535;
1061 grid_size
[1] = 65535;
1062 grid_size
[2] = 65535;
1064 return 3 * sizeof(uint64_t) ;
1066 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
1068 uint64_t *block_size
= ret
;
1069 if (rscreen
->chip_class
>= SI
&&
1070 ir_type
== PIPE_SHADER_IR_TGSI
) {
1071 block_size
[0] = 2048;
1072 block_size
[1] = 2048;
1073 block_size
[2] = 2048;
1075 block_size
[0] = 256;
1076 block_size
[1] = 256;
1077 block_size
[2] = 256;
1080 return 3 * sizeof(uint64_t);
1082 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
1084 uint64_t *max_threads_per_block
= ret
;
1085 if (rscreen
->chip_class
>= SI
&&
1086 ir_type
== PIPE_SHADER_IR_TGSI
)
1087 *max_threads_per_block
= 2048;
1089 *max_threads_per_block
= 256;
1091 return sizeof(uint64_t);
1092 case PIPE_COMPUTE_CAP_ADDRESS_BITS
:
1094 uint32_t *address_bits
= ret
;
1095 address_bits
[0] = 32;
1096 if (rscreen
->chip_class
>= SI
)
1097 address_bits
[0] = 64;
1099 return 1 * sizeof(uint32_t);
1101 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
1103 uint64_t *max_global_size
= ret
;
1104 uint64_t max_mem_alloc_size
;
1106 r600_get_compute_param(screen
, ir_type
,
1107 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
1108 &max_mem_alloc_size
);
1110 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
1111 * 1/4 of the MAX_GLOBAL_SIZE. Since the
1112 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
1113 * make sure we never report more than
1114 * 4 * MAX_MEM_ALLOC_SIZE.
1116 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
1117 MAX2(rscreen
->info
.gart_size
,
1118 rscreen
->info
.vram_size
));
1120 return sizeof(uint64_t);
1122 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
1124 uint64_t *max_local_size
= ret
;
1125 /* Value reported by the closed source driver. */
1126 *max_local_size
= 32768;
1128 return sizeof(uint64_t);
1130 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
1132 uint64_t *max_input_size
= ret
;
1133 /* Value reported by the closed source driver. */
1134 *max_input_size
= 1024;
1136 return sizeof(uint64_t);
1138 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
1140 uint64_t *max_mem_alloc_size
= ret
;
1142 *max_mem_alloc_size
= rscreen
->info
.max_alloc_size
;
1144 return sizeof(uint64_t);
1146 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
1148 uint32_t *max_clock_frequency
= ret
;
1149 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
1151 return sizeof(uint32_t);
1153 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
1155 uint32_t *max_compute_units
= ret
;
1156 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
1158 return sizeof(uint32_t);
1160 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
1162 uint32_t *images_supported
= ret
;
1163 *images_supported
= 0;
1165 return sizeof(uint32_t);
1166 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
1168 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
1170 uint32_t *subgroup_size
= ret
;
1171 *subgroup_size
= r600_wavefront_size(rscreen
->family
);
1173 return sizeof(uint32_t);
1174 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK
:
1176 uint64_t *max_variable_threads_per_block
= ret
;
1177 if (rscreen
->chip_class
>= SI
&&
1178 ir_type
== PIPE_SHADER_IR_TGSI
)
1179 *max_variable_threads_per_block
= SI_MAX_VARIABLE_THREADS_PER_BLOCK
;
1181 *max_variable_threads_per_block
= 0;
1183 return sizeof(uint64_t);
1186 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
1190 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
1192 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1194 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
1195 rscreen
->info
.clock_crystal_freq
;
1198 static void r600_fence_reference(struct pipe_screen
*screen
,
1199 struct pipe_fence_handle
**dst
,
1200 struct pipe_fence_handle
*src
)
1202 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
1203 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
1204 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
1206 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
1207 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
1208 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
1214 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
1215 struct pipe_context
*ctx
,
1216 struct pipe_fence_handle
*fence
,
1219 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
1220 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
1221 struct r600_common_context
*rctx
=
1222 ctx
? (struct r600_common_context
*)ctx
: NULL
;
1223 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
1226 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
1229 /* Recompute the timeout after waiting. */
1230 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1231 int64_t time
= os_time_get_nano();
1232 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1239 /* Flush the gfx IB if it hasn't been flushed yet. */
1241 rfence
->gfx_unflushed
.ctx
== rctx
&&
1242 rfence
->gfx_unflushed
.ib_index
== rctx
->num_gfx_cs_flushes
) {
1243 rctx
->gfx
.flush(rctx
, timeout
? 0 : RADEON_FLUSH_ASYNC
, NULL
);
1244 rfence
->gfx_unflushed
.ctx
= NULL
;
1249 /* Recompute the timeout after all that. */
1250 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
1251 int64_t time
= os_time_get_nano();
1252 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
1256 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
1259 static void r600_query_memory_info(struct pipe_screen
*screen
,
1260 struct pipe_memory_info
*info
)
1262 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1263 struct radeon_winsys
*ws
= rscreen
->ws
;
1264 unsigned vram_usage
, gtt_usage
;
1266 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
1267 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
1269 /* The real TTM memory usage is somewhat random, because:
1271 * 1) TTM delays freeing memory, because it can only free it after
1274 * 2) The memory usage can be really low if big VRAM evictions are
1275 * taking place, but the real usage is well above the size of VRAM.
1277 * Instead, return statistics of this process.
1279 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
1280 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
1282 info
->avail_device_memory
=
1283 vram_usage
<= info
->total_device_memory
?
1284 info
->total_device_memory
- vram_usage
: 0;
1285 info
->avail_staging_memory
=
1286 gtt_usage
<= info
->total_staging_memory
?
1287 info
->total_staging_memory
- gtt_usage
: 0;
1289 info
->device_memory_evicted
=
1290 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
1292 if (rscreen
->info
.drm_major
== 3 && rscreen
->info
.drm_minor
>= 4)
1293 info
->nr_device_memory_evictions
=
1294 ws
->query_value(ws
, RADEON_NUM_EVICTIONS
);
1296 /* Just return the number of evicted 64KB pages. */
1297 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
1300 struct pipe_resource
*r600_resource_create_common(struct pipe_screen
*screen
,
1301 const struct pipe_resource
*templ
)
1303 if (templ
->target
== PIPE_BUFFER
) {
1304 return r600_buffer_create(screen
, templ
, 256);
1306 return r600_texture_create(screen
, templ
);
1310 bool r600_common_screen_init(struct r600_common_screen
*rscreen
,
1311 struct radeon_winsys
*ws
)
1313 char llvm_string
[32] = {}, kernel_version
[128] = {};
1314 struct utsname uname_data
;
1316 ws
->query_info(ws
, &rscreen
->info
);
1318 if (uname(&uname_data
) == 0)
1319 snprintf(kernel_version
, sizeof(kernel_version
),
1320 " / %s", uname_data
.release
);
1322 if (HAVE_LLVM
> 0) {
1323 snprintf(llvm_string
, sizeof(llvm_string
),
1324 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
1325 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
1328 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
1329 "%s (DRM %i.%i.%i%s%s)",
1330 r600_get_chip_name(rscreen
), rscreen
->info
.drm_major
,
1331 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
1332 kernel_version
, llvm_string
);
1334 rscreen
->b
.get_name
= r600_get_name
;
1335 rscreen
->b
.get_vendor
= r600_get_vendor
;
1336 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
1337 rscreen
->b
.get_disk_shader_cache
= r600_get_disk_shader_cache
;
1338 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
1339 rscreen
->b
.get_paramf
= r600_get_paramf
;
1340 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
1341 rscreen
->b
.fence_finish
= r600_fence_finish
;
1342 rscreen
->b
.fence_reference
= r600_fence_reference
;
1343 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
1344 rscreen
->b
.resource_from_user_memory
= r600_buffer_from_user_memory
;
1345 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
1347 if (rscreen
->info
.has_uvd
) {
1348 rscreen
->b
.get_video_param
= rvid_get_video_param
;
1349 rscreen
->b
.is_video_format_supported
= rvid_is_format_supported
;
1351 rscreen
->b
.get_video_param
= r600_get_video_param
;
1352 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
1355 r600_init_screen_texture_functions(rscreen
);
1356 r600_init_screen_query_functions(rscreen
);
1359 rscreen
->family
= rscreen
->info
.family
;
1360 rscreen
->chip_class
= rscreen
->info
.chip_class
;
1361 rscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
1362 rscreen
->has_rbplus
= false;
1363 rscreen
->rbplus_allowed
= false;
1365 r600_disk_cache_create(rscreen
);
1367 slab_create_parent(&rscreen
->pool_transfers
, sizeof(struct r600_transfer
), 64);
1369 rscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1370 if (rscreen
->force_aniso
>= 0) {
1371 printf("radeon: Forcing anisotropy filter to %ix\n",
1372 /* round down to a power of two */
1373 1 << util_logbase2(rscreen
->force_aniso
));
1376 util_format_s3tc_init();
1377 (void) mtx_init(&rscreen
->aux_context_lock
, mtx_plain
);
1378 (void) mtx_init(&rscreen
->gpu_load_mutex
, mtx_plain
);
1380 if (rscreen
->debug_flags
& DBG_INFO
) {
1381 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
1382 printf("family = %i (%s)\n", rscreen
->info
.family
,
1383 r600_get_chip_name(rscreen
));
1384 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
1385 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
1386 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
1387 printf("vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_vis_size
, 1024*1024));
1388 printf("max_alloc_size = %i MB\n",
1389 (int)DIV_ROUND_UP(rscreen
->info
.max_alloc_size
, 1024*1024));
1390 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
1391 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
1392 printf("has_sdma = %i\n", rscreen
->info
.has_sdma
);
1393 printf("has_uvd = %i\n", rscreen
->info
.has_uvd
);
1394 printf("me_fw_version = %i\n", rscreen
->info
.me_fw_version
);
1395 printf("pfp_fw_version = %i\n", rscreen
->info
.pfp_fw_version
);
1396 printf("ce_fw_version = %i\n", rscreen
->info
.ce_fw_version
);
1397 printf("vce_fw_version = %i\n", rscreen
->info
.vce_fw_version
);
1398 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
1399 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
1400 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
1401 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
1402 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
1404 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
1405 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
1406 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
1407 printf("max_se = %i\n", rscreen
->info
.max_se
);
1408 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
1410 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
1411 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
1412 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
1413 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
1414 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
1415 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
1416 printf("enabled_rb_mask = 0x%x\n", rscreen
->info
.enabled_rb_mask
);
1421 void r600_destroy_common_screen(struct r600_common_screen
*rscreen
)
1423 r600_perfcounters_destroy(rscreen
);
1424 r600_gpu_load_kill_thread(rscreen
);
1426 mtx_destroy(&rscreen
->gpu_load_mutex
);
1427 mtx_destroy(&rscreen
->aux_context_lock
);
1428 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
1430 slab_destroy_parent(&rscreen
->pool_transfers
);
1432 disk_cache_destroy(rscreen
->disk_shader_cache
);
1433 rscreen
->ws
->destroy(rscreen
->ws
);
1437 bool r600_can_dump_shader(struct r600_common_screen
*rscreen
,
1440 switch (processor
) {
1441 case PIPE_SHADER_VERTEX
:
1442 return (rscreen
->debug_flags
& DBG_VS
) != 0;
1443 case PIPE_SHADER_TESS_CTRL
:
1444 return (rscreen
->debug_flags
& DBG_TCS
) != 0;
1445 case PIPE_SHADER_TESS_EVAL
:
1446 return (rscreen
->debug_flags
& DBG_TES
) != 0;
1447 case PIPE_SHADER_GEOMETRY
:
1448 return (rscreen
->debug_flags
& DBG_GS
) != 0;
1449 case PIPE_SHADER_FRAGMENT
:
1450 return (rscreen
->debug_flags
& DBG_PS
) != 0;
1451 case PIPE_SHADER_COMPUTE
:
1452 return (rscreen
->debug_flags
& DBG_CS
) != 0;
1458 bool r600_extra_shader_checks(struct r600_common_screen
*rscreen
, unsigned processor
)
1460 return (rscreen
->debug_flags
& DBG_CHECK_IR
) ||
1461 r600_can_dump_shader(rscreen
, processor
);
1464 void r600_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
1465 uint64_t offset
, uint64_t size
, unsigned value
)
1467 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
1469 mtx_lock(&rscreen
->aux_context_lock
);
1470 rctx
->dma_clear_buffer(&rctx
->b
, dst
, offset
, size
, value
);
1471 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
1472 mtx_unlock(&rscreen
->aux_context_lock
);