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>
46 struct r600_multi_fence
{
47 struct pipe_reference reference
;
48 struct pipe_fence_handle
*gfx
;
49 struct pipe_fence_handle
*sdma
;
53 * shader binary helpers.
55 void radeon_shader_binary_init(struct radeon_shader_binary
*b
)
57 memset(b
, 0, sizeof(*b
));
60 void radeon_shader_binary_clean(struct radeon_shader_binary
*b
)
67 FREE(b
->global_symbol_offsets
);
69 FREE(b
->disasm_string
);
70 FREE(b
->llvm_ir_string
);
77 void r600_draw_rectangle(struct blitter_context
*blitter
,
78 int x1
, int y1
, int x2
, int y2
, float depth
,
79 enum blitter_attrib_type type
,
80 const union pipe_color_union
*attrib
)
82 struct r600_common_context
*rctx
=
83 (struct r600_common_context
*)util_blitter_get_pipe(blitter
);
84 struct pipe_viewport_state viewport
;
85 struct pipe_resource
*buf
= NULL
;
89 if (type
== UTIL_BLITTER_ATTRIB_TEXCOORD
) {
90 util_blitter_draw_rectangle(blitter
, x1
, y1
, x2
, y2
, depth
, type
, attrib
);
94 /* Some operations (like color resolve on r6xx) don't work
95 * with the conventional primitive types.
96 * One that works is PT_RECTLIST, which we use here. */
99 viewport
.scale
[0] = 1.0f
;
100 viewport
.scale
[1] = 1.0f
;
101 viewport
.scale
[2] = 1.0f
;
102 viewport
.translate
[0] = 0.0f
;
103 viewport
.translate
[1] = 0.0f
;
104 viewport
.translate
[2] = 0.0f
;
105 rctx
->b
.set_viewport_states(&rctx
->b
, 0, 1, &viewport
);
107 /* Upload vertices. The hw rectangle has only 3 vertices,
108 * I guess the 4th one is derived from the first 3.
109 * The vertex specification should match u_blitter's vertex element state. */
110 u_upload_alloc(rctx
->uploader
, 0, sizeof(float) * 24, 256, &offset
, &buf
, (void**)&vb
);
130 memcpy(vb
+4, attrib
->f
, sizeof(float)*4);
131 memcpy(vb
+12, attrib
->f
, sizeof(float)*4);
132 memcpy(vb
+20, attrib
->f
, sizeof(float)*4);
136 util_draw_vertex_buffer(&rctx
->b
, NULL
, buf
, blitter
->vb_slot
, offset
,
137 R600_PRIM_RECTANGLE_LIST
, 3, 2);
138 pipe_resource_reference(&buf
, NULL
);
141 void r600_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
,
142 struct r600_resource
*dst
, struct r600_resource
*src
)
144 uint64_t vram
= 0, gtt
= 0;
147 if (dst
->domains
& RADEON_DOMAIN_VRAM
)
148 vram
+= dst
->buf
->size
;
149 else if (dst
->domains
& RADEON_DOMAIN_GTT
)
150 gtt
+= dst
->buf
->size
;
153 if (src
->domains
& RADEON_DOMAIN_VRAM
)
154 vram
+= src
->buf
->size
;
155 else if (src
->domains
& RADEON_DOMAIN_GTT
)
156 gtt
+= src
->buf
->size
;
159 /* Flush the GFX IB if DMA depends on it. */
160 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
162 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
163 RADEON_USAGE_READWRITE
)) ||
165 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
166 RADEON_USAGE_WRITE
))))
167 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
169 /* Flush if there's not enough space, or if the memory usage per IB
172 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
173 !ctx
->ws
->cs_memory_below_limit(ctx
->dma
.cs
, vram
, gtt
)) {
174 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
175 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
178 /* If GPUVM is not supported, the CS checker needs 2 entries
179 * in the buffer list per packet, which has to be done manually.
181 if (ctx
->screen
->info
.has_virtual_memory
) {
183 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
185 RADEON_PRIO_SDMA_BUFFER
);
187 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
189 RADEON_PRIO_SDMA_BUFFER
);
193 /* This is required to prevent read-after-write hazards. */
194 void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
196 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
198 /* done at the end of DMA calls, so increment this. */
199 rctx
->num_dma_calls
++;
201 /* IBs using too little memory are limited by the IB submission overhead.
202 * IBs using too much memory are limited by the kernel/TTM overhead.
203 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
205 * This heuristic makes sure that DMA requests are executed
206 * very soon after the call is made and lowers memory usage.
207 * It improves texture upload performance by keeping the DMA
208 * engine busy while uploads are being submitted.
210 if (rctx
->ws
->cs_query_memory_usage(rctx
->dma
.cs
) > 64 * 1024 * 1024) {
211 rctx
->dma
.flush(rctx
, RADEON_FLUSH_ASYNC
, NULL
);
215 r600_need_dma_space(rctx
, 1, NULL
, NULL
);
217 if (!radeon_emitted(cs
, 0)) /* empty queue */
220 /* NOP waits for idle on Evergreen and later. */
221 if (rctx
->chip_class
>= CIK
)
222 radeon_emit(cs
, 0x00000000); /* NOP */
223 else if (rctx
->chip_class
>= EVERGREEN
)
224 radeon_emit(cs
, 0xf0000000); /* NOP */
226 /* TODO: R600-R700 should use the FENCE packet.
227 * CS checker support is required. */
231 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
235 void r600_preflush_suspend_features(struct r600_common_context
*ctx
)
237 /* suspend queries */
238 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
239 r600_suspend_queries(ctx
);
241 ctx
->streamout
.suspended
= false;
242 if (ctx
->streamout
.begin_emitted
) {
243 r600_emit_streamout_end(ctx
);
244 ctx
->streamout
.suspended
= true;
248 void r600_postflush_resume_features(struct r600_common_context
*ctx
)
250 if (ctx
->streamout
.suspended
) {
251 ctx
->streamout
.append_bitmask
= ctx
->streamout
.enabled_mask
;
252 r600_streamout_buffers_dirty(ctx
);
256 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
257 r600_resume_queries(ctx
);
260 static void r600_flush_from_st(struct pipe_context
*ctx
,
261 struct pipe_fence_handle
**fence
,
264 struct pipe_screen
*screen
= ctx
->screen
;
265 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
267 struct pipe_fence_handle
*gfx_fence
= NULL
;
268 struct pipe_fence_handle
*sdma_fence
= NULL
;
270 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
271 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
272 if (flags
& PIPE_FLUSH_DEFERRED
)
273 rflags
|= RADEON_FLUSH_ASYNC
;
276 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
278 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
280 /* Both engines can signal out of order, so we need to keep both fences. */
281 if (gfx_fence
|| sdma_fence
) {
282 struct r600_multi_fence
*multi_fence
=
283 CALLOC_STRUCT(r600_multi_fence
);
287 multi_fence
->reference
.count
= 1;
288 multi_fence
->gfx
= gfx_fence
;
289 multi_fence
->sdma
= sdma_fence
;
291 screen
->fence_reference(screen
, fence
, NULL
);
292 *fence
= (struct pipe_fence_handle
*)multi_fence
;
296 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
297 struct pipe_fence_handle
**fence
)
299 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
300 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
301 struct radeon_saved_cs saved
;
303 (rctx
->screen
->debug_flags
& DBG_CHECK_VM
) &&
304 rctx
->check_vm_faults
;
306 if (!radeon_emitted(cs
, 0)) {
308 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
313 radeon_save_cs(rctx
->ws
, cs
, &saved
);
315 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
317 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
320 /* Use conservative timeout 800ms, after which we won't wait any
321 * longer and assume the GPU is hung.
323 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
325 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
326 radeon_clear_saved_cs(&saved
);
331 * Store a linearized copy of all chunks of \p cs together with the buffer
334 void radeon_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
335 struct radeon_saved_cs
*saved
)
340 /* Save the IB chunks. */
341 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
342 saved
->ib
= MALLOC(4 * saved
->num_dw
);
347 for (i
= 0; i
< cs
->num_prev
; ++i
) {
348 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
349 buf
+= cs
->prev
[i
].cdw
;
351 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
353 /* Save the buffer list. */
354 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
355 saved
->bo_list
= CALLOC(saved
->bo_count
,
356 sizeof(saved
->bo_list
[0]));
357 if (!saved
->bo_list
) {
361 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
366 fprintf(stderr
, "%s: out of memory\n", __func__
);
367 memset(saved
, 0, sizeof(*saved
));
370 void radeon_clear_saved_cs(struct radeon_saved_cs
*saved
)
373 FREE(saved
->bo_list
);
375 memset(saved
, 0, sizeof(*saved
));
378 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
380 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
381 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
382 RADEON_GPU_RESET_COUNTER
);
384 if (rctx
->gpu_reset_counter
== latest
)
385 return PIPE_NO_RESET
;
387 rctx
->gpu_reset_counter
= latest
;
388 return PIPE_UNKNOWN_CONTEXT_RESET
;
391 static void r600_set_debug_callback(struct pipe_context
*ctx
,
392 const struct pipe_debug_callback
*cb
)
394 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
399 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
402 bool r600_common_context_init(struct r600_common_context
*rctx
,
403 struct r600_common_screen
*rscreen
,
404 unsigned context_flags
)
406 util_slab_create(&rctx
->pool_transfers
,
407 sizeof(struct r600_transfer
), 64,
408 UTIL_SLAB_SINGLETHREADED
);
410 rctx
->screen
= rscreen
;
411 rctx
->ws
= rscreen
->ws
;
412 rctx
->family
= rscreen
->family
;
413 rctx
->chip_class
= rscreen
->chip_class
;
415 if (rscreen
->chip_class
>= CIK
)
416 rctx
->max_db
= MAX2(8, rscreen
->info
.num_render_backends
);
417 else if (rscreen
->chip_class
>= EVERGREEN
)
422 rctx
->b
.invalidate_resource
= r600_invalidate_resource
;
423 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
424 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
425 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
426 rctx
->b
.texture_subdata
= u_default_texture_subdata
;
427 rctx
->b
.memory_barrier
= r600_memory_barrier
;
428 rctx
->b
.flush
= r600_flush_from_st
;
429 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
431 /* evergreen_compute.c has a special codepath for global buffers.
432 * Everything else can use the direct path.
434 if ((rscreen
->chip_class
== EVERGREEN
|| rscreen
->chip_class
== CAYMAN
) &&
435 (context_flags
& PIPE_CONTEXT_COMPUTE_ONLY
))
436 rctx
->b
.buffer_subdata
= u_default_buffer_subdata
;
438 rctx
->b
.buffer_subdata
= r600_buffer_subdata
;
440 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
441 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
442 rctx
->gpu_reset_counter
=
443 rctx
->ws
->query_value(rctx
->ws
,
444 RADEON_GPU_RESET_COUNTER
);
447 LIST_INITHEAD(&rctx
->texture_buffers
);
449 r600_init_context_texture_functions(rctx
);
450 r600_init_viewport_functions(rctx
);
451 r600_streamout_init(rctx
);
452 r600_query_init(rctx
);
453 cayman_init_msaa(&rctx
->b
);
455 rctx
->allocator_zeroed_memory
=
456 u_suballocator_create(&rctx
->b
, rscreen
->info
.gart_page_size
,
457 0, PIPE_USAGE_DEFAULT
, true);
458 if (!rctx
->allocator_zeroed_memory
)
461 rctx
->uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
462 PIPE_BIND_INDEX_BUFFER
|
463 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
);
467 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
471 if (rscreen
->info
.has_sdma
&& !(rscreen
->debug_flags
& DBG_NO_ASYNC_DMA
)) {
472 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
475 rctx
->dma
.flush
= r600_flush_dma_ring
;
481 void r600_common_context_cleanup(struct r600_common_context
*rctx
)
485 /* Release DCC stats. */
486 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
487 assert(!rctx
->dcc_stats
[i
].query_active
);
489 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
490 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
491 rctx
->b
.destroy_query(&rctx
->b
,
492 rctx
->dcc_stats
[i
].ps_stats
[j
]);
494 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
498 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
500 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
502 rctx
->ws
->ctx_destroy(rctx
->ctx
);
504 if (rctx
->uploader
) {
505 u_upload_destroy(rctx
->uploader
);
508 util_slab_destroy(&rctx
->pool_transfers
);
510 if (rctx
->allocator_zeroed_memory
) {
511 u_suballocator_destroy(rctx
->allocator_zeroed_memory
);
513 rctx
->ws
->fence_reference(&rctx
->last_sdma_fence
, NULL
);
516 void r600_context_add_resource_size(struct pipe_context
*ctx
, struct pipe_resource
*r
)
518 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
519 struct r600_resource
*rr
= (struct r600_resource
*)r
;
526 * The idea is to compute a gross estimate of memory requirement of
527 * each draw call. After each draw call, memory will be precisely
528 * accounted. So the uncertainty is only on the current draw call.
529 * In practice this gave very good estimate (+/- 10% of the target
532 if (rr
->domains
& RADEON_DOMAIN_VRAM
)
533 rctx
->vram
+= rr
->buf
->size
;
534 else if (rr
->domains
& RADEON_DOMAIN_GTT
)
535 rctx
->gtt
+= rr
->buf
->size
;
542 static const struct debug_named_value common_debug_options
[] = {
544 { "tex", DBG_TEX
, "Print texture info" },
545 { "compute", DBG_COMPUTE
, "Print compute info" },
546 { "vm", DBG_VM
, "Print virtual addresses when creating resources" },
547 { "info", DBG_INFO
, "Print driver information" },
550 { "fs", DBG_FS
, "Print fetch shaders" },
551 { "vs", DBG_VS
, "Print vertex shaders" },
552 { "gs", DBG_GS
, "Print geometry shaders" },
553 { "ps", DBG_PS
, "Print pixel shaders" },
554 { "cs", DBG_CS
, "Print compute shaders" },
555 { "tcs", DBG_TCS
, "Print tessellation control shaders" },
556 { "tes", DBG_TES
, "Print tessellation evaluation shaders" },
557 { "noir", DBG_NO_IR
, "Don't print the LLVM IR"},
558 { "notgsi", DBG_NO_TGSI
, "Don't print the TGSI"},
559 { "noasm", DBG_NO_ASM
, "Don't print disassembled shaders"},
560 { "preoptir", DBG_PREOPT_IR
, "Print the LLVM IR before initial optimizations" },
562 { "testdma", DBG_TEST_DMA
, "Invoke SDMA tests and exit." },
565 { "nodma", DBG_NO_ASYNC_DMA
, "Disable asynchronous DMA" },
566 { "nohyperz", DBG_NO_HYPERZ
, "Disable Hyper-Z" },
567 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
568 { "noinvalrange", DBG_NO_DISCARD_RANGE
, "Disable handling of INVALIDATE_RANGE map flags" },
569 { "no2d", DBG_NO_2D_TILING
, "Disable 2D tiling" },
570 { "notiling", DBG_NO_TILING
, "Disable tiling" },
571 { "switch_on_eop", DBG_SWITCH_ON_EOP
, "Program WD/IA to switch on end-of-packet." },
572 { "forcedma", DBG_FORCE_DMA
, "Use asynchronous DMA for all operations when possible." },
573 { "precompile", DBG_PRECOMPILE
, "Compile one shader variant at shader creation." },
574 { "nowc", DBG_NO_WC
, "Disable GTT write combining" },
575 { "check_vm", DBG_CHECK_VM
, "Check VM faults and dump debug info." },
576 { "nodcc", DBG_NO_DCC
, "Disable DCC." },
577 { "nodccclear", DBG_NO_DCC_CLEAR
, "Disable DCC fast clear." },
578 { "norbplus", DBG_NO_RB_PLUS
, "Disable RB+ on Stoney." },
579 { "sisched", DBG_SI_SCHED
, "Enable LLVM SI Machine Instruction Scheduler." },
580 { "mono", DBG_MONOLITHIC_SHADERS
, "Use old-style monolithic shaders compiled on demand" },
581 { "noce", DBG_NO_CE
, "Disable the constant engine"},
582 { "unsafemath", DBG_UNSAFE_MATH
, "Enable unsafe math shader optimizations" },
583 { "nodccfb", DBG_NO_DCC_FB
, "Disable separate DCC on the main framebuffer" },
585 DEBUG_NAMED_VALUE_END
/* must be last */
588 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
593 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
598 static const char* r600_get_chip_name(struct r600_common_screen
*rscreen
)
600 switch (rscreen
->info
.family
) {
601 case CHIP_R600
: return "AMD R600";
602 case CHIP_RV610
: return "AMD RV610";
603 case CHIP_RV630
: return "AMD RV630";
604 case CHIP_RV670
: return "AMD RV670";
605 case CHIP_RV620
: return "AMD RV620";
606 case CHIP_RV635
: return "AMD RV635";
607 case CHIP_RS780
: return "AMD RS780";
608 case CHIP_RS880
: return "AMD RS880";
609 case CHIP_RV770
: return "AMD RV770";
610 case CHIP_RV730
: return "AMD RV730";
611 case CHIP_RV710
: return "AMD RV710";
612 case CHIP_RV740
: return "AMD RV740";
613 case CHIP_CEDAR
: return "AMD CEDAR";
614 case CHIP_REDWOOD
: return "AMD REDWOOD";
615 case CHIP_JUNIPER
: return "AMD JUNIPER";
616 case CHIP_CYPRESS
: return "AMD CYPRESS";
617 case CHIP_HEMLOCK
: return "AMD HEMLOCK";
618 case CHIP_PALM
: return "AMD PALM";
619 case CHIP_SUMO
: return "AMD SUMO";
620 case CHIP_SUMO2
: return "AMD SUMO2";
621 case CHIP_BARTS
: return "AMD BARTS";
622 case CHIP_TURKS
: return "AMD TURKS";
623 case CHIP_CAICOS
: return "AMD CAICOS";
624 case CHIP_CAYMAN
: return "AMD CAYMAN";
625 case CHIP_ARUBA
: return "AMD ARUBA";
626 case CHIP_TAHITI
: return "AMD TAHITI";
627 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
628 case CHIP_VERDE
: return "AMD CAPE VERDE";
629 case CHIP_OLAND
: return "AMD OLAND";
630 case CHIP_HAINAN
: return "AMD HAINAN";
631 case CHIP_BONAIRE
: return "AMD BONAIRE";
632 case CHIP_KAVERI
: return "AMD KAVERI";
633 case CHIP_KABINI
: return "AMD KABINI";
634 case CHIP_HAWAII
: return "AMD HAWAII";
635 case CHIP_MULLINS
: return "AMD MULLINS";
636 case CHIP_TONGA
: return "AMD TONGA";
637 case CHIP_ICELAND
: return "AMD ICELAND";
638 case CHIP_CARRIZO
: return "AMD CARRIZO";
639 case CHIP_FIJI
: return "AMD FIJI";
640 case CHIP_POLARIS10
: return "AMD POLARIS10";
641 case CHIP_POLARIS11
: return "AMD POLARIS11";
642 case CHIP_STONEY
: return "AMD STONEY";
643 default: return "AMD unknown";
647 static const char* r600_get_name(struct pipe_screen
* pscreen
)
649 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
651 return rscreen
->renderer_string
;
654 static float r600_get_paramf(struct pipe_screen
* pscreen
,
655 enum pipe_capf param
)
657 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
660 case PIPE_CAPF_MAX_LINE_WIDTH
:
661 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
662 case PIPE_CAPF_MAX_POINT_WIDTH
:
663 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
664 if (rscreen
->family
>= CHIP_CEDAR
)
668 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
670 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
672 case PIPE_CAPF_GUARD_BAND_LEFT
:
673 case PIPE_CAPF_GUARD_BAND_TOP
:
674 case PIPE_CAPF_GUARD_BAND_RIGHT
:
675 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
681 static int r600_get_video_param(struct pipe_screen
*screen
,
682 enum pipe_video_profile profile
,
683 enum pipe_video_entrypoint entrypoint
,
684 enum pipe_video_cap param
)
687 case PIPE_VIDEO_CAP_SUPPORTED
:
688 return vl_profile_supported(screen
, profile
, entrypoint
);
689 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
691 case PIPE_VIDEO_CAP_MAX_WIDTH
:
692 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
693 return vl_video_buffer_max_size(screen
);
694 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
695 return PIPE_FORMAT_NV12
;
696 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
698 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
700 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
702 case PIPE_VIDEO_CAP_MAX_LEVEL
:
703 return vl_level_supported(screen
, profile
);
709 const char *r600_get_llvm_processor_name(enum radeon_family family
)
752 case CHIP_TAHITI
: return "tahiti";
753 case CHIP_PITCAIRN
: return "pitcairn";
754 case CHIP_VERDE
: return "verde";
755 case CHIP_OLAND
: return "oland";
756 case CHIP_HAINAN
: return "hainan";
757 case CHIP_BONAIRE
: return "bonaire";
758 case CHIP_KABINI
: return "kabini";
759 case CHIP_KAVERI
: return "kaveri";
760 case CHIP_HAWAII
: return "hawaii";
763 case CHIP_TONGA
: return "tonga";
764 case CHIP_ICELAND
: return "iceland";
765 case CHIP_CARRIZO
: return "carrizo";
766 #if HAVE_LLVM <= 0x0307
767 case CHIP_FIJI
: return "tonga";
768 case CHIP_STONEY
: return "carrizo";
770 case CHIP_FIJI
: return "fiji";
771 case CHIP_STONEY
: return "stoney";
773 #if HAVE_LLVM <= 0x0308
774 case CHIP_POLARIS10
: return "tonga";
775 case CHIP_POLARIS11
: return "tonga";
777 case CHIP_POLARIS10
: return "polaris10";
778 case CHIP_POLARIS11
: return "polaris11";
784 static int r600_get_compute_param(struct pipe_screen
*screen
,
785 enum pipe_shader_ir ir_type
,
786 enum pipe_compute_cap param
,
789 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
791 //TODO: select these params by asic
793 case PIPE_COMPUTE_CAP_IR_TARGET
: {
796 if (rscreen
->family
<= CHIP_ARUBA
) {
801 switch(rscreen
->family
) {
802 /* Clang < 3.6 is missing Hainan in its list of
803 * GPUs, so we need to use the name of a similar GPU.
806 gpu
= r600_get_llvm_processor_name(rscreen
->family
);
810 sprintf(ret
, "%s-%s", gpu
, triple
);
812 /* +2 for dash and terminating NIL byte */
813 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
815 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
817 uint64_t *grid_dimension
= ret
;
818 grid_dimension
[0] = 3;
820 return 1 * sizeof(uint64_t);
822 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
824 uint64_t *grid_size
= ret
;
825 grid_size
[0] = 65535;
826 grid_size
[1] = 65535;
827 grid_size
[2] = 65535;
829 return 3 * sizeof(uint64_t) ;
831 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
833 uint64_t *block_size
= ret
;
834 if (rscreen
->chip_class
>= SI
&& HAVE_LLVM
>= 0x309 &&
835 ir_type
== PIPE_SHADER_IR_TGSI
) {
836 block_size
[0] = 2048;
837 block_size
[1] = 2048;
838 block_size
[2] = 2048;
845 return 3 * sizeof(uint64_t);
847 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
849 uint64_t *max_threads_per_block
= ret
;
850 if (rscreen
->chip_class
>= SI
&& HAVE_LLVM
>= 0x309 &&
851 ir_type
== PIPE_SHADER_IR_TGSI
)
852 *max_threads_per_block
= 2048;
854 *max_threads_per_block
= 256;
856 return sizeof(uint64_t);
858 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
860 uint64_t *max_global_size
= ret
;
861 uint64_t max_mem_alloc_size
;
863 r600_get_compute_param(screen
, ir_type
,
864 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
865 &max_mem_alloc_size
);
867 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
868 * 1/4 of the MAX_GLOBAL_SIZE. Since the
869 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
870 * make sure we never report more than
871 * 4 * MAX_MEM_ALLOC_SIZE.
873 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
874 MAX2(rscreen
->info
.gart_size
,
875 rscreen
->info
.vram_size
));
877 return sizeof(uint64_t);
879 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
881 uint64_t *max_local_size
= ret
;
882 /* Value reported by the closed source driver. */
883 *max_local_size
= 32768;
885 return sizeof(uint64_t);
887 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
889 uint64_t *max_input_size
= ret
;
890 /* Value reported by the closed source driver. */
891 *max_input_size
= 1024;
893 return sizeof(uint64_t);
895 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
897 uint64_t *max_mem_alloc_size
= ret
;
899 *max_mem_alloc_size
= rscreen
->info
.max_alloc_size
;
901 return sizeof(uint64_t);
903 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
905 uint32_t *max_clock_frequency
= ret
;
906 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
908 return sizeof(uint32_t);
910 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
912 uint32_t *max_compute_units
= ret
;
913 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
915 return sizeof(uint32_t);
917 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
919 uint32_t *images_supported
= ret
;
920 *images_supported
= 0;
922 return sizeof(uint32_t);
923 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
925 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
927 uint32_t *subgroup_size
= ret
;
928 *subgroup_size
= r600_wavefront_size(rscreen
->family
);
930 return sizeof(uint32_t);
933 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
937 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
939 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
941 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
942 rscreen
->info
.clock_crystal_freq
;
945 static void r600_fence_reference(struct pipe_screen
*screen
,
946 struct pipe_fence_handle
**dst
,
947 struct pipe_fence_handle
*src
)
949 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
950 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
951 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
953 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
954 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
955 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
961 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
962 struct pipe_fence_handle
*fence
,
965 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
966 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
967 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
970 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
973 /* Recompute the timeout after waiting. */
974 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
975 int64_t time
= os_time_get_nano();
976 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
983 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
986 static void r600_query_memory_info(struct pipe_screen
*screen
,
987 struct pipe_memory_info
*info
)
989 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
990 struct radeon_winsys
*ws
= rscreen
->ws
;
991 unsigned vram_usage
, gtt_usage
;
993 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
994 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
996 /* The real TTM memory usage is somewhat random, because:
998 * 1) TTM delays freeing memory, because it can only free it after
1001 * 2) The memory usage can be really low if big VRAM evictions are
1002 * taking place, but the real usage is well above the size of VRAM.
1004 * Instead, return statistics of this process.
1006 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
1007 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
1009 info
->avail_device_memory
=
1010 vram_usage
<= info
->total_device_memory
?
1011 info
->total_device_memory
- vram_usage
: 0;
1012 info
->avail_staging_memory
=
1013 gtt_usage
<= info
->total_staging_memory
?
1014 info
->total_staging_memory
- gtt_usage
: 0;
1016 info
->device_memory_evicted
=
1017 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
1018 /* Just return the number of evicted 64KB pages. */
1019 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
1022 struct pipe_resource
*r600_resource_create_common(struct pipe_screen
*screen
,
1023 const struct pipe_resource
*templ
)
1025 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1027 if (templ
->target
== PIPE_BUFFER
) {
1028 return r600_buffer_create(screen
, templ
,
1029 rscreen
->info
.gart_page_size
);
1031 return r600_texture_create(screen
, templ
);
1035 bool r600_common_screen_init(struct r600_common_screen
*rscreen
,
1036 struct radeon_winsys
*ws
)
1038 char llvm_string
[32] = {}, kernel_version
[128] = {};
1039 struct utsname uname_data
;
1041 ws
->query_info(ws
, &rscreen
->info
);
1043 if (uname(&uname_data
) == 0)
1044 snprintf(kernel_version
, sizeof(kernel_version
),
1045 " / %s", uname_data
.release
);
1048 snprintf(llvm_string
, sizeof(llvm_string
),
1049 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
1050 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
1053 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
1054 "%s (DRM %i.%i.%i%s%s)",
1055 r600_get_chip_name(rscreen
), rscreen
->info
.drm_major
,
1056 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
1057 kernel_version
, llvm_string
);
1059 rscreen
->b
.get_name
= r600_get_name
;
1060 rscreen
->b
.get_vendor
= r600_get_vendor
;
1061 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
1062 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
1063 rscreen
->b
.get_paramf
= r600_get_paramf
;
1064 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
1065 rscreen
->b
.fence_finish
= r600_fence_finish
;
1066 rscreen
->b
.fence_reference
= r600_fence_reference
;
1067 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
1068 rscreen
->b
.resource_from_user_memory
= r600_buffer_from_user_memory
;
1069 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
1071 if (rscreen
->info
.has_uvd
) {
1072 rscreen
->b
.get_video_param
= rvid_get_video_param
;
1073 rscreen
->b
.is_video_format_supported
= rvid_is_format_supported
;
1075 rscreen
->b
.get_video_param
= r600_get_video_param
;
1076 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
1079 r600_init_screen_texture_functions(rscreen
);
1080 r600_init_screen_query_functions(rscreen
);
1083 rscreen
->family
= rscreen
->info
.family
;
1084 rscreen
->chip_class
= rscreen
->info
.chip_class
;
1085 rscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
1087 rscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1088 if (rscreen
->force_aniso
>= 0) {
1089 printf("radeon: Forcing anisotropy filter to %ix\n",
1090 /* round down to a power of two */
1091 1 << util_logbase2(rscreen
->force_aniso
));
1094 util_format_s3tc_init();
1095 pipe_mutex_init(rscreen
->aux_context_lock
);
1096 pipe_mutex_init(rscreen
->gpu_load_mutex
);
1098 if (rscreen
->debug_flags
& DBG_INFO
) {
1099 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
1100 printf("family = %i (%s)\n", rscreen
->info
.family
,
1101 r600_get_chip_name(rscreen
));
1102 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
1103 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
1104 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
1105 printf("max_alloc_size = %i MB\n",
1106 (int)DIV_ROUND_UP(rscreen
->info
.max_alloc_size
, 1024*1024));
1107 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
1108 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
1109 printf("has_sdma = %i\n", rscreen
->info
.has_sdma
);
1110 printf("has_uvd = %i\n", rscreen
->info
.has_uvd
);
1111 printf("vce_fw_version = %i\n", rscreen
->info
.vce_fw_version
);
1112 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
1113 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
1114 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
1115 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
1116 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
1118 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
1119 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
1120 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
1121 printf("max_se = %i\n", rscreen
->info
.max_se
);
1122 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
1124 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
1125 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
1126 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
1127 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
1128 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
1129 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
1134 void r600_destroy_common_screen(struct r600_common_screen
*rscreen
)
1136 r600_perfcounters_destroy(rscreen
);
1137 r600_gpu_load_kill_thread(rscreen
);
1139 pipe_mutex_destroy(rscreen
->gpu_load_mutex
);
1140 pipe_mutex_destroy(rscreen
->aux_context_lock
);
1141 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
1143 rscreen
->ws
->destroy(rscreen
->ws
);
1147 bool r600_can_dump_shader(struct r600_common_screen
*rscreen
,
1150 switch (processor
) {
1151 case PIPE_SHADER_VERTEX
:
1152 return (rscreen
->debug_flags
& DBG_VS
) != 0;
1153 case PIPE_SHADER_TESS_CTRL
:
1154 return (rscreen
->debug_flags
& DBG_TCS
) != 0;
1155 case PIPE_SHADER_TESS_EVAL
:
1156 return (rscreen
->debug_flags
& DBG_TES
) != 0;
1157 case PIPE_SHADER_GEOMETRY
:
1158 return (rscreen
->debug_flags
& DBG_GS
) != 0;
1159 case PIPE_SHADER_FRAGMENT
:
1160 return (rscreen
->debug_flags
& DBG_PS
) != 0;
1161 case PIPE_SHADER_COMPUTE
:
1162 return (rscreen
->debug_flags
& DBG_CS
) != 0;
1168 void r600_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
1169 uint64_t offset
, uint64_t size
, unsigned value
,
1170 enum r600_coherency coher
)
1172 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
1174 pipe_mutex_lock(rscreen
->aux_context_lock
);
1175 rctx
->clear_buffer(&rctx
->b
, dst
, offset
, size
, value
, coher
);
1176 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
1177 pipe_mutex_unlock(rscreen
->aux_context_lock
);