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"
45 struct r600_multi_fence
{
46 struct pipe_reference reference
;
47 struct pipe_fence_handle
*gfx
;
48 struct pipe_fence_handle
*sdma
;
52 * shader binary helpers.
54 void radeon_shader_binary_init(struct radeon_shader_binary
*b
)
56 memset(b
, 0, sizeof(*b
));
59 void radeon_shader_binary_clean(struct radeon_shader_binary
*b
)
66 FREE(b
->global_symbol_offsets
);
68 FREE(b
->disasm_string
);
75 void r600_draw_rectangle(struct blitter_context
*blitter
,
76 int x1
, int y1
, int x2
, int y2
, float depth
,
77 enum blitter_attrib_type type
,
78 const union pipe_color_union
*attrib
)
80 struct r600_common_context
*rctx
=
81 (struct r600_common_context
*)util_blitter_get_pipe(blitter
);
82 struct pipe_viewport_state viewport
;
83 struct pipe_resource
*buf
= NULL
;
87 if (type
== UTIL_BLITTER_ATTRIB_TEXCOORD
) {
88 util_blitter_draw_rectangle(blitter
, x1
, y1
, x2
, y2
, depth
, type
, attrib
);
92 /* Some operations (like color resolve on r6xx) don't work
93 * with the conventional primitive types.
94 * One that works is PT_RECTLIST, which we use here. */
97 viewport
.scale
[0] = 1.0f
;
98 viewport
.scale
[1] = 1.0f
;
99 viewport
.scale
[2] = 1.0f
;
100 viewport
.translate
[0] = 0.0f
;
101 viewport
.translate
[1] = 0.0f
;
102 viewport
.translate
[2] = 0.0f
;
103 rctx
->b
.set_viewport_states(&rctx
->b
, 0, 1, &viewport
);
105 /* Upload vertices. The hw rectangle has only 3 vertices,
106 * I guess the 4th one is derived from the first 3.
107 * The vertex specification should match u_blitter's vertex element state. */
108 u_upload_alloc(rctx
->uploader
, 0, sizeof(float) * 24, 256, &offset
, &buf
, (void**)&vb
);
128 memcpy(vb
+4, attrib
->f
, sizeof(float)*4);
129 memcpy(vb
+12, attrib
->f
, sizeof(float)*4);
130 memcpy(vb
+20, attrib
->f
, sizeof(float)*4);
134 util_draw_vertex_buffer(&rctx
->b
, NULL
, buf
, blitter
->vb_slot
, offset
,
135 R600_PRIM_RECTANGLE_LIST
, 3, 2);
136 pipe_resource_reference(&buf
, NULL
);
139 void r600_need_dma_space(struct r600_common_context
*ctx
, unsigned num_dw
)
141 /* Flush the GFX IB if it's not empty. */
142 if (ctx
->gfx
.cs
->cdw
> ctx
->initial_gfx_cs_size
)
143 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
145 /* Flush if there's not enough space. */
146 if ((num_dw
+ ctx
->dma
.cs
->cdw
) > ctx
->dma
.cs
->max_dw
) {
147 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
148 assert((num_dw
+ ctx
->dma
.cs
->cdw
) <= ctx
->dma
.cs
->max_dw
);
152 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
156 void r600_preflush_suspend_features(struct r600_common_context
*ctx
)
158 /* suspend queries */
159 if (ctx
->num_cs_dw_nontimer_queries_suspend
) {
160 /* Since non-timer queries are suspended during blits,
161 * we have to guard against double-suspends. */
162 r600_suspend_nontimer_queries(ctx
);
163 ctx
->nontimer_queries_suspended_by_flush
= true;
165 if (!LIST_IS_EMPTY(&ctx
->active_timer_queries
))
166 r600_suspend_timer_queries(ctx
);
168 ctx
->streamout
.suspended
= false;
169 if (ctx
->streamout
.begin_emitted
) {
170 r600_emit_streamout_end(ctx
);
171 ctx
->streamout
.suspended
= true;
175 void r600_postflush_resume_features(struct r600_common_context
*ctx
)
177 if (ctx
->streamout
.suspended
) {
178 ctx
->streamout
.append_bitmask
= ctx
->streamout
.enabled_mask
;
179 r600_streamout_buffers_dirty(ctx
);
183 if (!LIST_IS_EMPTY(&ctx
->active_timer_queries
))
184 r600_resume_timer_queries(ctx
);
185 if (ctx
->nontimer_queries_suspended_by_flush
) {
186 ctx
->nontimer_queries_suspended_by_flush
= false;
187 r600_resume_nontimer_queries(ctx
);
191 static void r600_flush_from_st(struct pipe_context
*ctx
,
192 struct pipe_fence_handle
**fence
,
195 struct pipe_screen
*screen
= ctx
->screen
;
196 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
198 struct pipe_fence_handle
*gfx_fence
= NULL
;
199 struct pipe_fence_handle
*sdma_fence
= NULL
;
201 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
202 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
205 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
207 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
209 /* Both engines can signal out of order, so we need to keep both fences. */
210 if (gfx_fence
|| sdma_fence
) {
211 struct r600_multi_fence
*multi_fence
=
212 CALLOC_STRUCT(r600_multi_fence
);
216 multi_fence
->reference
.count
= 1;
217 multi_fence
->gfx
= gfx_fence
;
218 multi_fence
->sdma
= sdma_fence
;
220 screen
->fence_reference(screen
, fence
, NULL
);
221 *fence
= (struct pipe_fence_handle
*)multi_fence
;
225 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
226 struct pipe_fence_handle
**fence
)
228 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
229 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
232 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
234 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
237 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
239 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
240 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
241 RADEON_GPU_RESET_COUNTER
);
243 if (rctx
->gpu_reset_counter
== latest
)
244 return PIPE_NO_RESET
;
246 rctx
->gpu_reset_counter
= latest
;
247 return PIPE_UNKNOWN_CONTEXT_RESET
;
250 static void r600_set_debug_callback(struct pipe_context
*ctx
,
251 const struct pipe_debug_callback
*cb
)
253 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
258 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
261 bool r600_common_context_init(struct r600_common_context
*rctx
,
262 struct r600_common_screen
*rscreen
)
264 util_slab_create(&rctx
->pool_transfers
,
265 sizeof(struct r600_transfer
), 64,
266 UTIL_SLAB_SINGLETHREADED
);
268 rctx
->screen
= rscreen
;
269 rctx
->ws
= rscreen
->ws
;
270 rctx
->family
= rscreen
->family
;
271 rctx
->chip_class
= rscreen
->chip_class
;
273 if (rscreen
->chip_class
>= CIK
)
274 rctx
->max_db
= MAX2(8, rscreen
->info
.num_render_backends
);
275 else if (rscreen
->chip_class
>= EVERGREEN
)
280 rctx
->b
.invalidate_resource
= r600_invalidate_resource
;
281 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
282 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
283 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
284 rctx
->b
.transfer_inline_write
= u_default_transfer_inline_write
;
285 rctx
->b
.memory_barrier
= r600_memory_barrier
;
286 rctx
->b
.flush
= r600_flush_from_st
;
287 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
289 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
290 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
291 rctx
->gpu_reset_counter
=
292 rctx
->ws
->query_value(rctx
->ws
,
293 RADEON_GPU_RESET_COUNTER
);
296 LIST_INITHEAD(&rctx
->texture_buffers
);
298 r600_init_context_texture_functions(rctx
);
299 r600_streamout_init(rctx
);
300 r600_query_init(rctx
);
301 cayman_init_msaa(&rctx
->b
);
303 rctx
->allocator_so_filled_size
= u_suballocator_create(&rctx
->b
, 4096, 4,
304 0, PIPE_USAGE_DEFAULT
, TRUE
);
305 if (!rctx
->allocator_so_filled_size
)
308 rctx
->uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
309 PIPE_BIND_INDEX_BUFFER
|
310 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
);
314 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
318 if (rscreen
->info
.has_sdma
&& !(rscreen
->debug_flags
& DBG_NO_ASYNC_DMA
)) {
319 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
322 rctx
->dma
.flush
= r600_flush_dma_ring
;
328 void r600_common_context_cleanup(struct r600_common_context
*rctx
)
331 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
333 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
335 rctx
->ws
->ctx_destroy(rctx
->ctx
);
337 if (rctx
->uploader
) {
338 u_upload_destroy(rctx
->uploader
);
341 util_slab_destroy(&rctx
->pool_transfers
);
343 if (rctx
->allocator_so_filled_size
) {
344 u_suballocator_destroy(rctx
->allocator_so_filled_size
);
346 rctx
->ws
->fence_reference(&rctx
->last_sdma_fence
, NULL
);
349 void r600_context_add_resource_size(struct pipe_context
*ctx
, struct pipe_resource
*r
)
351 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
352 struct r600_resource
*rr
= (struct r600_resource
*)r
;
359 * The idea is to compute a gross estimate of memory requirement of
360 * each draw call. After each draw call, memory will be precisely
361 * accounted. So the uncertainty is only on the current draw call.
362 * In practice this gave very good estimate (+/- 10% of the target
365 if (rr
->domains
& RADEON_DOMAIN_GTT
) {
366 rctx
->gtt
+= rr
->buf
->size
;
368 if (rr
->domains
& RADEON_DOMAIN_VRAM
) {
369 rctx
->vram
+= rr
->buf
->size
;
377 static const struct debug_named_value common_debug_options
[] = {
379 { "tex", DBG_TEX
, "Print texture info" },
380 { "compute", DBG_COMPUTE
, "Print compute info" },
381 { "vm", DBG_VM
, "Print virtual addresses when creating resources" },
382 { "info", DBG_INFO
, "Print driver information" },
385 { "fs", DBG_FS
, "Print fetch shaders" },
386 { "vs", DBG_VS
, "Print vertex shaders" },
387 { "gs", DBG_GS
, "Print geometry shaders" },
388 { "ps", DBG_PS
, "Print pixel shaders" },
389 { "cs", DBG_CS
, "Print compute shaders" },
390 { "tcs", DBG_TCS
, "Print tessellation control shaders" },
391 { "tes", DBG_TES
, "Print tessellation evaluation shaders" },
392 { "noir", DBG_NO_IR
, "Don't print the LLVM IR"},
393 { "notgsi", DBG_NO_TGSI
, "Don't print the TGSI"},
394 { "noasm", DBG_NO_ASM
, "Don't print disassembled shaders"},
395 { "preoptir", DBG_PREOPT_IR
, "Print the LLVM IR before initial optimizations" },
398 { "nodma", DBG_NO_ASYNC_DMA
, "Disable asynchronous DMA" },
399 { "nohyperz", DBG_NO_HYPERZ
, "Disable Hyper-Z" },
400 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
401 { "noinvalrange", DBG_NO_DISCARD_RANGE
, "Disable handling of INVALIDATE_RANGE map flags" },
402 { "no2d", DBG_NO_2D_TILING
, "Disable 2D tiling" },
403 { "notiling", DBG_NO_TILING
, "Disable tiling" },
404 { "switch_on_eop", DBG_SWITCH_ON_EOP
, "Program WD/IA to switch on end-of-packet." },
405 { "forcedma", DBG_FORCE_DMA
, "Use asynchronous DMA for all operations when possible." },
406 { "precompile", DBG_PRECOMPILE
, "Compile one shader variant at shader creation." },
407 { "nowc", DBG_NO_WC
, "Disable GTT write combining" },
408 { "check_vm", DBG_CHECK_VM
, "Check VM faults and dump debug info." },
409 { "nodcc", DBG_NO_DCC
, "Disable DCC." },
410 { "nodccclear", DBG_NO_DCC_CLEAR
, "Disable DCC fast clear." },
411 { "norbplus", DBG_NO_RB_PLUS
, "Disable RB+ on Stoney." },
412 { "sisched", DBG_SI_SCHED
, "Enable LLVM SI Machine Instruction Scheduler." },
413 { "mono", DBG_MONOLITHIC_SHADERS
, "Use old-style monolithic shaders compiled on demand" },
415 DEBUG_NAMED_VALUE_END
/* must be last */
418 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
423 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
428 static const char* r600_get_chip_name(struct r600_common_screen
*rscreen
)
430 switch (rscreen
->info
.family
) {
431 case CHIP_R600
: return "AMD R600";
432 case CHIP_RV610
: return "AMD RV610";
433 case CHIP_RV630
: return "AMD RV630";
434 case CHIP_RV670
: return "AMD RV670";
435 case CHIP_RV620
: return "AMD RV620";
436 case CHIP_RV635
: return "AMD RV635";
437 case CHIP_RS780
: return "AMD RS780";
438 case CHIP_RS880
: return "AMD RS880";
439 case CHIP_RV770
: return "AMD RV770";
440 case CHIP_RV730
: return "AMD RV730";
441 case CHIP_RV710
: return "AMD RV710";
442 case CHIP_RV740
: return "AMD RV740";
443 case CHIP_CEDAR
: return "AMD CEDAR";
444 case CHIP_REDWOOD
: return "AMD REDWOOD";
445 case CHIP_JUNIPER
: return "AMD JUNIPER";
446 case CHIP_CYPRESS
: return "AMD CYPRESS";
447 case CHIP_HEMLOCK
: return "AMD HEMLOCK";
448 case CHIP_PALM
: return "AMD PALM";
449 case CHIP_SUMO
: return "AMD SUMO";
450 case CHIP_SUMO2
: return "AMD SUMO2";
451 case CHIP_BARTS
: return "AMD BARTS";
452 case CHIP_TURKS
: return "AMD TURKS";
453 case CHIP_CAICOS
: return "AMD CAICOS";
454 case CHIP_CAYMAN
: return "AMD CAYMAN";
455 case CHIP_ARUBA
: return "AMD ARUBA";
456 case CHIP_TAHITI
: return "AMD TAHITI";
457 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
458 case CHIP_VERDE
: return "AMD CAPE VERDE";
459 case CHIP_OLAND
: return "AMD OLAND";
460 case CHIP_HAINAN
: return "AMD HAINAN";
461 case CHIP_BONAIRE
: return "AMD BONAIRE";
462 case CHIP_KAVERI
: return "AMD KAVERI";
463 case CHIP_KABINI
: return "AMD KABINI";
464 case CHIP_HAWAII
: return "AMD HAWAII";
465 case CHIP_MULLINS
: return "AMD MULLINS";
466 case CHIP_TONGA
: return "AMD TONGA";
467 case CHIP_ICELAND
: return "AMD ICELAND";
468 case CHIP_CARRIZO
: return "AMD CARRIZO";
469 case CHIP_FIJI
: return "AMD FIJI";
470 case CHIP_POLARIS10
: return "AMD POLARIS10";
471 case CHIP_POLARIS11
: return "AMD POLARIS11";
472 case CHIP_STONEY
: return "AMD STONEY";
473 default: return "AMD unknown";
477 static const char* r600_get_name(struct pipe_screen
* pscreen
)
479 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
481 return rscreen
->renderer_string
;
484 static float r600_get_paramf(struct pipe_screen
* pscreen
,
485 enum pipe_capf param
)
487 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
490 case PIPE_CAPF_MAX_LINE_WIDTH
:
491 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
492 case PIPE_CAPF_MAX_POINT_WIDTH
:
493 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
494 if (rscreen
->family
>= CHIP_CEDAR
)
498 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
500 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
502 case PIPE_CAPF_GUARD_BAND_LEFT
:
503 case PIPE_CAPF_GUARD_BAND_TOP
:
504 case PIPE_CAPF_GUARD_BAND_RIGHT
:
505 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
511 static int r600_get_video_param(struct pipe_screen
*screen
,
512 enum pipe_video_profile profile
,
513 enum pipe_video_entrypoint entrypoint
,
514 enum pipe_video_cap param
)
517 case PIPE_VIDEO_CAP_SUPPORTED
:
518 return vl_profile_supported(screen
, profile
, entrypoint
);
519 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
521 case PIPE_VIDEO_CAP_MAX_WIDTH
:
522 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
523 return vl_video_buffer_max_size(screen
);
524 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
525 return PIPE_FORMAT_NV12
;
526 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
528 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
530 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
532 case PIPE_VIDEO_CAP_MAX_LEVEL
:
533 return vl_level_supported(screen
, profile
);
539 const char *r600_get_llvm_processor_name(enum radeon_family family
)
582 case CHIP_TAHITI
: return "tahiti";
583 case CHIP_PITCAIRN
: return "pitcairn";
584 case CHIP_VERDE
: return "verde";
585 case CHIP_OLAND
: return "oland";
586 case CHIP_HAINAN
: return "hainan";
587 case CHIP_BONAIRE
: return "bonaire";
588 case CHIP_KABINI
: return "kabini";
589 case CHIP_KAVERI
: return "kaveri";
590 case CHIP_HAWAII
: return "hawaii";
593 case CHIP_TONGA
: return "tonga";
594 case CHIP_ICELAND
: return "iceland";
595 case CHIP_CARRIZO
: return "carrizo";
596 #if HAVE_LLVM <= 0x0307
597 case CHIP_FIJI
: return "tonga";
598 case CHIP_STONEY
: return "carrizo";
600 case CHIP_FIJI
: return "fiji";
601 case CHIP_STONEY
: return "stoney";
603 #if HAVE_LLVM <= 0x0308
604 case CHIP_POLARIS10
: return "tonga";
605 case CHIP_POLARIS11
: return "tonga";
607 case CHIP_POLARIS10
: return "polaris10";
608 case CHIP_POLARIS11
: return "polaris11";
614 static int r600_get_compute_param(struct pipe_screen
*screen
,
615 enum pipe_compute_cap param
,
618 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
620 //TODO: select these params by asic
622 case PIPE_COMPUTE_CAP_IR_TARGET
: {
625 if (rscreen
->family
<= CHIP_ARUBA
) {
630 switch(rscreen
->family
) {
631 /* Clang < 3.6 is missing Hainan in its list of
632 * GPUs, so we need to use the name of a similar GPU.
635 gpu
= r600_get_llvm_processor_name(rscreen
->family
);
639 sprintf(ret
, "%s-%s", gpu
, triple
);
641 /* +2 for dash and terminating NIL byte */
642 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
644 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
646 uint64_t *grid_dimension
= ret
;
647 grid_dimension
[0] = 3;
649 return 1 * sizeof(uint64_t);
651 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
653 uint64_t *grid_size
= ret
;
654 grid_size
[0] = 65535;
655 grid_size
[1] = 65535;
658 return 3 * sizeof(uint64_t) ;
660 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
662 uint64_t *block_size
= ret
;
667 return 3 * sizeof(uint64_t);
669 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
671 uint64_t *max_threads_per_block
= ret
;
672 *max_threads_per_block
= 256;
674 return sizeof(uint64_t);
676 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
678 uint64_t *max_global_size
= ret
;
679 uint64_t max_mem_alloc_size
;
681 r600_get_compute_param(screen
,
682 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
683 &max_mem_alloc_size
);
685 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
686 * 1/4 of the MAX_GLOBAL_SIZE. Since the
687 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
688 * make sure we never report more than
689 * 4 * MAX_MEM_ALLOC_SIZE.
691 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
692 rscreen
->info
.gart_size
+
693 rscreen
->info
.vram_size
);
695 return sizeof(uint64_t);
697 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
699 uint64_t *max_local_size
= ret
;
700 /* Value reported by the closed source driver. */
701 *max_local_size
= 32768;
703 return sizeof(uint64_t);
705 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
707 uint64_t *max_input_size
= ret
;
708 /* Value reported by the closed source driver. */
709 *max_input_size
= 1024;
711 return sizeof(uint64_t);
713 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
715 uint64_t *max_mem_alloc_size
= ret
;
717 /* XXX: The limit in older kernels is 256 MB. We
718 * should add a query here for newer kernels.
720 *max_mem_alloc_size
= 256 * 1024 * 1024;
722 return sizeof(uint64_t);
724 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
726 uint32_t *max_clock_frequency
= ret
;
727 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
729 return sizeof(uint32_t);
731 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
733 uint32_t *max_compute_units
= ret
;
734 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
736 return sizeof(uint32_t);
738 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
740 uint32_t *images_supported
= ret
;
741 *images_supported
= 0;
743 return sizeof(uint32_t);
744 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
746 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
748 uint32_t *subgroup_size
= ret
;
749 *subgroup_size
= r600_wavefront_size(rscreen
->family
);
751 return sizeof(uint32_t);
754 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
758 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
760 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
762 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
763 rscreen
->info
.clock_crystal_freq
;
766 static void r600_fence_reference(struct pipe_screen
*screen
,
767 struct pipe_fence_handle
**dst
,
768 struct pipe_fence_handle
*src
)
770 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
771 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
772 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
774 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
775 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
776 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
782 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
783 struct pipe_fence_handle
*fence
,
786 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
787 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
788 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
791 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
794 /* Recompute the timeout after waiting. */
795 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
796 int64_t time
= os_time_get_nano();
797 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
804 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
807 static void r600_query_memory_info(struct pipe_screen
*screen
,
808 struct pipe_memory_info
*info
)
810 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
811 struct radeon_winsys
*ws
= rscreen
->ws
;
812 unsigned vram_usage
, gtt_usage
;
814 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
815 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
817 /* The real TTM memory usage is somewhat random, because:
819 * 1) TTM delays freeing memory, because it can only free it after
822 * 2) The memory usage can be really low if big VRAM evictions are
823 * taking place, but the real usage is well above the size of VRAM.
825 * Instead, return statistics of this process.
827 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
828 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
830 info
->avail_device_memory
=
831 vram_usage
<= info
->total_device_memory
?
832 info
->total_device_memory
- vram_usage
: 0;
833 info
->avail_staging_memory
=
834 gtt_usage
<= info
->total_staging_memory
?
835 info
->total_staging_memory
- gtt_usage
: 0;
837 info
->device_memory_evicted
=
838 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
839 /* Just return the number of evicted 64KB pages. */
840 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
843 struct pipe_resource
*r600_resource_create_common(struct pipe_screen
*screen
,
844 const struct pipe_resource
*templ
)
846 if (templ
->target
== PIPE_BUFFER
) {
847 return r600_buffer_create(screen
, templ
, 4096);
849 return r600_texture_create(screen
, templ
);
853 bool r600_common_screen_init(struct r600_common_screen
*rscreen
,
854 struct radeon_winsys
*ws
)
856 char llvm_string
[32] = {};
858 ws
->query_info(ws
, &rscreen
->info
);
861 snprintf(llvm_string
, sizeof(llvm_string
),
862 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
863 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
866 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
867 "%s (DRM %i.%i.%i%s)",
868 r600_get_chip_name(rscreen
), rscreen
->info
.drm_major
,
869 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
872 rscreen
->b
.get_name
= r600_get_name
;
873 rscreen
->b
.get_vendor
= r600_get_vendor
;
874 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
875 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
876 rscreen
->b
.get_paramf
= r600_get_paramf
;
877 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
878 rscreen
->b
.fence_finish
= r600_fence_finish
;
879 rscreen
->b
.fence_reference
= r600_fence_reference
;
880 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
881 rscreen
->b
.resource_from_user_memory
= r600_buffer_from_user_memory
;
882 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
884 if (rscreen
->info
.has_uvd
) {
885 rscreen
->b
.get_video_param
= rvid_get_video_param
;
886 rscreen
->b
.is_video_format_supported
= rvid_is_format_supported
;
888 rscreen
->b
.get_video_param
= r600_get_video_param
;
889 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
892 r600_init_screen_texture_functions(rscreen
);
893 r600_init_screen_query_functions(rscreen
);
896 rscreen
->family
= rscreen
->info
.family
;
897 rscreen
->chip_class
= rscreen
->info
.chip_class
;
898 rscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
900 util_format_s3tc_init();
901 pipe_mutex_init(rscreen
->aux_context_lock
);
902 pipe_mutex_init(rscreen
->gpu_load_mutex
);
904 if (rscreen
->debug_flags
& DBG_INFO
) {
905 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
906 printf("family = %i (%s)\n", rscreen
->info
.family
,
907 r600_get_chip_name(rscreen
));
908 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
909 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
910 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
911 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
912 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
913 printf("has_sdma = %i\n", rscreen
->info
.has_sdma
);
914 printf("has_uvd = %i\n", rscreen
->info
.has_uvd
);
915 printf("vce_fw_version = %i\n", rscreen
->info
.vce_fw_version
);
916 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
917 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
918 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
919 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
920 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
922 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
923 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
924 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
925 printf("max_se = %i\n", rscreen
->info
.max_se
);
926 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
928 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
929 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
930 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
931 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
932 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
933 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
934 printf("si_tile_mode_array_valid = %i\n", rscreen
->info
.si_tile_mode_array_valid
);
935 printf("cik_macrotile_mode_array_valid = %i\n", rscreen
->info
.cik_macrotile_mode_array_valid
);
940 void r600_destroy_common_screen(struct r600_common_screen
*rscreen
)
942 r600_perfcounters_destroy(rscreen
);
943 r600_gpu_load_kill_thread(rscreen
);
945 pipe_mutex_destroy(rscreen
->gpu_load_mutex
);
946 pipe_mutex_destroy(rscreen
->aux_context_lock
);
947 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
949 rscreen
->ws
->destroy(rscreen
->ws
);
953 bool r600_can_dump_shader(struct r600_common_screen
*rscreen
,
957 case TGSI_PROCESSOR_VERTEX
:
958 return (rscreen
->debug_flags
& DBG_VS
) != 0;
959 case TGSI_PROCESSOR_TESS_CTRL
:
960 return (rscreen
->debug_flags
& DBG_TCS
) != 0;
961 case TGSI_PROCESSOR_TESS_EVAL
:
962 return (rscreen
->debug_flags
& DBG_TES
) != 0;
963 case TGSI_PROCESSOR_GEOMETRY
:
964 return (rscreen
->debug_flags
& DBG_GS
) != 0;
965 case TGSI_PROCESSOR_FRAGMENT
:
966 return (rscreen
->debug_flags
& DBG_PS
) != 0;
967 case TGSI_PROCESSOR_COMPUTE
:
968 return (rscreen
->debug_flags
& DBG_CS
) != 0;
974 void r600_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
975 unsigned offset
, unsigned size
, unsigned value
,
978 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
980 pipe_mutex_lock(rscreen
->aux_context_lock
);
981 rctx
->clear_buffer(&rctx
->b
, dst
, offset
, size
, value
, is_framebuffer
);
982 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
983 pipe_mutex_unlock(rscreen
->aux_context_lock
);