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 vram
+= dst
->vram_usage
;
148 gtt
+= dst
->gart_usage
;
151 vram
+= src
->vram_usage
;
152 gtt
+= src
->gart_usage
;
155 /* Flush the GFX IB if DMA depends on it. */
156 if (radeon_emitted(ctx
->gfx
.cs
, ctx
->initial_gfx_cs_size
) &&
158 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, dst
->buf
,
159 RADEON_USAGE_READWRITE
)) ||
161 ctx
->ws
->cs_is_buffer_referenced(ctx
->gfx
.cs
, src
->buf
,
162 RADEON_USAGE_WRITE
))))
163 ctx
->gfx
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
165 /* Flush if there's not enough space, or if the memory usage per IB
168 if (!ctx
->ws
->cs_check_space(ctx
->dma
.cs
, num_dw
) ||
169 !radeon_cs_memory_below_limit(ctx
->screen
, ctx
->dma
.cs
, vram
, gtt
)) {
170 ctx
->dma
.flush(ctx
, RADEON_FLUSH_ASYNC
, NULL
);
171 assert((num_dw
+ ctx
->dma
.cs
->current
.cdw
) <= ctx
->dma
.cs
->current
.max_dw
);
174 /* If GPUVM is not supported, the CS checker needs 2 entries
175 * in the buffer list per packet, which has to be done manually.
177 if (ctx
->screen
->info
.has_virtual_memory
) {
179 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, dst
,
181 RADEON_PRIO_SDMA_BUFFER
);
183 radeon_add_to_buffer_list(ctx
, &ctx
->dma
, src
,
185 RADEON_PRIO_SDMA_BUFFER
);
189 /* This is required to prevent read-after-write hazards. */
190 void r600_dma_emit_wait_idle(struct r600_common_context
*rctx
)
192 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
194 /* done at the end of DMA calls, so increment this. */
195 rctx
->num_dma_calls
++;
197 /* IBs using too little memory are limited by the IB submission overhead.
198 * IBs using too much memory are limited by the kernel/TTM overhead.
199 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
201 * This heuristic makes sure that DMA requests are executed
202 * very soon after the call is made and lowers memory usage.
203 * It improves texture upload performance by keeping the DMA
204 * engine busy while uploads are being submitted.
206 if (cs
->used_vram
+ cs
->used_gart
> 64 * 1024 * 1024) {
207 rctx
->dma
.flush(rctx
, RADEON_FLUSH_ASYNC
, NULL
);
211 r600_need_dma_space(rctx
, 1, NULL
, NULL
);
213 if (!radeon_emitted(cs
, 0)) /* empty queue */
216 /* NOP waits for idle on Evergreen and later. */
217 if (rctx
->chip_class
>= CIK
)
218 radeon_emit(cs
, 0x00000000); /* NOP */
219 else if (rctx
->chip_class
>= EVERGREEN
)
220 radeon_emit(cs
, 0xf0000000); /* NOP */
222 /* TODO: R600-R700 should use the FENCE packet.
223 * CS checker support is required. */
227 static void r600_memory_barrier(struct pipe_context
*ctx
, unsigned flags
)
231 void r600_preflush_suspend_features(struct r600_common_context
*ctx
)
233 /* suspend queries */
234 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
235 r600_suspend_queries(ctx
);
237 ctx
->streamout
.suspended
= false;
238 if (ctx
->streamout
.begin_emitted
) {
239 r600_emit_streamout_end(ctx
);
240 ctx
->streamout
.suspended
= true;
244 void r600_postflush_resume_features(struct r600_common_context
*ctx
)
246 if (ctx
->streamout
.suspended
) {
247 ctx
->streamout
.append_bitmask
= ctx
->streamout
.enabled_mask
;
248 r600_streamout_buffers_dirty(ctx
);
252 if (!LIST_IS_EMPTY(&ctx
->active_queries
))
253 r600_resume_queries(ctx
);
256 static void r600_flush_from_st(struct pipe_context
*ctx
,
257 struct pipe_fence_handle
**fence
,
260 struct pipe_screen
*screen
= ctx
->screen
;
261 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
263 struct pipe_fence_handle
*gfx_fence
= NULL
;
264 struct pipe_fence_handle
*sdma_fence
= NULL
;
266 if (flags
& PIPE_FLUSH_END_OF_FRAME
)
267 rflags
|= RADEON_FLUSH_END_OF_FRAME
;
268 if (flags
& PIPE_FLUSH_DEFERRED
)
269 rflags
|= RADEON_FLUSH_ASYNC
;
272 rctx
->dma
.flush(rctx
, rflags
, fence
? &sdma_fence
: NULL
);
274 rctx
->gfx
.flush(rctx
, rflags
, fence
? &gfx_fence
: NULL
);
276 /* Both engines can signal out of order, so we need to keep both fences. */
277 if (gfx_fence
|| sdma_fence
) {
278 struct r600_multi_fence
*multi_fence
=
279 CALLOC_STRUCT(r600_multi_fence
);
283 multi_fence
->reference
.count
= 1;
284 multi_fence
->gfx
= gfx_fence
;
285 multi_fence
->sdma
= sdma_fence
;
287 screen
->fence_reference(screen
, fence
, NULL
);
288 *fence
= (struct pipe_fence_handle
*)multi_fence
;
292 static void r600_flush_dma_ring(void *ctx
, unsigned flags
,
293 struct pipe_fence_handle
**fence
)
295 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
296 struct radeon_winsys_cs
*cs
= rctx
->dma
.cs
;
297 struct radeon_saved_cs saved
;
299 (rctx
->screen
->debug_flags
& DBG_CHECK_VM
) &&
300 rctx
->check_vm_faults
;
302 if (!radeon_emitted(cs
, 0)) {
304 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
309 radeon_save_cs(rctx
->ws
, cs
, &saved
);
311 rctx
->ws
->cs_flush(cs
, flags
, &rctx
->last_sdma_fence
);
313 rctx
->ws
->fence_reference(fence
, rctx
->last_sdma_fence
);
316 /* Use conservative timeout 800ms, after which we won't wait any
317 * longer and assume the GPU is hung.
319 rctx
->ws
->fence_wait(rctx
->ws
, rctx
->last_sdma_fence
, 800*1000*1000);
321 rctx
->check_vm_faults(rctx
, &saved
, RING_DMA
);
322 radeon_clear_saved_cs(&saved
);
327 * Store a linearized copy of all chunks of \p cs together with the buffer
330 void radeon_save_cs(struct radeon_winsys
*ws
, struct radeon_winsys_cs
*cs
,
331 struct radeon_saved_cs
*saved
)
336 /* Save the IB chunks. */
337 saved
->num_dw
= cs
->prev_dw
+ cs
->current
.cdw
;
338 saved
->ib
= MALLOC(4 * saved
->num_dw
);
343 for (i
= 0; i
< cs
->num_prev
; ++i
) {
344 memcpy(buf
, cs
->prev
[i
].buf
, cs
->prev
[i
].cdw
* 4);
345 buf
+= cs
->prev
[i
].cdw
;
347 memcpy(buf
, cs
->current
.buf
, cs
->current
.cdw
* 4);
349 /* Save the buffer list. */
350 saved
->bo_count
= ws
->cs_get_buffer_list(cs
, NULL
);
351 saved
->bo_list
= CALLOC(saved
->bo_count
,
352 sizeof(saved
->bo_list
[0]));
353 if (!saved
->bo_list
) {
357 ws
->cs_get_buffer_list(cs
, saved
->bo_list
);
362 fprintf(stderr
, "%s: out of memory\n", __func__
);
363 memset(saved
, 0, sizeof(*saved
));
366 void radeon_clear_saved_cs(struct radeon_saved_cs
*saved
)
369 FREE(saved
->bo_list
);
371 memset(saved
, 0, sizeof(*saved
));
374 static enum pipe_reset_status
r600_get_reset_status(struct pipe_context
*ctx
)
376 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
377 unsigned latest
= rctx
->ws
->query_value(rctx
->ws
,
378 RADEON_GPU_RESET_COUNTER
);
380 if (rctx
->gpu_reset_counter
== latest
)
381 return PIPE_NO_RESET
;
383 rctx
->gpu_reset_counter
= latest
;
384 return PIPE_UNKNOWN_CONTEXT_RESET
;
387 static void r600_set_debug_callback(struct pipe_context
*ctx
,
388 const struct pipe_debug_callback
*cb
)
390 struct r600_common_context
*rctx
= (struct r600_common_context
*)ctx
;
395 memset(&rctx
->debug
, 0, sizeof(rctx
->debug
));
398 bool r600_common_context_init(struct r600_common_context
*rctx
,
399 struct r600_common_screen
*rscreen
,
400 unsigned context_flags
)
402 util_slab_create(&rctx
->pool_transfers
,
403 sizeof(struct r600_transfer
), 64,
404 UTIL_SLAB_SINGLETHREADED
);
406 rctx
->screen
= rscreen
;
407 rctx
->ws
= rscreen
->ws
;
408 rctx
->family
= rscreen
->family
;
409 rctx
->chip_class
= rscreen
->chip_class
;
411 if (rscreen
->chip_class
>= CIK
)
412 rctx
->max_db
= MAX2(8, rscreen
->info
.num_render_backends
);
413 else if (rscreen
->chip_class
>= EVERGREEN
)
418 rctx
->b
.invalidate_resource
= r600_invalidate_resource
;
419 rctx
->b
.transfer_map
= u_transfer_map_vtbl
;
420 rctx
->b
.transfer_flush_region
= u_transfer_flush_region_vtbl
;
421 rctx
->b
.transfer_unmap
= u_transfer_unmap_vtbl
;
422 rctx
->b
.texture_subdata
= u_default_texture_subdata
;
423 rctx
->b
.memory_barrier
= r600_memory_barrier
;
424 rctx
->b
.flush
= r600_flush_from_st
;
425 rctx
->b
.set_debug_callback
= r600_set_debug_callback
;
427 /* evergreen_compute.c has a special codepath for global buffers.
428 * Everything else can use the direct path.
430 if ((rscreen
->chip_class
== EVERGREEN
|| rscreen
->chip_class
== CAYMAN
) &&
431 (context_flags
& PIPE_CONTEXT_COMPUTE_ONLY
))
432 rctx
->b
.buffer_subdata
= u_default_buffer_subdata
;
434 rctx
->b
.buffer_subdata
= r600_buffer_subdata
;
436 if (rscreen
->info
.drm_major
== 2 && rscreen
->info
.drm_minor
>= 43) {
437 rctx
->b
.get_device_reset_status
= r600_get_reset_status
;
438 rctx
->gpu_reset_counter
=
439 rctx
->ws
->query_value(rctx
->ws
,
440 RADEON_GPU_RESET_COUNTER
);
443 LIST_INITHEAD(&rctx
->texture_buffers
);
445 r600_init_context_texture_functions(rctx
);
446 r600_init_viewport_functions(rctx
);
447 r600_streamout_init(rctx
);
448 r600_query_init(rctx
);
449 cayman_init_msaa(&rctx
->b
);
451 rctx
->allocator_zeroed_memory
=
452 u_suballocator_create(&rctx
->b
, rscreen
->info
.gart_page_size
,
453 0, PIPE_USAGE_DEFAULT
, true);
454 if (!rctx
->allocator_zeroed_memory
)
457 rctx
->uploader
= u_upload_create(&rctx
->b
, 1024 * 1024,
458 PIPE_BIND_INDEX_BUFFER
|
459 PIPE_BIND_CONSTANT_BUFFER
, PIPE_USAGE_STREAM
);
463 rctx
->ctx
= rctx
->ws
->ctx_create(rctx
->ws
);
467 if (rscreen
->info
.has_sdma
&& !(rscreen
->debug_flags
& DBG_NO_ASYNC_DMA
)) {
468 rctx
->dma
.cs
= rctx
->ws
->cs_create(rctx
->ctx
, RING_DMA
,
471 rctx
->dma
.flush
= r600_flush_dma_ring
;
474 p_atomic_inc(&rscreen
->num_contexts
);
478 void r600_common_context_cleanup(struct r600_common_context
*rctx
)
482 p_atomic_dec(&rctx
->screen
->num_contexts
);
484 /* Release DCC stats. */
485 for (i
= 0; i
< ARRAY_SIZE(rctx
->dcc_stats
); i
++) {
486 assert(!rctx
->dcc_stats
[i
].query_active
);
488 for (j
= 0; j
< ARRAY_SIZE(rctx
->dcc_stats
[i
].ps_stats
); j
++)
489 if (rctx
->dcc_stats
[i
].ps_stats
[j
])
490 rctx
->b
.destroy_query(&rctx
->b
,
491 rctx
->dcc_stats
[i
].ps_stats
[j
]);
493 r600_texture_reference(&rctx
->dcc_stats
[i
].tex
, NULL
);
497 rctx
->ws
->cs_destroy(rctx
->gfx
.cs
);
499 rctx
->ws
->cs_destroy(rctx
->dma
.cs
);
501 rctx
->ws
->ctx_destroy(rctx
->ctx
);
503 if (rctx
->uploader
) {
504 u_upload_destroy(rctx
->uploader
);
507 util_slab_destroy(&rctx
->pool_transfers
);
509 if (rctx
->allocator_zeroed_memory
) {
510 u_suballocator_destroy(rctx
->allocator_zeroed_memory
);
512 rctx
->ws
->fence_reference(&rctx
->last_gfx_fence
, NULL
);
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 rctx
->vram
+= rr
->vram_usage
;
533 rctx
->gtt
+= rr
->gart_usage
;
540 static const struct debug_named_value common_debug_options
[] = {
542 { "tex", DBG_TEX
, "Print texture info" },
543 { "compute", DBG_COMPUTE
, "Print compute info" },
544 { "vm", DBG_VM
, "Print virtual addresses when creating resources" },
545 { "info", DBG_INFO
, "Print driver information" },
548 { "fs", DBG_FS
, "Print fetch shaders" },
549 { "vs", DBG_VS
, "Print vertex shaders" },
550 { "gs", DBG_GS
, "Print geometry shaders" },
551 { "ps", DBG_PS
, "Print pixel shaders" },
552 { "cs", DBG_CS
, "Print compute shaders" },
553 { "tcs", DBG_TCS
, "Print tessellation control shaders" },
554 { "tes", DBG_TES
, "Print tessellation evaluation shaders" },
555 { "noir", DBG_NO_IR
, "Don't print the LLVM IR"},
556 { "notgsi", DBG_NO_TGSI
, "Don't print the TGSI"},
557 { "noasm", DBG_NO_ASM
, "Don't print disassembled shaders"},
558 { "preoptir", DBG_PREOPT_IR
, "Print the LLVM IR before initial optimizations" },
560 { "testdma", DBG_TEST_DMA
, "Invoke SDMA tests and exit." },
563 { "nodma", DBG_NO_ASYNC_DMA
, "Disable asynchronous DMA" },
564 { "nohyperz", DBG_NO_HYPERZ
, "Disable Hyper-Z" },
565 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
566 { "noinvalrange", DBG_NO_DISCARD_RANGE
, "Disable handling of INVALIDATE_RANGE map flags" },
567 { "no2d", DBG_NO_2D_TILING
, "Disable 2D tiling" },
568 { "notiling", DBG_NO_TILING
, "Disable tiling" },
569 { "switch_on_eop", DBG_SWITCH_ON_EOP
, "Program WD/IA to switch on end-of-packet." },
570 { "forcedma", DBG_FORCE_DMA
, "Use asynchronous DMA for all operations when possible." },
571 { "precompile", DBG_PRECOMPILE
, "Compile one shader variant at shader creation." },
572 { "nowc", DBG_NO_WC
, "Disable GTT write combining" },
573 { "check_vm", DBG_CHECK_VM
, "Check VM faults and dump debug info." },
574 { "nodcc", DBG_NO_DCC
, "Disable DCC." },
575 { "nodccclear", DBG_NO_DCC_CLEAR
, "Disable DCC fast clear." },
576 { "norbplus", DBG_NO_RB_PLUS
, "Disable RB+ on Stoney." },
577 { "sisched", DBG_SI_SCHED
, "Enable LLVM SI Machine Instruction Scheduler." },
578 { "mono", DBG_MONOLITHIC_SHADERS
, "Use old-style monolithic shaders compiled on demand" },
579 { "noce", DBG_NO_CE
, "Disable the constant engine"},
580 { "unsafemath", DBG_UNSAFE_MATH
, "Enable unsafe math shader optimizations" },
581 { "nodccfb", DBG_NO_DCC_FB
, "Disable separate DCC on the main framebuffer" },
583 DEBUG_NAMED_VALUE_END
/* must be last */
586 static const char* r600_get_vendor(struct pipe_screen
* pscreen
)
591 static const char* r600_get_device_vendor(struct pipe_screen
* pscreen
)
596 static const char* r600_get_chip_name(struct r600_common_screen
*rscreen
)
598 switch (rscreen
->info
.family
) {
599 case CHIP_R600
: return "AMD R600";
600 case CHIP_RV610
: return "AMD RV610";
601 case CHIP_RV630
: return "AMD RV630";
602 case CHIP_RV670
: return "AMD RV670";
603 case CHIP_RV620
: return "AMD RV620";
604 case CHIP_RV635
: return "AMD RV635";
605 case CHIP_RS780
: return "AMD RS780";
606 case CHIP_RS880
: return "AMD RS880";
607 case CHIP_RV770
: return "AMD RV770";
608 case CHIP_RV730
: return "AMD RV730";
609 case CHIP_RV710
: return "AMD RV710";
610 case CHIP_RV740
: return "AMD RV740";
611 case CHIP_CEDAR
: return "AMD CEDAR";
612 case CHIP_REDWOOD
: return "AMD REDWOOD";
613 case CHIP_JUNIPER
: return "AMD JUNIPER";
614 case CHIP_CYPRESS
: return "AMD CYPRESS";
615 case CHIP_HEMLOCK
: return "AMD HEMLOCK";
616 case CHIP_PALM
: return "AMD PALM";
617 case CHIP_SUMO
: return "AMD SUMO";
618 case CHIP_SUMO2
: return "AMD SUMO2";
619 case CHIP_BARTS
: return "AMD BARTS";
620 case CHIP_TURKS
: return "AMD TURKS";
621 case CHIP_CAICOS
: return "AMD CAICOS";
622 case CHIP_CAYMAN
: return "AMD CAYMAN";
623 case CHIP_ARUBA
: return "AMD ARUBA";
624 case CHIP_TAHITI
: return "AMD TAHITI";
625 case CHIP_PITCAIRN
: return "AMD PITCAIRN";
626 case CHIP_VERDE
: return "AMD CAPE VERDE";
627 case CHIP_OLAND
: return "AMD OLAND";
628 case CHIP_HAINAN
: return "AMD HAINAN";
629 case CHIP_BONAIRE
: return "AMD BONAIRE";
630 case CHIP_KAVERI
: return "AMD KAVERI";
631 case CHIP_KABINI
: return "AMD KABINI";
632 case CHIP_HAWAII
: return "AMD HAWAII";
633 case CHIP_MULLINS
: return "AMD MULLINS";
634 case CHIP_TONGA
: return "AMD TONGA";
635 case CHIP_ICELAND
: return "AMD ICELAND";
636 case CHIP_CARRIZO
: return "AMD CARRIZO";
637 case CHIP_FIJI
: return "AMD FIJI";
638 case CHIP_POLARIS10
: return "AMD POLARIS10";
639 case CHIP_POLARIS11
: return "AMD POLARIS11";
640 case CHIP_STONEY
: return "AMD STONEY";
641 default: return "AMD unknown";
645 static const char* r600_get_name(struct pipe_screen
* pscreen
)
647 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
649 return rscreen
->renderer_string
;
652 static float r600_get_paramf(struct pipe_screen
* pscreen
,
653 enum pipe_capf param
)
655 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)pscreen
;
658 case PIPE_CAPF_MAX_LINE_WIDTH
:
659 case PIPE_CAPF_MAX_LINE_WIDTH_AA
:
660 case PIPE_CAPF_MAX_POINT_WIDTH
:
661 case PIPE_CAPF_MAX_POINT_WIDTH_AA
:
662 if (rscreen
->family
>= CHIP_CEDAR
)
666 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY
:
668 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS
:
670 case PIPE_CAPF_GUARD_BAND_LEFT
:
671 case PIPE_CAPF_GUARD_BAND_TOP
:
672 case PIPE_CAPF_GUARD_BAND_RIGHT
:
673 case PIPE_CAPF_GUARD_BAND_BOTTOM
:
679 static int r600_get_video_param(struct pipe_screen
*screen
,
680 enum pipe_video_profile profile
,
681 enum pipe_video_entrypoint entrypoint
,
682 enum pipe_video_cap param
)
685 case PIPE_VIDEO_CAP_SUPPORTED
:
686 return vl_profile_supported(screen
, profile
, entrypoint
);
687 case PIPE_VIDEO_CAP_NPOT_TEXTURES
:
689 case PIPE_VIDEO_CAP_MAX_WIDTH
:
690 case PIPE_VIDEO_CAP_MAX_HEIGHT
:
691 return vl_video_buffer_max_size(screen
);
692 case PIPE_VIDEO_CAP_PREFERED_FORMAT
:
693 return PIPE_FORMAT_NV12
;
694 case PIPE_VIDEO_CAP_PREFERS_INTERLACED
:
696 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED
:
698 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE
:
700 case PIPE_VIDEO_CAP_MAX_LEVEL
:
701 return vl_level_supported(screen
, profile
);
707 const char *r600_get_llvm_processor_name(enum radeon_family family
)
750 case CHIP_TAHITI
: return "tahiti";
751 case CHIP_PITCAIRN
: return "pitcairn";
752 case CHIP_VERDE
: return "verde";
753 case CHIP_OLAND
: return "oland";
754 case CHIP_HAINAN
: return "hainan";
755 case CHIP_BONAIRE
: return "bonaire";
756 case CHIP_KABINI
: return "kabini";
757 case CHIP_KAVERI
: return "kaveri";
758 case CHIP_HAWAII
: return "hawaii";
761 case CHIP_TONGA
: return "tonga";
762 case CHIP_ICELAND
: return "iceland";
763 case CHIP_CARRIZO
: return "carrizo";
764 #if HAVE_LLVM <= 0x0307
765 case CHIP_FIJI
: return "tonga";
766 case CHIP_STONEY
: return "carrizo";
768 case CHIP_FIJI
: return "fiji";
769 case CHIP_STONEY
: return "stoney";
771 #if HAVE_LLVM <= 0x0308
772 case CHIP_POLARIS10
: return "tonga";
773 case CHIP_POLARIS11
: return "tonga";
775 case CHIP_POLARIS10
: return "polaris10";
776 case CHIP_POLARIS11
: return "polaris11";
782 static int r600_get_compute_param(struct pipe_screen
*screen
,
783 enum pipe_shader_ir ir_type
,
784 enum pipe_compute_cap param
,
787 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
789 //TODO: select these params by asic
791 case PIPE_COMPUTE_CAP_IR_TARGET
: {
794 if (rscreen
->family
<= CHIP_ARUBA
) {
799 switch(rscreen
->family
) {
800 /* Clang < 3.6 is missing Hainan in its list of
801 * GPUs, so we need to use the name of a similar GPU.
804 gpu
= r600_get_llvm_processor_name(rscreen
->family
);
808 sprintf(ret
, "%s-%s", gpu
, triple
);
810 /* +2 for dash and terminating NIL byte */
811 return (strlen(triple
) + strlen(gpu
) + 2) * sizeof(char);
813 case PIPE_COMPUTE_CAP_GRID_DIMENSION
:
815 uint64_t *grid_dimension
= ret
;
816 grid_dimension
[0] = 3;
818 return 1 * sizeof(uint64_t);
820 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE
:
822 uint64_t *grid_size
= ret
;
823 grid_size
[0] = 65535;
824 grid_size
[1] = 65535;
825 grid_size
[2] = 65535;
827 return 3 * sizeof(uint64_t) ;
829 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE
:
831 uint64_t *block_size
= ret
;
832 if (rscreen
->chip_class
>= SI
&& HAVE_LLVM
>= 0x309 &&
833 ir_type
== PIPE_SHADER_IR_TGSI
) {
834 block_size
[0] = 2048;
835 block_size
[1] = 2048;
836 block_size
[2] = 2048;
843 return 3 * sizeof(uint64_t);
845 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK
:
847 uint64_t *max_threads_per_block
= ret
;
848 if (rscreen
->chip_class
>= SI
&& HAVE_LLVM
>= 0x309 &&
849 ir_type
== PIPE_SHADER_IR_TGSI
)
850 *max_threads_per_block
= 2048;
852 *max_threads_per_block
= 256;
854 return sizeof(uint64_t);
856 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE
:
858 uint64_t *max_global_size
= ret
;
859 uint64_t max_mem_alloc_size
;
861 r600_get_compute_param(screen
, ir_type
,
862 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
,
863 &max_mem_alloc_size
);
865 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
866 * 1/4 of the MAX_GLOBAL_SIZE. Since the
867 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
868 * make sure we never report more than
869 * 4 * MAX_MEM_ALLOC_SIZE.
871 *max_global_size
= MIN2(4 * max_mem_alloc_size
,
872 MAX2(rscreen
->info
.gart_size
,
873 rscreen
->info
.vram_size
));
875 return sizeof(uint64_t);
877 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE
:
879 uint64_t *max_local_size
= ret
;
880 /* Value reported by the closed source driver. */
881 *max_local_size
= 32768;
883 return sizeof(uint64_t);
885 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE
:
887 uint64_t *max_input_size
= ret
;
888 /* Value reported by the closed source driver. */
889 *max_input_size
= 1024;
891 return sizeof(uint64_t);
893 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE
:
895 uint64_t *max_mem_alloc_size
= ret
;
897 *max_mem_alloc_size
= rscreen
->info
.max_alloc_size
;
899 return sizeof(uint64_t);
901 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY
:
903 uint32_t *max_clock_frequency
= ret
;
904 *max_clock_frequency
= rscreen
->info
.max_shader_clock
;
906 return sizeof(uint32_t);
908 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS
:
910 uint32_t *max_compute_units
= ret
;
911 *max_compute_units
= rscreen
->info
.num_good_compute_units
;
913 return sizeof(uint32_t);
915 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED
:
917 uint32_t *images_supported
= ret
;
918 *images_supported
= 0;
920 return sizeof(uint32_t);
921 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE
:
923 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE
:
925 uint32_t *subgroup_size
= ret
;
926 *subgroup_size
= r600_wavefront_size(rscreen
->family
);
928 return sizeof(uint32_t);
931 fprintf(stderr
, "unknown PIPE_COMPUTE_CAP %d\n", param
);
935 static uint64_t r600_get_timestamp(struct pipe_screen
*screen
)
937 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
939 return 1000000 * rscreen
->ws
->query_value(rscreen
->ws
, RADEON_TIMESTAMP
) /
940 rscreen
->info
.clock_crystal_freq
;
943 static void r600_fence_reference(struct pipe_screen
*screen
,
944 struct pipe_fence_handle
**dst
,
945 struct pipe_fence_handle
*src
)
947 struct radeon_winsys
*ws
= ((struct r600_common_screen
*)screen
)->ws
;
948 struct r600_multi_fence
**rdst
= (struct r600_multi_fence
**)dst
;
949 struct r600_multi_fence
*rsrc
= (struct r600_multi_fence
*)src
;
951 if (pipe_reference(&(*rdst
)->reference
, &rsrc
->reference
)) {
952 ws
->fence_reference(&(*rdst
)->gfx
, NULL
);
953 ws
->fence_reference(&(*rdst
)->sdma
, NULL
);
959 static boolean
r600_fence_finish(struct pipe_screen
*screen
,
960 struct pipe_fence_handle
*fence
,
963 struct radeon_winsys
*rws
= ((struct r600_common_screen
*)screen
)->ws
;
964 struct r600_multi_fence
*rfence
= (struct r600_multi_fence
*)fence
;
965 int64_t abs_timeout
= os_time_get_absolute_timeout(timeout
);
968 if (!rws
->fence_wait(rws
, rfence
->sdma
, timeout
))
971 /* Recompute the timeout after waiting. */
972 if (timeout
&& timeout
!= PIPE_TIMEOUT_INFINITE
) {
973 int64_t time
= os_time_get_nano();
974 timeout
= abs_timeout
> time
? abs_timeout
- time
: 0;
981 return rws
->fence_wait(rws
, rfence
->gfx
, timeout
);
984 static void r600_query_memory_info(struct pipe_screen
*screen
,
985 struct pipe_memory_info
*info
)
987 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
988 struct radeon_winsys
*ws
= rscreen
->ws
;
989 unsigned vram_usage
, gtt_usage
;
991 info
->total_device_memory
= rscreen
->info
.vram_size
/ 1024;
992 info
->total_staging_memory
= rscreen
->info
.gart_size
/ 1024;
994 /* The real TTM memory usage is somewhat random, because:
996 * 1) TTM delays freeing memory, because it can only free it after
999 * 2) The memory usage can be really low if big VRAM evictions are
1000 * taking place, but the real usage is well above the size of VRAM.
1002 * Instead, return statistics of this process.
1004 vram_usage
= ws
->query_value(ws
, RADEON_REQUESTED_VRAM_MEMORY
) / 1024;
1005 gtt_usage
= ws
->query_value(ws
, RADEON_REQUESTED_GTT_MEMORY
) / 1024;
1007 info
->avail_device_memory
=
1008 vram_usage
<= info
->total_device_memory
?
1009 info
->total_device_memory
- vram_usage
: 0;
1010 info
->avail_staging_memory
=
1011 gtt_usage
<= info
->total_staging_memory
?
1012 info
->total_staging_memory
- gtt_usage
: 0;
1014 info
->device_memory_evicted
=
1015 ws
->query_value(ws
, RADEON_NUM_BYTES_MOVED
) / 1024;
1016 /* Just return the number of evicted 64KB pages. */
1017 info
->nr_device_memory_evictions
= info
->device_memory_evicted
/ 64;
1020 struct pipe_resource
*r600_resource_create_common(struct pipe_screen
*screen
,
1021 const struct pipe_resource
*templ
)
1023 struct r600_common_screen
*rscreen
= (struct r600_common_screen
*)screen
;
1025 if (templ
->target
== PIPE_BUFFER
) {
1026 return r600_buffer_create(screen
, templ
,
1027 rscreen
->info
.gart_page_size
);
1029 return r600_texture_create(screen
, templ
);
1033 bool r600_common_screen_init(struct r600_common_screen
*rscreen
,
1034 struct radeon_winsys
*ws
)
1036 char llvm_string
[32] = {}, kernel_version
[128] = {};
1037 struct utsname uname_data
;
1039 ws
->query_info(ws
, &rscreen
->info
);
1041 if (uname(&uname_data
) == 0)
1042 snprintf(kernel_version
, sizeof(kernel_version
),
1043 " / %s", uname_data
.release
);
1046 snprintf(llvm_string
, sizeof(llvm_string
),
1047 ", LLVM %i.%i.%i", (HAVE_LLVM
>> 8) & 0xff,
1048 HAVE_LLVM
& 0xff, MESA_LLVM_VERSION_PATCH
);
1051 snprintf(rscreen
->renderer_string
, sizeof(rscreen
->renderer_string
),
1052 "%s (DRM %i.%i.%i%s%s)",
1053 r600_get_chip_name(rscreen
), rscreen
->info
.drm_major
,
1054 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
,
1055 kernel_version
, llvm_string
);
1057 rscreen
->b
.get_name
= r600_get_name
;
1058 rscreen
->b
.get_vendor
= r600_get_vendor
;
1059 rscreen
->b
.get_device_vendor
= r600_get_device_vendor
;
1060 rscreen
->b
.get_compute_param
= r600_get_compute_param
;
1061 rscreen
->b
.get_paramf
= r600_get_paramf
;
1062 rscreen
->b
.get_timestamp
= r600_get_timestamp
;
1063 rscreen
->b
.fence_finish
= r600_fence_finish
;
1064 rscreen
->b
.fence_reference
= r600_fence_reference
;
1065 rscreen
->b
.resource_destroy
= u_resource_destroy_vtbl
;
1066 rscreen
->b
.resource_from_user_memory
= r600_buffer_from_user_memory
;
1067 rscreen
->b
.query_memory_info
= r600_query_memory_info
;
1069 if (rscreen
->info
.has_uvd
) {
1070 rscreen
->b
.get_video_param
= rvid_get_video_param
;
1071 rscreen
->b
.is_video_format_supported
= rvid_is_format_supported
;
1073 rscreen
->b
.get_video_param
= r600_get_video_param
;
1074 rscreen
->b
.is_video_format_supported
= vl_video_buffer_is_format_supported
;
1077 r600_init_screen_texture_functions(rscreen
);
1078 r600_init_screen_query_functions(rscreen
);
1081 rscreen
->family
= rscreen
->info
.family
;
1082 rscreen
->chip_class
= rscreen
->info
.chip_class
;
1083 rscreen
->debug_flags
= debug_get_flags_option("R600_DEBUG", common_debug_options
, 0);
1085 rscreen
->force_aniso
= MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1086 if (rscreen
->force_aniso
>= 0) {
1087 printf("radeon: Forcing anisotropy filter to %ix\n",
1088 /* round down to a power of two */
1089 1 << util_logbase2(rscreen
->force_aniso
));
1092 util_format_s3tc_init();
1093 pipe_mutex_init(rscreen
->aux_context_lock
);
1094 pipe_mutex_init(rscreen
->gpu_load_mutex
);
1096 if (rscreen
->debug_flags
& DBG_INFO
) {
1097 printf("pci_id = 0x%x\n", rscreen
->info
.pci_id
);
1098 printf("family = %i (%s)\n", rscreen
->info
.family
,
1099 r600_get_chip_name(rscreen
));
1100 printf("chip_class = %i\n", rscreen
->info
.chip_class
);
1101 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.gart_size
, 1024*1024));
1102 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen
->info
.vram_size
, 1024*1024));
1103 printf("max_alloc_size = %i MB\n",
1104 (int)DIV_ROUND_UP(rscreen
->info
.max_alloc_size
, 1024*1024));
1105 printf("has_virtual_memory = %i\n", rscreen
->info
.has_virtual_memory
);
1106 printf("gfx_ib_pad_with_type2 = %i\n", rscreen
->info
.gfx_ib_pad_with_type2
);
1107 printf("has_sdma = %i\n", rscreen
->info
.has_sdma
);
1108 printf("has_uvd = %i\n", rscreen
->info
.has_uvd
);
1109 printf("vce_fw_version = %i\n", rscreen
->info
.vce_fw_version
);
1110 printf("vce_harvest_config = %i\n", rscreen
->info
.vce_harvest_config
);
1111 printf("clock_crystal_freq = %i\n", rscreen
->info
.clock_crystal_freq
);
1112 printf("drm = %i.%i.%i\n", rscreen
->info
.drm_major
,
1113 rscreen
->info
.drm_minor
, rscreen
->info
.drm_patchlevel
);
1114 printf("has_userptr = %i\n", rscreen
->info
.has_userptr
);
1116 printf("r600_max_quad_pipes = %i\n", rscreen
->info
.r600_max_quad_pipes
);
1117 printf("max_shader_clock = %i\n", rscreen
->info
.max_shader_clock
);
1118 printf("num_good_compute_units = %i\n", rscreen
->info
.num_good_compute_units
);
1119 printf("max_se = %i\n", rscreen
->info
.max_se
);
1120 printf("max_sh_per_se = %i\n", rscreen
->info
.max_sh_per_se
);
1122 printf("r600_gb_backend_map = %i\n", rscreen
->info
.r600_gb_backend_map
);
1123 printf("r600_gb_backend_map_valid = %i\n", rscreen
->info
.r600_gb_backend_map_valid
);
1124 printf("r600_num_banks = %i\n", rscreen
->info
.r600_num_banks
);
1125 printf("num_render_backends = %i\n", rscreen
->info
.num_render_backends
);
1126 printf("num_tile_pipes = %i\n", rscreen
->info
.num_tile_pipes
);
1127 printf("pipe_interleave_bytes = %i\n", rscreen
->info
.pipe_interleave_bytes
);
1132 void r600_destroy_common_screen(struct r600_common_screen
*rscreen
)
1134 r600_perfcounters_destroy(rscreen
);
1135 r600_gpu_load_kill_thread(rscreen
);
1137 pipe_mutex_destroy(rscreen
->gpu_load_mutex
);
1138 pipe_mutex_destroy(rscreen
->aux_context_lock
);
1139 rscreen
->aux_context
->destroy(rscreen
->aux_context
);
1141 rscreen
->ws
->destroy(rscreen
->ws
);
1145 bool r600_can_dump_shader(struct r600_common_screen
*rscreen
,
1148 switch (processor
) {
1149 case PIPE_SHADER_VERTEX
:
1150 return (rscreen
->debug_flags
& DBG_VS
) != 0;
1151 case PIPE_SHADER_TESS_CTRL
:
1152 return (rscreen
->debug_flags
& DBG_TCS
) != 0;
1153 case PIPE_SHADER_TESS_EVAL
:
1154 return (rscreen
->debug_flags
& DBG_TES
) != 0;
1155 case PIPE_SHADER_GEOMETRY
:
1156 return (rscreen
->debug_flags
& DBG_GS
) != 0;
1157 case PIPE_SHADER_FRAGMENT
:
1158 return (rscreen
->debug_flags
& DBG_PS
) != 0;
1159 case PIPE_SHADER_COMPUTE
:
1160 return (rscreen
->debug_flags
& DBG_CS
) != 0;
1166 void r600_screen_clear_buffer(struct r600_common_screen
*rscreen
, struct pipe_resource
*dst
,
1167 uint64_t offset
, uint64_t size
, unsigned value
,
1168 enum r600_coherency coher
)
1170 struct r600_common_context
*rctx
= (struct r600_common_context
*)rscreen
->aux_context
;
1172 pipe_mutex_lock(rscreen
->aux_context_lock
);
1173 rctx
->clear_buffer(&rctx
->b
, dst
, offset
, size
, value
, coher
);
1174 rscreen
->aux_context
->flush(rscreen
->aux_context
, NULL
, 0);
1175 pipe_mutex_unlock(rscreen
->aux_context_lock
);