1 /*
2 * Copyright 2013 Advanced Micro Devices, Inc.
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * SOFTWARE.
22 *
23 * Authors: Marek Olšák <maraeo@gmail.com>
24 *
25 */
26
27 #include "r600_pipe_common.h"
28 #include "r600_cs.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 "util/os_time.h"
36 #include "vl/vl_decoder.h"
37 #include "vl/vl_video_buffer.h"
38 #include "radeon_video.h"
39 #include <inttypes.h>
40 #include <sys/utsname.h>
41
42 #ifndef HAVE_LLVM
43 #define HAVE_LLVM 0
44 #endif
45
46 #if HAVE_LLVM
47 #include <llvm-c/TargetMachine.h>
48 #endif
49
50 #ifndef MESA_LLVM_VERSION_PATCH
51 #define MESA_LLVM_VERSION_PATCH 0
52 #endif
53
54 struct r600_multi_fence {
55 struct pipe_reference reference;
56 struct pipe_fence_handle *gfx;
57 struct pipe_fence_handle *sdma;
58
59 /* If the context wasn't flushed at fence creation, this is non-NULL. */
60 struct {
61 struct r600_common_context *ctx;
62 unsigned ib_index;
63 } gfx_unflushed;
64 };
65
66 /*
67 * shader binary helpers.
68 */
radeon_shader_binary_init(struct ac_shader_binary * b)69 void radeon_shader_binary_init(struct ac_shader_binary *b)
70 {
71 memset(b, 0, sizeof(*b));
72 }
73
radeon_shader_binary_clean(struct ac_shader_binary * b)74 void radeon_shader_binary_clean(struct ac_shader_binary *b)
75 {
76 if (!b)
77 return;
78 FREE(b->code);
79 FREE(b->config);
80 FREE(b->rodata);
81 FREE(b->global_symbol_offsets);
82 FREE(b->relocs);
83 FREE(b->disasm_string);
84 FREE(b->llvm_ir_string);
85 }
86
87 /*
88 * pipe_context
89 */
90
91 /**
92 * Write an EOP event.
93 *
94 * \param event EVENT_TYPE_*
95 * \param event_flags Optional cache flush flags (TC)
96 * \param data_sel 1 = fence, 3 = timestamp
97 * \param buf Buffer
98 * \param va GPU address
99 * \param old_value Previous fence value (for a bug workaround)
100 * \param new_value Fence value to write for this event.
101 */
r600_gfx_write_event_eop(struct r600_common_context * ctx,unsigned event,unsigned event_flags,unsigned data_sel,struct r600_resource * buf,uint64_t va,uint32_t new_fence,unsigned query_type)102 void r600_gfx_write_event_eop(struct r600_common_context *ctx,
103 unsigned event, unsigned event_flags,
104 unsigned data_sel,
105 struct r600_resource *buf, uint64_t va,
106 uint32_t new_fence, unsigned query_type)
107 {
108 struct radeon_winsys_cs *cs = ctx->gfx.cs;
109 unsigned op = EVENT_TYPE(event) |
110 EVENT_INDEX(5) |
111 event_flags;
112 unsigned sel = EOP_DATA_SEL(data_sel);
113
114 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
115 radeon_emit(cs, op);
116 radeon_emit(cs, va);
117 radeon_emit(cs, ((va >> 32) & 0xffff) | sel);
118 radeon_emit(cs, new_fence); /* immediate data */
119 radeon_emit(cs, 0); /* unused */
120
121 if (buf)
122 r600_emit_reloc(ctx, &ctx->gfx, buf, RADEON_USAGE_WRITE,
123 RADEON_PRIO_QUERY);
124 }
125
r600_gfx_write_fence_dwords(struct r600_common_screen * screen)126 unsigned r600_gfx_write_fence_dwords(struct r600_common_screen *screen)
127 {
128 unsigned dwords = 6;
129
130 if (!screen->info.has_virtual_memory)
131 dwords += 2;
132
133 return dwords;
134 }
135
r600_gfx_wait_fence(struct r600_common_context * ctx,uint64_t va,uint32_t ref,uint32_t mask)136 void r600_gfx_wait_fence(struct r600_common_context *ctx,
137 uint64_t va, uint32_t ref, uint32_t mask)
138 {
139 struct radeon_winsys_cs *cs = ctx->gfx.cs;
140
141 radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
142 radeon_emit(cs, WAIT_REG_MEM_EQUAL | WAIT_REG_MEM_MEM_SPACE(1));
143 radeon_emit(cs, va);
144 radeon_emit(cs, va >> 32);
145 radeon_emit(cs, ref); /* reference value */
146 radeon_emit(cs, mask); /* mask */
147 radeon_emit(cs, 4); /* poll interval */
148 }
149
r600_draw_rectangle(struct blitter_context * blitter,void * vertex_elements_cso,blitter_get_vs_func get_vs,int x1,int y1,int x2,int y2,float depth,unsigned num_instances,enum blitter_attrib_type type,const union blitter_attrib * attrib)150 void r600_draw_rectangle(struct blitter_context *blitter,
151 void *vertex_elements_cso,
152 blitter_get_vs_func get_vs,
153 int x1, int y1, int x2, int y2,
154 float depth, unsigned num_instances,
155 enum blitter_attrib_type type,
156 const union blitter_attrib *attrib)
157 {
158 struct r600_common_context *rctx =
159 (struct r600_common_context*)util_blitter_get_pipe(blitter);
160 struct pipe_viewport_state viewport;
161 struct pipe_resource *buf = NULL;
162 unsigned offset = 0;
163 float *vb;
164
165 rctx->b.bind_vertex_elements_state(&rctx->b, vertex_elements_cso);
166 rctx->b.bind_vs_state(&rctx->b, get_vs(blitter));
167
168 /* Some operations (like color resolve on r6xx) don't work
169 * with the conventional primitive types.
170 * One that works is PT_RECTLIST, which we use here. */
171
172 /* setup viewport */
173 viewport.scale[0] = 1.0f;
174 viewport.scale[1] = 1.0f;
175 viewport.scale[2] = 1.0f;
176 viewport.translate[0] = 0.0f;
177 viewport.translate[1] = 0.0f;
178 viewport.translate[2] = 0.0f;
179 rctx->b.set_viewport_states(&rctx->b, 0, 1, &viewport);
180
181 /* Upload vertices. The hw rectangle has only 3 vertices,
182 * The 4th one is derived from the first 3.
183 * The vertex specification should match u_blitter's vertex element state. */
184 u_upload_alloc(rctx->b.stream_uploader, 0, sizeof(float) * 24,
185 rctx->screen->info.tcc_cache_line_size,
186 &offset, &buf, (void**)&vb);
187 if (!buf)
188 return;
189
190 vb[0] = x1;
191 vb[1] = y1;
192 vb[2] = depth;
193 vb[3] = 1;
194
195 vb[8] = x1;
196 vb[9] = y2;
197 vb[10] = depth;
198 vb[11] = 1;
199
200 vb[16] = x2;
201 vb[17] = y1;
202 vb[18] = depth;
203 vb[19] = 1;
204
205 switch (type) {
206 case UTIL_BLITTER_ATTRIB_COLOR:
207 memcpy(vb+4, attrib->color, sizeof(float)*4);
208 memcpy(vb+12, attrib->color, sizeof(float)*4);
209 memcpy(vb+20, attrib->color, sizeof(float)*4);
210 break;
211 case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
212 case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
213 vb[6] = vb[14] = vb[22] = attrib->texcoord.z;
214 vb[7] = vb[15] = vb[23] = attrib->texcoord.w;
215 /* fall through */
216 vb[4] = attrib->texcoord.x1;
217 vb[5] = attrib->texcoord.y1;
218 vb[12] = attrib->texcoord.x1;
219 vb[13] = attrib->texcoord.y2;
220 vb[20] = attrib->texcoord.x2;
221 vb[21] = attrib->texcoord.y1;
222 break;
223 default:; /* Nothing to do. */
224 }
225
226 /* draw */
227 struct pipe_vertex_buffer vbuffer = {};
228 vbuffer.buffer.resource = buf;
229 vbuffer.stride = 2 * 4 * sizeof(float); /* vertex size */
230 vbuffer.buffer_offset = offset;
231
232 rctx->b.set_vertex_buffers(&rctx->b, blitter->vb_slot, 1, &vbuffer);
233 util_draw_arrays_instanced(&rctx->b, R600_PRIM_RECTANGLE_LIST, 0, 3,
234 0, num_instances);
235 pipe_resource_reference(&buf, NULL);
236 }
237
r600_dma_emit_wait_idle(struct r600_common_context * rctx)238 static void r600_dma_emit_wait_idle(struct r600_common_context *rctx)
239 {
240 struct radeon_winsys_cs *cs = rctx->dma.cs;
241
242 if (rctx->chip_class >= EVERGREEN)
243 radeon_emit(cs, 0xf0000000); /* NOP */
244 else {
245 /* TODO: R600-R700 should use the FENCE packet.
246 * CS checker support is required. */
247 }
248 }
249
r600_need_dma_space(struct r600_common_context * ctx,unsigned num_dw,struct r600_resource * dst,struct r600_resource * src)250 void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw,
251 struct r600_resource *dst, struct r600_resource *src)
252 {
253 uint64_t vram = ctx->dma.cs->used_vram;
254 uint64_t gtt = ctx->dma.cs->used_gart;
255
256 if (dst) {
257 vram += dst->vram_usage;
258 gtt += dst->gart_usage;
259 }
260 if (src) {
261 vram += src->vram_usage;
262 gtt += src->gart_usage;
263 }
264
265 /* Flush the GFX IB if DMA depends on it. */
266 if (radeon_emitted(ctx->gfx.cs, ctx->initial_gfx_cs_size) &&
267 ((dst &&
268 ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, dst->buf,
269 RADEON_USAGE_READWRITE)) ||
270 (src &&
271 ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, src->buf,
272 RADEON_USAGE_WRITE))))
273 ctx->gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
274
275 /* Flush if there's not enough space, or if the memory usage per IB
276 * is too large.
277 *
278 * IBs using too little memory are limited by the IB submission overhead.
279 * IBs using too much memory are limited by the kernel/TTM overhead.
280 * Too long IBs create CPU-GPU pipeline bubbles and add latency.
281 *
282 * This heuristic makes sure that DMA requests are executed
283 * very soon after the call is made and lowers memory usage.
284 * It improves texture upload performance by keeping the DMA
285 * engine busy while uploads are being submitted.
286 */
287 num_dw++; /* for emit_wait_idle below */
288 if (!ctx->ws->cs_check_space(ctx->dma.cs, num_dw) ||
289 ctx->dma.cs->used_vram + ctx->dma.cs->used_gart > 64 * 1024 * 1024 ||
290 !radeon_cs_memory_below_limit(ctx->screen, ctx->dma.cs, vram, gtt)) {
291 ctx->dma.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
292 assert((num_dw + ctx->dma.cs->current.cdw) <= ctx->dma.cs->current.max_dw);
293 }
294
295 /* Wait for idle if either buffer has been used in the IB before to
296 * prevent read-after-write hazards.
297 */
298 if ((dst &&
299 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs, dst->buf,
300 RADEON_USAGE_READWRITE)) ||
301 (src &&
302 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs, src->buf,
303 RADEON_USAGE_WRITE)))
304 r600_dma_emit_wait_idle(ctx);
305
306 /* If GPUVM is not supported, the CS checker needs 2 entries
307 * in the buffer list per packet, which has to be done manually.
308 */
309 if (ctx->screen->info.has_virtual_memory) {
310 if (dst)
311 radeon_add_to_buffer_list(ctx, &ctx->dma, dst,
312 RADEON_USAGE_WRITE,
313 RADEON_PRIO_SDMA_BUFFER);
314 if (src)
315 radeon_add_to_buffer_list(ctx, &ctx->dma, src,
316 RADEON_USAGE_READ,
317 RADEON_PRIO_SDMA_BUFFER);
318 }
319
320 /* this function is called before all DMA calls, so increment this. */
321 ctx->num_dma_calls++;
322 }
323
r600_preflush_suspend_features(struct r600_common_context * ctx)324 void r600_preflush_suspend_features(struct r600_common_context *ctx)
325 {
326 /* suspend queries */
327 if (!LIST_IS_EMPTY(&ctx->active_queries))
328 r600_suspend_queries(ctx);
329
330 ctx->streamout.suspended = false;
331 if (ctx->streamout.begin_emitted) {
332 r600_emit_streamout_end(ctx);
333 ctx->streamout.suspended = true;
334 }
335 }
336
r600_postflush_resume_features(struct r600_common_context * ctx)337 void r600_postflush_resume_features(struct r600_common_context *ctx)
338 {
339 if (ctx->streamout.suspended) {
340 ctx->streamout.append_bitmask = ctx->streamout.enabled_mask;
341 r600_streamout_buffers_dirty(ctx);
342 }
343
344 /* resume queries */
345 if (!LIST_IS_EMPTY(&ctx->active_queries))
346 r600_resume_queries(ctx);
347 }
348
r600_add_fence_dependency(struct r600_common_context * rctx,struct pipe_fence_handle * fence)349 static void r600_add_fence_dependency(struct r600_common_context *rctx,
350 struct pipe_fence_handle *fence)
351 {
352 struct radeon_winsys *ws = rctx->ws;
353
354 if (rctx->dma.cs)
355 ws->cs_add_fence_dependency(rctx->dma.cs, fence);
356 ws->cs_add_fence_dependency(rctx->gfx.cs, fence);
357 }
358
r600_fence_server_sync(struct pipe_context * ctx,struct pipe_fence_handle * fence)359 static void r600_fence_server_sync(struct pipe_context *ctx,
360 struct pipe_fence_handle *fence)
361 {
362 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
363 struct r600_multi_fence *rfence = (struct r600_multi_fence *)fence;
364
365 /* Only amdgpu needs to handle fence dependencies (for fence imports).
366 * radeon synchronizes all rings by default and will not implement
367 * fence imports.
368 */
369 if (rctx->screen->info.drm_major == 2)
370 return;
371
372 /* Only imported fences need to be handled by fence_server_sync,
373 * because the winsys handles synchronizations automatically for BOs
374 * within the process.
375 *
376 * Simply skip unflushed fences here, and the winsys will drop no-op
377 * dependencies (i.e. dependencies within the same ring).
378 */
379 if (rfence->gfx_unflushed.ctx)
380 return;
381
382 /* All unflushed commands will not start execution before
383 * this fence dependency is signalled.
384 *
385 * Should we flush the context to allow more GPU parallelism?
386 */
387 if (rfence->sdma)
388 r600_add_fence_dependency(rctx, rfence->sdma);
389 if (rfence->gfx)
390 r600_add_fence_dependency(rctx, rfence->gfx);
391 }
392
r600_flush_from_st(struct pipe_context * ctx,struct pipe_fence_handle ** fence,unsigned flags)393 static void r600_flush_from_st(struct pipe_context *ctx,
394 struct pipe_fence_handle **fence,
395 unsigned flags)
396 {
397 struct pipe_screen *screen = ctx->screen;
398 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
399 struct radeon_winsys *ws = rctx->ws;
400 struct pipe_fence_handle *gfx_fence = NULL;
401 struct pipe_fence_handle *sdma_fence = NULL;
402 bool deferred_fence = false;
403 unsigned rflags = PIPE_FLUSH_ASYNC;
404
405 if (flags & PIPE_FLUSH_END_OF_FRAME)
406 rflags |= PIPE_FLUSH_END_OF_FRAME;
407
408 /* DMA IBs are preambles to gfx IBs, therefore must be flushed first. */
409 if (rctx->dma.cs)
410 rctx->dma.flush(rctx, rflags, fence ? &sdma_fence : NULL);
411
412 if (!radeon_emitted(rctx->gfx.cs, rctx->initial_gfx_cs_size)) {
413 if (fence)
414 ws->fence_reference(&gfx_fence, rctx->last_gfx_fence);
415 if (!(flags & PIPE_FLUSH_DEFERRED))
416 ws->cs_sync_flush(rctx->gfx.cs);
417 } else {
418 /* Instead of flushing, create a deferred fence. Constraints:
419 * - The state tracker must allow a deferred flush.
420 * - The state tracker must request a fence.
421 * Thread safety in fence_finish must be ensured by the state tracker.
422 */
423 if (flags & PIPE_FLUSH_DEFERRED && fence) {
424 gfx_fence = rctx->ws->cs_get_next_fence(rctx->gfx.cs);
425 deferred_fence = true;
426 } else {
427 rctx->gfx.flush(rctx, rflags, fence ? &gfx_fence : NULL);
428 }
429 }
430
431 /* Both engines can signal out of order, so we need to keep both fences. */
432 if (fence) {
433 struct r600_multi_fence *multi_fence =
434 CALLOC_STRUCT(r600_multi_fence);
435 if (!multi_fence) {
436 ws->fence_reference(&sdma_fence, NULL);
437 ws->fence_reference(&gfx_fence, NULL);
438 goto finish;
439 }
440
441 multi_fence->reference.count = 1;
442 /* If both fences are NULL, fence_finish will always return true. */
443 multi_fence->gfx = gfx_fence;
444 multi_fence->sdma = sdma_fence;
445
446 if (deferred_fence) {
447 multi_fence->gfx_unflushed.ctx = rctx;
448 multi_fence->gfx_unflushed.ib_index = rctx->num_gfx_cs_flushes;
449 }
450
451 screen->fence_reference(screen, fence, NULL);
452 *fence = (struct pipe_fence_handle*)multi_fence;
453 }
454 finish:
455 if (!(flags & PIPE_FLUSH_DEFERRED)) {
456 if (rctx->dma.cs)
457 ws->cs_sync_flush(rctx->dma.cs);
458 ws->cs_sync_flush(rctx->gfx.cs);
459 }
460 }
461
r600_flush_dma_ring(void * ctx,unsigned flags,struct pipe_fence_handle ** fence)462 static void r600_flush_dma_ring(void *ctx, unsigned flags,
463 struct pipe_fence_handle **fence)
464 {
465 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
466 struct radeon_winsys_cs *cs = rctx->dma.cs;
467 struct radeon_saved_cs saved;
468 bool check_vm =
469 (rctx->screen->debug_flags & DBG_CHECK_VM) &&
470 rctx->check_vm_faults;
471
472 if (!radeon_emitted(cs, 0)) {
473 if (fence)
474 rctx->ws->fence_reference(fence, rctx->last_sdma_fence);
475 return;
476 }
477
478 if (check_vm)
479 radeon_save_cs(rctx->ws, cs, &saved, true);
480
481 rctx->ws->cs_flush(cs, flags, &rctx->last_sdma_fence);
482 if (fence)
483 rctx->ws->fence_reference(fence, rctx->last_sdma_fence);
484
485 if (check_vm) {
486 /* Use conservative timeout 800ms, after which we won't wait any
487 * longer and assume the GPU is hung.
488 */
489 rctx->ws->fence_wait(rctx->ws, rctx->last_sdma_fence, 800*1000*1000);
490
491 rctx->check_vm_faults(rctx, &saved, RING_DMA);
492 radeon_clear_saved_cs(&saved);
493 }
494 }
495
496 /**
497 * Store a linearized copy of all chunks of \p cs together with the buffer
498 * list in \p saved.
499 */
radeon_save_cs(struct radeon_winsys * ws,struct radeon_winsys_cs * cs,struct radeon_saved_cs * saved,bool get_buffer_list)500 void radeon_save_cs(struct radeon_winsys *ws, struct radeon_winsys_cs *cs,
501 struct radeon_saved_cs *saved, bool get_buffer_list)
502 {
503 uint32_t *buf;
504 unsigned i;
505
506 /* Save the IB chunks. */
507 saved->num_dw = cs->prev_dw + cs->current.cdw;
508 saved->ib = MALLOC(4 * saved->num_dw);
509 if (!saved->ib)
510 goto oom;
511
512 buf = saved->ib;
513 for (i = 0; i < cs->num_prev; ++i) {
514 memcpy(buf, cs->prev[i].buf, cs->prev[i].cdw * 4);
515 buf += cs->prev[i].cdw;
516 }
517 memcpy(buf, cs->current.buf, cs->current.cdw * 4);
518
519 if (!get_buffer_list)
520 return;
521
522 /* Save the buffer list. */
523 saved->bo_count = ws->cs_get_buffer_list(cs, NULL);
524 saved->bo_list = CALLOC(saved->bo_count,
525 sizeof(saved->bo_list[0]));
526 if (!saved->bo_list) {
527 FREE(saved->ib);
528 goto oom;
529 }
530 ws->cs_get_buffer_list(cs, saved->bo_list);
531
532 return;
533
534 oom:
535 fprintf(stderr, "%s: out of memory\n", __func__);
536 memset(saved, 0, sizeof(*saved));
537 }
538
radeon_clear_saved_cs(struct radeon_saved_cs * saved)539 void radeon_clear_saved_cs(struct radeon_saved_cs *saved)
540 {
541 FREE(saved->ib);
542 FREE(saved->bo_list);
543
544 memset(saved, 0, sizeof(*saved));
545 }
546
r600_get_reset_status(struct pipe_context * ctx)547 static enum pipe_reset_status r600_get_reset_status(struct pipe_context *ctx)
548 {
549 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
550 unsigned latest = rctx->ws->query_value(rctx->ws,
551 RADEON_GPU_RESET_COUNTER);
552
553 if (rctx->gpu_reset_counter == latest)
554 return PIPE_NO_RESET;
555
556 rctx->gpu_reset_counter = latest;
557 return PIPE_UNKNOWN_CONTEXT_RESET;
558 }
559
r600_set_debug_callback(struct pipe_context * ctx,const struct pipe_debug_callback * cb)560 static void r600_set_debug_callback(struct pipe_context *ctx,
561 const struct pipe_debug_callback *cb)
562 {
563 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
564
565 if (cb)
566 rctx->debug = *cb;
567 else
568 memset(&rctx->debug, 0, sizeof(rctx->debug));
569 }
570
r600_set_device_reset_callback(struct pipe_context * ctx,const struct pipe_device_reset_callback * cb)571 static void r600_set_device_reset_callback(struct pipe_context *ctx,
572 const struct pipe_device_reset_callback *cb)
573 {
574 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
575
576 if (cb)
577 rctx->device_reset_callback = *cb;
578 else
579 memset(&rctx->device_reset_callback, 0,
580 sizeof(rctx->device_reset_callback));
581 }
582
r600_check_device_reset(struct r600_common_context * rctx)583 bool r600_check_device_reset(struct r600_common_context *rctx)
584 {
585 enum pipe_reset_status status;
586
587 if (!rctx->device_reset_callback.reset)
588 return false;
589
590 if (!rctx->b.get_device_reset_status)
591 return false;
592
593 status = rctx->b.get_device_reset_status(&rctx->b);
594 if (status == PIPE_NO_RESET)
595 return false;
596
597 rctx->device_reset_callback.reset(rctx->device_reset_callback.data, status);
598 return true;
599 }
600
r600_dma_clear_buffer_fallback(struct pipe_context * ctx,struct pipe_resource * dst,uint64_t offset,uint64_t size,unsigned value)601 static void r600_dma_clear_buffer_fallback(struct pipe_context *ctx,
602 struct pipe_resource *dst,
603 uint64_t offset, uint64_t size,
604 unsigned value)
605 {
606 struct r600_common_context *rctx = (struct r600_common_context *)ctx;
607
608 rctx->clear_buffer(ctx, dst, offset, size, value, R600_COHERENCY_NONE);
609 }
610
r600_resource_commit(struct pipe_context * pctx,struct pipe_resource * resource,unsigned level,struct pipe_box * box,bool commit)611 static bool r600_resource_commit(struct pipe_context *pctx,
612 struct pipe_resource *resource,
613 unsigned level, struct pipe_box *box,
614 bool commit)
615 {
616 struct r600_common_context *ctx = (struct r600_common_context *)pctx;
617 struct r600_resource *res = r600_resource(resource);
618
619 /*
620 * Since buffer commitment changes cannot be pipelined, we need to
621 * (a) flush any pending commands that refer to the buffer we're about
622 * to change, and
623 * (b) wait for threaded submit to finish, including those that were
624 * triggered by some other, earlier operation.
625 */
626 if (radeon_emitted(ctx->gfx.cs, ctx->initial_gfx_cs_size) &&
627 ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs,
628 res->buf, RADEON_USAGE_READWRITE)) {
629 ctx->gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
630 }
631 if (radeon_emitted(ctx->dma.cs, 0) &&
632 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs,
633 res->buf, RADEON_USAGE_READWRITE)) {
634 ctx->dma.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
635 }
636
637 ctx->ws->cs_sync_flush(ctx->dma.cs);
638 ctx->ws->cs_sync_flush(ctx->gfx.cs);
639
640 assert(resource->target == PIPE_BUFFER);
641
642 return ctx->ws->buffer_commit(res->buf, box->x, box->width, commit);
643 }
644
r600_common_context_init(struct r600_common_context * rctx,struct r600_common_screen * rscreen,unsigned context_flags)645 bool r600_common_context_init(struct r600_common_context *rctx,
646 struct r600_common_screen *rscreen,
647 unsigned context_flags)
648 {
649 slab_create_child(&rctx->pool_transfers, &rscreen->pool_transfers);
650 slab_create_child(&rctx->pool_transfers_unsync, &rscreen->pool_transfers);
651
652 rctx->screen = rscreen;
653 rctx->ws = rscreen->ws;
654 rctx->family = rscreen->family;
655 rctx->chip_class = rscreen->chip_class;
656
657 rctx->b.invalidate_resource = r600_invalidate_resource;
658 rctx->b.resource_commit = r600_resource_commit;
659 rctx->b.transfer_map = u_transfer_map_vtbl;
660 rctx->b.transfer_flush_region = u_transfer_flush_region_vtbl;
661 rctx->b.transfer_unmap = u_transfer_unmap_vtbl;
662 rctx->b.texture_subdata = u_default_texture_subdata;
663 rctx->b.flush = r600_flush_from_st;
664 rctx->b.set_debug_callback = r600_set_debug_callback;
665 rctx->b.fence_server_sync = r600_fence_server_sync;
666 rctx->dma_clear_buffer = r600_dma_clear_buffer_fallback;
667
668 /* evergreen_compute.c has a special codepath for global buffers.
669 * Everything else can use the direct path.
670 */
671 if ((rscreen->chip_class == EVERGREEN || rscreen->chip_class == CAYMAN) &&
672 (context_flags & PIPE_CONTEXT_COMPUTE_ONLY))
673 rctx->b.buffer_subdata = u_default_buffer_subdata;
674 else
675 rctx->b.buffer_subdata = r600_buffer_subdata;
676
677 if (rscreen->info.drm_major == 2 && rscreen->info.drm_minor >= 43) {
678 rctx->b.get_device_reset_status = r600_get_reset_status;
679 rctx->gpu_reset_counter =
680 rctx->ws->query_value(rctx->ws,
681 RADEON_GPU_RESET_COUNTER);
682 }
683
684 rctx->b.set_device_reset_callback = r600_set_device_reset_callback;
685
686 r600_init_context_texture_functions(rctx);
687 r600_init_viewport_functions(rctx);
688 r600_streamout_init(rctx);
689 r600_query_init(rctx);
690 cayman_init_msaa(&rctx->b);
691
692 rctx->allocator_zeroed_memory =
693 u_suballocator_create(&rctx->b, rscreen->info.gart_page_size,
694 0, PIPE_USAGE_DEFAULT, 0, true);
695 if (!rctx->allocator_zeroed_memory)
696 return false;
697
698 rctx->b.stream_uploader = u_upload_create(&rctx->b, 1024 * 1024,
699 0, PIPE_USAGE_STREAM, 0);
700 if (!rctx->b.stream_uploader)
701 return false;
702
703 rctx->b.const_uploader = u_upload_create(&rctx->b, 128 * 1024,
704 0, PIPE_USAGE_DEFAULT, 0);
705 if (!rctx->b.const_uploader)
706 return false;
707
708 rctx->ctx = rctx->ws->ctx_create(rctx->ws);
709 if (!rctx->ctx)
710 return false;
711
712 if (rscreen->info.num_sdma_rings && !(rscreen->debug_flags & DBG_NO_ASYNC_DMA)) {
713 rctx->dma.cs = rctx->ws->cs_create(rctx->ctx, RING_DMA,
714 r600_flush_dma_ring,
715 rctx);
716 rctx->dma.flush = r600_flush_dma_ring;
717 }
718
719 return true;
720 }
721
r600_common_context_cleanup(struct r600_common_context * rctx)722 void r600_common_context_cleanup(struct r600_common_context *rctx)
723 {
724 if (rctx->query_result_shader)
725 rctx->b.delete_compute_state(&rctx->b, rctx->query_result_shader);
726
727 if (rctx->gfx.cs)
728 rctx->ws->cs_destroy(rctx->gfx.cs);
729 if (rctx->dma.cs)
730 rctx->ws->cs_destroy(rctx->dma.cs);
731 if (rctx->ctx)
732 rctx->ws->ctx_destroy(rctx->ctx);
733
734 if (rctx->b.stream_uploader)
735 u_upload_destroy(rctx->b.stream_uploader);
736 if (rctx->b.const_uploader)
737 u_upload_destroy(rctx->b.const_uploader);
738
739 slab_destroy_child(&rctx->pool_transfers);
740 slab_destroy_child(&rctx->pool_transfers_unsync);
741
742 if (rctx->allocator_zeroed_memory) {
743 u_suballocator_destroy(rctx->allocator_zeroed_memory);
744 }
745 rctx->ws->fence_reference(&rctx->last_gfx_fence, NULL);
746 rctx->ws->fence_reference(&rctx->last_sdma_fence, NULL);
747 r600_resource_reference(&rctx->eop_bug_scratch, NULL);
748 }
749
750 /*
751 * pipe_screen
752 */
753
754 static const struct debug_named_value common_debug_options[] = {
755 /* logging */
756 { "tex", DBG_TEX, "Print texture info" },
757 { "nir", DBG_NIR, "Enable experimental NIR shaders" },
758 { "compute", DBG_COMPUTE, "Print compute info" },
759 { "vm", DBG_VM, "Print virtual addresses when creating resources" },
760 { "info", DBG_INFO, "Print driver information" },
761
762 /* shaders */
763 { "fs", DBG_FS, "Print fetch shaders" },
764 { "vs", DBG_VS, "Print vertex shaders" },
765 { "gs", DBG_GS, "Print geometry shaders" },
766 { "ps", DBG_PS, "Print pixel shaders" },
767 { "cs", DBG_CS, "Print compute shaders" },
768 { "tcs", DBG_TCS, "Print tessellation control shaders" },
769 { "tes", DBG_TES, "Print tessellation evaluation shaders" },
770 { "noir", DBG_NO_IR, "Don't print the LLVM IR"},
771 { "notgsi", DBG_NO_TGSI, "Don't print the TGSI"},
772 { "noasm", DBG_NO_ASM, "Don't print disassembled shaders"},
773 { "preoptir", DBG_PREOPT_IR, "Print the LLVM IR before initial optimizations" },
774 { "checkir", DBG_CHECK_IR, "Enable additional sanity checks on shader IR" },
775 { "nooptvariant", DBG_NO_OPT_VARIANT, "Disable compiling optimized shader variants." },
776
777 { "testdma", DBG_TEST_DMA, "Invoke SDMA tests and exit." },
778 { "testvmfaultcp", DBG_TEST_VMFAULT_CP, "Invoke a CP VM fault test and exit." },
779 { "testvmfaultsdma", DBG_TEST_VMFAULT_SDMA, "Invoke a SDMA VM fault test and exit." },
780 { "testvmfaultshader", DBG_TEST_VMFAULT_SHADER, "Invoke a shader VM fault test and exit." },
781
782 /* features */
783 { "nodma", DBG_NO_ASYNC_DMA, "Disable asynchronous DMA" },
784 { "nohyperz", DBG_NO_HYPERZ, "Disable Hyper-Z" },
785 /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
786 { "noinvalrange", DBG_NO_DISCARD_RANGE, "Disable handling of INVALIDATE_RANGE map flags" },
787 { "no2d", DBG_NO_2D_TILING, "Disable 2D tiling" },
788 { "notiling", DBG_NO_TILING, "Disable tiling" },
789 { "switch_on_eop", DBG_SWITCH_ON_EOP, "Program WD/IA to switch on end-of-packet." },
790 { "forcedma", DBG_FORCE_DMA, "Use asynchronous DMA for all operations when possible." },
791 { "precompile", DBG_PRECOMPILE, "Compile one shader variant at shader creation." },
792 { "nowc", DBG_NO_WC, "Disable GTT write combining" },
793 { "check_vm", DBG_CHECK_VM, "Check VM faults and dump debug info." },
794 { "unsafemath", DBG_UNSAFE_MATH, "Enable unsafe math shader optimizations" },
795
796 DEBUG_NAMED_VALUE_END /* must be last */
797 };
798
r600_get_vendor(struct pipe_screen * pscreen)799 static const char* r600_get_vendor(struct pipe_screen* pscreen)
800 {
801 return "X.Org";
802 }
803
r600_get_device_vendor(struct pipe_screen * pscreen)804 static const char* r600_get_device_vendor(struct pipe_screen* pscreen)
805 {
806 return "AMD";
807 }
808
r600_get_marketing_name(struct radeon_winsys * ws)809 static const char *r600_get_marketing_name(struct radeon_winsys *ws)
810 {
811 if (!ws->get_chip_name)
812 return NULL;
813 return ws->get_chip_name(ws);
814 }
815
r600_get_family_name(const struct r600_common_screen * rscreen)816 static const char *r600_get_family_name(const struct r600_common_screen *rscreen)
817 {
818 switch (rscreen->info.family) {
819 case CHIP_R600: return "AMD R600";
820 case CHIP_RV610: return "AMD RV610";
821 case CHIP_RV630: return "AMD RV630";
822 case CHIP_RV670: return "AMD RV670";
823 case CHIP_RV620: return "AMD RV620";
824 case CHIP_RV635: return "AMD RV635";
825 case CHIP_RS780: return "AMD RS780";
826 case CHIP_RS880: return "AMD RS880";
827 case CHIP_RV770: return "AMD RV770";
828 case CHIP_RV730: return "AMD RV730";
829 case CHIP_RV710: return "AMD RV710";
830 case CHIP_RV740: return "AMD RV740";
831 case CHIP_CEDAR: return "AMD CEDAR";
832 case CHIP_REDWOOD: return "AMD REDWOOD";
833 case CHIP_JUNIPER: return "AMD JUNIPER";
834 case CHIP_CYPRESS: return "AMD CYPRESS";
835 case CHIP_HEMLOCK: return "AMD HEMLOCK";
836 case CHIP_PALM: return "AMD PALM";
837 case CHIP_SUMO: return "AMD SUMO";
838 case CHIP_SUMO2: return "AMD SUMO2";
839 case CHIP_BARTS: return "AMD BARTS";
840 case CHIP_TURKS: return "AMD TURKS";
841 case CHIP_CAICOS: return "AMD CAICOS";
842 case CHIP_CAYMAN: return "AMD CAYMAN";
843 case CHIP_ARUBA: return "AMD ARUBA";
844 default: return "AMD unknown";
845 }
846 }
847
r600_disk_cache_create(struct r600_common_screen * rscreen)848 static void r600_disk_cache_create(struct r600_common_screen *rscreen)
849 {
850 /* Don't use the cache if shader dumping is enabled. */
851 if (rscreen->debug_flags & DBG_ALL_SHADERS)
852 return;
853
854 uint32_t mesa_timestamp;
855 if (disk_cache_get_function_timestamp(r600_disk_cache_create,
856 &mesa_timestamp)) {
857 char *timestamp_str;
858 int res = -1;
859
860 res = asprintf(×tamp_str, "%u",mesa_timestamp);
861 if (res != -1) {
862 /* These flags affect shader compilation. */
863 uint64_t shader_debug_flags =
864 rscreen->debug_flags &
865 (DBG_FS_CORRECT_DERIVS_AFTER_KILL |
866 DBG_UNSAFE_MATH);
867
868 rscreen->disk_shader_cache =
869 disk_cache_create(r600_get_family_name(rscreen),
870 timestamp_str,
871 shader_debug_flags);
872 free(timestamp_str);
873 }
874 }
875 }
876
r600_get_disk_shader_cache(struct pipe_screen * pscreen)877 static struct disk_cache *r600_get_disk_shader_cache(struct pipe_screen *pscreen)
878 {
879 struct r600_common_screen *rscreen = (struct r600_common_screen*)pscreen;
880 return rscreen->disk_shader_cache;
881 }
882
r600_get_name(struct pipe_screen * pscreen)883 static const char* r600_get_name(struct pipe_screen* pscreen)
884 {
885 struct r600_common_screen *rscreen = (struct r600_common_screen*)pscreen;
886
887 return rscreen->renderer_string;
888 }
889
r600_get_paramf(struct pipe_screen * pscreen,enum pipe_capf param)890 static float r600_get_paramf(struct pipe_screen* pscreen,
891 enum pipe_capf param)
892 {
893 struct r600_common_screen *rscreen = (struct r600_common_screen *)pscreen;
894
895 switch (param) {
896 case PIPE_CAPF_MAX_LINE_WIDTH:
897 case PIPE_CAPF_MAX_LINE_WIDTH_AA:
898 case PIPE_CAPF_MAX_POINT_WIDTH:
899 case PIPE_CAPF_MAX_POINT_WIDTH_AA:
900 if (rscreen->family >= CHIP_CEDAR)
901 return 16384.0f;
902 else
903 return 8192.0f;
904 case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
905 return 16.0f;
906 case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
907 return 16.0f;
908 case PIPE_CAPF_GUARD_BAND_LEFT:
909 case PIPE_CAPF_GUARD_BAND_TOP:
910 case PIPE_CAPF_GUARD_BAND_RIGHT:
911 case PIPE_CAPF_GUARD_BAND_BOTTOM:
912 return 0.0f;
913 }
914 return 0.0f;
915 }
916
r600_get_video_param(struct pipe_screen * screen,enum pipe_video_profile profile,enum pipe_video_entrypoint entrypoint,enum pipe_video_cap param)917 static int r600_get_video_param(struct pipe_screen *screen,
918 enum pipe_video_profile profile,
919 enum pipe_video_entrypoint entrypoint,
920 enum pipe_video_cap param)
921 {
922 switch (param) {
923 case PIPE_VIDEO_CAP_SUPPORTED:
924 return vl_profile_supported(screen, profile, entrypoint);
925 case PIPE_VIDEO_CAP_NPOT_TEXTURES:
926 return 1;
927 case PIPE_VIDEO_CAP_MAX_WIDTH:
928 case PIPE_VIDEO_CAP_MAX_HEIGHT:
929 return vl_video_buffer_max_size(screen);
930 case PIPE_VIDEO_CAP_PREFERED_FORMAT:
931 return PIPE_FORMAT_NV12;
932 case PIPE_VIDEO_CAP_PREFERS_INTERLACED:
933 return false;
934 case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED:
935 return false;
936 case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE:
937 return true;
938 case PIPE_VIDEO_CAP_MAX_LEVEL:
939 return vl_level_supported(screen, profile);
940 default:
941 return 0;
942 }
943 }
944
r600_get_llvm_processor_name(enum radeon_family family)945 const char *r600_get_llvm_processor_name(enum radeon_family family)
946 {
947 switch (family) {
948 case CHIP_R600:
949 case CHIP_RV630:
950 case CHIP_RV635:
951 case CHIP_RV670:
952 return "r600";
953 case CHIP_RV610:
954 case CHIP_RV620:
955 case CHIP_RS780:
956 case CHIP_RS880:
957 return "rs880";
958 case CHIP_RV710:
959 return "rv710";
960 case CHIP_RV730:
961 return "rv730";
962 case CHIP_RV740:
963 case CHIP_RV770:
964 return "rv770";
965 case CHIP_PALM:
966 case CHIP_CEDAR:
967 return "cedar";
968 case CHIP_SUMO:
969 case CHIP_SUMO2:
970 return "sumo";
971 case CHIP_REDWOOD:
972 return "redwood";
973 case CHIP_JUNIPER:
974 return "juniper";
975 case CHIP_HEMLOCK:
976 case CHIP_CYPRESS:
977 return "cypress";
978 case CHIP_BARTS:
979 return "barts";
980 case CHIP_TURKS:
981 return "turks";
982 case CHIP_CAICOS:
983 return "caicos";
984 case CHIP_CAYMAN:
985 case CHIP_ARUBA:
986 return "cayman";
987
988 default:
989 return "";
990 }
991 }
992
get_max_threads_per_block(struct r600_common_screen * screen,enum pipe_shader_ir ir_type)993 static unsigned get_max_threads_per_block(struct r600_common_screen *screen,
994 enum pipe_shader_ir ir_type)
995 {
996 if (ir_type != PIPE_SHADER_IR_TGSI)
997 return 256;
998 if (screen->chip_class >= EVERGREEN)
999 return 2048;
1000 return 256;
1001 }
1002
r600_get_compute_param(struct pipe_screen * screen,enum pipe_shader_ir ir_type,enum pipe_compute_cap param,void * ret)1003 static int r600_get_compute_param(struct pipe_screen *screen,
1004 enum pipe_shader_ir ir_type,
1005 enum pipe_compute_cap param,
1006 void *ret)
1007 {
1008 struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
1009
1010 //TODO: select these params by asic
1011 switch (param) {
1012 case PIPE_COMPUTE_CAP_IR_TARGET: {
1013 const char *gpu;
1014 const char *triple = "r600--";
1015 gpu = r600_get_llvm_processor_name(rscreen->family);
1016 if (ret) {
1017 sprintf(ret, "%s-%s", gpu, triple);
1018 }
1019 /* +2 for dash and terminating NIL byte */
1020 return (strlen(triple) + strlen(gpu) + 2) * sizeof(char);
1021 }
1022 case PIPE_COMPUTE_CAP_GRID_DIMENSION:
1023 if (ret) {
1024 uint64_t *grid_dimension = ret;
1025 grid_dimension[0] = 3;
1026 }
1027 return 1 * sizeof(uint64_t);
1028
1029 case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
1030 if (ret) {
1031 uint64_t *grid_size = ret;
1032 grid_size[0] = 65535;
1033 grid_size[1] = 65535;
1034 grid_size[2] = 65535;
1035 }
1036 return 3 * sizeof(uint64_t) ;
1037
1038 case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
1039 if (ret) {
1040 uint64_t *block_size = ret;
1041 unsigned threads_per_block = get_max_threads_per_block(rscreen, ir_type);
1042 block_size[0] = threads_per_block;
1043 block_size[1] = threads_per_block;
1044 block_size[2] = threads_per_block;
1045 }
1046 return 3 * sizeof(uint64_t);
1047
1048 case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
1049 if (ret) {
1050 uint64_t *max_threads_per_block = ret;
1051 *max_threads_per_block = get_max_threads_per_block(rscreen, ir_type);
1052 }
1053 return sizeof(uint64_t);
1054 case PIPE_COMPUTE_CAP_ADDRESS_BITS:
1055 if (ret) {
1056 uint32_t *address_bits = ret;
1057 address_bits[0] = 32;
1058 }
1059 return 1 * sizeof(uint32_t);
1060
1061 case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
1062 if (ret) {
1063 uint64_t *max_global_size = ret;
1064 uint64_t max_mem_alloc_size;
1065
1066 r600_get_compute_param(screen, ir_type,
1067 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
1068 &max_mem_alloc_size);
1069
1070 /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
1071 * 1/4 of the MAX_GLOBAL_SIZE. Since the
1072 * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
1073 * make sure we never report more than
1074 * 4 * MAX_MEM_ALLOC_SIZE.
1075 */
1076 *max_global_size = MIN2(4 * max_mem_alloc_size,
1077 MAX2(rscreen->info.gart_size,
1078 rscreen->info.vram_size));
1079 }
1080 return sizeof(uint64_t);
1081
1082 case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
1083 if (ret) {
1084 uint64_t *max_local_size = ret;
1085 /* Value reported by the closed source driver. */
1086 *max_local_size = 32768;
1087 }
1088 return sizeof(uint64_t);
1089
1090 case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
1091 if (ret) {
1092 uint64_t *max_input_size = ret;
1093 /* Value reported by the closed source driver. */
1094 *max_input_size = 1024;
1095 }
1096 return sizeof(uint64_t);
1097
1098 case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
1099 if (ret) {
1100 uint64_t *max_mem_alloc_size = ret;
1101
1102 *max_mem_alloc_size = rscreen->info.max_alloc_size;
1103 }
1104 return sizeof(uint64_t);
1105
1106 case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
1107 if (ret) {
1108 uint32_t *max_clock_frequency = ret;
1109 *max_clock_frequency = rscreen->info.max_shader_clock;
1110 }
1111 return sizeof(uint32_t);
1112
1113 case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
1114 if (ret) {
1115 uint32_t *max_compute_units = ret;
1116 *max_compute_units = rscreen->info.num_good_compute_units;
1117 }
1118 return sizeof(uint32_t);
1119
1120 case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
1121 if (ret) {
1122 uint32_t *images_supported = ret;
1123 *images_supported = 0;
1124 }
1125 return sizeof(uint32_t);
1126 case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
1127 break; /* unused */
1128 case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
1129 if (ret) {
1130 uint32_t *subgroup_size = ret;
1131 *subgroup_size = r600_wavefront_size(rscreen->family);
1132 }
1133 return sizeof(uint32_t);
1134 case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
1135 if (ret) {
1136 uint64_t *max_variable_threads_per_block = ret;
1137 *max_variable_threads_per_block = 0;
1138 }
1139 return sizeof(uint64_t);
1140 }
1141
1142 fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
1143 return 0;
1144 }
1145
r600_get_timestamp(struct pipe_screen * screen)1146 static uint64_t r600_get_timestamp(struct pipe_screen *screen)
1147 {
1148 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
1149
1150 return 1000000 * rscreen->ws->query_value(rscreen->ws, RADEON_TIMESTAMP) /
1151 rscreen->info.clock_crystal_freq;
1152 }
1153
r600_fence_reference(struct pipe_screen * screen,struct pipe_fence_handle ** dst,struct pipe_fence_handle * src)1154 static void r600_fence_reference(struct pipe_screen *screen,
1155 struct pipe_fence_handle **dst,
1156 struct pipe_fence_handle *src)
1157 {
1158 struct radeon_winsys *ws = ((struct r600_common_screen*)screen)->ws;
1159 struct r600_multi_fence **rdst = (struct r600_multi_fence **)dst;
1160 struct r600_multi_fence *rsrc = (struct r600_multi_fence *)src;
1161
1162 if (pipe_reference(&(*rdst)->reference, &rsrc->reference)) {
1163 ws->fence_reference(&(*rdst)->gfx, NULL);
1164 ws->fence_reference(&(*rdst)->sdma, NULL);
1165 FREE(*rdst);
1166 }
1167 *rdst = rsrc;
1168 }
1169
r600_fence_finish(struct pipe_screen * screen,struct pipe_context * ctx,struct pipe_fence_handle * fence,uint64_t timeout)1170 static boolean r600_fence_finish(struct pipe_screen *screen,
1171 struct pipe_context *ctx,
1172 struct pipe_fence_handle *fence,
1173 uint64_t timeout)
1174 {
1175 struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
1176 struct r600_multi_fence *rfence = (struct r600_multi_fence *)fence;
1177 struct r600_common_context *rctx;
1178 int64_t abs_timeout = os_time_get_absolute_timeout(timeout);
1179
1180 ctx = threaded_context_unwrap_sync(ctx);
1181 rctx = ctx ? (struct r600_common_context*)ctx : NULL;
1182
1183 if (rfence->sdma) {
1184 if (!rws->fence_wait(rws, rfence->sdma, timeout))
1185 return false;
1186
1187 /* Recompute the timeout after waiting. */
1188 if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
1189 int64_t time = os_time_get_nano();
1190 timeout = abs_timeout > time ? abs_timeout - time : 0;
1191 }
1192 }
1193
1194 if (!rfence->gfx)
1195 return true;
1196
1197 /* Flush the gfx IB if it hasn't been flushed yet. */
1198 if (rctx &&
1199 rfence->gfx_unflushed.ctx == rctx &&
1200 rfence->gfx_unflushed.ib_index == rctx->num_gfx_cs_flushes) {
1201 rctx->gfx.flush(rctx, timeout ? 0 : PIPE_FLUSH_ASYNC, NULL);
1202 rfence->gfx_unflushed.ctx = NULL;
1203
1204 if (!timeout)
1205 return false;
1206
1207 /* Recompute the timeout after all that. */
1208 if (timeout && timeout != PIPE_TIMEOUT_INFINITE) {
1209 int64_t time = os_time_get_nano();
1210 timeout = abs_timeout > time ? abs_timeout - time : 0;
1211 }
1212 }
1213
1214 return rws->fence_wait(rws, rfence->gfx, timeout);
1215 }
1216
r600_query_memory_info(struct pipe_screen * screen,struct pipe_memory_info * info)1217 static void r600_query_memory_info(struct pipe_screen *screen,
1218 struct pipe_memory_info *info)
1219 {
1220 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
1221 struct radeon_winsys *ws = rscreen->ws;
1222 unsigned vram_usage, gtt_usage;
1223
1224 info->total_device_memory = rscreen->info.vram_size / 1024;
1225 info->total_staging_memory = rscreen->info.gart_size / 1024;
1226
1227 /* The real TTM memory usage is somewhat random, because:
1228 *
1229 * 1) TTM delays freeing memory, because it can only free it after
1230 * fences expire.
1231 *
1232 * 2) The memory usage can be really low if big VRAM evictions are
1233 * taking place, but the real usage is well above the size of VRAM.
1234 *
1235 * Instead, return statistics of this process.
1236 */
1237 vram_usage = ws->query_value(ws, RADEON_REQUESTED_VRAM_MEMORY) / 1024;
1238 gtt_usage = ws->query_value(ws, RADEON_REQUESTED_GTT_MEMORY) / 1024;
1239
1240 info->avail_device_memory =
1241 vram_usage <= info->total_device_memory ?
1242 info->total_device_memory - vram_usage : 0;
1243 info->avail_staging_memory =
1244 gtt_usage <= info->total_staging_memory ?
1245 info->total_staging_memory - gtt_usage : 0;
1246
1247 info->device_memory_evicted =
1248 ws->query_value(ws, RADEON_NUM_BYTES_MOVED) / 1024;
1249
1250 if (rscreen->info.drm_major == 3 && rscreen->info.drm_minor >= 4)
1251 info->nr_device_memory_evictions =
1252 ws->query_value(ws, RADEON_NUM_EVICTIONS);
1253 else
1254 /* Just return the number of evicted 64KB pages. */
1255 info->nr_device_memory_evictions = info->device_memory_evicted / 64;
1256 }
1257
r600_resource_create_common(struct pipe_screen * screen,const struct pipe_resource * templ)1258 struct pipe_resource *r600_resource_create_common(struct pipe_screen *screen,
1259 const struct pipe_resource *templ)
1260 {
1261 if (templ->target == PIPE_BUFFER) {
1262 return r600_buffer_create(screen, templ, 256);
1263 } else {
1264 return r600_texture_create(screen, templ);
1265 }
1266 }
1267
r600_common_screen_init(struct r600_common_screen * rscreen,struct radeon_winsys * ws)1268 bool r600_common_screen_init(struct r600_common_screen *rscreen,
1269 struct radeon_winsys *ws)
1270 {
1271 char family_name[32] = {}, llvm_string[32] = {}, kernel_version[128] = {};
1272 struct utsname uname_data;
1273 const char *chip_name;
1274
1275 ws->query_info(ws, &rscreen->info);
1276 rscreen->ws = ws;
1277
1278 if ((chip_name = r600_get_marketing_name(ws)))
1279 snprintf(family_name, sizeof(family_name), "%s / ",
1280 r600_get_family_name(rscreen) + 4);
1281 else
1282 chip_name = r600_get_family_name(rscreen);
1283
1284 if (uname(&uname_data) == 0)
1285 snprintf(kernel_version, sizeof(kernel_version),
1286 " / %s", uname_data.release);
1287
1288 if (HAVE_LLVM > 0) {
1289 snprintf(llvm_string, sizeof(llvm_string),
1290 ", LLVM %i.%i.%i", (HAVE_LLVM >> 8) & 0xff,
1291 HAVE_LLVM & 0xff, MESA_LLVM_VERSION_PATCH);
1292 }
1293
1294 snprintf(rscreen->renderer_string, sizeof(rscreen->renderer_string),
1295 "%s (%sDRM %i.%i.%i%s%s)",
1296 chip_name, family_name, rscreen->info.drm_major,
1297 rscreen->info.drm_minor, rscreen->info.drm_patchlevel,
1298 kernel_version, llvm_string);
1299
1300 rscreen->b.get_name = r600_get_name;
1301 rscreen->b.get_vendor = r600_get_vendor;
1302 rscreen->b.get_device_vendor = r600_get_device_vendor;
1303 rscreen->b.get_disk_shader_cache = r600_get_disk_shader_cache;
1304 rscreen->b.get_compute_param = r600_get_compute_param;
1305 rscreen->b.get_paramf = r600_get_paramf;
1306 rscreen->b.get_timestamp = r600_get_timestamp;
1307 rscreen->b.fence_finish = r600_fence_finish;
1308 rscreen->b.fence_reference = r600_fence_reference;
1309 rscreen->b.resource_destroy = u_resource_destroy_vtbl;
1310 rscreen->b.resource_from_user_memory = r600_buffer_from_user_memory;
1311 rscreen->b.query_memory_info = r600_query_memory_info;
1312
1313 if (rscreen->info.has_hw_decode) {
1314 rscreen->b.get_video_param = rvid_get_video_param;
1315 rscreen->b.is_video_format_supported = rvid_is_format_supported;
1316 } else {
1317 rscreen->b.get_video_param = r600_get_video_param;
1318 rscreen->b.is_video_format_supported = vl_video_buffer_is_format_supported;
1319 }
1320
1321 r600_init_screen_texture_functions(rscreen);
1322 r600_init_screen_query_functions(rscreen);
1323
1324 rscreen->family = rscreen->info.family;
1325 rscreen->chip_class = rscreen->info.chip_class;
1326 rscreen->debug_flags |= debug_get_flags_option("R600_DEBUG", common_debug_options, 0);
1327
1328 r600_disk_cache_create(rscreen);
1329
1330 slab_create_parent(&rscreen->pool_transfers, sizeof(struct r600_transfer), 64);
1331
1332 rscreen->force_aniso = MIN2(16, debug_get_num_option("R600_TEX_ANISO", -1));
1333 if (rscreen->force_aniso >= 0) {
1334 printf("radeon: Forcing anisotropy filter to %ix\n",
1335 /* round down to a power of two */
1336 1 << util_logbase2(rscreen->force_aniso));
1337 }
1338
1339 (void) mtx_init(&rscreen->aux_context_lock, mtx_plain);
1340 (void) mtx_init(&rscreen->gpu_load_mutex, mtx_plain);
1341
1342 if (rscreen->debug_flags & DBG_INFO) {
1343 printf("pci (domain:bus:dev.func): %04x:%02x:%02x.%x\n",
1344 rscreen->info.pci_domain, rscreen->info.pci_bus,
1345 rscreen->info.pci_dev, rscreen->info.pci_func);
1346 printf("pci_id = 0x%x\n", rscreen->info.pci_id);
1347 printf("family = %i (%s)\n", rscreen->info.family,
1348 r600_get_family_name(rscreen));
1349 printf("chip_class = %i\n", rscreen->info.chip_class);
1350 printf("pte_fragment_size = %u\n", rscreen->info.pte_fragment_size);
1351 printf("gart_page_size = %u\n", rscreen->info.gart_page_size);
1352 printf("gart_size = %i MB\n", (int)DIV_ROUND_UP(rscreen->info.gart_size, 1024*1024));
1353 printf("vram_size = %i MB\n", (int)DIV_ROUND_UP(rscreen->info.vram_size, 1024*1024));
1354 printf("vram_vis_size = %i MB\n", (int)DIV_ROUND_UP(rscreen->info.vram_vis_size, 1024*1024));
1355 printf("max_alloc_size = %i MB\n",
1356 (int)DIV_ROUND_UP(rscreen->info.max_alloc_size, 1024*1024));
1357 printf("min_alloc_size = %u\n", rscreen->info.min_alloc_size);
1358 printf("has_dedicated_vram = %u\n", rscreen->info.has_dedicated_vram);
1359 printf("has_virtual_memory = %i\n", rscreen->info.has_virtual_memory);
1360 printf("gfx_ib_pad_with_type2 = %i\n", rscreen->info.gfx_ib_pad_with_type2);
1361 printf("has_hw_decode = %u\n", rscreen->info.has_hw_decode);
1362 printf("num_sdma_rings = %i\n", rscreen->info.num_sdma_rings);
1363 printf("num_compute_rings = %u\n", rscreen->info.num_compute_rings);
1364 printf("uvd_fw_version = %u\n", rscreen->info.uvd_fw_version);
1365 printf("vce_fw_version = %u\n", rscreen->info.vce_fw_version);
1366 printf("me_fw_version = %i\n", rscreen->info.me_fw_version);
1367 printf("pfp_fw_version = %i\n", rscreen->info.pfp_fw_version);
1368 printf("ce_fw_version = %i\n", rscreen->info.ce_fw_version);
1369 printf("vce_harvest_config = %i\n", rscreen->info.vce_harvest_config);
1370 printf("clock_crystal_freq = %i\n", rscreen->info.clock_crystal_freq);
1371 printf("tcc_cache_line_size = %u\n", rscreen->info.tcc_cache_line_size);
1372 printf("drm = %i.%i.%i\n", rscreen->info.drm_major,
1373 rscreen->info.drm_minor, rscreen->info.drm_patchlevel);
1374 printf("has_userptr = %i\n", rscreen->info.has_userptr);
1375 printf("has_syncobj = %u\n", rscreen->info.has_syncobj);
1376
1377 printf("r600_max_quad_pipes = %i\n", rscreen->info.r600_max_quad_pipes);
1378 printf("max_shader_clock = %i\n", rscreen->info.max_shader_clock);
1379 printf("num_good_compute_units = %i\n", rscreen->info.num_good_compute_units);
1380 printf("max_se = %i\n", rscreen->info.max_se);
1381 printf("max_sh_per_se = %i\n", rscreen->info.max_sh_per_se);
1382
1383 printf("r600_gb_backend_map = %i\n", rscreen->info.r600_gb_backend_map);
1384 printf("r600_gb_backend_map_valid = %i\n", rscreen->info.r600_gb_backend_map_valid);
1385 printf("r600_num_banks = %i\n", rscreen->info.r600_num_banks);
1386 printf("num_render_backends = %i\n", rscreen->info.num_render_backends);
1387 printf("num_tile_pipes = %i\n", rscreen->info.num_tile_pipes);
1388 printf("pipe_interleave_bytes = %i\n", rscreen->info.pipe_interleave_bytes);
1389 printf("enabled_rb_mask = 0x%x\n", rscreen->info.enabled_rb_mask);
1390 printf("max_alignment = %u\n", (unsigned)rscreen->info.max_alignment);
1391 }
1392 return true;
1393 }
1394
r600_destroy_common_screen(struct r600_common_screen * rscreen)1395 void r600_destroy_common_screen(struct r600_common_screen *rscreen)
1396 {
1397 r600_perfcounters_destroy(rscreen);
1398 r600_gpu_load_kill_thread(rscreen);
1399
1400 mtx_destroy(&rscreen->gpu_load_mutex);
1401 mtx_destroy(&rscreen->aux_context_lock);
1402 rscreen->aux_context->destroy(rscreen->aux_context);
1403
1404 slab_destroy_parent(&rscreen->pool_transfers);
1405
1406 disk_cache_destroy(rscreen->disk_shader_cache);
1407 rscreen->ws->destroy(rscreen->ws);
1408 FREE(rscreen);
1409 }
1410
r600_can_dump_shader(struct r600_common_screen * rscreen,unsigned processor)1411 bool r600_can_dump_shader(struct r600_common_screen *rscreen,
1412 unsigned processor)
1413 {
1414 return rscreen->debug_flags & (1 << processor);
1415 }
1416
r600_extra_shader_checks(struct r600_common_screen * rscreen,unsigned processor)1417 bool r600_extra_shader_checks(struct r600_common_screen *rscreen, unsigned processor)
1418 {
1419 return (rscreen->debug_flags & DBG_CHECK_IR) ||
1420 r600_can_dump_shader(rscreen, processor);
1421 }
1422
r600_screen_clear_buffer(struct r600_common_screen * rscreen,struct pipe_resource * dst,uint64_t offset,uint64_t size,unsigned value)1423 void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
1424 uint64_t offset, uint64_t size, unsigned value)
1425 {
1426 struct r600_common_context *rctx = (struct r600_common_context*)rscreen->aux_context;
1427
1428 mtx_lock(&rscreen->aux_context_lock);
1429 rctx->dma_clear_buffer(&rctx->b, dst, offset, size, value);
1430 rscreen->aux_context->flush(rscreen->aux_context, NULL, 0);
1431 mtx_unlock(&rscreen->aux_context_lock);
1432 }
1433