Merge remote-tracking branch 'public/master' into vulkan
[mesa.git] / src / gallium / drivers / radeon / r600_buffer_common.c
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Marek Olšák
25 */
26
27 #include "r600_cs.h"
28 #include "util/u_memory.h"
29 #include "util/u_upload_mgr.h"
30 #include <inttypes.h>
31 #include <stdio.h>
32
33 boolean r600_rings_is_buffer_referenced(struct r600_common_context *ctx,
34 struct pb_buffer *buf,
35 enum radeon_bo_usage usage)
36 {
37 if (ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs, buf, usage)) {
38 return TRUE;
39 }
40 if (ctx->dma.cs && ctx->dma.cs->cdw &&
41 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs, buf, usage)) {
42 return TRUE;
43 }
44 return FALSE;
45 }
46
47 void *r600_buffer_map_sync_with_rings(struct r600_common_context *ctx,
48 struct r600_resource *resource,
49 unsigned usage)
50 {
51 enum radeon_bo_usage rusage = RADEON_USAGE_READWRITE;
52 bool busy = false;
53
54 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
55 return ctx->ws->buffer_map(resource->buf, NULL, usage);
56 }
57
58 if (!(usage & PIPE_TRANSFER_WRITE)) {
59 /* have to wait for the last write */
60 rusage = RADEON_USAGE_WRITE;
61 }
62
63 if (ctx->gfx.cs->cdw != ctx->initial_gfx_cs_size &&
64 ctx->ws->cs_is_buffer_referenced(ctx->gfx.cs,
65 resource->buf, rusage)) {
66 if (usage & PIPE_TRANSFER_DONTBLOCK) {
67 ctx->gfx.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
68 return NULL;
69 } else {
70 ctx->gfx.flush(ctx, 0, NULL);
71 busy = true;
72 }
73 }
74 if (ctx->dma.cs &&
75 ctx->dma.cs->cdw &&
76 ctx->ws->cs_is_buffer_referenced(ctx->dma.cs,
77 resource->buf, rusage)) {
78 if (usage & PIPE_TRANSFER_DONTBLOCK) {
79 ctx->dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
80 return NULL;
81 } else {
82 ctx->dma.flush(ctx, 0, NULL);
83 busy = true;
84 }
85 }
86
87 if (busy || !ctx->ws->buffer_wait(resource->buf, 0, rusage)) {
88 if (usage & PIPE_TRANSFER_DONTBLOCK) {
89 return NULL;
90 } else {
91 /* We will be wait for the GPU. Wait for any offloaded
92 * CS flush to complete to avoid busy-waiting in the winsys. */
93 ctx->ws->cs_sync_flush(ctx->gfx.cs);
94 if (ctx->dma.cs)
95 ctx->ws->cs_sync_flush(ctx->dma.cs);
96 }
97 }
98
99 /* Setting the CS to NULL will prevent doing checks we have done already. */
100 return ctx->ws->buffer_map(resource->buf, NULL, usage);
101 }
102
103 bool r600_init_resource(struct r600_common_screen *rscreen,
104 struct r600_resource *res,
105 uint64_t size, unsigned alignment,
106 bool use_reusable_pool)
107 {
108 struct r600_texture *rtex = (struct r600_texture*)res;
109 struct pb_buffer *old_buf, *new_buf;
110 enum radeon_bo_flag flags = 0;
111
112 switch (res->b.b.usage) {
113 case PIPE_USAGE_STREAM:
114 flags = RADEON_FLAG_GTT_WC;
115 /* fall through */
116 case PIPE_USAGE_STAGING:
117 /* Transfers are likely to occur more often with these resources. */
118 res->domains = RADEON_DOMAIN_GTT;
119 break;
120 case PIPE_USAGE_DYNAMIC:
121 /* Older kernels didn't always flush the HDP cache before
122 * CS execution
123 */
124 if (rscreen->info.drm_major == 2 &&
125 rscreen->info.drm_minor < 40) {
126 res->domains = RADEON_DOMAIN_GTT;
127 flags |= RADEON_FLAG_GTT_WC;
128 break;
129 }
130 flags |= RADEON_FLAG_CPU_ACCESS;
131 /* fall through */
132 case PIPE_USAGE_DEFAULT:
133 case PIPE_USAGE_IMMUTABLE:
134 default:
135 /* Not listing GTT here improves performance in some apps. */
136 res->domains = RADEON_DOMAIN_VRAM;
137 flags |= RADEON_FLAG_GTT_WC;
138 break;
139 }
140
141 if (res->b.b.target == PIPE_BUFFER &&
142 res->b.b.flags & (PIPE_RESOURCE_FLAG_MAP_PERSISTENT |
143 PIPE_RESOURCE_FLAG_MAP_COHERENT)) {
144 /* Use GTT for all persistent mappings with older kernels,
145 * because they didn't always flush the HDP cache before CS
146 * execution.
147 *
148 * Write-combined CPU mappings are fine, the kernel ensures all CPU
149 * writes finish before the GPU executes a command stream.
150 */
151 if (rscreen->info.drm_major == 2 &&
152 rscreen->info.drm_minor < 40)
153 res->domains = RADEON_DOMAIN_GTT;
154 else if (res->domains & RADEON_DOMAIN_VRAM)
155 flags |= RADEON_FLAG_CPU_ACCESS;
156 }
157
158 /* Tiled textures are unmappable. Always put them in VRAM. */
159 if (res->b.b.target != PIPE_BUFFER &&
160 rtex->surface.level[0].mode >= RADEON_SURF_MODE_1D) {
161 res->domains = RADEON_DOMAIN_VRAM;
162 flags &= ~RADEON_FLAG_CPU_ACCESS;
163 flags |= RADEON_FLAG_NO_CPU_ACCESS |
164 RADEON_FLAG_GTT_WC;
165 }
166
167 /* If VRAM is just stolen system memory, allow both VRAM and GTT,
168 * whichever has free space. If a buffer is evicted from VRAM to GTT,
169 * it will stay there.
170 */
171 if (!rscreen->info.has_dedicated_vram &&
172 res->domains == RADEON_DOMAIN_VRAM)
173 res->domains = RADEON_DOMAIN_VRAM_GTT;
174
175 if (rscreen->debug_flags & DBG_NO_WC)
176 flags &= ~RADEON_FLAG_GTT_WC;
177
178 /* Allocate a new resource. */
179 new_buf = rscreen->ws->buffer_create(rscreen->ws, size, alignment,
180 use_reusable_pool,
181 res->domains, flags);
182 if (!new_buf) {
183 return false;
184 }
185
186 /* Replace the pointer such that if res->buf wasn't NULL, it won't be
187 * NULL. This should prevent crashes with multiple contexts using
188 * the same buffer where one of the contexts invalidates it while
189 * the others are using it. */
190 old_buf = res->buf;
191 res->buf = new_buf; /* should be atomic */
192
193 if (rscreen->info.has_virtual_memory)
194 res->gpu_address = rscreen->ws->buffer_get_virtual_address(res->buf);
195 else
196 res->gpu_address = 0;
197
198 pb_reference(&old_buf, NULL);
199
200 util_range_set_empty(&res->valid_buffer_range);
201 res->TC_L2_dirty = false;
202
203 if (rscreen->debug_flags & DBG_VM && res->b.b.target == PIPE_BUFFER) {
204 fprintf(stderr, "VM start=0x%"PRIX64" end=0x%"PRIX64" | Buffer %"PRIu64" bytes\n",
205 res->gpu_address, res->gpu_address + res->buf->size,
206 res->buf->size);
207 }
208 return true;
209 }
210
211 static void r600_buffer_destroy(struct pipe_screen *screen,
212 struct pipe_resource *buf)
213 {
214 struct r600_resource *rbuffer = r600_resource(buf);
215
216 util_range_destroy(&rbuffer->valid_buffer_range);
217 pb_reference(&rbuffer->buf, NULL);
218 FREE(rbuffer);
219 }
220
221 static bool
222 r600_invalidate_buffer(struct r600_common_context *rctx,
223 struct r600_resource *rbuffer)
224 {
225 /* Shared buffers can't be reallocated. */
226 if (rbuffer->is_shared)
227 return false;
228
229 /* In AMD_pinned_memory, the user pointer association only gets
230 * broken when the buffer is explicitly re-allocated.
231 */
232 if (rctx->ws->buffer_is_user_ptr(rbuffer->buf))
233 return false;
234
235 /* Check if mapping this buffer would cause waiting for the GPU. */
236 if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
237 !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
238 rctx->invalidate_buffer(&rctx->b, &rbuffer->b.b);
239 } else {
240 util_range_set_empty(&rbuffer->valid_buffer_range);
241 }
242
243 return true;
244 }
245
246 void r600_invalidate_resource(struct pipe_context *ctx,
247 struct pipe_resource *resource)
248 {
249 struct r600_common_context *rctx = (struct r600_common_context*)ctx;
250 struct r600_resource *rbuffer = r600_resource(resource);
251
252 /* We currently only do anyting here for buffers */
253 if (resource->target == PIPE_BUFFER)
254 (void)r600_invalidate_buffer(rctx, rbuffer);
255 }
256
257 static void *r600_buffer_get_transfer(struct pipe_context *ctx,
258 struct pipe_resource *resource,
259 unsigned level,
260 unsigned usage,
261 const struct pipe_box *box,
262 struct pipe_transfer **ptransfer,
263 void *data, struct r600_resource *staging,
264 unsigned offset)
265 {
266 struct r600_common_context *rctx = (struct r600_common_context*)ctx;
267 struct r600_transfer *transfer = util_slab_alloc(&rctx->pool_transfers);
268
269 transfer->transfer.resource = resource;
270 transfer->transfer.level = level;
271 transfer->transfer.usage = usage;
272 transfer->transfer.box = *box;
273 transfer->transfer.stride = 0;
274 transfer->transfer.layer_stride = 0;
275 transfer->offset = offset;
276 transfer->staging = staging;
277 *ptransfer = &transfer->transfer;
278 return data;
279 }
280
281 static bool r600_can_dma_copy_buffer(struct r600_common_context *rctx,
282 unsigned dstx, unsigned srcx, unsigned size)
283 {
284 bool dword_aligned = !(dstx % 4) && !(srcx % 4) && !(size % 4);
285
286 return rctx->screen->has_cp_dma ||
287 (dword_aligned && (rctx->dma.cs ||
288 rctx->screen->has_streamout));
289
290 }
291
292 static void *r600_buffer_transfer_map(struct pipe_context *ctx,
293 struct pipe_resource *resource,
294 unsigned level,
295 unsigned usage,
296 const struct pipe_box *box,
297 struct pipe_transfer **ptransfer)
298 {
299 struct r600_common_context *rctx = (struct r600_common_context*)ctx;
300 struct r600_common_screen *rscreen = (struct r600_common_screen*)ctx->screen;
301 struct r600_resource *rbuffer = r600_resource(resource);
302 uint8_t *data;
303
304 assert(box->x + box->width <= resource->width0);
305
306 /* See if the buffer range being mapped has never been initialized,
307 * in which case it can be mapped unsynchronized. */
308 if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
309 usage & PIPE_TRANSFER_WRITE &&
310 !rbuffer->is_shared &&
311 !util_ranges_intersect(&rbuffer->valid_buffer_range, box->x, box->x + box->width)) {
312 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
313 }
314
315 /* If discarding the entire range, discard the whole resource instead. */
316 if (usage & PIPE_TRANSFER_DISCARD_RANGE &&
317 box->x == 0 && box->width == resource->width0) {
318 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
319 }
320
321 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE &&
322 !(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
323 assert(usage & PIPE_TRANSFER_WRITE);
324
325 if (r600_invalidate_buffer(rctx, rbuffer)) {
326 /* At this point, the buffer is always idle. */
327 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
328 } else {
329 /* Fall back to a temporary buffer. */
330 usage |= PIPE_TRANSFER_DISCARD_RANGE;
331 }
332 }
333
334 if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
335 !(usage & (PIPE_TRANSFER_UNSYNCHRONIZED |
336 PIPE_TRANSFER_PERSISTENT)) &&
337 !(rscreen->debug_flags & DBG_NO_DISCARD_RANGE) &&
338 r600_can_dma_copy_buffer(rctx, box->x, 0, box->width)) {
339 assert(usage & PIPE_TRANSFER_WRITE);
340
341 /* Check if mapping this buffer would cause waiting for the GPU. */
342 if (r600_rings_is_buffer_referenced(rctx, rbuffer->buf, RADEON_USAGE_READWRITE) ||
343 !rctx->ws->buffer_wait(rbuffer->buf, 0, RADEON_USAGE_READWRITE)) {
344 /* Do a wait-free write-only transfer using a temporary buffer. */
345 unsigned offset;
346 struct r600_resource *staging = NULL;
347
348 u_upload_alloc(rctx->uploader, 0, box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT),
349 256, &offset, (struct pipe_resource**)&staging, (void**)&data);
350
351 if (staging) {
352 data += box->x % R600_MAP_BUFFER_ALIGNMENT;
353 return r600_buffer_get_transfer(ctx, resource, level, usage, box,
354 ptransfer, data, staging, offset);
355 }
356 } else {
357 /* At this point, the buffer is always idle (we checked it above). */
358 usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
359 }
360 }
361 /* Using a staging buffer in GTT for larger reads is much faster. */
362 else if ((usage & PIPE_TRANSFER_READ) &&
363 !(usage & (PIPE_TRANSFER_WRITE |
364 PIPE_TRANSFER_PERSISTENT)) &&
365 rbuffer->domains == RADEON_DOMAIN_VRAM &&
366 r600_can_dma_copy_buffer(rctx, 0, box->x, box->width)) {
367 struct r600_resource *staging;
368
369 staging = (struct r600_resource*) pipe_buffer_create(
370 ctx->screen, PIPE_BIND_TRANSFER_READ, PIPE_USAGE_STAGING,
371 box->width + (box->x % R600_MAP_BUFFER_ALIGNMENT));
372 if (staging) {
373 /* Copy the VRAM buffer to the staging buffer. */
374 rctx->dma_copy(ctx, &staging->b.b, 0,
375 box->x % R600_MAP_BUFFER_ALIGNMENT,
376 0, 0, resource, level, box);
377
378 data = r600_buffer_map_sync_with_rings(rctx, staging, PIPE_TRANSFER_READ);
379 if (!data) {
380 pipe_resource_reference((struct pipe_resource **)&staging, NULL);
381 return NULL;
382 }
383 data += box->x % R600_MAP_BUFFER_ALIGNMENT;
384
385 return r600_buffer_get_transfer(ctx, resource, level, usage, box,
386 ptransfer, data, staging, 0);
387 }
388 }
389
390 data = r600_buffer_map_sync_with_rings(rctx, rbuffer, usage);
391 if (!data) {
392 return NULL;
393 }
394 data += box->x;
395
396 return r600_buffer_get_transfer(ctx, resource, level, usage, box,
397 ptransfer, data, NULL, 0);
398 }
399
400 static void r600_buffer_do_flush_region(struct pipe_context *ctx,
401 struct pipe_transfer *transfer,
402 const struct pipe_box *box)
403 {
404 struct r600_common_context *rctx = (struct r600_common_context*)ctx;
405 struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
406 struct r600_resource *rbuffer = r600_resource(transfer->resource);
407
408 if (rtransfer->staging) {
409 struct pipe_resource *dst, *src;
410 unsigned soffset;
411 struct pipe_box dma_box;
412
413 dst = transfer->resource;
414 src = &rtransfer->staging->b.b;
415 soffset = rtransfer->offset + box->x % R600_MAP_BUFFER_ALIGNMENT;
416
417 u_box_1d(soffset, box->width, &dma_box);
418
419 /* Copy the staging buffer into the original one. */
420 rctx->dma_copy(ctx, dst, 0, box->x, 0, 0, src, 0, &dma_box);
421 }
422
423 util_range_add(&rbuffer->valid_buffer_range, box->x,
424 box->x + box->width);
425 }
426
427 static void r600_buffer_flush_region(struct pipe_context *ctx,
428 struct pipe_transfer *transfer,
429 const struct pipe_box *rel_box)
430 {
431 if (transfer->usage & (PIPE_TRANSFER_WRITE |
432 PIPE_TRANSFER_FLUSH_EXPLICIT)) {
433 struct pipe_box box;
434
435 u_box_1d(transfer->box.x + rel_box->x, rel_box->width, &box);
436 r600_buffer_do_flush_region(ctx, transfer, &box);
437 }
438 }
439
440 static void r600_buffer_transfer_unmap(struct pipe_context *ctx,
441 struct pipe_transfer *transfer)
442 {
443 struct r600_common_context *rctx = (struct r600_common_context*)ctx;
444 struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
445
446 if (transfer->usage & PIPE_TRANSFER_WRITE &&
447 !(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT))
448 r600_buffer_do_flush_region(ctx, transfer, &transfer->box);
449
450 if (rtransfer->staging)
451 pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
452
453 util_slab_free(&rctx->pool_transfers, transfer);
454 }
455
456 static const struct u_resource_vtbl r600_buffer_vtbl =
457 {
458 NULL, /* get_handle */
459 r600_buffer_destroy, /* resource_destroy */
460 r600_buffer_transfer_map, /* transfer_map */
461 r600_buffer_flush_region, /* transfer_flush_region */
462 r600_buffer_transfer_unmap, /* transfer_unmap */
463 NULL /* transfer_inline_write */
464 };
465
466 static struct r600_resource *
467 r600_alloc_buffer_struct(struct pipe_screen *screen,
468 const struct pipe_resource *templ)
469 {
470 struct r600_resource *rbuffer;
471
472 rbuffer = MALLOC_STRUCT(r600_resource);
473
474 rbuffer->b.b = *templ;
475 pipe_reference_init(&rbuffer->b.b.reference, 1);
476 rbuffer->b.b.screen = screen;
477 rbuffer->b.vtbl = &r600_buffer_vtbl;
478 rbuffer->buf = NULL;
479 rbuffer->TC_L2_dirty = false;
480 rbuffer->is_shared = false;
481 util_range_init(&rbuffer->valid_buffer_range);
482 return rbuffer;
483 }
484
485 struct pipe_resource *r600_buffer_create(struct pipe_screen *screen,
486 const struct pipe_resource *templ,
487 unsigned alignment)
488 {
489 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
490 struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
491
492 if (!r600_init_resource(rscreen, rbuffer, templ->width0, alignment, TRUE)) {
493 FREE(rbuffer);
494 return NULL;
495 }
496 return &rbuffer->b.b;
497 }
498
499 struct pipe_resource *r600_aligned_buffer_create(struct pipe_screen *screen,
500 unsigned bind,
501 unsigned usage,
502 unsigned size,
503 unsigned alignment)
504 {
505 struct pipe_resource buffer;
506
507 memset(&buffer, 0, sizeof buffer);
508 buffer.target = PIPE_BUFFER;
509 buffer.format = PIPE_FORMAT_R8_UNORM;
510 buffer.bind = bind;
511 buffer.usage = usage;
512 buffer.flags = 0;
513 buffer.width0 = size;
514 buffer.height0 = 1;
515 buffer.depth0 = 1;
516 buffer.array_size = 1;
517 return r600_buffer_create(screen, &buffer, alignment);
518 }
519
520 struct pipe_resource *
521 r600_buffer_from_user_memory(struct pipe_screen *screen,
522 const struct pipe_resource *templ,
523 void *user_memory)
524 {
525 struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
526 struct radeon_winsys *ws = rscreen->ws;
527 struct r600_resource *rbuffer = r600_alloc_buffer_struct(screen, templ);
528
529 rbuffer->domains = RADEON_DOMAIN_GTT;
530 util_range_add(&rbuffer->valid_buffer_range, 0, templ->width0);
531
532 /* Convert a user pointer to a buffer. */
533 rbuffer->buf = ws->buffer_from_ptr(ws, user_memory, templ->width0);
534 if (!rbuffer->buf) {
535 FREE(rbuffer);
536 return NULL;
537 }
538
539 if (rscreen->info.has_virtual_memory)
540 rbuffer->gpu_address =
541 ws->buffer_get_virtual_address(rbuffer->buf);
542 else
543 rbuffer->gpu_address = 0;
544
545 return &rbuffer->b.b;
546 }