i965: Separate gen < 8 and gen >= 8 paths explicitly in wrap_mode()
[mesa.git] / src / mesa / drivers / dri / i965 / intel_buffer_objects.c
1 /*
2 * Copyright 2003 VMware, Inc.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial portions
15 * of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
21 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
22 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
23 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 /**
27 * @file intel_buffer_objects.c
28 *
29 * This provides core GL buffer object functionality.
30 */
31
32 #include "main/imports.h"
33 #include "main/mtypes.h"
34 #include "main/macros.h"
35 #include "main/bufferobj.h"
36
37 #include "brw_context.h"
38 #include "intel_blit.h"
39 #include "intel_buffer_objects.h"
40 #include "intel_batchbuffer.h"
41
42 static void
43 mark_buffer_gpu_usage(struct intel_buffer_object *intel_obj,
44 uint32_t offset, uint32_t size)
45 {
46 intel_obj->gpu_active_start = MIN2(intel_obj->gpu_active_start, offset);
47 intel_obj->gpu_active_end = MAX2(intel_obj->gpu_active_end, offset + size);
48 }
49
50 static void
51 mark_buffer_inactive(struct intel_buffer_object *intel_obj)
52 {
53 intel_obj->gpu_active_start = ~0;
54 intel_obj->gpu_active_end = 0;
55 }
56
57 /** Allocates a new brw_bo to store the data for the buffer object. */
58 static void
59 alloc_buffer_object(struct brw_context *brw,
60 struct intel_buffer_object *intel_obj)
61 {
62 intel_obj->buffer = brw_bo_alloc(brw->bufmgr, "bufferobj",
63 intel_obj->Base.Size, 64);
64
65 /* the buffer might be bound as a uniform buffer, need to update it
66 */
67 if (intel_obj->Base.UsageHistory & USAGE_UNIFORM_BUFFER)
68 brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
69 if (intel_obj->Base.UsageHistory & USAGE_SHADER_STORAGE_BUFFER)
70 brw->ctx.NewDriverState |= BRW_NEW_UNIFORM_BUFFER;
71 if (intel_obj->Base.UsageHistory & USAGE_TEXTURE_BUFFER)
72 brw->ctx.NewDriverState |= BRW_NEW_TEXTURE_BUFFER;
73 if (intel_obj->Base.UsageHistory & USAGE_ATOMIC_COUNTER_BUFFER)
74 brw->ctx.NewDriverState |= BRW_NEW_ATOMIC_BUFFER;
75
76 mark_buffer_inactive(intel_obj);
77 }
78
79 static void
80 release_buffer(struct intel_buffer_object *intel_obj)
81 {
82 brw_bo_unreference(intel_obj->buffer);
83 intel_obj->buffer = NULL;
84 }
85
86 /**
87 * The NewBufferObject() driver hook.
88 *
89 * Allocates a new intel_buffer_object structure and initializes it.
90 *
91 * There is some duplication between mesa's bufferobjects and our
92 * bufmgr buffers. Both have an integer handle and a hashtable to
93 * lookup an opaque structure. It would be nice if the handles and
94 * internal structure where somehow shared.
95 */
96 static struct gl_buffer_object *
97 brw_new_buffer_object(struct gl_context * ctx, GLuint name)
98 {
99 struct intel_buffer_object *obj = CALLOC_STRUCT(intel_buffer_object);
100 if (!obj) {
101 _mesa_error_no_memory(__func__);
102 }
103
104 _mesa_initialize_buffer_object(ctx, &obj->Base, name);
105
106 obj->buffer = NULL;
107
108 return &obj->Base;
109 }
110
111 /**
112 * The DeleteBuffer() driver hook.
113 *
114 * Deletes a single OpenGL buffer object. Used by glDeleteBuffers().
115 */
116 static void
117 brw_delete_buffer(struct gl_context * ctx, struct gl_buffer_object *obj)
118 {
119 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
120
121 assert(intel_obj);
122
123 /* Buffer objects are automatically unmapped when deleting according
124 * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
125 * (though it does if you call glDeleteBuffers)
126 */
127 _mesa_buffer_unmap_all_mappings(ctx, obj);
128
129 brw_bo_unreference(intel_obj->buffer);
130 _mesa_delete_buffer_object(ctx, obj);
131 }
132
133
134 /**
135 * The BufferData() driver hook.
136 *
137 * Implements glBufferData(), which recreates a buffer object's data store
138 * and populates it with the given data, if present.
139 *
140 * Any data that was previously stored in the buffer object is lost.
141 *
142 * \return true for success, false if out of memory
143 */
144 static GLboolean
145 brw_buffer_data(struct gl_context *ctx,
146 GLenum target,
147 GLsizeiptrARB size,
148 const GLvoid *data,
149 GLenum usage,
150 GLbitfield storageFlags,
151 struct gl_buffer_object *obj)
152 {
153 struct brw_context *brw = brw_context(ctx);
154 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
155
156 /* Part of the ABI, but this function doesn't use it.
157 */
158 (void) target;
159
160 intel_obj->Base.Size = size;
161 intel_obj->Base.Usage = usage;
162 intel_obj->Base.StorageFlags = storageFlags;
163
164 assert(!obj->Mappings[MAP_USER].Pointer); /* Mesa should have unmapped it */
165 assert(!obj->Mappings[MAP_INTERNAL].Pointer);
166
167 if (intel_obj->buffer != NULL)
168 release_buffer(intel_obj);
169
170 if (size != 0) {
171 alloc_buffer_object(brw, intel_obj);
172 if (!intel_obj->buffer)
173 return false;
174
175 if (data != NULL)
176 brw_bo_subdata(intel_obj->buffer, 0, size, data);
177 }
178
179 return true;
180 }
181
182
183 /**
184 * The BufferSubData() driver hook.
185 *
186 * Implements glBufferSubData(), which replaces a portion of the data in a
187 * buffer object.
188 *
189 * If the data range specified by (size + offset) extends beyond the end of
190 * the buffer or if data is NULL, no copy is performed.
191 */
192 static void
193 brw_buffer_subdata(struct gl_context *ctx,
194 GLintptrARB offset,
195 GLsizeiptrARB size,
196 const GLvoid *data,
197 struct gl_buffer_object *obj)
198 {
199 struct brw_context *brw = brw_context(ctx);
200 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
201 bool busy;
202
203 if (size == 0)
204 return;
205
206 assert(intel_obj);
207
208 /* See if we can unsynchronized write the data into the user's BO. This
209 * avoids GPU stalls in unfortunately common user patterns (uploading
210 * sequentially into a BO, with draw calls in between each upload).
211 *
212 * Once we've hit this path, we mark this GL BO as preferring stalling to
213 * blits, so that we can hopefully hit this path again in the future
214 * (otherwise, an app that might occasionally stall but mostly not will end
215 * up with blitting all the time, at the cost of bandwidth)
216 */
217 if (offset + size <= intel_obj->gpu_active_start ||
218 intel_obj->gpu_active_end <= offset) {
219 void *map = brw_bo_map(brw, intel_obj->buffer, MAP_WRITE | MAP_ASYNC);
220 memcpy(map + offset, data, size);
221 brw_bo_unmap(intel_obj->buffer);
222
223 if (intel_obj->gpu_active_end > intel_obj->gpu_active_start)
224 intel_obj->prefer_stall_to_blit = true;
225 return;
226 }
227
228 busy =
229 brw_bo_busy(intel_obj->buffer) ||
230 brw_batch_references(&brw->batch, intel_obj->buffer);
231
232 if (busy) {
233 if (size == intel_obj->Base.Size) {
234 /* Replace the current busy bo so the subdata doesn't stall. */
235 brw_bo_unreference(intel_obj->buffer);
236 alloc_buffer_object(brw, intel_obj);
237 } else if (!intel_obj->prefer_stall_to_blit) {
238 perf_debug("Using a blit copy to avoid stalling on "
239 "glBufferSubData(%ld, %ld) (%ldkb) to a busy "
240 "(%d-%d) buffer object.\n",
241 (long)offset, (long)offset + size, (long)(size/1024),
242 intel_obj->gpu_active_start,
243 intel_obj->gpu_active_end);
244 struct brw_bo *temp_bo =
245 brw_bo_alloc(brw->bufmgr, "subdata temp", size, 64);
246
247 brw_bo_subdata(temp_bo, 0, size, data);
248
249 intel_emit_linear_blit(brw,
250 intel_obj->buffer, offset,
251 temp_bo, 0,
252 size);
253
254 brw_bo_unreference(temp_bo);
255 return;
256 } else {
257 perf_debug("Stalling on glBufferSubData(%ld, %ld) (%ldkb) to a busy "
258 "(%d-%d) buffer object. Use glMapBufferRange() to "
259 "avoid this.\n",
260 (long)offset, (long)offset + size, (long)(size/1024),
261 intel_obj->gpu_active_start,
262 intel_obj->gpu_active_end);
263 intel_batchbuffer_flush(brw);
264 }
265 }
266
267 brw_bo_subdata(intel_obj->buffer, offset, size, data);
268 mark_buffer_inactive(intel_obj);
269 }
270
271
272 /**
273 * The GetBufferSubData() driver hook.
274 *
275 * Implements glGetBufferSubData(), which copies a subrange of a buffer
276 * object into user memory.
277 */
278 static void
279 brw_get_buffer_subdata(struct gl_context *ctx,
280 GLintptrARB offset,
281 GLsizeiptrARB size,
282 GLvoid *data,
283 struct gl_buffer_object *obj)
284 {
285 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
286 struct brw_context *brw = brw_context(ctx);
287
288 assert(intel_obj);
289 if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
290 intel_batchbuffer_flush(brw);
291 }
292 brw_bo_get_subdata(intel_obj->buffer, offset, size, data);
293
294 mark_buffer_inactive(intel_obj);
295 }
296
297
298 /**
299 * The MapBufferRange() driver hook.
300 *
301 * This implements both glMapBufferRange() and glMapBuffer().
302 *
303 * The goal of this extension is to allow apps to accumulate their rendering
304 * at the same time as they accumulate their buffer object. Without it,
305 * you'd end up blocking on execution of rendering every time you mapped
306 * the buffer to put new data in.
307 *
308 * We support it in 3 ways: If unsynchronized, then don't bother
309 * flushing the batchbuffer before mapping the buffer, which can save blocking
310 * in many cases. If we would still block, and they allow the whole buffer
311 * to be invalidated, then just allocate a new buffer to replace the old one.
312 * If not, and we'd block, and they allow the subrange of the buffer to be
313 * invalidated, then we can make a new little BO, let them write into that,
314 * and blit it into the real BO at unmap time.
315 */
316 static void *
317 brw_map_buffer_range(struct gl_context *ctx,
318 GLintptr offset, GLsizeiptr length,
319 GLbitfield access, struct gl_buffer_object *obj,
320 gl_map_buffer_index index)
321 {
322 struct brw_context *brw = brw_context(ctx);
323 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
324
325 assert(intel_obj);
326
327 STATIC_ASSERT(GL_MAP_UNSYNCHRONIZED_BIT == MAP_ASYNC);
328 STATIC_ASSERT(GL_MAP_WRITE_BIT == MAP_WRITE);
329 STATIC_ASSERT(GL_MAP_READ_BIT == MAP_READ);
330 STATIC_ASSERT(GL_MAP_PERSISTENT_BIT == MAP_PERSISTENT);
331 STATIC_ASSERT(GL_MAP_COHERENT_BIT == MAP_COHERENT);
332 assert((access & MAP_INTERNAL_MASK) == 0);
333
334 /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
335 * internally uses our functions directly.
336 */
337 obj->Mappings[index].Offset = offset;
338 obj->Mappings[index].Length = length;
339 obj->Mappings[index].AccessFlags = access;
340
341 if (intel_obj->buffer == NULL) {
342 obj->Mappings[index].Pointer = NULL;
343 return NULL;
344 }
345
346 /* If the access is synchronized (like a normal buffer mapping), then get
347 * things flushed out so the later mapping syncs appropriately through GEM.
348 * If the user doesn't care about existing buffer contents and mapping would
349 * cause us to block, then throw out the old buffer.
350 *
351 * If they set INVALIDATE_BUFFER, we can pitch the current contents to
352 * achieve the required synchronization.
353 */
354 if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
355 if (brw_batch_references(&brw->batch, intel_obj->buffer)) {
356 if (access & GL_MAP_INVALIDATE_BUFFER_BIT) {
357 brw_bo_unreference(intel_obj->buffer);
358 alloc_buffer_object(brw, intel_obj);
359 } else {
360 perf_debug("Stalling on the GPU for mapping a busy buffer "
361 "object\n");
362 intel_batchbuffer_flush(brw);
363 }
364 } else if (brw_bo_busy(intel_obj->buffer) &&
365 (access & GL_MAP_INVALIDATE_BUFFER_BIT)) {
366 brw_bo_unreference(intel_obj->buffer);
367 alloc_buffer_object(brw, intel_obj);
368 }
369 }
370
371 /* If the user is mapping a range of an active buffer object but
372 * doesn't require the current contents of that range, make a new
373 * BO, and we'll copy what they put in there out at unmap or
374 * FlushRange time.
375 *
376 * That is, unless they're looking for a persistent mapping -- we would
377 * need to do blits in the MemoryBarrier call, and it's easier to just do a
378 * GPU stall and do a mapping.
379 */
380 if (!(access & (GL_MAP_UNSYNCHRONIZED_BIT | GL_MAP_PERSISTENT_BIT)) &&
381 (access & GL_MAP_INVALIDATE_RANGE_BIT) &&
382 brw_bo_busy(intel_obj->buffer)) {
383 /* Ensure that the base alignment of the allocation meets the alignment
384 * guarantees the driver has advertised to the application.
385 */
386 const unsigned alignment = ctx->Const.MinMapBufferAlignment;
387
388 intel_obj->map_extra[index] = (uintptr_t) offset % alignment;
389 intel_obj->range_map_bo[index] = brw_bo_alloc(brw->bufmgr,
390 "BO blit temp",
391 length +
392 intel_obj->map_extra[index],
393 alignment);
394 void *map = brw_bo_map(brw, intel_obj->range_map_bo[index], access);
395 obj->Mappings[index].Pointer = map + intel_obj->map_extra[index];
396 return obj->Mappings[index].Pointer;
397 }
398
399 void *map = brw_bo_map(brw, intel_obj->buffer, access);
400 if (!(access & GL_MAP_UNSYNCHRONIZED_BIT)) {
401 mark_buffer_inactive(intel_obj);
402 }
403
404 obj->Mappings[index].Pointer = map + offset;
405 return obj->Mappings[index].Pointer;
406 }
407
408 /**
409 * The FlushMappedBufferRange() driver hook.
410 *
411 * Implements glFlushMappedBufferRange(), which signifies that modifications
412 * have been made to a range of a mapped buffer, and it should be flushed.
413 *
414 * This is only used for buffers mapped with GL_MAP_FLUSH_EXPLICIT_BIT.
415 *
416 * Ideally we'd use a BO to avoid taking up cache space for the temporary
417 * data, but FlushMappedBufferRange may be followed by further writes to
418 * the pointer, so we would have to re-map after emitting our blit, which
419 * would defeat the point.
420 */
421 static void
422 brw_flush_mapped_buffer_range(struct gl_context *ctx,
423 GLintptr offset, GLsizeiptr length,
424 struct gl_buffer_object *obj,
425 gl_map_buffer_index index)
426 {
427 struct brw_context *brw = brw_context(ctx);
428 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
429
430 assert(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT);
431
432 /* If we gave a direct mapping of the buffer instead of using a temporary,
433 * then there's nothing to do.
434 */
435 if (intel_obj->range_map_bo[index] == NULL)
436 return;
437
438 if (length == 0)
439 return;
440
441 /* Note that we're not unmapping our buffer while executing the blit. We
442 * need to have a mapping still at the end of this call, since the user
443 * gets to make further modifications and glFlushMappedBufferRange() calls.
444 * This is safe, because:
445 *
446 * - On LLC platforms, we're using a CPU mapping that's coherent with the
447 * GPU (except for the render caches), so the kernel doesn't need to do
448 * any flushing work for us except for what happens at batch exec time
449 * anyway.
450 *
451 * - On non-LLC platforms, we're using a GTT mapping that writes directly
452 * to system memory (except for the chipset cache that gets flushed at
453 * batch exec time).
454 *
455 * In both cases we don't need to stall for the previous blit to complete
456 * so we can re-map (and we definitely don't want to, since that would be
457 * slow): If the user edits a part of their buffer that's previously been
458 * blitted, then our lack of synchoronization is fine, because either
459 * they'll get some too-new data in the first blit and not do another blit
460 * of that area (but in that case the results are undefined), or they'll do
461 * another blit of that area and the complete newer data will land the
462 * second time.
463 */
464 intel_emit_linear_blit(brw,
465 intel_obj->buffer,
466 obj->Mappings[index].Offset + offset,
467 intel_obj->range_map_bo[index],
468 intel_obj->map_extra[index] + offset,
469 length);
470 mark_buffer_gpu_usage(intel_obj,
471 obj->Mappings[index].Offset + offset,
472 length);
473 }
474
475
476 /**
477 * The UnmapBuffer() driver hook.
478 *
479 * Implements glUnmapBuffer().
480 */
481 static GLboolean
482 brw_unmap_buffer(struct gl_context *ctx,
483 struct gl_buffer_object *obj,
484 gl_map_buffer_index index)
485 {
486 struct brw_context *brw = brw_context(ctx);
487 struct intel_buffer_object *intel_obj = intel_buffer_object(obj);
488
489 assert(intel_obj);
490 assert(obj->Mappings[index].Pointer);
491 if (intel_obj->range_map_bo[index] != NULL) {
492 brw_bo_unmap(intel_obj->range_map_bo[index]);
493
494 if (!(obj->Mappings[index].AccessFlags & GL_MAP_FLUSH_EXPLICIT_BIT)) {
495 intel_emit_linear_blit(brw,
496 intel_obj->buffer, obj->Mappings[index].Offset,
497 intel_obj->range_map_bo[index],
498 intel_obj->map_extra[index],
499 obj->Mappings[index].Length);
500 mark_buffer_gpu_usage(intel_obj, obj->Mappings[index].Offset,
501 obj->Mappings[index].Length);
502 }
503
504 /* Since we've emitted some blits to buffers that will (likely) be used
505 * in rendering operations in other cache domains in this batch, emit a
506 * flush. Once again, we wish for a domain tracker in libdrm to cover
507 * usage inside of a batchbuffer.
508 */
509 brw_emit_mi_flush(brw);
510
511 brw_bo_unreference(intel_obj->range_map_bo[index]);
512 intel_obj->range_map_bo[index] = NULL;
513 } else if (intel_obj->buffer != NULL) {
514 brw_bo_unmap(intel_obj->buffer);
515 }
516 obj->Mappings[index].Pointer = NULL;
517 obj->Mappings[index].Offset = 0;
518 obj->Mappings[index].Length = 0;
519
520 return true;
521 }
522
523 /**
524 * Gets a pointer to the object's BO, and marks the given range as being used
525 * on the GPU.
526 *
527 * Anywhere that uses buffer objects in the pipeline should be using this to
528 * mark the range of the buffer that is being accessed by the pipeline.
529 */
530 struct brw_bo *
531 intel_bufferobj_buffer(struct brw_context *brw,
532 struct intel_buffer_object *intel_obj,
533 uint32_t offset, uint32_t size)
534 {
535 /* This is needed so that things like transform feedback and texture buffer
536 * objects that need a BO but don't want to check that they exist for
537 * draw-time validation can just always get a BO from a GL buffer object.
538 */
539 if (intel_obj->buffer == NULL)
540 alloc_buffer_object(brw, intel_obj);
541
542 mark_buffer_gpu_usage(intel_obj, offset, size);
543
544 return intel_obj->buffer;
545 }
546
547 /**
548 * The CopyBufferSubData() driver hook.
549 *
550 * Implements glCopyBufferSubData(), which copies a portion of one buffer
551 * object's data to another. Independent source and destination offsets
552 * are allowed.
553 */
554 static void
555 brw_copy_buffer_subdata(struct gl_context *ctx,
556 struct gl_buffer_object *src,
557 struct gl_buffer_object *dst,
558 GLintptr read_offset, GLintptr write_offset,
559 GLsizeiptr size)
560 {
561 struct brw_context *brw = brw_context(ctx);
562 struct intel_buffer_object *intel_src = intel_buffer_object(src);
563 struct intel_buffer_object *intel_dst = intel_buffer_object(dst);
564 struct brw_bo *src_bo, *dst_bo;
565
566 if (size == 0)
567 return;
568
569 dst_bo = intel_bufferobj_buffer(brw, intel_dst, write_offset, size);
570 src_bo = intel_bufferobj_buffer(brw, intel_src, read_offset, size);
571
572 intel_emit_linear_blit(brw,
573 dst_bo, write_offset,
574 src_bo, read_offset, size);
575
576 /* Since we've emitted some blits to buffers that will (likely) be used
577 * in rendering operations in other cache domains in this batch, emit a
578 * flush. Once again, we wish for a domain tracker in libdrm to cover
579 * usage inside of a batchbuffer.
580 */
581 brw_emit_mi_flush(brw);
582 }
583
584 void
585 intelInitBufferObjectFuncs(struct dd_function_table *functions)
586 {
587 functions->NewBufferObject = brw_new_buffer_object;
588 functions->DeleteBuffer = brw_delete_buffer;
589 functions->BufferData = brw_buffer_data;
590 functions->BufferSubData = brw_buffer_subdata;
591 functions->GetBufferSubData = brw_get_buffer_subdata;
592 functions->MapBufferRange = brw_map_buffer_range;
593 functions->FlushMappedBufferRange = brw_flush_mapped_buffer_range;
594 functions->UnmapBuffer = brw_unmap_buffer;
595 functions->CopyBufferSubData = brw_copy_buffer_subdata;
596 }