1 /**************************************************************************
3 * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
29 #include "main/imports.h"
30 #include "main/mtypes.h"
31 #include "main/macros.h"
32 #include "main/bufferobj.h"
34 #include "intel_context.h"
35 #include "intel_blit.h"
36 #include "intel_buffer_objects.h"
37 #include "intel_batchbuffer.h"
38 #include "intel_regions.h"
41 intel_bufferobj_unmap(GLcontext
* ctx
,
42 GLenum target
, struct gl_buffer_object
*obj
);
44 /** Allocates a new dri_bo to store the data for the buffer object. */
46 intel_bufferobj_alloc_buffer(struct intel_context
*intel
,
47 struct intel_buffer_object
*intel_obj
)
49 intel_obj
->buffer
= dri_bo_alloc(intel
->bufmgr
, "bufferobj",
50 intel_obj
->Base
.Size
, 64);
54 * There is some duplication between mesa's bufferobjects and our
55 * bufmgr buffers. Both have an integer handle and a hashtable to
56 * lookup an opaque structure. It would be nice if the handles and
57 * internal structure where somehow shared.
59 static struct gl_buffer_object
*
60 intel_bufferobj_alloc(GLcontext
* ctx
, GLuint name
, GLenum target
)
62 struct intel_buffer_object
*obj
= CALLOC_STRUCT(intel_buffer_object
);
64 _mesa_initialize_buffer_object(&obj
->Base
, name
, target
);
71 /* Break the COW tie to the region. The region gets to keep the data.
74 intel_bufferobj_release_region(struct intel_context
*intel
,
75 struct intel_buffer_object
*intel_obj
)
77 assert(intel_obj
->region
->buffer
== intel_obj
->buffer
);
78 intel_obj
->region
->pbo
= NULL
;
79 intel_obj
->region
= NULL
;
81 dri_bo_unreference(intel_obj
->buffer
);
82 intel_obj
->buffer
= NULL
;
85 /* Break the COW tie to the region. Both the pbo and the region end
86 * up with a copy of the data.
89 intel_bufferobj_cow(struct intel_context
*intel
,
90 struct intel_buffer_object
*intel_obj
)
92 assert(intel_obj
->region
);
93 intel_region_cow(intel
, intel_obj
->region
);
98 * Deallocate/free a vertex/pixel buffer object.
99 * Called via glDeleteBuffersARB().
102 intel_bufferobj_free(GLcontext
* ctx
, struct gl_buffer_object
*obj
)
104 struct intel_context
*intel
= intel_context(ctx
);
105 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
109 /* Buffer objects are automatically unmapped when deleting according
110 * to the spec, but Mesa doesn't do UnmapBuffer for us at context destroy
111 * (though it does if you call glDeleteBuffers)
114 intel_bufferobj_unmap(ctx
, 0, obj
);
116 _mesa_free(intel_obj
->sys_buffer
);
117 if (intel_obj
->region
) {
118 intel_bufferobj_release_region(intel
, intel_obj
);
120 else if (intel_obj
->buffer
) {
121 dri_bo_unreference(intel_obj
->buffer
);
124 _mesa_free(intel_obj
);
130 * Allocate space for and store data in a buffer object. Any data that was
131 * previously stored in the buffer object is lost. If data is NULL,
132 * memory will be allocated, but no copy will occur.
133 * Called via ctx->Driver.BufferData().
134 * \return GL_TRUE for success, GL_FALSE if out of memory
137 intel_bufferobj_data(GLcontext
* ctx
,
141 GLenum usage
, struct gl_buffer_object
*obj
)
143 struct intel_context
*intel
= intel_context(ctx
);
144 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
146 intel_obj
->Base
.Size
= size
;
147 intel_obj
->Base
.Usage
= usage
;
149 assert(!obj
->Pointer
); /* Mesa should have unmapped it */
151 if (intel_obj
->region
)
152 intel_bufferobj_release_region(intel
, intel_obj
);
154 if (intel_obj
->buffer
!= NULL
) {
155 dri_bo_unreference(intel_obj
->buffer
);
156 intel_obj
->buffer
= NULL
;
158 _mesa_free(intel_obj
->sys_buffer
);
159 intel_obj
->sys_buffer
= NULL
;
163 /* On pre-965, stick VBOs in system memory, as we're always doing swtnl
164 * with their contents anyway.
166 if (target
== GL_ARRAY_BUFFER
|| target
== GL_ELEMENT_ARRAY_BUFFER
) {
167 intel_obj
->sys_buffer
= _mesa_malloc(size
);
168 if (intel_obj
->sys_buffer
!= NULL
) {
170 memcpy(intel_obj
->sys_buffer
, data
, size
);
175 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
176 if (!intel_obj
->buffer
)
180 dri_bo_subdata(intel_obj
->buffer
, 0, size
, data
);
188 * Replace data in a subrange of buffer object. If the data range
189 * specified by size + offset extends beyond the end of the buffer or
190 * if data is NULL, no copy is performed.
191 * Called via glBufferSubDataARB().
194 intel_bufferobj_subdata(GLcontext
* ctx
,
198 const GLvoid
* data
, struct gl_buffer_object
*obj
)
200 struct intel_context
*intel
= intel_context(ctx
);
201 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
205 if (intel_obj
->region
)
206 intel_bufferobj_cow(intel
, intel_obj
);
208 if (intel_obj
->sys_buffer
)
209 memcpy((char *)intel_obj
->sys_buffer
+ offset
, data
, size
);
211 /* Flush any existing batchbuffer that might reference this data. */
212 if (drm_intel_bo_references(intel
->batch
->buf
, intel_obj
->buffer
))
215 dri_bo_subdata(intel_obj
->buffer
, offset
, size
, data
);
221 * Called via glGetBufferSubDataARB().
224 intel_bufferobj_get_subdata(GLcontext
* ctx
,
228 GLvoid
* data
, struct gl_buffer_object
*obj
)
230 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
233 if (intel_obj
->sys_buffer
)
234 memcpy(data
, (char *)intel_obj
->sys_buffer
+ offset
, size
);
236 dri_bo_get_subdata(intel_obj
->buffer
, offset
, size
, data
);
242 * Called via glMapBufferARB().
245 intel_bufferobj_map(GLcontext
* ctx
,
247 GLenum access
, struct gl_buffer_object
*obj
)
249 struct intel_context
*intel
= intel_context(ctx
);
250 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
251 GLboolean read_only
= (access
== GL_READ_ONLY_ARB
);
252 GLboolean write_only
= (access
== GL_WRITE_ONLY_ARB
);
256 if (intel_obj
->sys_buffer
) {
257 obj
->Pointer
= intel_obj
->sys_buffer
;
261 /* Flush any existing batchbuffer that might reference this data. */
262 if (drm_intel_bo_references(intel
->batch
->buf
, intel_obj
->buffer
))
265 if (intel_obj
->region
)
266 intel_bufferobj_cow(intel
, intel_obj
);
268 if (intel_obj
->buffer
== NULL
) {
273 if (write_only
&& intel
->intelScreen
->kernel_exec_fencing
) {
274 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
275 intel_obj
->mapped_gtt
= GL_TRUE
;
277 drm_intel_bo_map(intel_obj
->buffer
, !read_only
);
278 intel_obj
->mapped_gtt
= GL_FALSE
;
281 obj
->Pointer
= intel_obj
->buffer
->virtual;
282 obj
->Length
= obj
->Size
;
289 * Called via glMapBufferRange().
291 * The goal of this extension is to allow apps to accumulate their rendering
292 * at the same time as they accumulate their buffer object. Without it,
293 * you'd end up blocking on execution of rendering every time you mapped
294 * the buffer to put new data in.
296 * We support it in 3 ways: If unsynchronized, then don't bother
297 * flushing the batchbuffer before mapping the buffer, which can save blocking
298 * in many cases. If we would still block, and they allow the whole buffer
299 * to be invalidated, then just allocate a new buffer to replace the old one.
300 * If not, and we'd block, and they allow the subrange of the buffer to be
301 * invalidated, then we can make a new little BO, let them write into that,
302 * and blit it into the real BO at unmap time.
305 intel_bufferobj_map_range(GLcontext
* ctx
,
306 GLenum target
, GLintptr offset
, GLsizeiptr length
,
307 GLbitfield access
, struct gl_buffer_object
*obj
)
309 struct intel_context
*intel
= intel_context(ctx
);
310 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
314 /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
315 * internally uses our functions directly.
317 obj
->Offset
= offset
;
318 obj
->Length
= length
;
319 obj
->AccessFlags
= access
;
321 if (intel_obj
->sys_buffer
) {
322 obj
->Pointer
= intel_obj
->sys_buffer
+ offset
;
326 if (intel_obj
->region
)
327 intel_bufferobj_cow(intel
, intel_obj
);
329 /* If the mapping is synchronized with other GL operations, flush
330 * the batchbuffer so that GEM knows about the buffer access for later
333 if (!(access
& GL_MAP_UNSYNCHRONIZED_BIT
) &&
334 drm_intel_bo_references(intel
->batch
->buf
, intel_obj
->buffer
))
337 if (intel_obj
->buffer
== NULL
) {
342 /* If the user doesn't care about existing buffer contents and mapping
343 * would cause us to block, then throw out the old buffer.
345 if (!(access
& GL_MAP_UNSYNCHRONIZED_BIT
) &&
346 (access
& GL_MAP_INVALIDATE_BUFFER_BIT
) &&
347 drm_intel_bo_busy(intel_obj
->buffer
)) {
348 drm_intel_bo_unreference(intel_obj
->buffer
);
349 intel_obj
->buffer
= dri_bo_alloc(intel
->bufmgr
, "bufferobj",
350 intel_obj
->Base
.Size
, 64);
353 /* If the user is mapping a range of an active buffer object but
354 * doesn't require the current contents of that range, make a new
355 * BO, and we'll copy what they put in there out at unmap or
358 if ((access
& GL_MAP_INVALIDATE_RANGE_BIT
) &&
359 drm_intel_bo_busy(intel_obj
->buffer
)) {
360 if (access
& GL_MAP_FLUSH_EXPLICIT_BIT
) {
361 intel_obj
->range_map_buffer
= _mesa_malloc(length
);
362 obj
->Pointer
= intel_obj
->range_map_buffer
;
364 intel_obj
->range_map_bo
= drm_intel_bo_alloc(intel
->bufmgr
,
367 if (!(access
& GL_MAP_READ_BIT
) &&
368 intel
->intelScreen
->kernel_exec_fencing
) {
369 drm_intel_gem_bo_map_gtt(intel_obj
->range_map_bo
);
370 intel_obj
->mapped_gtt
= GL_TRUE
;
372 drm_intel_bo_map(intel_obj
->range_map_bo
,
373 (access
& GL_MAP_WRITE_BIT
) != 0);
374 intel_obj
->mapped_gtt
= GL_FALSE
;
376 obj
->Pointer
= intel_obj
->range_map_bo
->virtual;
381 if (!(access
& GL_MAP_READ_BIT
) &&
382 intel
->intelScreen
->kernel_exec_fencing
) {
383 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
384 intel_obj
->mapped_gtt
= GL_TRUE
;
386 drm_intel_bo_map(intel_obj
->buffer
, (access
& GL_MAP_WRITE_BIT
) != 0);
387 intel_obj
->mapped_gtt
= GL_FALSE
;
390 obj
->Pointer
= intel_obj
->buffer
->virtual + offset
;
394 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
395 * data, but FlushMappedBufferRange may be followed by further writes to
396 * the pointer, so we would have to re-map after emitting our blit, which
397 * would defeat the point.
400 intel_bufferobj_flush_mapped_range(GLcontext
*ctx
, GLenum target
,
401 GLintptr offset
, GLsizeiptr length
,
402 struct gl_buffer_object
*obj
)
404 struct intel_context
*intel
= intel_context(ctx
);
405 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
406 drm_intel_bo
*temp_bo
;
408 /* Unless we're in the range map using a temporary system buffer,
409 * there's no work to do.
411 if (intel_obj
->range_map_buffer
== NULL
)
414 temp_bo
= drm_intel_bo_alloc(intel
->bufmgr
, "range map flush", length
, 64);
416 drm_intel_bo_subdata(temp_bo
, 0, length
, intel_obj
->range_map_buffer
);
418 intel_emit_linear_blit(intel
,
419 intel_obj
->buffer
, obj
->Offset
+ offset
,
423 drm_intel_bo_unreference(temp_bo
);
428 * Called via glUnmapBuffer().
431 intel_bufferobj_unmap(GLcontext
* ctx
,
432 GLenum target
, struct gl_buffer_object
*obj
)
434 struct intel_context
*intel
= intel_context(ctx
);
435 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
438 assert(obj
->Pointer
);
439 if (intel_obj
->sys_buffer
!= NULL
) {
440 /* always keep the mapping around. */
441 } else if (intel_obj
->range_map_buffer
!= NULL
) {
442 /* Since we've emitted some blits to buffers that will (likely) be used
443 * in rendering operations in other cache domains in this batch, emit a
444 * flush. Once again, we wish for a domain tracker in libdrm to cover
445 * usage inside of a batchbuffer.
447 intel_batchbuffer_emit_mi_flush(intel
->batch
);
448 free(intel_obj
->range_map_buffer
);
449 intel_obj
->range_map_buffer
= NULL
;
450 } else if (intel_obj
->range_map_bo
!= NULL
) {
451 if (intel_obj
->mapped_gtt
) {
452 drm_intel_gem_bo_unmap_gtt(intel_obj
->range_map_bo
);
454 drm_intel_bo_unmap(intel_obj
->range_map_bo
);
457 intel_emit_linear_blit(intel
,
458 intel_obj
->buffer
, obj
->Offset
,
459 intel_obj
->range_map_bo
, 0,
462 /* Since we've emitted some blits to buffers that will (likely) be used
463 * in rendering operations in other cache domains in this batch, emit a
464 * flush. Once again, we wish for a domain tracker in libdrm to cover
465 * usage inside of a batchbuffer.
467 intel_batchbuffer_emit_mi_flush(intel
->batch
);
469 drm_intel_bo_unreference(intel_obj
->range_map_bo
);
470 intel_obj
->range_map_bo
= NULL
;
471 } else if (intel_obj
->buffer
!= NULL
) {
472 if (intel_obj
->mapped_gtt
) {
473 drm_intel_gem_bo_unmap_gtt(intel_obj
->buffer
);
475 drm_intel_bo_unmap(intel_obj
->buffer
);
486 intel_bufferobj_buffer(struct intel_context
*intel
,
487 struct intel_buffer_object
*intel_obj
, GLuint flag
)
489 if (intel_obj
->region
) {
490 if (flag
== INTEL_WRITE_PART
)
491 intel_bufferobj_cow(intel
, intel_obj
);
492 else if (flag
== INTEL_WRITE_FULL
) {
493 intel_bufferobj_release_region(intel
, intel_obj
);
494 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
498 if (intel_obj
->buffer
== NULL
) {
499 void *sys_buffer
= intel_obj
->sys_buffer
;
501 /* only one of buffer and sys_buffer could be non-NULL */
502 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
503 intel_obj
->sys_buffer
= NULL
;
505 intel_bufferobj_subdata(&intel
->ctx
,
508 intel_obj
->Base
.Size
,
511 _mesa_free(sys_buffer
);
512 intel_obj
->sys_buffer
= NULL
;
515 return intel_obj
->buffer
;
519 intel_bufferobj_copy_subdata(GLcontext
*ctx
,
520 struct gl_buffer_object
*src
,
521 struct gl_buffer_object
*dst
,
522 GLintptr read_offset
, GLintptr write_offset
,
525 struct intel_context
*intel
= intel_context(ctx
);
526 struct intel_buffer_object
*intel_src
= intel_buffer_object(src
);
527 struct intel_buffer_object
*intel_dst
= intel_buffer_object(dst
);
528 drm_intel_bo
*src_bo
, *dst_bo
;
533 /* If we're in system memory, just map and memcpy. */
534 if (intel_src
->sys_buffer
|| intel_dst
->sys_buffer
) {
535 /* The same buffer may be used, but note that regions copied may
539 char *ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
541 memcpy(ptr
+ write_offset
, ptr
+ read_offset
, size
);
542 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
547 src_ptr
= intel_bufferobj_map(ctx
, GL_COPY_READ_BUFFER
,
549 dst_ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
552 memcpy(dst_ptr
+ write_offset
, src_ptr
+ read_offset
, size
);
554 intel_bufferobj_unmap(ctx
, GL_COPY_READ_BUFFER
, src
);
555 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
559 /* Otherwise, we have real BOs, so blit them. */
561 dst_bo
= intel_bufferobj_buffer(intel
, intel_dst
, INTEL_WRITE_PART
);
562 src_bo
= intel_bufferobj_buffer(intel
, intel_src
, INTEL_READ
);
564 intel_emit_linear_blit(intel
,
565 dst_bo
, write_offset
,
566 src_bo
, read_offset
, size
);
568 /* Since we've emitted some blits to buffers that will (likely) be used
569 * in rendering operations in other cache domains in this batch, emit a
570 * flush. Once again, we wish for a domain tracker in libdrm to cover
571 * usage inside of a batchbuffer.
573 intel_batchbuffer_emit_mi_flush(intel
->batch
);
577 intelInitBufferObjectFuncs(struct dd_function_table
*functions
)
579 functions
->NewBufferObject
= intel_bufferobj_alloc
;
580 functions
->DeleteBuffer
= intel_bufferobj_free
;
581 functions
->BufferData
= intel_bufferobj_data
;
582 functions
->BufferSubData
= intel_bufferobj_subdata
;
583 functions
->GetBufferSubData
= intel_bufferobj_get_subdata
;
584 functions
->MapBuffer
= intel_bufferobj_map
;
585 functions
->MapBufferRange
= intel_bufferobj_map_range
;
586 functions
->FlushMappedBufferRange
= intel_bufferobj_flush_mapped_range
;
587 functions
->UnmapBuffer
= intel_bufferobj_unmap
;
588 functions
->CopyBufferSubData
= intel_bufferobj_copy_subdata
;