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. */
214 dri_bo_subdata(intel_obj
->buffer
, offset
, size
, data
);
220 * Called via glGetBufferSubDataARB().
223 intel_bufferobj_get_subdata(GLcontext
* ctx
,
227 GLvoid
* data
, struct gl_buffer_object
*obj
)
229 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
232 dri_bo_get_subdata(intel_obj
->buffer
, offset
, size
, data
);
238 * Called via glMapBufferARB().
241 intel_bufferobj_map(GLcontext
* ctx
,
243 GLenum access
, struct gl_buffer_object
*obj
)
245 struct intel_context
*intel
= intel_context(ctx
);
246 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
247 GLboolean read_only
= (access
== GL_READ_ONLY_ARB
);
248 GLboolean write_only
= (access
== GL_WRITE_ONLY_ARB
);
252 if (intel_obj
->sys_buffer
) {
253 obj
->Pointer
= intel_obj
->sys_buffer
;
257 /* Flush any existing batchbuffer that might have written to this
262 if (intel_obj
->region
)
263 intel_bufferobj_cow(intel
, intel_obj
);
265 if (intel_obj
->buffer
== NULL
) {
270 if (write_only
&& intel
->intelScreen
->kernel_exec_fencing
) {
271 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
272 intel_obj
->mapped_gtt
= GL_TRUE
;
274 drm_intel_bo_map(intel_obj
->buffer
, !read_only
);
275 intel_obj
->mapped_gtt
= GL_FALSE
;
278 obj
->Pointer
= intel_obj
->buffer
->virtual;
279 obj
->Length
= obj
->Size
;
286 * Called via glMapBufferRange().
288 * The goal of this extension is to allow apps to accumulate their rendering
289 * at the same time as they accumulate their buffer object. Without it,
290 * you'd end up blocking on execution of rendering every time you mapped
291 * the buffer to put new data in.
293 * We support it in 3 ways: If unsynchronized, then don't bother
294 * flushing the batchbuffer before mapping the buffer, which can save blocking
295 * in many cases. If we would still block, and they allow the whole buffer
296 * to be invalidated, then just allocate a new buffer to replace the old one.
297 * If not, and we'd block, and they allow the subrange of the buffer to be
298 * invalidated, then we can make a new little BO, let them write into that,
299 * and blit it into the real BO at unmap time.
302 intel_bufferobj_map_range(GLcontext
* ctx
,
303 GLenum target
, GLintptr offset
, GLsizeiptr length
,
304 GLbitfield access
, struct gl_buffer_object
*obj
)
306 struct intel_context
*intel
= intel_context(ctx
);
307 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
311 /* _mesa_MapBufferRange (GL entrypoint) sets these, but the vbo module also
312 * internally uses our functions directly.
314 obj
->Offset
= offset
;
315 obj
->Length
= length
;
316 obj
->AccessFlags
= access
;
318 if (intel_obj
->sys_buffer
) {
319 obj
->Pointer
= intel_obj
->sys_buffer
+ offset
;
323 if (intel_obj
->region
)
324 intel_bufferobj_cow(intel
, intel_obj
);
326 /* If the mapping is synchronized with other GL operations, flush
327 * the batchbuffer so that GEM knows about the buffer access for later
330 if (!(access
& GL_MAP_UNSYNCHRONIZED_BIT
))
333 if (intel_obj
->buffer
== NULL
) {
338 /* If the user doesn't care about existing buffer contents and mapping
339 * would cause us to block, then throw out the old buffer.
341 if (!(access
& GL_MAP_UNSYNCHRONIZED_BIT
) &&
342 (access
& GL_MAP_INVALIDATE_BUFFER_BIT
) &&
343 drm_intel_bo_busy(intel_obj
->buffer
)) {
344 drm_intel_bo_unreference(intel_obj
->buffer
);
345 intel_obj
->buffer
= dri_bo_alloc(intel
->bufmgr
, "bufferobj",
346 intel_obj
->Base
.Size
, 64);
349 /* If the user is mapping a range of an active buffer object but
350 * doesn't require the current contents of that range, make a new
351 * BO, and we'll copy what they put in there out at unmap or
354 if ((access
& GL_MAP_INVALIDATE_RANGE_BIT
) &&
355 drm_intel_bo_busy(intel_obj
->buffer
)) {
356 if (access
& GL_MAP_FLUSH_EXPLICIT_BIT
) {
357 intel_obj
->range_map_buffer
= _mesa_malloc(length
);
358 obj
->Pointer
= intel_obj
->range_map_buffer
;
360 intel_obj
->range_map_bo
= drm_intel_bo_alloc(intel
->bufmgr
,
363 if (!(access
& GL_MAP_READ_BIT
) &&
364 intel
->intelScreen
->kernel_exec_fencing
) {
365 drm_intel_gem_bo_map_gtt(intel_obj
->range_map_bo
);
366 intel_obj
->mapped_gtt
= GL_TRUE
;
368 drm_intel_bo_map(intel_obj
->range_map_bo
,
369 (access
& GL_MAP_WRITE_BIT
) != 0);
370 intel_obj
->mapped_gtt
= GL_FALSE
;
372 obj
->Pointer
= intel_obj
->range_map_bo
->virtual;
377 if (!(access
& GL_MAP_READ_BIT
) &&
378 intel
->intelScreen
->kernel_exec_fencing
) {
379 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
380 intel_obj
->mapped_gtt
= GL_TRUE
;
382 drm_intel_bo_map(intel_obj
->buffer
, (access
& GL_MAP_WRITE_BIT
) != 0);
383 intel_obj
->mapped_gtt
= GL_FALSE
;
386 obj
->Pointer
= intel_obj
->buffer
->virtual + offset
;
390 /* Ideally we'd use a BO to avoid taking up cache space for the temporary
391 * data, but FlushMappedBufferRange may be followed by further writes to
392 * the pointer, so we would have to re-map after emitting our blit, which
393 * would defeat the point.
396 intel_bufferobj_flush_mapped_range(GLcontext
*ctx
, GLenum target
,
397 GLintptr offset
, GLsizeiptr length
,
398 struct gl_buffer_object
*obj
)
400 struct intel_context
*intel
= intel_context(ctx
);
401 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
402 drm_intel_bo
*temp_bo
;
404 /* Unless we're in the range map using a temporary system buffer,
405 * there's no work to do.
407 if (intel_obj
->range_map_buffer
== NULL
)
410 temp_bo
= drm_intel_bo_alloc(intel
->bufmgr
, "range map flush", length
, 64);
412 drm_intel_bo_subdata(temp_bo
, 0, length
, intel_obj
->range_map_buffer
);
414 intel_emit_linear_blit(intel
,
415 intel_obj
->buffer
, obj
->Offset
+ offset
,
419 drm_intel_bo_unreference(temp_bo
);
424 * Called via glUnmapBuffer().
427 intel_bufferobj_unmap(GLcontext
* ctx
,
428 GLenum target
, struct gl_buffer_object
*obj
)
430 struct intel_context
*intel
= intel_context(ctx
);
431 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
434 assert(obj
->Pointer
);
435 if (intel_obj
->sys_buffer
!= NULL
) {
436 /* always keep the mapping around. */
437 } else if (intel_obj
->range_map_buffer
!= NULL
) {
438 /* Since we've emitted some blits to buffers that will (likely) be used
439 * in rendering operations in other cache domains in this batch, emit a
440 * flush. Once again, we wish for a domain tracker in libdrm to cover
441 * usage inside of a batchbuffer.
443 intel_batchbuffer_emit_mi_flush(intel
->batch
);
444 free(intel_obj
->range_map_buffer
);
445 intel_obj
->range_map_buffer
= NULL
;
446 } else if (intel_obj
->range_map_bo
!= NULL
) {
447 if (intel_obj
->mapped_gtt
) {
448 drm_intel_gem_bo_unmap_gtt(intel_obj
->range_map_bo
);
450 drm_intel_bo_unmap(intel_obj
->range_map_bo
);
453 intel_emit_linear_blit(intel
,
454 intel_obj
->buffer
, obj
->Offset
,
455 intel_obj
->range_map_bo
, 0,
458 /* Since we've emitted some blits to buffers that will (likely) be used
459 * in rendering operations in other cache domains in this batch, emit a
460 * flush. Once again, we wish for a domain tracker in libdrm to cover
461 * usage inside of a batchbuffer.
463 intel_batchbuffer_emit_mi_flush(intel
->batch
);
465 drm_intel_bo_unreference(intel_obj
->range_map_bo
);
466 intel_obj
->range_map_bo
= NULL
;
467 } else if (intel_obj
->buffer
!= NULL
) {
468 if (intel_obj
->mapped_gtt
) {
469 drm_intel_gem_bo_unmap_gtt(intel_obj
->buffer
);
471 drm_intel_bo_unmap(intel_obj
->buffer
);
482 intel_bufferobj_buffer(struct intel_context
*intel
,
483 struct intel_buffer_object
*intel_obj
, GLuint flag
)
485 if (intel_obj
->region
) {
486 if (flag
== INTEL_WRITE_PART
)
487 intel_bufferobj_cow(intel
, intel_obj
);
488 else if (flag
== INTEL_WRITE_FULL
) {
489 intel_bufferobj_release_region(intel
, intel_obj
);
490 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
494 if (intel_obj
->buffer
== NULL
) {
495 void *sys_buffer
= intel_obj
->sys_buffer
;
497 /* only one of buffer and sys_buffer could be non-NULL */
498 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
499 intel_obj
->sys_buffer
= NULL
;
501 intel_bufferobj_subdata(&intel
->ctx
,
504 intel_obj
->Base
.Size
,
507 _mesa_free(sys_buffer
);
508 intel_obj
->sys_buffer
= NULL
;
511 return intel_obj
->buffer
;
515 intel_bufferobj_copy_subdata(GLcontext
*ctx
,
516 struct gl_buffer_object
*src
,
517 struct gl_buffer_object
*dst
,
518 GLintptr read_offset
, GLintptr write_offset
,
521 struct intel_context
*intel
= intel_context(ctx
);
522 struct intel_buffer_object
*intel_src
= intel_buffer_object(src
);
523 struct intel_buffer_object
*intel_dst
= intel_buffer_object(dst
);
524 drm_intel_bo
*src_bo
, *dst_bo
;
529 /* If we're in system memory, just map and memcpy. */
530 if (intel_src
->sys_buffer
|| intel_dst
->sys_buffer
) {
531 /* The same buffer may be used, but note that regions copied may
535 char *ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
537 memcpy(ptr
+ write_offset
, ptr
+ read_offset
, size
);
538 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
543 src_ptr
= intel_bufferobj_map(ctx
, GL_COPY_READ_BUFFER
,
545 dst_ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
548 memcpy(dst_ptr
+ write_offset
, src_ptr
+ read_offset
, size
);
550 intel_bufferobj_unmap(ctx
, GL_COPY_READ_BUFFER
, src
);
551 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
555 /* Otherwise, we have real BOs, so blit them. */
557 dst_bo
= intel_bufferobj_buffer(intel
, intel_dst
, INTEL_WRITE_PART
);
558 src_bo
= intel_bufferobj_buffer(intel
, intel_src
, INTEL_READ
);
560 intel_emit_linear_blit(intel
,
561 dst_bo
, write_offset
,
562 src_bo
, read_offset
, size
);
564 /* Since we've emitted some blits to buffers that will (likely) be used
565 * in rendering operations in other cache domains in this batch, emit a
566 * flush. Once again, we wish for a domain tracker in libdrm to cover
567 * usage inside of a batchbuffer.
569 intel_batchbuffer_emit_mi_flush(intel
->batch
);
573 intelInitBufferObjectFuncs(struct dd_function_table
*functions
)
575 functions
->NewBufferObject
= intel_bufferobj_alloc
;
576 functions
->DeleteBuffer
= intel_bufferobj_free
;
577 functions
->BufferData
= intel_bufferobj_data
;
578 functions
->BufferSubData
= intel_bufferobj_subdata
;
579 functions
->GetBufferSubData
= intel_bufferobj_get_subdata
;
580 functions
->MapBuffer
= intel_bufferobj_map
;
581 functions
->MapBufferRange
= intel_bufferobj_map_range
;
582 functions
->FlushMappedBufferRange
= intel_bufferobj_flush_mapped_range
;
583 functions
->UnmapBuffer
= intel_bufferobj_unmap
;
584 functions
->CopyBufferSubData
= intel_bufferobj_copy_subdata
;