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 glBufferDataARB().
136 intel_bufferobj_data(GLcontext
* ctx
,
140 GLenum usage
, struct gl_buffer_object
*obj
)
142 struct intel_context
*intel
= intel_context(ctx
);
143 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
145 intel_obj
->Base
.Size
= size
;
146 intel_obj
->Base
.Usage
= usage
;
148 assert(!obj
->Pointer
); /* Mesa should have unmapped it */
150 if (intel_obj
->region
)
151 intel_bufferobj_release_region(intel
, intel_obj
);
153 if (intel_obj
->buffer
!= NULL
) {
154 dri_bo_unreference(intel_obj
->buffer
);
155 intel_obj
->buffer
= NULL
;
157 _mesa_free(intel_obj
->sys_buffer
);
158 intel_obj
->sys_buffer
= NULL
;
162 /* On pre-965, stick VBOs in system memory, as we're always doing swtnl
163 * with their contents anyway.
165 if (target
== GL_ARRAY_BUFFER
|| target
== GL_ELEMENT_ARRAY_BUFFER
) {
166 intel_obj
->sys_buffer
= _mesa_malloc(size
);
167 if (intel_obj
->sys_buffer
!= NULL
) {
169 memcpy(intel_obj
->sys_buffer
, data
, size
);
174 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
177 dri_bo_subdata(intel_obj
->buffer
, 0, size
, data
);
183 * Replace data in a subrange of buffer object. If the data range
184 * specified by size + offset extends beyond the end of the buffer or
185 * if data is NULL, no copy is performed.
186 * Called via glBufferSubDataARB().
189 intel_bufferobj_subdata(GLcontext
* ctx
,
193 const GLvoid
* data
, struct gl_buffer_object
*obj
)
195 struct intel_context
*intel
= intel_context(ctx
);
196 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
200 if (intel_obj
->region
)
201 intel_bufferobj_cow(intel
, intel_obj
);
203 if (intel_obj
->sys_buffer
)
204 memcpy((char *)intel_obj
->sys_buffer
+ offset
, data
, size
);
206 dri_bo_subdata(intel_obj
->buffer
, offset
, size
, data
);
211 * Called via glGetBufferSubDataARB().
214 intel_bufferobj_get_subdata(GLcontext
* ctx
,
218 GLvoid
* data
, struct gl_buffer_object
*obj
)
220 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
223 dri_bo_get_subdata(intel_obj
->buffer
, offset
, size
, data
);
229 * Called via glMapBufferARB().
232 intel_bufferobj_map(GLcontext
* ctx
,
234 GLenum access
, struct gl_buffer_object
*obj
)
236 struct intel_context
*intel
= intel_context(ctx
);
237 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
238 GLboolean read_only
= (access
== GL_READ_ONLY_ARB
);
239 GLboolean write_only
= (access
== GL_WRITE_ONLY_ARB
);
243 if (intel_obj
->sys_buffer
) {
244 obj
->Pointer
= intel_obj
->sys_buffer
;
248 /* Flush any existing batchbuffer that might have written to this
253 if (intel_obj
->region
)
254 intel_bufferobj_cow(intel
, intel_obj
);
256 if (intel_obj
->buffer
== NULL
) {
261 if (write_only
&& intel
->intelScreen
->kernel_exec_fencing
) {
262 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
263 intel_obj
->mapped_gtt
= GL_TRUE
;
265 drm_intel_bo_map(intel_obj
->buffer
, !read_only
);
266 intel_obj
->mapped_gtt
= GL_FALSE
;
269 obj
->Pointer
= intel_obj
->buffer
->virtual;
274 * Called via glMapBufferRange().
276 * The goal of this extension is to allow apps to accumulate their rendering
277 * at the same time as they accumulate their buffer object. Without it,
278 * you'd end up blocking on execution of rendering every time you mapped
279 * the buffer to put new data in.
281 * We support it in 3 ways: If unsynchronized, then don't bother
282 * flushing the batchbuffer before mapping the buffer, which can save blocking
283 * in many cases. If we would still block, and they allow the whole buffer
284 * to be invalidated, then just allocate a new buffer to replace the old one.
285 * If not, and we'd block, and they allow the subrange of the buffer to be
286 * invalidated, then we can make a new little BO, let them write into that,
287 * and blit it into the real BO at unmap time.
290 intel_bufferobj_map_range(GLcontext
* ctx
,
291 GLenum target
, GLsizei offset
, GLsizeiptr length
,
292 GLbitfield access
, struct gl_buffer_object
*obj
)
294 struct intel_context
*intel
= intel_context(ctx
);
295 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
299 if (intel_obj
->sys_buffer
) {
300 obj
->Pointer
= intel_obj
->sys_buffer
+ offset
;
304 if (intel_obj
->region
)
305 intel_bufferobj_cow(intel
, intel_obj
);
307 /* If the mapping is synchronized with other GL operations, flush
308 * the batchbuffer so that GEM knows about the buffer access for later
311 if ((access
& GL_MAP_WRITE_BIT
) && !(access
& GL_MAP_UNSYNCHRONIZED_BIT
))
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
->buffer
== NULL
) {
326 /* If the user doesn't care about existing buffer contents and mapping
327 * would cause us to block, then throw out the old buffer.
329 if (!(access
& GL_MAP_UNSYNCHRONIZED_BIT
) &&
330 (access
& GL_MAP_INVALIDATE_BUFFER_BIT
) &&
331 drm_intel_bo_busy(intel_obj
->buffer
)) {
332 drm_intel_bo_unreference(intel_obj
->buffer
);
333 intel_obj
->buffer
= dri_bo_alloc(intel
->bufmgr
, "bufferobj",
334 intel_obj
->Base
.Size
, 64);
337 /* If the user is mapping a range of an active buffer object but
338 * doesn't require the current contents of that range, make a new
339 * BO, and we'll copy what they put in there out at unmap or
342 if ((access
& GL_MAP_INVALIDATE_RANGE_BIT
) &&
343 drm_intel_bo_busy(intel_obj
->buffer
)) {
344 intel_obj
->range_map_bo
= drm_intel_bo_alloc(intel
->bufmgr
,
347 if (!(access
& GL_MAP_READ_BIT
) &&
348 intel
->intelScreen
->kernel_exec_fencing
) {
349 drm_intel_gem_bo_map_gtt(intel_obj
->range_map_bo
);
350 intel_obj
->mapped_gtt
= GL_TRUE
;
352 drm_intel_bo_map(intel_obj
->range_map_bo
,
353 (access
& GL_MAP_WRITE_BIT
) != 0);
354 intel_obj
->mapped_gtt
= GL_FALSE
;
356 obj
->Pointer
= intel_obj
->range_map_bo
->virtual;
360 if (!(access
& GL_MAP_READ_BIT
) &&
361 intel
->intelScreen
->kernel_exec_fencing
) {
362 drm_intel_gem_bo_map_gtt(intel_obj
->buffer
);
363 intel_obj
->mapped_gtt
= GL_TRUE
;
365 drm_intel_bo_map(intel_obj
->buffer
, (access
& GL_MAP_WRITE_BIT
) != 0);
366 intel_obj
->mapped_gtt
= GL_FALSE
;
369 obj
->Pointer
= intel_obj
->buffer
->virtual + offset
;
375 * Called via glUnmapBuffer().
378 intel_bufferobj_unmap(GLcontext
* ctx
,
379 GLenum target
, struct gl_buffer_object
*obj
)
381 struct intel_context
*intel
= intel_context(ctx
);
382 struct intel_buffer_object
*intel_obj
= intel_buffer_object(obj
);
385 assert(obj
->Pointer
);
386 if (intel_obj
->sys_buffer
!= NULL
) {
387 /* always keep the mapping around. */
388 } else if (intel_obj
->range_map_bo
!= NULL
) {
389 if (intel_obj
->mapped_gtt
) {
390 drm_intel_gem_bo_unmap_gtt(intel_obj
->range_map_bo
);
392 drm_intel_bo_unmap(intel_obj
->range_map_bo
);
395 /* We ignore the FLUSH_EXPLICIT bit and the calls associated with it.
396 * It would be a small win to support that, but for now we just copy
397 * the whole mapped range into place.
399 intel_emit_linear_blit(intel
,
400 intel_obj
->buffer
, obj
->Offset
,
401 intel_obj
->range_map_bo
, 0,
404 /* Since we've emitted some blits to buffers that will (likely) be used
405 * in rendering operations in other cache domains in this batch, emit a
406 * flush. Once again, we wish for a domain tracker in libdrm to cover
407 * usage inside of a batchbuffer.
409 intel_batchbuffer_emit_mi_flush(intel
->batch
);
411 drm_intel_bo_unreference(intel_obj
->range_map_bo
);
412 intel_obj
->range_map_bo
= NULL
;
413 } else if (intel_obj
->buffer
!= NULL
) {
414 if (intel_obj
->mapped_gtt
) {
415 drm_intel_gem_bo_unmap_gtt(intel_obj
->buffer
);
417 drm_intel_bo_unmap(intel_obj
->buffer
);
426 intel_bufferobj_buffer(struct intel_context
*intel
,
427 struct intel_buffer_object
*intel_obj
, GLuint flag
)
429 if (intel_obj
->region
) {
430 if (flag
== INTEL_WRITE_PART
)
431 intel_bufferobj_cow(intel
, intel_obj
);
432 else if (flag
== INTEL_WRITE_FULL
) {
433 intel_bufferobj_release_region(intel
, intel_obj
);
434 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
438 if (intel_obj
->buffer
== NULL
) {
439 void *sys_buffer
= intel_obj
->sys_buffer
;
441 /* only one of buffer and sys_buffer could be non-NULL */
442 intel_bufferobj_alloc_buffer(intel
, intel_obj
);
443 intel_obj
->sys_buffer
= NULL
;
445 intel_bufferobj_subdata(&intel
->ctx
,
448 intel_obj
->Base
.Size
,
451 _mesa_free(sys_buffer
);
452 intel_obj
->sys_buffer
= NULL
;
455 return intel_obj
->buffer
;
459 intel_bufferobj_copy_subdata(GLcontext
*ctx
,
460 struct gl_buffer_object
*src
,
461 struct gl_buffer_object
*dst
,
462 GLintptr read_offset
, GLintptr write_offset
,
465 struct intel_context
*intel
= intel_context(ctx
);
466 struct intel_buffer_object
*intel_src
= intel_buffer_object(src
);
467 struct intel_buffer_object
*intel_dst
= intel_buffer_object(dst
);
468 drm_intel_bo
*src_bo
, *dst_bo
;
473 /* If we're in system memory, just map and memcpy. */
474 if (intel_src
->sys_buffer
|| intel_dst
->sys_buffer
) {
475 /* The same buffer may be used, but note that regions copied may
479 char *ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
481 memcpy(ptr
+ write_offset
, ptr
+ read_offset
, size
);
482 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
487 src_ptr
= intel_bufferobj_map(ctx
, GL_COPY_READ_BUFFER
,
489 dst_ptr
= intel_bufferobj_map(ctx
, GL_COPY_WRITE_BUFFER
,
492 memcpy(dst_ptr
+ write_offset
, src_ptr
+ read_offset
, size
);
494 intel_bufferobj_unmap(ctx
, GL_COPY_READ_BUFFER
, src
);
495 intel_bufferobj_unmap(ctx
, GL_COPY_WRITE_BUFFER
, dst
);
499 /* Otherwise, we have real BOs, so blit them. */
501 dst_bo
= intel_bufferobj_buffer(intel
, intel_dst
, INTEL_WRITE_PART
);
502 src_bo
= intel_bufferobj_buffer(intel
, intel_src
, INTEL_READ
);
504 intel_emit_linear_blit(intel
,
505 dst_bo
, write_offset
,
506 src_bo
, read_offset
, size
);
508 /* Since we've emitted some blits to buffers that will (likely) be used
509 * in rendering operations in other cache domains in this batch, emit a
510 * flush. Once again, we wish for a domain tracker in libdrm to cover
511 * usage inside of a batchbuffer.
513 intel_batchbuffer_emit_mi_flush(intel
->batch
);
517 intelInitBufferObjectFuncs(struct dd_function_table
*functions
)
519 functions
->NewBufferObject
= intel_bufferobj_alloc
;
520 functions
->DeleteBuffer
= intel_bufferobj_free
;
521 functions
->BufferData
= intel_bufferobj_data
;
522 functions
->BufferSubData
= intel_bufferobj_subdata
;
523 functions
->GetBufferSubData
= intel_bufferobj_get_subdata
;
524 functions
->MapBuffer
= intel_bufferobj_map
;
525 functions
->MapBufferRange
= intel_bufferobj_map_range
;
526 functions
->UnmapBuffer
= intel_bufferobj_unmap
;
527 functions
->CopyBufferSubData
= intel_bufferobj_copy_subdata
;