1 /**************************************************************************
3 * Copyright 2006 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 **************************************************************************/
28 /* Provide additional functionality on top of bufmgr buffers:
29 * - 2d semantics and blit operations
30 * - refcounting of buffers for multiple images in a buffer.
31 * - refcounting of buffer mappings.
32 * - some logic for moving the buffers to the best memory pools for
35 * Most of this is to make it easier to implement the fixed-layout
36 * mipmap tree required by intel hardware in the face of GL's
37 * programming interface where each image can be specifed in random
38 * order and it isn't clear what layout the tree should have until the
42 #include <sys/ioctl.h>
45 #include "intel_context.h"
46 #include "intel_regions.h"
47 #include "intel_blit.h"
48 #include "intel_buffer_objects.h"
49 #include "intel_bufmgr.h"
50 #include "intel_batchbuffer.h"
52 #define FILE_DEBUG_FLAG DEBUG_REGION
54 /* This should be set to the maximum backtrace size desired.
55 * Set it to 0 to disable backtrace debugging.
57 #define DEBUG_BACKTRACE_SIZE 0
59 #if DEBUG_BACKTRACE_SIZE == 0
60 /* Use the standard debug output */
61 #define _DBG(...) DBG(__VA_ARGS__)
63 /* Use backtracing debug output */
64 #define _DBG(...) {debug_backtrace(); DBG(__VA_ARGS__);}
66 /* Backtracing debug support */
72 void *trace
[DEBUG_BACKTRACE_SIZE
];
73 char **strings
= NULL
;
77 traceSize
= backtrace(trace
, DEBUG_BACKTRACE_SIZE
);
78 strings
= backtrace_symbols(trace
, traceSize
);
79 if (strings
== NULL
) {
84 /* Spit out all the strings with a colon separator. Ignore
85 * the first, since we don't really care about the call
86 * to debug_backtrace() itself. Skip until the final "/" in
87 * the trace to avoid really long lines.
89 for (i
= 1; i
< traceSize
; i
++) {
90 char *p
= strings
[i
], *slash
= strings
[i
];
100 /* Free up the memory, and we're done */
108 /* XXX: Thread safety?
111 intel_region_map(struct intel_context
*intel
, struct intel_region
*region
)
113 intelFlush(&intel
->ctx
);
115 _DBG("%s %p\n", __FUNCTION__
, region
);
116 if (!region
->map_refcount
++) {
118 intel_region_cow(intel
, region
);
120 if (region
->tiling
!= I915_TILING_NONE
&&
121 intel
->intelScreen
->kernel_exec_fencing
)
122 drm_intel_gem_bo_map_gtt(region
->buffer
);
124 dri_bo_map(region
->buffer
, GL_TRUE
);
125 region
->map
= region
->buffer
->virtual;
132 intel_region_unmap(struct intel_context
*intel
, struct intel_region
*region
)
134 _DBG("%s %p\n", __FUNCTION__
, region
);
135 if (!--region
->map_refcount
) {
136 if (region
->tiling
!= I915_TILING_NONE
&&
137 intel
->intelScreen
->kernel_exec_fencing
)
138 drm_intel_gem_bo_unmap_gtt(region
->buffer
);
140 dri_bo_unmap(region
->buffer
);
145 static struct intel_region
*
146 intel_region_alloc_internal(struct intel_context
*intel
,
148 GLuint width
, GLuint height
, GLuint pitch
,
151 struct intel_region
*region
;
153 if (buffer
== NULL
) {
154 _DBG("%s <-- NULL\n", __FUNCTION__
);
158 region
= calloc(sizeof(*region
), 1);
160 region
->width
= width
;
161 region
->height
= height
;
162 region
->pitch
= pitch
;
163 region
->refcount
= 1;
164 region
->buffer
= buffer
;
166 /* Default to no tiling */
167 region
->tiling
= I915_TILING_NONE
;
168 region
->bit_6_swizzle
= I915_BIT_6_SWIZZLE_NONE
;
170 _DBG("%s <-- %p\n", __FUNCTION__
, region
);
174 struct intel_region
*
175 intel_region_alloc(struct intel_context
*intel
,
177 GLuint cpp
, GLuint width
, GLuint height
, GLuint pitch
,
178 GLboolean expect_accelerated_upload
)
181 struct intel_region
*region
;
183 /* If we're tiled, our allocations are in 8 or 32-row blocks, so
184 * failure to align our height means that we won't allocate enough pages.
186 * If we're untiled, we still have to align to 2 rows high because the
187 * data port accesses 2x2 blocks even if the bottom row isn't to be
188 * rendered, so failure to align means we could walk off the end of the
191 if (tiling
== I915_TILING_X
)
192 height
= ALIGN(height
, 8);
193 else if (tiling
== I915_TILING_Y
)
194 height
= ALIGN(height
, 32);
196 height
= ALIGN(height
, 2);
198 /* If we're untiled, we have to align to 2 rows high because the
199 * data port accesses 2x2 blocks even if the bottom row isn't to be
200 * rendered, so failure to align means we could walk off the end of the
203 height
= ALIGN(height
, 2);
205 if (expect_accelerated_upload
) {
206 buffer
= drm_intel_bo_alloc_for_render(intel
->bufmgr
, "region",
207 pitch
* cpp
* height
, 64);
209 buffer
= drm_intel_bo_alloc(intel
->bufmgr
, "region",
210 pitch
* cpp
* height
, 64);
213 region
= intel_region_alloc_internal(intel
, cpp
, width
, height
,
216 if (tiling
!= I915_TILING_NONE
) {
217 assert(((pitch
* cpp
) & 127) == 0);
218 drm_intel_bo_set_tiling(buffer
, &tiling
, pitch
* cpp
);
219 drm_intel_bo_get_tiling(buffer
, ®ion
->tiling
, ®ion
->bit_6_swizzle
);
225 struct intel_region
*
226 intel_region_alloc_for_handle(struct intel_context
*intel
,
228 GLuint width
, GLuint height
, GLuint pitch
,
229 GLuint handle
, const char *name
)
231 struct intel_region
*region
;
235 buffer
= intel_bo_gem_create_from_name(intel
->bufmgr
, name
, handle
);
237 region
= intel_region_alloc_internal(intel
, cpp
,
238 width
, height
, pitch
, buffer
);
242 ret
= dri_bo_get_tiling(region
->buffer
, ®ion
->tiling
,
243 ®ion
->bit_6_swizzle
);
245 fprintf(stderr
, "Couldn't get tiling of buffer %d (%s): %s\n",
246 handle
, name
, strerror(-ret
));
247 intel_region_release(®ion
);
255 intel_region_reference(struct intel_region
**dst
, struct intel_region
*src
)
258 _DBG("%s %p %d\n", __FUNCTION__
, src
, src
->refcount
);
260 assert(*dst
== NULL
);
268 intel_region_release(struct intel_region
**region_handle
)
270 struct intel_region
*region
= *region_handle
;
272 if (region
== NULL
) {
273 _DBG("%s NULL\n", __FUNCTION__
);
277 _DBG("%s %p %d\n", __FUNCTION__
, region
, region
->refcount
- 1);
279 ASSERT(region
->refcount
> 0);
282 if (region
->refcount
== 0) {
283 assert(region
->map_refcount
== 0);
286 region
->pbo
->region
= NULL
;
288 dri_bo_unreference(region
->buffer
);
292 *region_handle
= NULL
;
296 * XXX Move this into core Mesa?
299 _mesa_copy_rect(GLubyte
* dst
,
307 GLuint src_pitch
, GLuint src_x
, GLuint src_y
)
315 dst
+= dst_y
* dst_pitch
;
316 src
+= src_y
* dst_pitch
;
319 if (width
== dst_pitch
&& width
== src_pitch
)
320 memcpy(dst
, src
, height
* width
);
322 for (i
= 0; i
< height
; i
++) {
323 memcpy(dst
, src
, width
);
331 /* Upload data to a rectangular sub-region. Lots of choices how to do this:
333 * - memcpy by span to current destination
334 * - upload data as new buffer and blit
336 * Currently always memcpy.
339 intel_region_data(struct intel_context
*intel
,
340 struct intel_region
*dst
,
342 GLuint dstx
, GLuint dsty
,
343 const void *src
, GLuint src_pitch
,
344 GLuint srcx
, GLuint srcy
, GLuint width
, GLuint height
)
346 _DBG("%s\n", __FUNCTION__
);
353 dsty
== 0 && width
== dst
->pitch
&& height
== dst
->height
)
354 intel_region_release_pbo(intel
, dst
);
356 intel_region_cow(intel
, dst
);
359 _mesa_copy_rect(intel_region_map(intel
, dst
) + dst_offset
,
362 dstx
, dsty
, width
, height
, src
, src_pitch
, srcx
, srcy
);
364 intel_region_unmap(intel
, dst
);
367 /* Copy rectangular sub-regions. Need better logic about when to
368 * push buffers into AGP - will currently do so whenever possible.
371 intel_region_copy(struct intel_context
*intel
,
372 struct intel_region
*dst
,
374 GLuint dstx
, GLuint dsty
,
375 struct intel_region
*src
,
377 GLuint srcx
, GLuint srcy
, GLuint width
, GLuint height
,
380 _DBG("%s\n", __FUNCTION__
);
387 dsty
== 0 && width
== dst
->pitch
&& height
== dst
->height
)
388 intel_region_release_pbo(intel
, dst
);
390 intel_region_cow(intel
, dst
);
393 assert(src
->cpp
== dst
->cpp
);
395 return intelEmitCopyBlit(intel
,
397 src
->pitch
, src
->buffer
, src_offset
, src
->tiling
,
398 dst
->pitch
, dst
->buffer
, dst_offset
, dst
->tiling
,
399 srcx
, srcy
, dstx
, dsty
, width
, height
,
403 /* Attach to a pbo, discarding our data. Effectively zero-copy upload
407 intel_region_attach_pbo(struct intel_context
*intel
,
408 struct intel_region
*region
,
409 struct intel_buffer_object
*pbo
)
413 if (region
->pbo
== pbo
)
416 _DBG("%s %p %p\n", __FUNCTION__
, region
, pbo
);
418 /* If there is already a pbo attached, break the cow tie now.
419 * Don't call intel_region_release_pbo() as that would
420 * unnecessarily allocate a new buffer we would have to immediately
424 region
->pbo
->region
= NULL
;
428 if (region
->buffer
) {
429 dri_bo_unreference(region
->buffer
);
430 region
->buffer
= NULL
;
433 /* make sure pbo has a buffer of its own */
434 buffer
= intel_bufferobj_buffer(intel
, pbo
, INTEL_WRITE_FULL
);
437 region
->pbo
->region
= region
;
438 dri_bo_reference(buffer
);
439 region
->buffer
= buffer
;
440 region
->tiling
= I915_TILING_NONE
;
444 /* Break the COW tie to the pbo and allocate a new buffer.
445 * The pbo gets to keep the data.
448 intel_region_release_pbo(struct intel_context
*intel
,
449 struct intel_region
*region
)
451 _DBG("%s %p\n", __FUNCTION__
, region
);
452 assert(region
->buffer
== region
->pbo
->buffer
);
453 region
->pbo
->region
= NULL
;
455 dri_bo_unreference(region
->buffer
);
456 region
->buffer
= NULL
;
458 region
->buffer
= dri_bo_alloc(intel
->bufmgr
, "region",
459 region
->pitch
* region
->cpp
* region
->height
,
463 /* Break the COW tie to the pbo. Both the pbo and the region end up
464 * with a copy of the data.
467 intel_region_cow(struct intel_context
*intel
, struct intel_region
*region
)
469 struct intel_buffer_object
*pbo
= region
->pbo
;
472 intel_region_release_pbo(intel
, region
);
474 assert(region
->cpp
* region
->pitch
* region
->height
== pbo
->Base
.Size
);
476 _DBG("%s %p (%d bytes)\n", __FUNCTION__
, region
, pbo
->Base
.Size
);
478 /* Now blit from the texture buffer to the new buffer:
481 ok
= intelEmitCopyBlit(intel
,
483 region
->pitch
, pbo
->buffer
, 0, region
->tiling
,
484 region
->pitch
, region
->buffer
, 0, region
->tiling
,
486 region
->pitch
, region
->height
,
492 intel_region_buffer(struct intel_context
*intel
,
493 struct intel_region
*region
, GLuint flag
)
496 if (flag
== INTEL_WRITE_PART
)
497 intel_region_cow(intel
, region
);
498 else if (flag
== INTEL_WRITE_FULL
)
499 intel_region_release_pbo(intel
, region
);
502 return region
->buffer
;