svga: Don't fake DXT compression ability.
[mesa.git] / src / gallium / drivers / svga / svga_resource_texture.c
1 /**********************************************************
2 * Copyright 2008-2009 VMware, Inc. All rights reserved.
3 *
4 * Permission is hereby granted, free of charge, to any person
5 * obtaining a copy of this software and associated documentation
6 * files (the "Software"), to deal in the Software without
7 * restriction, including without limitation the rights to use, copy,
8 * modify, merge, publish, distribute, sublicense, and/or sell copies
9 * of the Software, and to permit persons to whom the Software is
10 * furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be
13 * included in all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 **********************************************************/
25
26 #include "svga_cmd.h"
27
28 #include "pipe/p_state.h"
29 #include "pipe/p_defines.h"
30 #include "util/u_inlines.h"
31 #include "os/os_thread.h"
32 #include "util/u_format.h"
33 #include "util/u_math.h"
34 #include "util/u_memory.h"
35
36 #include "svga_screen.h"
37 #include "svga_context.h"
38 #include "svga_resource_texture.h"
39 #include "svga_resource_buffer.h"
40 #include "svga_sampler_view.h"
41 #include "svga_winsys.h"
42 #include "svga_debug.h"
43
44
45 /* XXX: This isn't a real hardware flag, but just a hack for kernel to
46 * know about primary surfaces. Find a better way to accomplish this.
47 */
48 #define SVGA3D_SURFACE_HINT_SCANOUT (1 << 9)
49
50
51 static unsigned int
52 svga_texture_is_referenced( struct pipe_context *pipe,
53 struct pipe_resource *texture,
54 unsigned level, int layer)
55 {
56 struct svga_texture *tex = svga_texture(texture);
57 struct svga_screen *ss = svga_screen(pipe->screen);
58
59 /**
60 * The screen does not cache texture writes.
61 */
62
63 if (!tex->handle || ss->sws->surface_is_flushed(ss->sws, tex->handle))
64 return PIPE_UNREFERENCED;
65
66 /**
67 * sws->surface_is_flushed() does not distinguish between read references
68 * and write references. So assume a reference is both.
69 */
70
71 return PIPE_REFERENCED_FOR_READ | PIPE_REFERENCED_FOR_WRITE;
72 }
73
74
75
76 /*
77 * Helper function and arrays
78 */
79
80 SVGA3dSurfaceFormat
81 svga_translate_format(enum pipe_format format)
82 {
83 switch(format) {
84
85 case PIPE_FORMAT_B8G8R8A8_UNORM:
86 return SVGA3D_A8R8G8B8;
87 case PIPE_FORMAT_B8G8R8X8_UNORM:
88 return SVGA3D_X8R8G8B8;
89
90 /* Required for GL2.1:
91 */
92 case PIPE_FORMAT_B8G8R8A8_SRGB:
93 return SVGA3D_A8R8G8B8;
94
95 case PIPE_FORMAT_B5G6R5_UNORM:
96 return SVGA3D_R5G6B5;
97 case PIPE_FORMAT_B5G5R5A1_UNORM:
98 return SVGA3D_A1R5G5B5;
99 case PIPE_FORMAT_B4G4R4A4_UNORM:
100 return SVGA3D_A4R4G4B4;
101
102
103 /* XXX: Doesn't seem to work properly.
104 case PIPE_FORMAT_Z32_UNORM:
105 return SVGA3D_Z_D32;
106 */
107 case PIPE_FORMAT_Z16_UNORM:
108 return SVGA3D_Z_D16;
109 case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
110 return SVGA3D_Z_D24S8;
111 case PIPE_FORMAT_X8Z24_UNORM:
112 return SVGA3D_Z_D24X8;
113
114 case PIPE_FORMAT_A8_UNORM:
115 return SVGA3D_ALPHA8;
116 case PIPE_FORMAT_L8_UNORM:
117 return SVGA3D_LUMINANCE8;
118
119 case PIPE_FORMAT_DXT1_RGB:
120 case PIPE_FORMAT_DXT1_RGBA:
121 return SVGA3D_DXT1;
122 case PIPE_FORMAT_DXT3_RGBA:
123 return SVGA3D_DXT3;
124 case PIPE_FORMAT_DXT5_RGBA:
125 return SVGA3D_DXT5;
126
127 default:
128 return SVGA3D_FORMAT_INVALID;
129 }
130 }
131
132
133 SVGA3dSurfaceFormat
134 svga_translate_format_render(enum pipe_format format)
135 {
136 switch(format) {
137 case PIPE_FORMAT_B8G8R8A8_UNORM:
138 case PIPE_FORMAT_B8G8R8X8_UNORM:
139 case PIPE_FORMAT_B5G5R5A1_UNORM:
140 case PIPE_FORMAT_B4G4R4A4_UNORM:
141 case PIPE_FORMAT_B5G6R5_UNORM:
142 case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
143 case PIPE_FORMAT_X8Z24_UNORM:
144 case PIPE_FORMAT_Z32_UNORM:
145 case PIPE_FORMAT_Z16_UNORM:
146 case PIPE_FORMAT_L8_UNORM:
147 return svga_translate_format(format);
148
149 default:
150 return SVGA3D_FORMAT_INVALID;
151 }
152 }
153
154
155 static INLINE void
156 svga_transfer_dma_band(struct svga_context *svga,
157 struct svga_transfer *st,
158 SVGA3dTransferType transfer,
159 unsigned y, unsigned h, unsigned srcy)
160 {
161 struct svga_texture *texture = svga_texture(st->base.resource);
162 SVGA3dCopyBox box;
163 enum pipe_error ret;
164
165 box.x = st->base.box.x;
166 box.y = y;
167 box.z = st->base.box.z;
168 box.w = st->base.box.width;
169 box.h = h;
170 box.d = 1;
171 box.srcx = 0;
172 box.srcy = srcy;
173 box.srcz = 0;
174
175 if (st->base.resource->target == PIPE_TEXTURE_CUBE) {
176 st->face = st->base.box.z;
177 box.z = 0;
178 }
179 else
180 st->face = 0;
181
182 SVGA_DBG(DEBUG_DMA, "dma %s sid %p, face %u, (%u, %u, %u) - (%u, %u, %u), %ubpp\n",
183 transfer == SVGA3D_WRITE_HOST_VRAM ? "to" : "from",
184 texture->handle,
185 st->face,
186 st->base.box.x,
187 y,
188 box.z,
189 st->base.box.x + st->base.box.width,
190 y + h,
191 box.z + 1,
192 util_format_get_blocksize(texture->b.b.format) * 8 /
193 (util_format_get_blockwidth(texture->b.b.format)*util_format_get_blockheight(texture->b.b.format)));
194
195 ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1);
196 if(ret != PIPE_OK) {
197 svga_context_flush(svga, NULL);
198 ret = SVGA3D_SurfaceDMA(svga->swc, st, transfer, &box, 1);
199 assert(ret == PIPE_OK);
200 }
201 }
202
203
204 static INLINE void
205 svga_transfer_dma(struct svga_context *svga,
206 struct svga_transfer *st,
207 SVGA3dTransferType transfer)
208 {
209 struct svga_texture *texture = svga_texture(st->base.resource);
210 struct svga_screen *screen = svga_screen(texture->b.b.screen);
211 struct svga_winsys_screen *sws = screen->sws;
212 struct pipe_fence_handle *fence = NULL;
213
214 if (transfer == SVGA3D_READ_HOST_VRAM) {
215 SVGA_DBG(DEBUG_PERF, "%s: readback transfer\n", __FUNCTION__);
216 }
217
218
219 if(!st->swbuf) {
220 /* Do the DMA transfer in a single go */
221
222 svga_transfer_dma_band(svga, st, transfer, st->base.box.y, st->base.box.height, 0);
223
224 if(transfer == SVGA3D_READ_HOST_VRAM) {
225 svga_context_flush(svga, &fence);
226 sws->fence_finish(sws, fence, 0);
227 sws->fence_reference(sws, &fence, NULL);
228 }
229 }
230 else {
231 unsigned y, h, srcy;
232 unsigned blockheight = util_format_get_blockheight(st->base.resource->format);
233 h = st->hw_nblocksy * blockheight;
234 srcy = 0;
235 for(y = 0; y < st->base.box.height; y += h) {
236 unsigned offset, length;
237 void *hw, *sw;
238
239 if (y + h > st->base.box.height)
240 h = st->base.box.height - y;
241
242 /* Transfer band must be aligned to pixel block boundaries */
243 assert(y % blockheight == 0);
244 assert(h % blockheight == 0);
245
246 offset = y * st->base.stride / blockheight;
247 length = h * st->base.stride / blockheight;
248
249 sw = (uint8_t *)st->swbuf + offset;
250
251 if(transfer == SVGA3D_WRITE_HOST_VRAM) {
252 /* Wait for the previous DMAs to complete */
253 /* TODO: keep one DMA (at half the size) in the background */
254 if(y) {
255 svga_context_flush(svga, &fence);
256 sws->fence_finish(sws, fence, 0);
257 sws->fence_reference(sws, &fence, NULL);
258 }
259
260 hw = sws->buffer_map(sws, st->hwbuf, PIPE_TRANSFER_WRITE);
261 assert(hw);
262 if(hw) {
263 memcpy(hw, sw, length);
264 sws->buffer_unmap(sws, st->hwbuf);
265 }
266 }
267
268 svga_transfer_dma_band(svga, st, transfer, y, h, srcy);
269
270 if(transfer == SVGA3D_READ_HOST_VRAM) {
271 svga_context_flush(svga, &fence);
272 sws->fence_finish(sws, fence, 0);
273
274 hw = sws->buffer_map(sws, st->hwbuf, PIPE_TRANSFER_READ);
275 assert(hw);
276 if(hw) {
277 memcpy(sw, hw, length);
278 sws->buffer_unmap(sws, st->hwbuf);
279 }
280 }
281 }
282 }
283 }
284
285
286
287
288
289 static boolean
290 svga_texture_get_handle(struct pipe_screen *screen,
291 struct pipe_resource *texture,
292 struct winsys_handle *whandle)
293 {
294 struct svga_winsys_screen *sws = svga_winsys_screen(texture->screen);
295 unsigned stride;
296
297 assert(svga_texture(texture)->key.cachable == 0);
298 svga_texture(texture)->key.cachable = 0;
299 stride = util_format_get_nblocksx(texture->format, texture->width0) *
300 util_format_get_blocksize(texture->format);
301 return sws->surface_get_handle(sws, svga_texture(texture)->handle, stride, whandle);
302 }
303
304
305 static void
306 svga_texture_destroy(struct pipe_screen *screen,
307 struct pipe_resource *pt)
308 {
309 struct svga_screen *ss = svga_screen(screen);
310 struct svga_texture *tex = (struct svga_texture *)pt;
311
312 ss->texture_timestamp++;
313
314 svga_sampler_view_reference(&tex->cached_view, NULL);
315
316 /*
317 DBG("%s deleting %p\n", __FUNCTION__, (void *) tex);
318 */
319 SVGA_DBG(DEBUG_DMA, "unref sid %p (texture)\n", tex->handle);
320 svga_screen_surface_destroy(ss, &tex->key, &tex->handle);
321
322 FREE(tex);
323 }
324
325
326
327
328
329
330
331 /* XXX: Still implementing this as if it was a screen function, but
332 * can now modify it to queue transfers on the context.
333 */
334 static struct pipe_transfer *
335 svga_texture_get_transfer(struct pipe_context *pipe,
336 struct pipe_resource *texture,
337 unsigned level,
338 unsigned usage,
339 const struct pipe_box *box)
340 {
341 struct svga_context *svga = svga_context(pipe);
342 struct svga_screen *ss = svga_screen(pipe->screen);
343 struct svga_winsys_screen *sws = ss->sws;
344 struct svga_transfer *st;
345 unsigned nblocksx = util_format_get_nblocksx(texture->format, box->width);
346 unsigned nblocksy = util_format_get_nblocksy(texture->format, box->height);
347
348 /* We can't map texture storage directly */
349 if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
350 return NULL;
351
352 assert(box->depth == 1);
353 st = CALLOC_STRUCT(svga_transfer);
354 if (!st)
355 return NULL;
356
357 pipe_resource_reference(&st->base.resource, texture);
358 st->base.level = level;
359 st->base.usage = usage;
360 st->base.box = *box;
361 st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
362 st->base.layer_stride = 0;
363
364 st->hw_nblocksy = nblocksy;
365
366 st->hwbuf = svga_winsys_buffer_create(svga,
367 1,
368 0,
369 st->hw_nblocksy*st->base.stride);
370 while(!st->hwbuf && (st->hw_nblocksy /= 2)) {
371 st->hwbuf = svga_winsys_buffer_create(svga,
372 1,
373 0,
374 st->hw_nblocksy*st->base.stride);
375 }
376
377 if(!st->hwbuf)
378 goto no_hwbuf;
379
380 if(st->hw_nblocksy < nblocksy) {
381 /* We couldn't allocate a hardware buffer big enough for the transfer,
382 * so allocate regular malloc memory instead */
383 if (0) {
384 debug_printf("%s: failed to allocate %u KB of DMA, "
385 "splitting into %u x %u KB DMA transfers\n",
386 __FUNCTION__,
387 (nblocksy*st->base.stride + 1023)/1024,
388 (nblocksy + st->hw_nblocksy - 1)/st->hw_nblocksy,
389 (st->hw_nblocksy*st->base.stride + 1023)/1024);
390 }
391
392 st->swbuf = MALLOC(nblocksy*st->base.stride);
393 if(!st->swbuf)
394 goto no_swbuf;
395 }
396
397 if (usage & PIPE_TRANSFER_READ)
398 svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM);
399
400 return &st->base;
401
402 no_swbuf:
403 sws->buffer_destroy(sws, st->hwbuf);
404 no_hwbuf:
405 FREE(st);
406 return NULL;
407 }
408
409
410 /* XXX: Still implementing this as if it was a screen function, but
411 * can now modify it to queue transfers on the context.
412 */
413 static void *
414 svga_texture_transfer_map( struct pipe_context *pipe,
415 struct pipe_transfer *transfer )
416 {
417 struct svga_screen *ss = svga_screen(pipe->screen);
418 struct svga_winsys_screen *sws = ss->sws;
419 struct svga_transfer *st = svga_transfer(transfer);
420
421 if(st->swbuf)
422 return st->swbuf;
423 else
424 /* The wait for read transfers already happened when svga_transfer_dma
425 * was called. */
426 return sws->buffer_map(sws, st->hwbuf, transfer->usage);
427 }
428
429
430 /* XXX: Still implementing this as if it was a screen function, but
431 * can now modify it to queue transfers on the context.
432 */
433 static void
434 svga_texture_transfer_unmap(struct pipe_context *pipe,
435 struct pipe_transfer *transfer)
436 {
437 struct svga_screen *ss = svga_screen(pipe->screen);
438 struct svga_winsys_screen *sws = ss->sws;
439 struct svga_transfer *st = svga_transfer(transfer);
440
441 if(!st->swbuf)
442 sws->buffer_unmap(sws, st->hwbuf);
443 }
444
445
446 static void
447 svga_texture_transfer_destroy(struct pipe_context *pipe,
448 struct pipe_transfer *transfer)
449 {
450 struct svga_context *svga = svga_context(pipe);
451 struct svga_texture *tex = svga_texture(transfer->resource);
452 struct svga_screen *ss = svga_screen(pipe->screen);
453 struct svga_winsys_screen *sws = ss->sws;
454 struct svga_transfer *st = svga_transfer(transfer);
455
456 if (st->base.usage & PIPE_TRANSFER_WRITE) {
457 svga_transfer_dma(svga, st, SVGA3D_WRITE_HOST_VRAM);
458 ss->texture_timestamp++;
459 tex->view_age[transfer->level] = ++(tex->age);
460 if (transfer->resource->target == PIPE_TEXTURE_CUBE)
461 tex->defined[transfer->box.z][transfer->level] = TRUE;
462 else
463 tex->defined[0][transfer->level] = TRUE;
464 }
465
466 pipe_resource_reference(&st->base.resource, NULL);
467 FREE(st->swbuf);
468 sws->buffer_destroy(sws, st->hwbuf);
469 FREE(st);
470 }
471
472
473
474
475
476 struct u_resource_vtbl svga_texture_vtbl =
477 {
478 svga_texture_get_handle, /* get_handle */
479 svga_texture_destroy, /* resource_destroy */
480 svga_texture_is_referenced, /* is_resource_referenced */
481 svga_texture_get_transfer, /* get_transfer */
482 svga_texture_transfer_destroy, /* transfer_destroy */
483 svga_texture_transfer_map, /* transfer_map */
484 u_default_transfer_flush_region, /* transfer_flush_region */
485 svga_texture_transfer_unmap, /* transfer_unmap */
486 u_default_transfer_inline_write /* transfer_inline_write */
487 };
488
489
490
491
492 struct pipe_resource *
493 svga_texture_create(struct pipe_screen *screen,
494 const struct pipe_resource *template)
495 {
496 struct svga_screen *svgascreen = svga_screen(screen);
497 struct svga_texture *tex = CALLOC_STRUCT(svga_texture);
498
499 if (!tex)
500 goto error1;
501
502 tex->b.b = *template;
503 tex->b.vtbl = &svga_texture_vtbl;
504 pipe_reference_init(&tex->b.b.reference, 1);
505 tex->b.b.screen = screen;
506
507 assert(template->last_level < SVGA_MAX_TEXTURE_LEVELS);
508 if(template->last_level >= SVGA_MAX_TEXTURE_LEVELS)
509 goto error2;
510
511 tex->key.flags = 0;
512 tex->key.size.width = template->width0;
513 tex->key.size.height = template->height0;
514 tex->key.size.depth = template->depth0;
515
516 if(template->target == PIPE_TEXTURE_CUBE) {
517 tex->key.flags |= SVGA3D_SURFACE_CUBEMAP;
518 tex->key.numFaces = 6;
519 }
520 else {
521 tex->key.numFaces = 1;
522 }
523
524 tex->key.cachable = 1;
525
526 if (template->bind & PIPE_BIND_SAMPLER_VIEW)
527 tex->key.flags |= SVGA3D_SURFACE_HINT_TEXTURE;
528
529 if (template->bind & PIPE_BIND_DISPLAY_TARGET) {
530 tex->key.cachable = 0;
531 }
532
533 if (template->bind & PIPE_BIND_SHARED) {
534 tex->key.cachable = 0;
535 }
536
537 if (template->bind & PIPE_BIND_SCANOUT) {
538 tex->key.flags |= SVGA3D_SURFACE_HINT_SCANOUT;
539 tex->key.cachable = 0;
540 }
541
542 /*
543 * XXX: Never pass the SVGA3D_SURFACE_HINT_RENDERTARGET hint. Mesa cannot
544 * know beforehand whether a texture will be used as a rendertarget or not
545 * and it always requests PIPE_BIND_RENDER_TARGET, therefore
546 * passing the SVGA3D_SURFACE_HINT_RENDERTARGET here defeats its purpose.
547 */
548 #if 0
549 if((template->bind & PIPE_BIND_RENDER_TARGET) &&
550 !util_format_is_s3tc(template->format))
551 tex->key.flags |= SVGA3D_SURFACE_HINT_RENDERTARGET;
552 #endif
553
554 if(template->bind & PIPE_BIND_DEPTH_STENCIL)
555 tex->key.flags |= SVGA3D_SURFACE_HINT_DEPTHSTENCIL;
556
557 tex->key.numMipLevels = template->last_level + 1;
558
559 tex->key.format = svga_translate_format(template->format);
560 if(tex->key.format == SVGA3D_FORMAT_INVALID)
561 goto error2;
562
563 SVGA_DBG(DEBUG_DMA, "surface_create for texture\n", tex->handle);
564 tex->handle = svga_screen_surface_create(svgascreen, &tex->key);
565 if (tex->handle)
566 SVGA_DBG(DEBUG_DMA, " --> got sid %p (texture)\n", tex->handle);
567
568 return &tex->b.b;
569
570 error2:
571 FREE(tex);
572 error1:
573 return NULL;
574 }
575
576
577
578
579 struct pipe_resource *
580 svga_texture_from_handle(struct pipe_screen *screen,
581 const struct pipe_resource *template,
582 struct winsys_handle *whandle)
583 {
584 struct svga_winsys_screen *sws = svga_winsys_screen(screen);
585 struct svga_winsys_surface *srf;
586 struct svga_texture *tex;
587 enum SVGA3dSurfaceFormat format = 0;
588 assert(screen);
589
590 /* Only supports one type */
591 if ((template->target != PIPE_TEXTURE_2D &&
592 template->target != PIPE_TEXTURE_RECT) ||
593 template->last_level != 0 ||
594 template->depth0 != 1) {
595 return NULL;
596 }
597
598 srf = sws->surface_from_handle(sws, whandle, &format);
599
600 if (!srf)
601 return NULL;
602
603 if (svga_translate_format(template->format) != format) {
604 unsigned f1 = svga_translate_format(template->format);
605 unsigned f2 = format;
606
607 /* It's okay for XRGB and ARGB or depth with/out stencil to get mixed up */
608 if ( !( (f1 == SVGA3D_X8R8G8B8 && f2 == SVGA3D_A8R8G8B8) ||
609 (f1 == SVGA3D_A8R8G8B8 && f2 == SVGA3D_X8R8G8B8) ||
610 (f1 == SVGA3D_Z_D24X8 && f2 == SVGA3D_Z_D24S8) ) ) {
611 debug_printf("%s wrong format %u != %u\n", __FUNCTION__, f1, f2);
612 return NULL;
613 }
614 }
615
616 tex = CALLOC_STRUCT(svga_texture);
617 if (!tex)
618 return NULL;
619
620 tex->b.b = *template;
621 tex->b.vtbl = &svga_texture_vtbl;
622 pipe_reference_init(&tex->b.b.reference, 1);
623 tex->b.b.screen = screen;
624
625 if (format == SVGA3D_X8R8G8B8)
626 tex->b.b.format = PIPE_FORMAT_B8G8R8X8_UNORM;
627 else if (format == SVGA3D_A8R8G8B8)
628 tex->b.b.format = PIPE_FORMAT_B8G8R8A8_UNORM;
629 else {
630 /* ?? */
631 }
632
633 SVGA_DBG(DEBUG_DMA, "wrap surface sid %p\n", srf);
634
635 tex->key.cachable = 0;
636 tex->handle = srf;
637
638 return &tex->b.b;
639 }
640