gallium: split transfer_inline_write into buffer and texture callbacks
[mesa.git] / src / gallium / drivers / freedreno / freedreno_resource.c
1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3 /*
4 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 * Rob Clark <robclark@freedesktop.org>
27 */
28
29 #include "util/u_format.h"
30 #include "util/u_format_rgtc.h"
31 #include "util/u_format_zs.h"
32 #include "util/u_inlines.h"
33 #include "util/u_transfer.h"
34 #include "util/u_string.h"
35 #include "util/u_surface.h"
36
37 #include "freedreno_resource.h"
38 #include "freedreno_screen.h"
39 #include "freedreno_surface.h"
40 #include "freedreno_context.h"
41 #include "freedreno_query_hw.h"
42 #include "freedreno_util.h"
43
44 #include <errno.h>
45
46 /* XXX this should go away, needed for 'struct winsys_handle' */
47 #include "state_tracker/drm_driver.h"
48
49 static bool
50 pending(struct fd_resource *rsc, enum fd_resource_status status)
51 {
52 return (rsc->status & status) ||
53 (rsc->stencil && (rsc->stencil->status & status));
54 }
55
56 static void
57 fd_invalidate_resource(struct fd_context *ctx, struct pipe_resource *prsc)
58 {
59 int i;
60
61 /* Go through the entire state and see if the resource is bound
62 * anywhere. If it is, mark the relevant state as dirty. This is called on
63 * realloc_bo.
64 */
65
66 /* Constbufs */
67 for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS && !(ctx->dirty & FD_DIRTY_CONSTBUF); i++) {
68 if (ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer == prsc)
69 ctx->dirty |= FD_DIRTY_CONSTBUF;
70 if (ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer == prsc)
71 ctx->dirty |= FD_DIRTY_CONSTBUF;
72 }
73
74 /* VBOs */
75 for (i = 0; i < ctx->vtx.vertexbuf.count && !(ctx->dirty & FD_DIRTY_VTXBUF); i++) {
76 if (ctx->vtx.vertexbuf.vb[i].buffer == prsc)
77 ctx->dirty |= FD_DIRTY_VTXBUF;
78 }
79
80 /* Index buffer */
81 if (ctx->indexbuf.buffer == prsc)
82 ctx->dirty |= FD_DIRTY_INDEXBUF;
83
84 /* Textures */
85 for (i = 0; i < ctx->verttex.num_textures && !(ctx->dirty & FD_DIRTY_VERTTEX); i++) {
86 if (ctx->verttex.textures[i] && (ctx->verttex.textures[i]->texture == prsc))
87 ctx->dirty |= FD_DIRTY_VERTTEX;
88 }
89 for (i = 0; i < ctx->fragtex.num_textures && !(ctx->dirty & FD_DIRTY_FRAGTEX); i++) {
90 if (ctx->fragtex.textures[i] && (ctx->fragtex.textures[i]->texture == prsc))
91 ctx->dirty |= FD_DIRTY_FRAGTEX;
92 }
93 }
94
95 static void
96 realloc_bo(struct fd_resource *rsc, uint32_t size)
97 {
98 struct fd_screen *screen = fd_screen(rsc->base.b.screen);
99 uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
100 DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */
101
102 /* if we start using things other than write-combine,
103 * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
104 */
105
106 if (rsc->bo)
107 fd_bo_del(rsc->bo);
108
109 rsc->bo = fd_bo_new(screen->dev, size, flags);
110 rsc->timestamp = 0;
111 rsc->status = 0;
112 rsc->pending_ctx = NULL;
113 list_delinit(&rsc->list);
114 util_range_set_empty(&rsc->valid_buffer_range);
115 }
116
117 static unsigned
118 fd_resource_layer_offset(struct fd_resource *rsc,
119 struct fd_resource_slice *slice,
120 unsigned layer)
121 {
122 if (rsc->layer_first)
123 return layer * rsc->layer_size;
124 else
125 return layer * slice->size0;
126 }
127
128 static void
129 fd_resource_flush_z32s8(struct fd_transfer *trans, const struct pipe_box *box)
130 {
131 struct fd_resource *rsc = fd_resource(trans->base.resource);
132 struct fd_resource_slice *slice = fd_resource_slice(rsc, trans->base.level);
133 struct fd_resource_slice *sslice = fd_resource_slice(rsc->stencil, trans->base.level);
134 enum pipe_format format = trans->base.resource->format;
135
136 float *depth = fd_bo_map(rsc->bo) + slice->offset +
137 fd_resource_layer_offset(rsc, slice, trans->base.box.z) +
138 (trans->base.box.y + box->y) * slice->pitch * 4 + (trans->base.box.x + box->x) * 4;
139 uint8_t *stencil = fd_bo_map(rsc->stencil->bo) + sslice->offset +
140 fd_resource_layer_offset(rsc->stencil, sslice, trans->base.box.z) +
141 (trans->base.box.y + box->y) * sslice->pitch + trans->base.box.x + box->x;
142
143 if (format != PIPE_FORMAT_X32_S8X24_UINT)
144 util_format_z32_float_s8x24_uint_unpack_z_float(
145 depth, slice->pitch * 4,
146 trans->staging, trans->base.stride,
147 box->width, box->height);
148
149 util_format_z32_float_s8x24_uint_unpack_s_8uint(
150 stencil, sslice->pitch,
151 trans->staging, trans->base.stride,
152 box->width, box->height);
153 }
154
155 static void
156 fd_resource_flush_rgtc(struct fd_transfer *trans, const struct pipe_box *box)
157 {
158 struct fd_resource *rsc = fd_resource(trans->base.resource);
159 struct fd_resource_slice *slice = fd_resource_slice(rsc, trans->base.level);
160 enum pipe_format format = trans->base.resource->format;
161
162 uint8_t *data = fd_bo_map(rsc->bo) + slice->offset +
163 fd_resource_layer_offset(rsc, slice, trans->base.box.z) +
164 ((trans->base.box.y + box->y) * slice->pitch +
165 trans->base.box.x + box->x) * rsc->cpp;
166
167 uint8_t *source = trans->staging +
168 util_format_get_nblocksy(format, box->y) * trans->base.stride +
169 util_format_get_stride(format, box->x);
170
171 switch (format) {
172 case PIPE_FORMAT_RGTC1_UNORM:
173 case PIPE_FORMAT_RGTC1_SNORM:
174 case PIPE_FORMAT_LATC1_UNORM:
175 case PIPE_FORMAT_LATC1_SNORM:
176 util_format_rgtc1_unorm_unpack_rgba_8unorm(
177 data, slice->pitch * rsc->cpp,
178 source, trans->base.stride,
179 box->width, box->height);
180 break;
181 case PIPE_FORMAT_RGTC2_UNORM:
182 case PIPE_FORMAT_RGTC2_SNORM:
183 case PIPE_FORMAT_LATC2_UNORM:
184 case PIPE_FORMAT_LATC2_SNORM:
185 util_format_rgtc2_unorm_unpack_rgba_8unorm(
186 data, slice->pitch * rsc->cpp,
187 source, trans->base.stride,
188 box->width, box->height);
189 break;
190 default:
191 assert(!"Unexpected format\n");
192 break;
193 }
194 }
195
196 static void
197 fd_resource_flush(struct fd_transfer *trans, const struct pipe_box *box)
198 {
199 enum pipe_format format = trans->base.resource->format;
200
201 switch (format) {
202 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
203 case PIPE_FORMAT_X32_S8X24_UINT:
204 fd_resource_flush_z32s8(trans, box);
205 break;
206 case PIPE_FORMAT_RGTC1_UNORM:
207 case PIPE_FORMAT_RGTC1_SNORM:
208 case PIPE_FORMAT_RGTC2_UNORM:
209 case PIPE_FORMAT_RGTC2_SNORM:
210 case PIPE_FORMAT_LATC1_UNORM:
211 case PIPE_FORMAT_LATC1_SNORM:
212 case PIPE_FORMAT_LATC2_UNORM:
213 case PIPE_FORMAT_LATC2_SNORM:
214 fd_resource_flush_rgtc(trans, box);
215 break;
216 default:
217 assert(!"Unexpected staging transfer type");
218 break;
219 }
220 }
221
222 static void fd_resource_transfer_flush_region(struct pipe_context *pctx,
223 struct pipe_transfer *ptrans,
224 const struct pipe_box *box)
225 {
226 struct fd_resource *rsc = fd_resource(ptrans->resource);
227 struct fd_transfer *trans = fd_transfer(ptrans);
228
229 if (ptrans->resource->target == PIPE_BUFFER)
230 util_range_add(&rsc->valid_buffer_range,
231 ptrans->box.x + box->x,
232 ptrans->box.x + box->x + box->width);
233
234 if (trans->staging)
235 fd_resource_flush(trans, box);
236 }
237
238 static void
239 fd_resource_transfer_unmap(struct pipe_context *pctx,
240 struct pipe_transfer *ptrans)
241 {
242 struct fd_context *ctx = fd_context(pctx);
243 struct fd_resource *rsc = fd_resource(ptrans->resource);
244 struct fd_transfer *trans = fd_transfer(ptrans);
245
246 if (trans->staging && !(ptrans->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
247 struct pipe_box box;
248 u_box_2d(0, 0, ptrans->box.width, ptrans->box.height, &box);
249 fd_resource_flush(trans, &box);
250 }
251
252 if (!(ptrans->usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
253 fd_bo_cpu_fini(rsc->bo);
254 if (rsc->stencil)
255 fd_bo_cpu_fini(rsc->stencil->bo);
256 }
257
258 util_range_add(&rsc->valid_buffer_range,
259 ptrans->box.x,
260 ptrans->box.x + ptrans->box.width);
261
262 pipe_resource_reference(&ptrans->resource, NULL);
263 util_slab_free(&ctx->transfer_pool, ptrans);
264
265 free(trans->staging);
266 }
267
268 static void *
269 fd_resource_transfer_map(struct pipe_context *pctx,
270 struct pipe_resource *prsc,
271 unsigned level, unsigned usage,
272 const struct pipe_box *box,
273 struct pipe_transfer **pptrans)
274 {
275 struct fd_context *ctx = fd_context(pctx);
276 struct fd_resource *rsc = fd_resource(prsc);
277 struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
278 struct fd_transfer *trans;
279 struct pipe_transfer *ptrans;
280 enum pipe_format format = prsc->format;
281 uint32_t op = 0;
282 uint32_t offset;
283 char *buf;
284 int ret = 0;
285
286 DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc, level, usage,
287 box->width, box->height, box->x, box->y);
288
289 ptrans = util_slab_alloc(&ctx->transfer_pool);
290 if (!ptrans)
291 return NULL;
292
293 /* util_slab_alloc() doesn't zero: */
294 trans = fd_transfer(ptrans);
295 memset(trans, 0, sizeof(*trans));
296
297 pipe_resource_reference(&ptrans->resource, prsc);
298 ptrans->level = level;
299 ptrans->usage = usage;
300 ptrans->box = *box;
301 ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->cpp;
302 ptrans->layer_stride = rsc->layer_first ? rsc->layer_size : slice->size0;
303
304 if (usage & PIPE_TRANSFER_READ)
305 op |= DRM_FREEDRENO_PREP_READ;
306
307 if (usage & PIPE_TRANSFER_WRITE)
308 op |= DRM_FREEDRENO_PREP_WRITE;
309
310 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
311 realloc_bo(rsc, fd_bo_size(rsc->bo));
312 if (rsc->stencil)
313 realloc_bo(rsc->stencil, fd_bo_size(rsc->stencil->bo));
314 fd_invalidate_resource(ctx, prsc);
315 } else if ((usage & PIPE_TRANSFER_WRITE) &&
316 prsc->target == PIPE_BUFFER &&
317 !util_ranges_intersect(&rsc->valid_buffer_range,
318 box->x, box->x + box->width)) {
319 /* We are trying to write to a previously uninitialized range. No need
320 * to wait.
321 */
322 } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
323 /* If the GPU is writing to the resource, or if it is reading from the
324 * resource and we're trying to write to it, flush the renders.
325 */
326 if (((ptrans->usage & PIPE_TRANSFER_WRITE) &&
327 pending(rsc, FD_PENDING_READ | FD_PENDING_WRITE)) ||
328 pending(rsc, FD_PENDING_WRITE))
329 fd_context_render(pctx);
330
331 /* The GPU keeps track of how the various bo's are being used, and
332 * will wait if necessary for the proper operation to have
333 * completed.
334 */
335 ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
336 if (ret)
337 goto fail;
338 }
339
340 buf = fd_bo_map(rsc->bo);
341 if (!buf)
342 goto fail;
343
344 offset = slice->offset +
345 box->y / util_format_get_blockheight(format) * ptrans->stride +
346 box->x / util_format_get_blockwidth(format) * rsc->cpp +
347 fd_resource_layer_offset(rsc, slice, box->z);
348
349 if (prsc->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT ||
350 prsc->format == PIPE_FORMAT_X32_S8X24_UINT) {
351 assert(trans->base.box.depth == 1);
352
353 trans->base.stride = trans->base.box.width * rsc->cpp * 2;
354 trans->staging = malloc(trans->base.stride * trans->base.box.height);
355 if (!trans->staging)
356 goto fail;
357
358 /* if we're not discarding the whole range (or resource), we must copy
359 * the real data in.
360 */
361 if (!(usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
362 PIPE_TRANSFER_DISCARD_RANGE))) {
363 struct fd_resource_slice *sslice =
364 fd_resource_slice(rsc->stencil, level);
365 void *sbuf = fd_bo_map(rsc->stencil->bo);
366 if (!sbuf)
367 goto fail;
368
369 float *depth = (float *)(buf + slice->offset +
370 fd_resource_layer_offset(rsc, slice, box->z) +
371 box->y * slice->pitch * 4 + box->x * 4);
372 uint8_t *stencil = sbuf + sslice->offset +
373 fd_resource_layer_offset(rsc->stencil, sslice, box->z) +
374 box->y * sslice->pitch + box->x;
375
376 if (format != PIPE_FORMAT_X32_S8X24_UINT)
377 util_format_z32_float_s8x24_uint_pack_z_float(
378 trans->staging, trans->base.stride,
379 depth, slice->pitch * 4,
380 box->width, box->height);
381
382 util_format_z32_float_s8x24_uint_pack_s_8uint(
383 trans->staging, trans->base.stride,
384 stencil, sslice->pitch,
385 box->width, box->height);
386 }
387
388 buf = trans->staging;
389 offset = 0;
390 } else if (rsc->internal_format != format &&
391 util_format_description(format)->layout == UTIL_FORMAT_LAYOUT_RGTC) {
392 assert(trans->base.box.depth == 1);
393
394 trans->base.stride = util_format_get_stride(
395 format, trans->base.box.width);
396 trans->staging = malloc(
397 util_format_get_2d_size(format, trans->base.stride,
398 trans->base.box.height));
399 if (!trans->staging)
400 goto fail;
401
402 /* if we're not discarding the whole range (or resource), we must copy
403 * the real data in.
404 */
405 if (!(usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
406 PIPE_TRANSFER_DISCARD_RANGE))) {
407 uint8_t *rgba8 = (uint8_t *)buf + slice->offset +
408 fd_resource_layer_offset(rsc, slice, box->z) +
409 box->y * slice->pitch * rsc->cpp + box->x * rsc->cpp;
410
411 switch (format) {
412 case PIPE_FORMAT_RGTC1_UNORM:
413 case PIPE_FORMAT_RGTC1_SNORM:
414 case PIPE_FORMAT_LATC1_UNORM:
415 case PIPE_FORMAT_LATC1_SNORM:
416 util_format_rgtc1_unorm_pack_rgba_8unorm(
417 trans->staging, trans->base.stride,
418 rgba8, slice->pitch * rsc->cpp,
419 box->width, box->height);
420 break;
421 case PIPE_FORMAT_RGTC2_UNORM:
422 case PIPE_FORMAT_RGTC2_SNORM:
423 case PIPE_FORMAT_LATC2_UNORM:
424 case PIPE_FORMAT_LATC2_SNORM:
425 util_format_rgtc2_unorm_pack_rgba_8unorm(
426 trans->staging, trans->base.stride,
427 rgba8, slice->pitch * rsc->cpp,
428 box->width, box->height);
429 break;
430 default:
431 assert(!"Unexpected format");
432 break;
433 }
434 }
435
436 buf = trans->staging;
437 offset = 0;
438 }
439
440 *pptrans = ptrans;
441
442 return buf + offset;
443
444 fail:
445 fd_resource_transfer_unmap(pctx, ptrans);
446 return NULL;
447 }
448
449 static void
450 fd_resource_destroy(struct pipe_screen *pscreen,
451 struct pipe_resource *prsc)
452 {
453 struct fd_resource *rsc = fd_resource(prsc);
454 if (rsc->bo)
455 fd_bo_del(rsc->bo);
456 list_delinit(&rsc->list);
457 util_range_destroy(&rsc->valid_buffer_range);
458 FREE(rsc);
459 }
460
461 static boolean
462 fd_resource_get_handle(struct pipe_screen *pscreen,
463 struct pipe_resource *prsc,
464 struct winsys_handle *handle)
465 {
466 struct fd_resource *rsc = fd_resource(prsc);
467
468 return fd_screen_bo_get_handle(pscreen, rsc->bo,
469 rsc->slices[0].pitch * rsc->cpp, handle);
470 }
471
472
473 static const struct u_resource_vtbl fd_resource_vtbl = {
474 .resource_get_handle = fd_resource_get_handle,
475 .resource_destroy = fd_resource_destroy,
476 .transfer_map = fd_resource_transfer_map,
477 .transfer_flush_region = fd_resource_transfer_flush_region,
478 .transfer_unmap = fd_resource_transfer_unmap,
479 };
480
481 static uint32_t
482 setup_slices(struct fd_resource *rsc, uint32_t alignment, enum pipe_format format)
483 {
484 struct pipe_resource *prsc = &rsc->base.b;
485 enum util_format_layout layout = util_format_description(format)->layout;
486 uint32_t level, size = 0;
487 uint32_t width = prsc->width0;
488 uint32_t height = prsc->height0;
489 uint32_t depth = prsc->depth0;
490 /* in layer_first layout, the level (slice) contains just one
491 * layer (since in fact the layer contains the slices)
492 */
493 uint32_t layers_in_level = rsc->layer_first ? 1 : prsc->array_size;
494
495 for (level = 0; level <= prsc->last_level; level++) {
496 struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
497 uint32_t blocks;
498
499 if (layout == UTIL_FORMAT_LAYOUT_ASTC)
500 slice->pitch = width =
501 util_align_npot(width, 32 * util_format_get_blockwidth(format));
502 else
503 slice->pitch = width = align(width, 32);
504 slice->offset = size;
505 blocks = util_format_get_nblocks(format, width, height);
506 /* 1d array and 2d array textures must all have the same layer size
507 * for each miplevel on a3xx. 3d textures can have different layer
508 * sizes for high levels, but the hw auto-sizer is buggy (or at least
509 * different than what this code does), so as soon as the layer size
510 * range gets into range, we stop reducing it.
511 */
512 if (prsc->target == PIPE_TEXTURE_3D && (
513 level == 1 ||
514 (level > 1 && rsc->slices[level - 1].size0 > 0xf000)))
515 slice->size0 = align(blocks * rsc->cpp, alignment);
516 else if (level == 0 || rsc->layer_first || alignment == 1)
517 slice->size0 = align(blocks * rsc->cpp, alignment);
518 else
519 slice->size0 = rsc->slices[level - 1].size0;
520
521 size += slice->size0 * depth * layers_in_level;
522
523 width = u_minify(width, 1);
524 height = u_minify(height, 1);
525 depth = u_minify(depth, 1);
526 }
527
528 return size;
529 }
530
531 static uint32_t
532 slice_alignment(struct pipe_screen *pscreen, const struct pipe_resource *tmpl)
533 {
534 /* on a3xx, 2d array and 3d textures seem to want their
535 * layers aligned to page boundaries:
536 */
537 switch (tmpl->target) {
538 case PIPE_TEXTURE_3D:
539 case PIPE_TEXTURE_1D_ARRAY:
540 case PIPE_TEXTURE_2D_ARRAY:
541 return 4096;
542 default:
543 return 1;
544 }
545 }
546
547 /**
548 * Create a new texture object, using the given template info.
549 */
550 static struct pipe_resource *
551 fd_resource_create(struct pipe_screen *pscreen,
552 const struct pipe_resource *tmpl)
553 {
554 struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
555 struct pipe_resource *prsc = &rsc->base.b;
556 enum pipe_format format = tmpl->format;
557 uint32_t size, alignment;
558
559 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
560 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
561 tmpl->target, util_format_name(format),
562 tmpl->width0, tmpl->height0, tmpl->depth0,
563 tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
564 tmpl->usage, tmpl->bind, tmpl->flags);
565
566 if (!rsc)
567 return NULL;
568
569 *prsc = *tmpl;
570
571 pipe_reference_init(&prsc->reference, 1);
572 list_inithead(&rsc->list);
573 prsc->screen = pscreen;
574
575 util_range_init(&rsc->valid_buffer_range);
576
577 rsc->base.vtbl = &fd_resource_vtbl;
578
579 if (format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
580 format = PIPE_FORMAT_Z32_FLOAT;
581 else if (fd_screen(pscreen)->gpu_id < 400 &&
582 util_format_description(format)->layout == UTIL_FORMAT_LAYOUT_RGTC)
583 format = PIPE_FORMAT_R8G8B8A8_UNORM;
584 rsc->internal_format = format;
585 rsc->cpp = util_format_get_blocksize(format);
586
587 assert(rsc->cpp);
588
589 alignment = slice_alignment(pscreen, tmpl);
590 if (is_a4xx(fd_screen(pscreen))) {
591 switch (tmpl->target) {
592 case PIPE_TEXTURE_3D:
593 rsc->layer_first = false;
594 break;
595 default:
596 rsc->layer_first = true;
597 alignment = 1;
598 break;
599 }
600 }
601
602 size = setup_slices(rsc, alignment, format);
603
604 if (rsc->layer_first) {
605 rsc->layer_size = align(size, 4096);
606 size = rsc->layer_size * prsc->array_size;
607 }
608
609 realloc_bo(rsc, size);
610 if (!rsc->bo)
611 goto fail;
612
613 /* There is no native Z32F_S8 sampling or rendering format, so this must
614 * be emulated via two separate textures. The depth texture still keeps
615 * its Z32F_S8 format though, and we also keep a reference to a separate
616 * S8 texture.
617 */
618 if (tmpl->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT) {
619 struct pipe_resource stencil = *tmpl;
620 stencil.format = PIPE_FORMAT_S8_UINT;
621 rsc->stencil = fd_resource(fd_resource_create(pscreen, &stencil));
622 if (!rsc->stencil)
623 goto fail;
624 }
625
626 return prsc;
627 fail:
628 fd_resource_destroy(pscreen, prsc);
629 return NULL;
630 }
631
632 /**
633 * Create a texture from a winsys_handle. The handle is often created in
634 * another process by first creating a pipe texture and then calling
635 * resource_get_handle.
636 */
637 static struct pipe_resource *
638 fd_resource_from_handle(struct pipe_screen *pscreen,
639 const struct pipe_resource *tmpl,
640 struct winsys_handle *handle, unsigned usage)
641 {
642 struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
643 struct fd_resource_slice *slice = &rsc->slices[0];
644 struct pipe_resource *prsc = &rsc->base.b;
645
646 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
647 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
648 tmpl->target, util_format_name(tmpl->format),
649 tmpl->width0, tmpl->height0, tmpl->depth0,
650 tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
651 tmpl->usage, tmpl->bind, tmpl->flags);
652
653 if (!rsc)
654 return NULL;
655
656 *prsc = *tmpl;
657
658 pipe_reference_init(&prsc->reference, 1);
659 list_inithead(&rsc->list);
660 prsc->screen = pscreen;
661
662 util_range_init(&rsc->valid_buffer_range);
663
664 rsc->bo = fd_screen_bo_from_handle(pscreen, handle, &slice->pitch);
665 if (!rsc->bo)
666 goto fail;
667
668 rsc->base.vtbl = &fd_resource_vtbl;
669 rsc->cpp = util_format_get_blocksize(tmpl->format);
670 slice->pitch /= rsc->cpp;
671 slice->offset = handle->offset;
672
673 assert(rsc->cpp);
674
675 return prsc;
676
677 fail:
678 fd_resource_destroy(pscreen, prsc);
679 return NULL;
680 }
681
682 static void fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond);
683 static void fd_blitter_pipe_end(struct fd_context *ctx);
684
685 /**
686 * _copy_region using pipe (3d engine)
687 */
688 static bool
689 fd_blitter_pipe_copy_region(struct fd_context *ctx,
690 struct pipe_resource *dst,
691 unsigned dst_level,
692 unsigned dstx, unsigned dsty, unsigned dstz,
693 struct pipe_resource *src,
694 unsigned src_level,
695 const struct pipe_box *src_box)
696 {
697 /* not until we allow rendertargets to be buffers */
698 if (dst->target == PIPE_BUFFER || src->target == PIPE_BUFFER)
699 return false;
700
701 if (!util_blitter_is_copy_supported(ctx->blitter, dst, src))
702 return false;
703
704 fd_blitter_pipe_begin(ctx, false);
705 util_blitter_copy_texture(ctx->blitter,
706 dst, dst_level, dstx, dsty, dstz,
707 src, src_level, src_box);
708 fd_blitter_pipe_end(ctx);
709
710 return true;
711 }
712
713 /**
714 * Copy a block of pixels from one resource to another.
715 * The resource must be of the same format.
716 * Resources with nr_samples > 1 are not allowed.
717 */
718 static void
719 fd_resource_copy_region(struct pipe_context *pctx,
720 struct pipe_resource *dst,
721 unsigned dst_level,
722 unsigned dstx, unsigned dsty, unsigned dstz,
723 struct pipe_resource *src,
724 unsigned src_level,
725 const struct pipe_box *src_box)
726 {
727 struct fd_context *ctx = fd_context(pctx);
728
729 /* TODO if we have 2d core, or other DMA engine that could be used
730 * for simple copies and reasonably easily synchronized with the 3d
731 * core, this is where we'd plug it in..
732 */
733
734 /* try blit on 3d pipe: */
735 if (fd_blitter_pipe_copy_region(ctx,
736 dst, dst_level, dstx, dsty, dstz,
737 src, src_level, src_box))
738 return;
739
740 /* else fallback to pure sw: */
741 util_resource_copy_region(pctx,
742 dst, dst_level, dstx, dsty, dstz,
743 src, src_level, src_box);
744 }
745
746 bool
747 fd_render_condition_check(struct pipe_context *pctx)
748 {
749 struct fd_context *ctx = fd_context(pctx);
750
751 if (!ctx->cond_query)
752 return true;
753
754 union pipe_query_result res = { 0 };
755 bool wait =
756 ctx->cond_mode != PIPE_RENDER_COND_NO_WAIT &&
757 ctx->cond_mode != PIPE_RENDER_COND_BY_REGION_NO_WAIT;
758
759 if (pctx->get_query_result(pctx, ctx->cond_query, wait, &res))
760 return (bool)res.u64 != ctx->cond_cond;
761
762 return true;
763 }
764
765 /**
766 * Optimal hardware path for blitting pixels.
767 * Scaling, format conversion, up- and downsampling (resolve) are allowed.
768 */
769 static void
770 fd_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
771 {
772 struct fd_context *ctx = fd_context(pctx);
773 struct pipe_blit_info info = *blit_info;
774
775 if (info.src.resource->nr_samples > 1 &&
776 info.dst.resource->nr_samples <= 1 &&
777 !util_format_is_depth_or_stencil(info.src.resource->format) &&
778 !util_format_is_pure_integer(info.src.resource->format)) {
779 DBG("color resolve unimplemented");
780 return;
781 }
782
783 if (info.render_condition_enable && !fd_render_condition_check(pctx))
784 return;
785
786 if (util_try_blit_via_copy_region(pctx, &info)) {
787 return; /* done */
788 }
789
790 if (info.mask & PIPE_MASK_S) {
791 DBG("cannot blit stencil, skipping");
792 info.mask &= ~PIPE_MASK_S;
793 }
794
795 if (!util_blitter_is_blit_supported(ctx->blitter, &info)) {
796 DBG("blit unsupported %s -> %s",
797 util_format_short_name(info.src.resource->format),
798 util_format_short_name(info.dst.resource->format));
799 return;
800 }
801
802 fd_blitter_pipe_begin(ctx, info.render_condition_enable);
803 util_blitter_blit(ctx->blitter, &info);
804 fd_blitter_pipe_end(ctx);
805 }
806
807 static void
808 fd_blitter_pipe_begin(struct fd_context *ctx, bool render_cond)
809 {
810 util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
811 util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
812 util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
813 util_blitter_save_so_targets(ctx->blitter, ctx->streamout.num_targets,
814 ctx->streamout.targets);
815 util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer);
816 util_blitter_save_viewport(ctx->blitter, &ctx->viewport);
817 util_blitter_save_scissor(ctx->blitter, &ctx->scissor);
818 util_blitter_save_fragment_shader(ctx->blitter, ctx->prog.fp);
819 util_blitter_save_blend(ctx->blitter, ctx->blend);
820 util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa);
821 util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref);
822 util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask);
823 util_blitter_save_framebuffer(ctx->blitter, &ctx->framebuffer);
824 util_blitter_save_fragment_sampler_states(ctx->blitter,
825 ctx->fragtex.num_samplers,
826 (void **)ctx->fragtex.samplers);
827 util_blitter_save_fragment_sampler_views(ctx->blitter,
828 ctx->fragtex.num_textures, ctx->fragtex.textures);
829 if (!render_cond)
830 util_blitter_save_render_condition(ctx->blitter,
831 ctx->cond_query, ctx->cond_cond, ctx->cond_mode);
832
833 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_BLIT);
834 }
835
836 static void
837 fd_blitter_pipe_end(struct fd_context *ctx)
838 {
839 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
840 }
841
842 static void
843 fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
844 {
845 struct fd_resource *rsc = fd_resource(prsc);
846
847 if (pending(rsc, FD_PENDING_WRITE | FD_PENDING_READ))
848 fd_context_render(pctx);
849 }
850
851 void
852 fd_resource_screen_init(struct pipe_screen *pscreen)
853 {
854 pscreen->resource_create = fd_resource_create;
855 pscreen->resource_from_handle = fd_resource_from_handle;
856 pscreen->resource_get_handle = u_resource_get_handle_vtbl;
857 pscreen->resource_destroy = u_resource_destroy_vtbl;
858 }
859
860 void
861 fd_resource_context_init(struct pipe_context *pctx)
862 {
863 pctx->transfer_map = u_transfer_map_vtbl;
864 pctx->transfer_flush_region = u_transfer_flush_region_vtbl;
865 pctx->transfer_unmap = u_transfer_unmap_vtbl;
866 pctx->buffer_subdata = u_default_buffer_subdata;
867 pctx->texture_subdata = u_default_texture_subdata;
868 pctx->create_surface = fd_create_surface;
869 pctx->surface_destroy = fd_surface_destroy;
870 pctx->resource_copy_region = fd_resource_copy_region;
871 pctx->blit = fd_blit;
872 pctx->flush_resource = fd_flush_resource;
873 }