freedreno/a3xx: fix 3d texture layout
[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_inlines.h"
31 #include "util/u_transfer.h"
32 #include "util/u_string.h"
33 #include "util/u_surface.h"
34
35 #include "freedreno_resource.h"
36 #include "freedreno_screen.h"
37 #include "freedreno_surface.h"
38 #include "freedreno_context.h"
39 #include "freedreno_query_hw.h"
40 #include "freedreno_util.h"
41
42 #include <errno.h>
43
44 static void
45 realloc_bo(struct fd_resource *rsc, uint32_t size)
46 {
47 struct fd_screen *screen = fd_screen(rsc->base.b.screen);
48 uint32_t flags = DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
49 DRM_FREEDRENO_GEM_TYPE_KMEM; /* TODO */
50
51 /* if we start using things other than write-combine,
52 * be sure to check for PIPE_RESOURCE_FLAG_MAP_COHERENT
53 */
54
55 if (rsc->bo)
56 fd_bo_del(rsc->bo);
57
58 rsc->bo = fd_bo_new(screen->dev, size, flags);
59 rsc->timestamp = 0;
60 rsc->dirty = false;
61 }
62
63 static void fd_resource_transfer_flush_region(struct pipe_context *pctx,
64 struct pipe_transfer *ptrans,
65 const struct pipe_box *box)
66 {
67 struct fd_context *ctx = fd_context(pctx);
68 struct fd_resource *rsc = fd_resource(ptrans->resource);
69
70 if (rsc->dirty)
71 fd_context_render(pctx);
72
73 if (rsc->timestamp) {
74 fd_pipe_wait(ctx->screen->pipe, rsc->timestamp);
75 rsc->timestamp = 0;
76 }
77 }
78
79 static void
80 fd_resource_transfer_unmap(struct pipe_context *pctx,
81 struct pipe_transfer *ptrans)
82 {
83 struct fd_context *ctx = fd_context(pctx);
84 struct fd_resource *rsc = fd_resource(ptrans->resource);
85 if (!(ptrans->usage & PIPE_TRANSFER_UNSYNCHRONIZED))
86 fd_bo_cpu_fini(rsc->bo);
87 pipe_resource_reference(&ptrans->resource, NULL);
88 util_slab_free(&ctx->transfer_pool, ptrans);
89 }
90
91 static void *
92 fd_resource_transfer_map(struct pipe_context *pctx,
93 struct pipe_resource *prsc,
94 unsigned level, unsigned usage,
95 const struct pipe_box *box,
96 struct pipe_transfer **pptrans)
97 {
98 struct fd_context *ctx = fd_context(pctx);
99 struct fd_resource *rsc = fd_resource(prsc);
100 struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
101 struct pipe_transfer *ptrans;
102 enum pipe_format format = prsc->format;
103 uint32_t op = 0;
104 uint32_t offset;
105 char *buf;
106 int ret = 0;
107
108 DBG("prsc=%p, level=%u, usage=%x", prsc, level, usage);
109
110 ptrans = util_slab_alloc(&ctx->transfer_pool);
111 if (!ptrans)
112 return NULL;
113
114 /* util_slab_alloc() doesn't zero: */
115 memset(ptrans, 0, sizeof(*ptrans));
116
117 pipe_resource_reference(&ptrans->resource, prsc);
118 ptrans->level = level;
119 ptrans->usage = usage;
120 ptrans->box = *box;
121 ptrans->stride = slice->pitch * rsc->cpp;
122 ptrans->layer_stride = slice->size0;
123
124 if (usage & PIPE_TRANSFER_READ)
125 op |= DRM_FREEDRENO_PREP_READ;
126
127 if (usage & PIPE_TRANSFER_WRITE)
128 op |= DRM_FREEDRENO_PREP_WRITE;
129
130 /* some state trackers (at least XA) don't do this.. */
131 if (!(usage & (PIPE_TRANSFER_FLUSH_EXPLICIT | PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)))
132 fd_resource_transfer_flush_region(pctx, ptrans, box);
133
134 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
135 realloc_bo(rsc, fd_bo_size(rsc->bo));
136 } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
137 ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
138 if (ret)
139 goto fail;
140 }
141
142 buf = fd_bo_map(rsc->bo);
143 if (!buf) {
144 fd_resource_transfer_unmap(pctx, ptrans);
145 return NULL;
146 }
147
148 *pptrans = ptrans;
149
150 if (rsc->layer_first) {
151 offset = slice->offset +
152 box->y / util_format_get_blockheight(format) * ptrans->stride +
153 box->x / util_format_get_blockwidth(format) * rsc->cpp +
154 box->z * rsc->layer_size;
155 } else {
156 offset = slice->offset +
157 box->y / util_format_get_blockheight(format) * ptrans->stride +
158 box->x / util_format_get_blockwidth(format) * rsc->cpp +
159 box->z * slice->size0;
160 }
161
162 return buf + offset;
163
164 fail:
165 fd_resource_transfer_unmap(pctx, ptrans);
166 return NULL;
167 }
168
169 static void
170 fd_resource_destroy(struct pipe_screen *pscreen,
171 struct pipe_resource *prsc)
172 {
173 struct fd_resource *rsc = fd_resource(prsc);
174 if (rsc->bo)
175 fd_bo_del(rsc->bo);
176 FREE(rsc);
177 }
178
179 static boolean
180 fd_resource_get_handle(struct pipe_screen *pscreen,
181 struct pipe_resource *prsc,
182 struct winsys_handle *handle)
183 {
184 struct fd_resource *rsc = fd_resource(prsc);
185
186 return fd_screen_bo_get_handle(pscreen, rsc->bo,
187 rsc->slices[0].pitch * rsc->cpp, handle);
188 }
189
190
191 static const struct u_resource_vtbl fd_resource_vtbl = {
192 .resource_get_handle = fd_resource_get_handle,
193 .resource_destroy = fd_resource_destroy,
194 .transfer_map = fd_resource_transfer_map,
195 .transfer_flush_region = fd_resource_transfer_flush_region,
196 .transfer_unmap = fd_resource_transfer_unmap,
197 .transfer_inline_write = u_default_transfer_inline_write,
198 };
199
200 static uint32_t
201 setup_slices(struct fd_resource *rsc, uint32_t alignment)
202 {
203 struct pipe_resource *prsc = &rsc->base.b;
204 uint32_t level, size = 0;
205 uint32_t width = prsc->width0;
206 uint32_t height = prsc->height0;
207 uint32_t depth = prsc->depth0;
208 /* in layer_first layout, the level (slice) contains just one
209 * layer (since in fact the layer contains the slices)
210 */
211 uint32_t layers_in_level = rsc->layer_first ? 1 : prsc->array_size;
212
213 for (level = 0; level <= prsc->last_level; level++) {
214 struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
215
216 slice->pitch = width = align(width, 32);
217 slice->offset = size;
218 /* 1d array and 2d array textures must all have the same layer size
219 * for each miplevel on a3xx. 3d textures can have different layer
220 * sizes for high levels, but the hw auto-sizer is buggy (or at least
221 * different than what this code does), so as soon as the layer size
222 * range gets into range, we stop reducing it.
223 */
224 if (prsc->target == PIPE_TEXTURE_3D && (
225 level == 1 ||
226 (level > 1 && rsc->slices[level - 1].size0 > 0xf000)))
227 slice->size0 = align(slice->pitch * height * rsc->cpp, alignment);
228 else if (level == 0 || rsc->layer_first || alignment == 1)
229 slice->size0 = align(slice->pitch * height * rsc->cpp, alignment);
230 else
231 slice->size0 = rsc->slices[level - 1].size0;
232
233 size += slice->size0 * depth * layers_in_level;
234
235 width = u_minify(width, 1);
236 height = u_minify(height, 1);
237 depth = u_minify(depth, 1);
238 }
239
240 return size;
241 }
242
243 static uint32_t
244 slice_alignment(struct pipe_screen *pscreen, const struct pipe_resource *tmpl)
245 {
246 /* on a3xx, 2d array and 3d textures seem to want their
247 * layers aligned to page boundaries:
248 */
249 switch (tmpl->target) {
250 case PIPE_TEXTURE_3D:
251 case PIPE_TEXTURE_1D_ARRAY:
252 case PIPE_TEXTURE_2D_ARRAY:
253 return 4096;
254 default:
255 return 1;
256 }
257 }
258
259 /**
260 * Create a new texture object, using the given template info.
261 */
262 static struct pipe_resource *
263 fd_resource_create(struct pipe_screen *pscreen,
264 const struct pipe_resource *tmpl)
265 {
266 struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
267 struct pipe_resource *prsc = &rsc->base.b;
268 uint32_t size;
269
270 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
271 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
272 tmpl->target, util_format_name(tmpl->format),
273 tmpl->width0, tmpl->height0, tmpl->depth0,
274 tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
275 tmpl->usage, tmpl->bind, tmpl->flags);
276
277 if (!rsc)
278 return NULL;
279
280 *prsc = *tmpl;
281
282 pipe_reference_init(&prsc->reference, 1);
283 prsc->screen = pscreen;
284
285 rsc->base.vtbl = &fd_resource_vtbl;
286 rsc->cpp = util_format_get_blocksize(tmpl->format);
287
288 assert(rsc->cpp);
289
290 if (is_a4xx(fd_screen(pscreen))) {
291 switch (tmpl->target) {
292 case PIPE_TEXTURE_3D:
293 /* TODO 3D_ARRAY? */
294 rsc->layer_first = false;
295 break;
296 default:
297 rsc->layer_first = true;
298 break;
299 }
300 }
301
302 size = setup_slices(rsc, slice_alignment(pscreen, tmpl));
303
304 if (rsc->layer_first) {
305 rsc->layer_size = align(size, 4096);
306 size = rsc->layer_size * prsc->array_size;
307 }
308
309 realloc_bo(rsc, size);
310 if (!rsc->bo)
311 goto fail;
312
313 return prsc;
314 fail:
315 fd_resource_destroy(pscreen, prsc);
316 return NULL;
317 }
318
319 /**
320 * Create a texture from a winsys_handle. The handle is often created in
321 * another process by first creating a pipe texture and then calling
322 * resource_get_handle.
323 */
324 static struct pipe_resource *
325 fd_resource_from_handle(struct pipe_screen *pscreen,
326 const struct pipe_resource *tmpl,
327 struct winsys_handle *handle)
328 {
329 struct fd_resource *rsc = CALLOC_STRUCT(fd_resource);
330 struct fd_resource_slice *slice = &rsc->slices[0];
331 struct pipe_resource *prsc = &rsc->base.b;
332
333 DBG("target=%d, format=%s, %ux%ux%u, array_size=%u, last_level=%u, "
334 "nr_samples=%u, usage=%u, bind=%x, flags=%x",
335 tmpl->target, util_format_name(tmpl->format),
336 tmpl->width0, tmpl->height0, tmpl->depth0,
337 tmpl->array_size, tmpl->last_level, tmpl->nr_samples,
338 tmpl->usage, tmpl->bind, tmpl->flags);
339
340 if (!rsc)
341 return NULL;
342
343 *prsc = *tmpl;
344
345 pipe_reference_init(&prsc->reference, 1);
346 prsc->screen = pscreen;
347
348 rsc->bo = fd_screen_bo_from_handle(pscreen, handle, &slice->pitch);
349 if (!rsc->bo)
350 goto fail;
351
352 rsc->base.vtbl = &fd_resource_vtbl;
353 rsc->cpp = util_format_get_blocksize(tmpl->format);
354 slice->pitch /= rsc->cpp;
355
356 assert(rsc->cpp);
357
358 return prsc;
359
360 fail:
361 fd_resource_destroy(pscreen, prsc);
362 return NULL;
363 }
364
365 static void fd_blitter_pipe_begin(struct fd_context *ctx);
366 static void fd_blitter_pipe_end(struct fd_context *ctx);
367
368 /**
369 * _copy_region using pipe (3d engine)
370 */
371 static bool
372 fd_blitter_pipe_copy_region(struct fd_context *ctx,
373 struct pipe_resource *dst,
374 unsigned dst_level,
375 unsigned dstx, unsigned dsty, unsigned dstz,
376 struct pipe_resource *src,
377 unsigned src_level,
378 const struct pipe_box *src_box)
379 {
380 /* not until we allow rendertargets to be buffers */
381 if (dst->target == PIPE_BUFFER || src->target == PIPE_BUFFER)
382 return false;
383
384 if (!util_blitter_is_copy_supported(ctx->blitter, dst, src))
385 return false;
386
387 fd_blitter_pipe_begin(ctx);
388 util_blitter_copy_texture(ctx->blitter,
389 dst, dst_level, dstx, dsty, dstz,
390 src, src_level, src_box);
391 fd_blitter_pipe_end(ctx);
392
393 return true;
394 }
395
396 /**
397 * Copy a block of pixels from one resource to another.
398 * The resource must be of the same format.
399 * Resources with nr_samples > 1 are not allowed.
400 */
401 static void
402 fd_resource_copy_region(struct pipe_context *pctx,
403 struct pipe_resource *dst,
404 unsigned dst_level,
405 unsigned dstx, unsigned dsty, unsigned dstz,
406 struct pipe_resource *src,
407 unsigned src_level,
408 const struct pipe_box *src_box)
409 {
410 struct fd_context *ctx = fd_context(pctx);
411
412 /* TODO if we have 2d core, or other DMA engine that could be used
413 * for simple copies and reasonably easily synchronized with the 3d
414 * core, this is where we'd plug it in..
415 */
416
417 /* try blit on 3d pipe: */
418 if (fd_blitter_pipe_copy_region(ctx,
419 dst, dst_level, dstx, dsty, dstz,
420 src, src_level, src_box))
421 return;
422
423 /* else fallback to pure sw: */
424 util_resource_copy_region(pctx,
425 dst, dst_level, dstx, dsty, dstz,
426 src, src_level, src_box);
427 }
428
429 /**
430 * Optimal hardware path for blitting pixels.
431 * Scaling, format conversion, up- and downsampling (resolve) are allowed.
432 */
433 static void
434 fd_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
435 {
436 struct fd_context *ctx = fd_context(pctx);
437 struct pipe_blit_info info = *blit_info;
438
439 if (info.src.resource->nr_samples > 1 &&
440 info.dst.resource->nr_samples <= 1 &&
441 !util_format_is_depth_or_stencil(info.src.resource->format) &&
442 !util_format_is_pure_integer(info.src.resource->format)) {
443 DBG("color resolve unimplemented");
444 return;
445 }
446
447 if (util_try_blit_via_copy_region(pctx, &info)) {
448 return; /* done */
449 }
450
451 if (info.mask & PIPE_MASK_S) {
452 DBG("cannot blit stencil, skipping");
453 info.mask &= ~PIPE_MASK_S;
454 }
455
456 if (!util_blitter_is_blit_supported(ctx->blitter, &info)) {
457 DBG("blit unsupported %s -> %s",
458 util_format_short_name(info.src.resource->format),
459 util_format_short_name(info.dst.resource->format));
460 return;
461 }
462
463 fd_blitter_pipe_begin(ctx);
464 util_blitter_blit(ctx->blitter, &info);
465 fd_blitter_pipe_end(ctx);
466 }
467
468 static void
469 fd_blitter_pipe_begin(struct fd_context *ctx)
470 {
471 util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
472 util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
473 util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
474 util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer);
475 util_blitter_save_viewport(ctx->blitter, &ctx->viewport);
476 util_blitter_save_scissor(ctx->blitter, &ctx->scissor);
477 util_blitter_save_fragment_shader(ctx->blitter, ctx->prog.fp);
478 util_blitter_save_blend(ctx->blitter, ctx->blend);
479 util_blitter_save_depth_stencil_alpha(ctx->blitter, ctx->zsa);
480 util_blitter_save_stencil_ref(ctx->blitter, &ctx->stencil_ref);
481 util_blitter_save_sample_mask(ctx->blitter, ctx->sample_mask);
482 util_blitter_save_framebuffer(ctx->blitter, &ctx->framebuffer);
483 util_blitter_save_fragment_sampler_states(ctx->blitter,
484 ctx->fragtex.num_samplers,
485 (void **)ctx->fragtex.samplers);
486 util_blitter_save_fragment_sampler_views(ctx->blitter,
487 ctx->fragtex.num_textures, ctx->fragtex.textures);
488
489 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_BLIT);
490 }
491
492 static void
493 fd_blitter_pipe_end(struct fd_context *ctx)
494 {
495 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
496 }
497
498 static void
499 fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
500 {
501 struct fd_resource *rsc = fd_resource(prsc);
502
503 if (rsc->dirty)
504 fd_context_render(pctx);
505 }
506
507 void
508 fd_resource_screen_init(struct pipe_screen *pscreen)
509 {
510 pscreen->resource_create = fd_resource_create;
511 pscreen->resource_from_handle = fd_resource_from_handle;
512 pscreen->resource_get_handle = u_resource_get_handle_vtbl;
513 pscreen->resource_destroy = u_resource_destroy_vtbl;
514 }
515
516 void
517 fd_resource_context_init(struct pipe_context *pctx)
518 {
519 pctx->transfer_map = u_transfer_map_vtbl;
520 pctx->transfer_flush_region = u_transfer_flush_region_vtbl;
521 pctx->transfer_unmap = u_transfer_unmap_vtbl;
522 pctx->transfer_inline_write = u_transfer_inline_write_vtbl;
523 pctx->create_surface = fd_create_surface;
524 pctx->surface_destroy = fd_surface_destroy;
525 pctx->resource_copy_region = fd_resource_copy_region;
526 pctx->blit = fd_blit;
527 pctx->flush_resource = fd_flush_resource;
528 }