nv50: account for pesky prefetch in size calculation of linear textures
[mesa.git] / src / gallium / drivers / nv50 / nv50_miptree.c
1 /*
2 * Copyright 2008 Ben Skeggs
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23 #include "pipe/p_state.h"
24 #include "pipe/p_defines.h"
25 #include "util/u_inlines.h"
26 #include "util/u_format.h"
27
28 #include "nv50_context.h"
29 #include "nv50_resource.h"
30
31 uint32_t
32 nv50_tex_choose_tile_dims_helper(unsigned nx, unsigned ny, unsigned nz)
33 {
34 uint32_t tile_mode = 0x000;
35
36 if (ny > 64) tile_mode = 0x040; /* height 128 tiles */
37 else
38 if (ny > 32) tile_mode = 0x030; /* height 64 tiles */
39 else
40 if (ny > 16) tile_mode = 0x020; /* height 32 tiles */
41 else
42 if (ny > 8) tile_mode = 0x010; /* height 16 tiles */
43
44 if (nz == 1)
45 return tile_mode;
46 else
47 if (tile_mode > 0x020)
48 tile_mode = 0x020;
49
50 if (nz > 16 && tile_mode < 0x020)
51 return tile_mode | 0x500; /* depth 32 tiles */
52 if (nz > 8) return tile_mode | 0x400; /* depth 16 tiles */
53 if (nz > 4) return tile_mode | 0x300; /* depth 8 tiles */
54 if (nz > 2) return tile_mode | 0x200; /* depth 4 tiles */
55
56 return tile_mode | 0x100;
57 }
58
59 static uint32_t
60 nv50_tex_choose_tile_dims(unsigned nx, unsigned ny, unsigned nz)
61 {
62 return nv50_tex_choose_tile_dims_helper(nx, ny * 2, nz);
63 }
64
65 static uint32_t
66 nv50_mt_choose_storage_type(struct nv50_miptree *mt, boolean compressed)
67 {
68 const unsigned ms = mt->ms_x + mt->ms_y;
69
70 uint32_t tile_flags;
71
72 if (unlikely(mt->base.base.flags & NOUVEAU_RESOURCE_FLAG_LINEAR))
73 return 0;
74 if (unlikely(mt->base.base.bind & PIPE_BIND_CURSOR))
75 return 0;
76
77 switch (mt->base.base.format) {
78 case PIPE_FORMAT_Z16_UNORM:
79 tile_flags = 0x6c + ms;
80 break;
81 case PIPE_FORMAT_S8_UINT_Z24_UNORM:
82 tile_flags = 0x18 + ms;
83 break;
84 case PIPE_FORMAT_Z24X8_UNORM:
85 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
86 tile_flags = 0x128 + ms;
87 break;
88 case PIPE_FORMAT_Z32_FLOAT:
89 tile_flags = 0x40 + ms;
90 break;
91 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
92 tile_flags = 0x60 + ms;
93 break;
94 default:
95 switch (util_format_get_blocksizebits(mt->base.base.format)) {
96 case 128:
97 assert(ms < 3);
98 tile_flags = 0x74;
99 break;
100 case 64:
101 switch (ms) {
102 case 2: tile_flags = 0xfc; break;
103 case 3: tile_flags = 0xfd; break;
104 default:
105 tile_flags = 0x70;
106 break;
107 }
108 break;
109 case 32:
110 if (mt->base.base.bind & PIPE_BIND_SCANOUT) {
111 assert(ms == 0);
112 tile_flags = 0x7a;
113 } else {
114 switch (ms) {
115 case 2: tile_flags = 0xf8; break;
116 case 3: tile_flags = 0xf9; break;
117 default:
118 tile_flags = 0x70;
119 break;
120 }
121 }
122 break;
123 case 16:
124 case 8:
125 tile_flags = 0x70;
126 break;
127 default:
128 return 0;
129 }
130 if (mt->base.base.bind & PIPE_BIND_CURSOR)
131 tile_flags = 0;
132 }
133
134 if (!compressed)
135 tile_flags &= ~0x180;
136
137 return tile_flags;
138 }
139
140 void
141 nv50_miptree_destroy(struct pipe_screen *pscreen, struct pipe_resource *pt)
142 {
143 struct nv50_miptree *mt = nv50_miptree(pt);
144
145 nouveau_bo_ref(NULL, &mt->base.bo);
146
147 nouveau_fence_ref(NULL, &mt->base.fence);
148 nouveau_fence_ref(NULL, &mt->base.fence_wr);
149
150 NOUVEAU_DRV_STAT(nouveau_screen(pscreen), tex_obj_current_count, -1);
151 NOUVEAU_DRV_STAT(nouveau_screen(pscreen), tex_obj_current_bytes,
152 -(uint64_t)mt->total_size);
153
154 FREE(mt);
155 }
156
157 boolean
158 nv50_miptree_get_handle(struct pipe_screen *pscreen,
159 struct pipe_resource *pt,
160 struct winsys_handle *whandle)
161 {
162 struct nv50_miptree *mt = nv50_miptree(pt);
163 unsigned stride;
164
165 if (!mt || !mt->base.bo)
166 return FALSE;
167
168 stride = mt->level[0].pitch;
169
170 return nouveau_screen_bo_get_handle(pscreen,
171 mt->base.bo,
172 stride,
173 whandle);
174 }
175
176 const struct u_resource_vtbl nv50_miptree_vtbl =
177 {
178 nv50_miptree_get_handle, /* get_handle */
179 nv50_miptree_destroy, /* resource_destroy */
180 nv50_miptree_transfer_map, /* transfer_map */
181 u_default_transfer_flush_region, /* transfer_flush_region */
182 nv50_miptree_transfer_unmap, /* transfer_unmap */
183 u_default_transfer_inline_write /* transfer_inline_write */
184 };
185
186 static INLINE boolean
187 nv50_miptree_init_ms_mode(struct nv50_miptree *mt)
188 {
189 switch (mt->base.base.nr_samples) {
190 case 8:
191 mt->ms_mode = NV50_3D_MULTISAMPLE_MODE_MS8;
192 mt->ms_x = 2;
193 mt->ms_y = 1;
194 break;
195 case 4:
196 mt->ms_mode = NV50_3D_MULTISAMPLE_MODE_MS4;
197 mt->ms_x = 1;
198 mt->ms_y = 1;
199 break;
200 case 2:
201 mt->ms_mode = NV50_3D_MULTISAMPLE_MODE_MS2;
202 mt->ms_x = 1;
203 break;
204 case 1:
205 case 0:
206 mt->ms_mode = NV50_3D_MULTISAMPLE_MODE_MS1;
207 break;
208 default:
209 NOUVEAU_ERR("invalid nr_samples: %u\n", mt->base.base.nr_samples);
210 return FALSE;
211 }
212 return TRUE;
213 }
214
215 boolean
216 nv50_miptree_init_layout_linear(struct nv50_miptree *mt)
217 {
218 struct pipe_resource *pt = &mt->base.base;
219 const unsigned blocksize = util_format_get_blocksize(pt->format);
220 unsigned h = pt->height0;
221
222 if (util_format_is_depth_or_stencil(pt->format))
223 return FALSE;
224
225 if ((pt->last_level > 0) || (pt->depth0 > 1) || (pt->array_size > 1))
226 return FALSE;
227 if (mt->ms_x | mt->ms_y)
228 return FALSE;
229
230 mt->level[0].pitch = align(pt->width0 * blocksize, 64);
231
232 /* Account for very generous prefetch (allocate size as if tiled). */
233 h = MAX2(h, 8);
234 h = util_next_power_of_two(h);
235
236 mt->total_size = mt->level[0].pitch * h;
237
238 return TRUE;
239 }
240
241 static void
242 nv50_miptree_init_layout_tiled(struct nv50_miptree *mt)
243 {
244 struct pipe_resource *pt = &mt->base.base;
245 unsigned w, h, d, l;
246 const unsigned blocksize = util_format_get_blocksize(pt->format);
247
248 mt->layout_3d = pt->target == PIPE_TEXTURE_3D;
249
250 w = pt->width0 << mt->ms_x;
251 h = pt->height0 << mt->ms_y;
252
253 /* For 3D textures, a mipmap is spanned by all the layers, for array
254 * textures and cube maps, each layer contains its own mipmaps.
255 */
256 d = mt->layout_3d ? pt->depth0 : 1;
257
258 for (l = 0; l <= pt->last_level; ++l) {
259 struct nv50_miptree_level *lvl = &mt->level[l];
260 unsigned tsx, tsy, tsz;
261 unsigned nbx = util_format_get_nblocksx(pt->format, w);
262 unsigned nby = util_format_get_nblocksy(pt->format, h);
263
264 lvl->offset = mt->total_size;
265
266 lvl->tile_mode = nv50_tex_choose_tile_dims(nbx, nby, d);
267
268 tsx = NV50_TILE_SIZE_X(lvl->tile_mode); /* x is tile row pitch in bytes */
269 tsy = NV50_TILE_SIZE_Y(lvl->tile_mode);
270 tsz = NV50_TILE_SIZE_Z(lvl->tile_mode);
271
272 lvl->pitch = align(nbx * blocksize, tsx);
273
274 mt->total_size += lvl->pitch * align(nby, tsy) * align(d, tsz);
275
276 w = u_minify(w, 1);
277 h = u_minify(h, 1);
278 d = u_minify(d, 1);
279 }
280
281 if (pt->array_size > 1) {
282 mt->layer_stride = align(mt->total_size,
283 NV50_TILE_SIZE(mt->level[0].tile_mode));
284 mt->total_size = mt->layer_stride * pt->array_size;
285 }
286 }
287
288 struct pipe_resource *
289 nv50_miptree_create(struct pipe_screen *pscreen,
290 const struct pipe_resource *templ)
291 {
292 struct nouveau_device *dev = nouveau_screen(pscreen)->device;
293 struct nv50_miptree *mt = CALLOC_STRUCT(nv50_miptree);
294 struct pipe_resource *pt = &mt->base.base;
295 int ret;
296 union nouveau_bo_config bo_config;
297 uint32_t bo_flags;
298
299 if (!mt)
300 return NULL;
301
302 mt->base.vtbl = &nv50_miptree_vtbl;
303 *pt = *templ;
304 pipe_reference_init(&pt->reference, 1);
305 pt->screen = pscreen;
306
307 bo_config.nv50.memtype = nv50_mt_choose_storage_type(mt, TRUE);
308
309 if (!nv50_miptree_init_ms_mode(mt)) {
310 FREE(mt);
311 return NULL;
312 }
313
314 if (bo_config.nv50.memtype != 0) {
315 nv50_miptree_init_layout_tiled(mt);
316 } else
317 if (!nv50_miptree_init_layout_linear(mt)) {
318 FREE(mt);
319 return NULL;
320 }
321 bo_config.nv50.tile_mode = mt->level[0].tile_mode;
322
323 bo_flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_NOSNOOP;
324 if (mt->base.base.bind & (PIPE_BIND_CURSOR | PIPE_BIND_DISPLAY_TARGET))
325 bo_flags |= NOUVEAU_BO_CONTIG;
326
327 ret = nouveau_bo_new(dev, bo_flags, 4096, mt->total_size, &bo_config,
328 &mt->base.bo);
329 if (ret) {
330 FREE(mt);
331 return NULL;
332 }
333 mt->base.domain = NOUVEAU_BO_VRAM;
334 mt->base.address = mt->base.bo->offset;
335
336 return pt;
337 }
338
339 struct pipe_resource *
340 nv50_miptree_from_handle(struct pipe_screen *pscreen,
341 const struct pipe_resource *templ,
342 struct winsys_handle *whandle)
343 {
344 struct nv50_miptree *mt;
345 unsigned stride;
346
347 /* only supports 2D, non-mipmapped textures for the moment */
348 if ((templ->target != PIPE_TEXTURE_2D &&
349 templ->target != PIPE_TEXTURE_RECT) ||
350 templ->last_level != 0 ||
351 templ->depth0 != 1 ||
352 templ->array_size > 1)
353 return NULL;
354
355 mt = CALLOC_STRUCT(nv50_miptree);
356 if (!mt)
357 return NULL;
358
359 mt->base.bo = nouveau_screen_bo_from_handle(pscreen, whandle, &stride);
360 if (mt->base.bo == NULL) {
361 FREE(mt);
362 return NULL;
363 }
364 mt->base.domain = NOUVEAU_BO_VRAM;
365 mt->base.address = mt->base.bo->offset;
366
367 mt->base.base = *templ;
368 mt->base.vtbl = &nv50_miptree_vtbl;
369 pipe_reference_init(&mt->base.base.reference, 1);
370 mt->base.base.screen = pscreen;
371 mt->level[0].pitch = stride;
372 mt->level[0].offset = 0;
373 mt->level[0].tile_mode = mt->base.bo->config.nv50.tile_mode;
374
375 /* no need to adjust bo reference count */
376 return &mt->base.base;
377 }
378
379
380 /* Offset of zslice @z from start of level @l. */
381 INLINE unsigned
382 nv50_mt_zslice_offset(const struct nv50_miptree *mt, unsigned l, unsigned z)
383 {
384 const struct pipe_resource *pt = &mt->base.base;
385
386 unsigned tds = NV50_TILE_SHIFT_Z(mt->level[l].tile_mode);
387 unsigned ths = NV50_TILE_SHIFT_Y(mt->level[l].tile_mode);
388
389 unsigned nby = util_format_get_nblocksy(pt->format,
390 u_minify(pt->height0, l));
391
392 /* to next 2D tile slice within a 3D tile */
393 unsigned stride_2d = NV50_TILE_SIZE_2D(mt->level[l].tile_mode);
394
395 /* to slice in the next (in z direction) 3D tile */
396 unsigned stride_3d = (align(nby, (1 << ths)) * mt->level[l].pitch) << tds;
397
398 return (z & ((1 << tds) - 1)) * stride_2d + (z >> tds) * stride_3d;
399 }
400
401 /* Surface functions.
402 */
403
404 struct nv50_surface *
405 nv50_surface_from_miptree(struct nv50_miptree *mt,
406 const struct pipe_surface *templ)
407 {
408 struct pipe_surface *ps;
409 struct nv50_surface *ns = CALLOC_STRUCT(nv50_surface);
410 if (!ns)
411 return NULL;
412 ps = &ns->base;
413
414 pipe_reference_init(&ps->reference, 1);
415 pipe_resource_reference(&ps->texture, &mt->base.base);
416
417 ps->format = templ->format;
418 ps->writable = templ->writable;
419 ps->u.tex.level = templ->u.tex.level;
420 ps->u.tex.first_layer = templ->u.tex.first_layer;
421 ps->u.tex.last_layer = templ->u.tex.last_layer;
422
423 ns->width = u_minify(mt->base.base.width0, ps->u.tex.level);
424 ns->height = u_minify(mt->base.base.height0, ps->u.tex.level);
425 ns->depth = ps->u.tex.last_layer - ps->u.tex.first_layer + 1;
426 ns->offset = mt->level[templ->u.tex.level].offset;
427
428 /* comment says there are going to be removed, but they're used by the st */
429 ps->width = ns->width;
430 ps->height = ns->height;
431
432 ns->width <<= mt->ms_x;
433 ns->height <<= mt->ms_y;
434
435 return ns;
436 }
437
438 struct pipe_surface *
439 nv50_miptree_surface_new(struct pipe_context *pipe,
440 struct pipe_resource *pt,
441 const struct pipe_surface *templ)
442 {
443 struct nv50_miptree *mt = nv50_miptree(pt);
444 struct nv50_surface *ns = nv50_surface_from_miptree(mt, templ);
445 if (!ns)
446 return NULL;
447 ns->base.context = pipe;
448
449 if (ns->base.u.tex.first_layer) {
450 const unsigned l = ns->base.u.tex.level;
451 const unsigned z = ns->base.u.tex.first_layer;
452
453 if (mt->layout_3d) {
454 ns->offset += nv50_mt_zslice_offset(mt, l, z);
455
456 /* TODO: switch to depth 1 tiles; but actually this shouldn't happen */
457 if (ns->depth > 1 &&
458 (z & (NV50_TILE_SIZE_Z(mt->level[l].tile_mode) - 1)))
459 NOUVEAU_ERR("Creating unsupported 3D surface !\n");
460 } else {
461 ns->offset += mt->layer_stride * z;
462 }
463 }
464
465 return &ns->base;
466 }