u_vbuf.c revision e0773da1e897164ed7597437070e32b867734ee5
1/**************************************************************************
2 *
3 * Copyright 2011 Marek Olšák <maraeo@gmail.com>
4 * All Rights Reserved.
5 *
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:
13 *
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
16 * of the Software.
17 *
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 AUTHORS 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.
25 *
26 **************************************************************************/
27
28#include "util/u_vbuf.h"
29
30#include "util/u_dump.h"
31#include "util/u_format.h"
32#include "util/u_inlines.h"
33#include "util/u_memory.h"
34#include "util/u_upload_mgr.h"
35#include "translate/translate.h"
36#include "translate/translate_cache.h"
37#include "cso_cache/cso_cache.h"
38#include "cso_cache/cso_hash.h"
39
40struct u_vbuf_elements {
41   unsigned count;
42   struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS];
43
44   unsigned src_format_size[PIPE_MAX_ATTRIBS];
45
46   /* If (velem[i].src_format != native_format[i]), the vertex buffer
47    * referenced by the vertex element cannot be used for rendering and
48    * its vertex data must be translated to native_format[i]. */
49   enum pipe_format native_format[PIPE_MAX_ATTRIBS];
50   unsigned native_format_size[PIPE_MAX_ATTRIBS];
51
52   /* This might mean two things:
53    * - src_format != native_format, as discussed above.
54    * - src_offset % 4 != 0 (if the caps don't allow such an offset). */
55   boolean incompatible_layout;
56   /* Per-element flags. */
57   boolean incompatible_layout_elem[PIPE_MAX_ATTRIBS];
58
59   void *driver_cso;
60};
61
62enum {
63   VB_VERTEX = 0,
64   VB_INSTANCE = 1,
65   VB_CONST = 2,
66   VB_NUM = 3
67};
68
69struct u_vbuf {
70   struct u_vbuf_caps caps;
71
72   struct pipe_context *pipe;
73   struct translate_cache *translate_cache;
74   struct cso_cache *cso_cache;
75   struct u_upload_mgr *uploader;
76
77   /* This is what was set in set_vertex_buffers.
78    * May contain user buffers. */
79   struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
80   unsigned nr_vertex_buffers;
81
82   /* Saved vertex buffers. */
83   struct pipe_vertex_buffer vertex_buffer_saved[PIPE_MAX_ATTRIBS];
84   unsigned nr_vertex_buffers_saved;
85
86   /* Vertex buffers for the driver.
87    * There are no user buffers. */
88   struct pipe_vertex_buffer real_vertex_buffer[PIPE_MAX_ATTRIBS];
89   int nr_real_vertex_buffers;
90   boolean vertex_buffers_dirty;
91
92   /* The index buffer. */
93   struct pipe_index_buffer index_buffer;
94
95   /* Vertex elements. */
96   struct u_vbuf_elements *ve, *ve_saved;
97
98   /* Vertex elements used for the translate fallback. */
99   struct pipe_vertex_element fallback_velems[PIPE_MAX_ATTRIBS];
100   /* If non-NULL, this is a vertex element state used for the translate
101    * fallback and therefore used for rendering too. */
102   boolean using_translate;
103   /* The vertex buffer slot index where translated vertices have been
104    * stored in. */
105   unsigned fallback_vbs[VB_NUM];
106
107   /* Whether there is any user buffer. */
108   boolean any_user_vbs;
109   /* Whether there is a buffer with a non-native layout. */
110   boolean incompatible_vb_layout;
111   /* Per-buffer flags. */
112   boolean incompatible_vb[PIPE_MAX_ATTRIBS];
113};
114
115static void *
116u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count,
117                              const struct pipe_vertex_element *attribs);
118static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso);
119
120
121void u_vbuf_get_caps(struct pipe_screen *screen, struct u_vbuf_caps *caps)
122{
123   caps->format_fixed32 =
124      screen->is_format_supported(screen, PIPE_FORMAT_R32_FIXED, PIPE_BUFFER,
125                                  0, PIPE_BIND_VERTEX_BUFFER);
126
127   caps->format_float16 =
128      screen->is_format_supported(screen, PIPE_FORMAT_R16_FLOAT, PIPE_BUFFER,
129                                  0, PIPE_BIND_VERTEX_BUFFER);
130
131   caps->format_float64 =
132      screen->is_format_supported(screen, PIPE_FORMAT_R64_FLOAT, PIPE_BUFFER,
133                                  0, PIPE_BIND_VERTEX_BUFFER);
134
135   caps->format_norm32 =
136      screen->is_format_supported(screen, PIPE_FORMAT_R32_UNORM, PIPE_BUFFER,
137                                  0, PIPE_BIND_VERTEX_BUFFER) &&
138      screen->is_format_supported(screen, PIPE_FORMAT_R32_SNORM, PIPE_BUFFER,
139                                  0, PIPE_BIND_VERTEX_BUFFER);
140
141   caps->format_scaled32 =
142      screen->is_format_supported(screen, PIPE_FORMAT_R32_USCALED, PIPE_BUFFER,
143                                  0, PIPE_BIND_VERTEX_BUFFER) &&
144      screen->is_format_supported(screen, PIPE_FORMAT_R32_SSCALED, PIPE_BUFFER,
145                                  0, PIPE_BIND_VERTEX_BUFFER);
146
147   caps->fetch_dword_unaligned =
148      !screen->get_param(screen,
149                        PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY) &&
150      !screen->get_param(screen,
151                        PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY) &&
152      !screen->get_param(screen,
153                        PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY);
154
155   caps->user_vertex_buffers =
156      screen->get_param(screen, PIPE_CAP_USER_VERTEX_BUFFERS);
157}
158
159struct u_vbuf *
160u_vbuf_create(struct pipe_context *pipe,
161              struct u_vbuf_caps *caps)
162{
163   struct u_vbuf *mgr = CALLOC_STRUCT(u_vbuf);
164
165   mgr->caps = *caps;
166   mgr->pipe = pipe;
167   mgr->cso_cache = cso_cache_create();
168   mgr->translate_cache = translate_cache_create();
169   memset(mgr->fallback_vbs, ~0, sizeof(mgr->fallback_vbs));
170
171   mgr->uploader = u_upload_create(pipe, 1024 * 1024, 4,
172                                   PIPE_BIND_VERTEX_BUFFER);
173
174   return mgr;
175}
176
177/* u_vbuf uses its own caching for vertex elements, because it needs to keep
178 * its own preprocessed state per vertex element CSO. */
179static struct u_vbuf_elements *
180u_vbuf_set_vertex_elements_internal(struct u_vbuf *mgr, unsigned count,
181                                    const struct pipe_vertex_element *states)
182{
183   struct pipe_context *pipe = mgr->pipe;
184   unsigned key_size, hash_key;
185   struct cso_hash_iter iter;
186   struct u_vbuf_elements *ve;
187   struct cso_velems_state velems_state;
188
189   /* need to include the count into the stored state data too. */
190   key_size = sizeof(struct pipe_vertex_element) * count + sizeof(unsigned);
191   velems_state.count = count;
192   memcpy(velems_state.velems, states,
193          sizeof(struct pipe_vertex_element) * count);
194   hash_key = cso_construct_key((void*)&velems_state, key_size);
195   iter = cso_find_state_template(mgr->cso_cache, hash_key, CSO_VELEMENTS,
196                                  (void*)&velems_state, key_size);
197
198   if (cso_hash_iter_is_null(iter)) {
199      struct cso_velements *cso = MALLOC_STRUCT(cso_velements);
200      memcpy(&cso->state, &velems_state, key_size);
201      cso->data = u_vbuf_create_vertex_elements(mgr, count, states);
202      cso->delete_state = (cso_state_callback)u_vbuf_delete_vertex_elements;
203      cso->context = (void*)mgr;
204
205      iter = cso_insert_state(mgr->cso_cache, hash_key, CSO_VELEMENTS, cso);
206      ve = cso->data;
207   } else {
208      ve = ((struct cso_velements *)cso_hash_iter_data(iter))->data;
209   }
210
211   assert(ve);
212   pipe->bind_vertex_elements_state(pipe, ve->driver_cso);
213   return ve;
214}
215
216void u_vbuf_set_vertex_elements(struct u_vbuf *mgr, unsigned count,
217                               const struct pipe_vertex_element *states)
218{
219   mgr->ve = u_vbuf_set_vertex_elements_internal(mgr, count, states);
220}
221
222void u_vbuf_destroy(struct u_vbuf *mgr)
223{
224   unsigned i;
225
226   for (i = 0; i < mgr->nr_vertex_buffers; i++) {
227      pipe_resource_reference(&mgr->vertex_buffer[i].buffer, NULL);
228   }
229   for (i = 0; i < mgr->nr_real_vertex_buffers; i++) {
230      pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL);
231   }
232
233   translate_cache_destroy(mgr->translate_cache);
234   u_upload_destroy(mgr->uploader);
235   cso_cache_delete(mgr->cso_cache);
236   FREE(mgr);
237}
238
239static void
240u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key,
241                         unsigned vb_mask, unsigned out_vb,
242                         int start_vertex, unsigned num_vertices,
243                         int start_index, unsigned num_indices, int min_index,
244                         boolean unroll_indices)
245{
246   struct translate *tr;
247   struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0};
248   struct pipe_resource *out_buffer = NULL;
249   uint8_t *out_map;
250   unsigned i, out_offset;
251
252   /* Get a translate object. */
253   tr = translate_cache_find(mgr->translate_cache, key);
254
255   /* Map buffers we want to translate. */
256   for (i = 0; i < mgr->nr_vertex_buffers; i++) {
257      if (vb_mask & (1 << i)) {
258         struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[i];
259         unsigned offset = vb->buffer_offset + vb->stride * start_vertex;
260         uint8_t *map;
261
262         if (vb->buffer->user_ptr) {
263            map = vb->buffer->user_ptr + offset;
264         } else {
265            unsigned size = vb->stride ? num_vertices * vb->stride
266                                       : sizeof(double)*4;
267
268            if (offset+size > vb->buffer->width0) {
269               size = vb->buffer->width0 - offset;
270            }
271
272            map = pipe_buffer_map_range(mgr->pipe, vb->buffer, offset, size,
273                                        PIPE_TRANSFER_READ, &vb_transfer[i]);
274         }
275
276         /* Subtract min_index so that indexing with the index buffer works. */
277         if (unroll_indices) {
278            map -= vb->stride * min_index;
279         }
280
281         tr->set_buffer(tr, i, map, vb->stride, ~0);
282      }
283   }
284
285   /* Translate. */
286   if (unroll_indices) {
287      struct pipe_index_buffer *ib = &mgr->index_buffer;
288      struct pipe_transfer *transfer = NULL;
289      unsigned offset = ib->offset + start_index * ib->index_size;
290      uint8_t *map;
291
292      assert(ib->buffer && ib->index_size);
293
294      if (ib->buffer->user_ptr) {
295         map = ib->buffer->user_ptr + offset;
296      } else {
297         map = pipe_buffer_map_range(mgr->pipe, ib->buffer, offset,
298                                     num_indices * ib->index_size,
299                                     PIPE_TRANSFER_READ, &transfer);
300      }
301
302      /* Create and map the output buffer. */
303      u_upload_alloc(mgr->uploader, 0,
304                     key->output_stride * num_indices,
305                     &out_offset, &out_buffer,
306                     (void**)&out_map);
307
308      switch (ib->index_size) {
309      case 4:
310         tr->run_elts(tr, (unsigned*)map, num_indices, 0, out_map);
311         break;
312      case 2:
313         tr->run_elts16(tr, (uint16_t*)map, num_indices, 0, out_map);
314         break;
315      case 1:
316         tr->run_elts8(tr, map, num_indices, 0, out_map);
317         break;
318      }
319
320      if (transfer) {
321         pipe_buffer_unmap(mgr->pipe, transfer);
322      }
323   } else {
324      /* Create and map the output buffer. */
325      u_upload_alloc(mgr->uploader,
326                     key->output_stride * start_vertex,
327                     key->output_stride * num_vertices,
328                     &out_offset, &out_buffer,
329                     (void**)&out_map);
330
331      out_offset -= key->output_stride * start_vertex;
332
333      tr->run(tr, 0, num_vertices, 0, out_map);
334   }
335
336   /* Unmap all buffers. */
337   for (i = 0; i < mgr->nr_vertex_buffers; i++) {
338      if (vb_transfer[i]) {
339         pipe_buffer_unmap(mgr->pipe, vb_transfer[i]);
340      }
341   }
342
343   /* Setup the new vertex buffer. */
344   mgr->real_vertex_buffer[out_vb].buffer_offset = out_offset;
345   mgr->real_vertex_buffer[out_vb].stride = key->output_stride;
346
347   /* Move the buffer reference. */
348   pipe_resource_reference(
349      &mgr->real_vertex_buffer[out_vb].buffer, NULL);
350   mgr->real_vertex_buffer[out_vb].buffer = out_buffer;
351}
352
353static boolean
354u_vbuf_translate_find_free_vb_slots(struct u_vbuf *mgr,
355                                    unsigned mask[VB_NUM])
356{
357   unsigned i, type;
358   unsigned nr = mgr->ve->count;
359   boolean used_vb[PIPE_MAX_ATTRIBS] = {0};
360   unsigned fallback_vbs[VB_NUM];
361
362   memset(fallback_vbs, ~0, sizeof(fallback_vbs));
363
364   /* Mark used vertex buffers as... used. */
365   for (i = 0; i < nr; i++) {
366      if (!mgr->ve->incompatible_layout_elem[i]) {
367         unsigned index = mgr->ve->ve[i].vertex_buffer_index;
368
369         if (!mgr->incompatible_vb[index]) {
370            used_vb[index] = TRUE;
371         }
372      }
373   }
374
375   /* Find free slots for each type if needed. */
376   i = 0;
377   for (type = 0; type < VB_NUM; type++) {
378      if (mask[type]) {
379         for (; i < PIPE_MAX_ATTRIBS; i++) {
380            if (!used_vb[i]) {
381               /*printf("found slot=%i for type=%i\n", i, type);*/
382               fallback_vbs[type] = i;
383               i++;
384               if (i > mgr->nr_real_vertex_buffers) {
385                  mgr->nr_real_vertex_buffers = i;
386               }
387               break;
388            }
389         }
390         if (i == PIPE_MAX_ATTRIBS) {
391            /* fail, reset the number to its original value */
392            mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers;
393            return FALSE;
394         }
395      }
396   }
397
398   memcpy(mgr->fallback_vbs, fallback_vbs, sizeof(fallback_vbs));
399   return TRUE;
400}
401
402static boolean
403u_vbuf_translate_begin(struct u_vbuf *mgr,
404                       int start_vertex, unsigned num_vertices,
405                       int start_instance, unsigned num_instances,
406                       int start_index, unsigned num_indices, int min_index,
407                       boolean unroll_indices)
408{
409   unsigned mask[VB_NUM] = {0};
410   struct translate_key key[VB_NUM];
411   unsigned elem_index[VB_NUM][PIPE_MAX_ATTRIBS]; /* ... into key.elements */
412   unsigned i, type;
413
414   int start[VB_NUM] = {
415      start_vertex,     /* VERTEX */
416      start_instance,   /* INSTANCE */
417      0                 /* CONST */
418   };
419
420   unsigned num[VB_NUM] = {
421      num_vertices,     /* VERTEX */
422      num_instances,    /* INSTANCE */
423      1                 /* CONST */
424   };
425
426   memset(key, 0, sizeof(key));
427   memset(elem_index, ~0, sizeof(elem_index));
428
429   /* See if there are vertex attribs of each type to translate and
430    * which ones. */
431   for (i = 0; i < mgr->ve->count; i++) {
432      unsigned vb_index = mgr->ve->ve[i].vertex_buffer_index;
433
434      if (!mgr->vertex_buffer[vb_index].stride) {
435         if (!mgr->ve->incompatible_layout_elem[i] &&
436             !mgr->incompatible_vb[vb_index]) {
437            continue;
438         }
439         mask[VB_CONST] |= 1 << vb_index;
440      } else if (mgr->ve->ve[i].instance_divisor) {
441         if (!mgr->ve->incompatible_layout_elem[i] &&
442             !mgr->incompatible_vb[vb_index]) {
443            continue;
444         }
445         mask[VB_INSTANCE] |= 1 << vb_index;
446      } else {
447         if (!unroll_indices &&
448             !mgr->ve->incompatible_layout_elem[i] &&
449             !mgr->incompatible_vb[vb_index]) {
450            continue;
451         }
452         mask[VB_VERTEX] |= 1 << vb_index;
453      }
454   }
455
456   assert(mask[VB_VERTEX] || mask[VB_INSTANCE] || mask[VB_CONST]);
457
458   /* Find free vertex buffer slots. */
459   if (!u_vbuf_translate_find_free_vb_slots(mgr, mask)) {
460      return FALSE;
461   }
462
463   /* Initialize the translate keys. */
464   for (i = 0; i < mgr->ve->count; i++) {
465      struct translate_key *k;
466      struct translate_element *te;
467      unsigned bit, vb_index = mgr->ve->ve[i].vertex_buffer_index;
468      bit = 1 << vb_index;
469
470      if (!mgr->ve->incompatible_layout_elem[i] &&
471          !mgr->incompatible_vb[vb_index] &&
472          (!unroll_indices || !(mask[VB_VERTEX] & bit))) {
473         continue;
474      }
475
476      /* Set type to what we will translate.
477       * Whether vertex, instance, or constant attribs. */
478      for (type = 0; type < VB_NUM; type++) {
479         if (mask[type] & bit) {
480            break;
481         }
482      }
483      assert(type < VB_NUM);
484      assert(translate_is_output_format_supported(mgr->ve->native_format[i]));
485      /*printf("velem=%i type=%i\n", i, type);*/
486
487      /* Add the vertex element. */
488      k = &key[type];
489      elem_index[type][i] = k->nr_elements;
490
491      te = &k->element[k->nr_elements];
492      te->type = TRANSLATE_ELEMENT_NORMAL;
493      te->instance_divisor = 0;
494      te->input_buffer = vb_index;
495      te->input_format = mgr->ve->ve[i].src_format;
496      te->input_offset = mgr->ve->ve[i].src_offset;
497      te->output_format = mgr->ve->native_format[i];
498      te->output_offset = k->output_stride;
499
500      k->output_stride += mgr->ve->native_format_size[i];
501      k->nr_elements++;
502   }
503
504   /* Translate buffers. */
505   for (type = 0; type < VB_NUM; type++) {
506      if (key[type].nr_elements) {
507         u_vbuf_translate_buffers(mgr, &key[type], mask[type],
508                                  mgr->fallback_vbs[type],
509                                  start[type], num[type],
510                                  start_index, num_indices, min_index,
511                                  unroll_indices && type == VB_VERTEX);
512
513         /* Fixup the stride for constant attribs. */
514         if (type == VB_CONST) {
515            mgr->real_vertex_buffer[mgr->fallback_vbs[VB_CONST]].stride = 0;
516         }
517      }
518   }
519
520   /* Setup new vertex elements. */
521   for (i = 0; i < mgr->ve->count; i++) {
522      for (type = 0; type < VB_NUM; type++) {
523         if (elem_index[type][i] < key[type].nr_elements) {
524            struct translate_element *te = &key[type].element[elem_index[type][i]];
525            mgr->fallback_velems[i].instance_divisor = mgr->ve->ve[i].instance_divisor;
526            mgr->fallback_velems[i].src_format = te->output_format;
527            mgr->fallback_velems[i].src_offset = te->output_offset;
528            mgr->fallback_velems[i].vertex_buffer_index = mgr->fallback_vbs[type];
529
530            /* elem_index[type][i] can only be set for one type. */
531            assert(type > VB_INSTANCE || elem_index[type+1][i] == ~0);
532            assert(type > VB_VERTEX   || elem_index[type+2][i] == ~0);
533            break;
534         }
535      }
536      /* No translating, just copy the original vertex element over. */
537      if (type == VB_NUM) {
538         memcpy(&mgr->fallback_velems[i], &mgr->ve->ve[i],
539                sizeof(struct pipe_vertex_element));
540      }
541   }
542
543   u_vbuf_set_vertex_elements_internal(mgr, mgr->ve->count,
544                                       mgr->fallback_velems);
545   mgr->using_translate = TRUE;
546   return TRUE;
547}
548
549static void u_vbuf_translate_end(struct u_vbuf *mgr)
550{
551   unsigned i;
552
553   /* Restore vertex elements. */
554   mgr->pipe->bind_vertex_elements_state(mgr->pipe, mgr->ve->driver_cso);
555   mgr->using_translate = FALSE;
556
557   /* Unreference the now-unused VBOs. */
558   for (i = 0; i < VB_NUM; i++) {
559      unsigned vb = mgr->fallback_vbs[i];
560      if (vb != ~0) {
561         pipe_resource_reference(&mgr->real_vertex_buffer[vb].buffer, NULL);
562         mgr->fallback_vbs[i] = ~0;
563      }
564   }
565   mgr->nr_real_vertex_buffers = mgr->nr_vertex_buffers;
566}
567
568#define FORMAT_REPLACE(what, withwhat) \
569    case PIPE_FORMAT_##what: format = PIPE_FORMAT_##withwhat; break
570
571static void *
572u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count,
573                              const struct pipe_vertex_element *attribs)
574{
575   struct pipe_context *pipe = mgr->pipe;
576   unsigned i;
577   struct pipe_vertex_element native_attribs[PIPE_MAX_ATTRIBS];
578   struct u_vbuf_elements *ve = CALLOC_STRUCT(u_vbuf_elements);
579
580   ve->count = count;
581
582   memcpy(ve->ve, attribs, sizeof(struct pipe_vertex_element) * count);
583   memcpy(native_attribs, attribs, sizeof(struct pipe_vertex_element) * count);
584
585   /* Set the best native format in case the original format is not
586    * supported. */
587   for (i = 0; i < count; i++) {
588      enum pipe_format format = ve->ve[i].src_format;
589
590      ve->src_format_size[i] = util_format_get_blocksize(format);
591
592      /* Choose a native format.
593       * For now we don't care about the alignment, that's going to
594       * be sorted out later. */
595      if (!mgr->caps.format_fixed32) {
596         switch (format) {
597            FORMAT_REPLACE(R32_FIXED,           R32_FLOAT);
598            FORMAT_REPLACE(R32G32_FIXED,        R32G32_FLOAT);
599            FORMAT_REPLACE(R32G32B32_FIXED,     R32G32B32_FLOAT);
600            FORMAT_REPLACE(R32G32B32A32_FIXED,  R32G32B32A32_FLOAT);
601            default:;
602         }
603      }
604      if (!mgr->caps.format_float16) {
605         switch (format) {
606            FORMAT_REPLACE(R16_FLOAT,           R32_FLOAT);
607            FORMAT_REPLACE(R16G16_FLOAT,        R32G32_FLOAT);
608            FORMAT_REPLACE(R16G16B16_FLOAT,     R32G32B32_FLOAT);
609            FORMAT_REPLACE(R16G16B16A16_FLOAT,  R32G32B32A32_FLOAT);
610            default:;
611         }
612      }
613      if (!mgr->caps.format_float64) {
614         switch (format) {
615            FORMAT_REPLACE(R64_FLOAT,           R32_FLOAT);
616            FORMAT_REPLACE(R64G64_FLOAT,        R32G32_FLOAT);
617            FORMAT_REPLACE(R64G64B64_FLOAT,     R32G32B32_FLOAT);
618            FORMAT_REPLACE(R64G64B64A64_FLOAT,  R32G32B32A32_FLOAT);
619            default:;
620         }
621      }
622      if (!mgr->caps.format_norm32) {
623         switch (format) {
624            FORMAT_REPLACE(R32_UNORM,           R32_FLOAT);
625            FORMAT_REPLACE(R32G32_UNORM,        R32G32_FLOAT);
626            FORMAT_REPLACE(R32G32B32_UNORM,     R32G32B32_FLOAT);
627            FORMAT_REPLACE(R32G32B32A32_UNORM,  R32G32B32A32_FLOAT);
628            FORMAT_REPLACE(R32_SNORM,           R32_FLOAT);
629            FORMAT_REPLACE(R32G32_SNORM,        R32G32_FLOAT);
630            FORMAT_REPLACE(R32G32B32_SNORM,     R32G32B32_FLOAT);
631            FORMAT_REPLACE(R32G32B32A32_SNORM,  R32G32B32A32_FLOAT);
632            default:;
633         }
634      }
635      if (!mgr->caps.format_scaled32) {
636         switch (format) {
637            FORMAT_REPLACE(R32_USCALED,         R32_FLOAT);
638            FORMAT_REPLACE(R32G32_USCALED,      R32G32_FLOAT);
639            FORMAT_REPLACE(R32G32B32_USCALED,   R32G32B32_FLOAT);
640            FORMAT_REPLACE(R32G32B32A32_USCALED,R32G32B32A32_FLOAT);
641            FORMAT_REPLACE(R32_SSCALED,         R32_FLOAT);
642            FORMAT_REPLACE(R32G32_SSCALED,      R32G32_FLOAT);
643            FORMAT_REPLACE(R32G32B32_SSCALED,   R32G32B32_FLOAT);
644            FORMAT_REPLACE(R32G32B32A32_SSCALED,R32G32B32A32_FLOAT);
645            default:;
646         }
647      }
648
649      native_attribs[i].src_format = format;
650      ve->native_format[i] = format;
651      ve->native_format_size[i] =
652            util_format_get_blocksize(ve->native_format[i]);
653
654      ve->incompatible_layout_elem[i] =
655            ve->ve[i].src_format != ve->native_format[i] ||
656            (!mgr->caps.fetch_dword_unaligned && ve->ve[i].src_offset % 4 != 0);
657      ve->incompatible_layout =
658            ve->incompatible_layout ||
659            ve->incompatible_layout_elem[i];
660   }
661
662   /* Align the formats to the size of DWORD if needed. */
663   if (!mgr->caps.fetch_dword_unaligned) {
664      for (i = 0; i < count; i++) {
665         ve->native_format_size[i] = align(ve->native_format_size[i], 4);
666      }
667   }
668
669   ve->driver_cso =
670      pipe->create_vertex_elements_state(pipe, count, native_attribs);
671   return ve;
672}
673
674static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso)
675{
676   struct pipe_context *pipe = mgr->pipe;
677   struct u_vbuf_elements *ve = cso;
678
679   pipe->delete_vertex_elements_state(pipe, ve->driver_cso);
680   FREE(ve);
681}
682
683void u_vbuf_set_vertex_buffers(struct u_vbuf *mgr, unsigned count,
684                               const struct pipe_vertex_buffer *bufs)
685{
686   unsigned i;
687
688   mgr->any_user_vbs = FALSE;
689   mgr->incompatible_vb_layout = FALSE;
690   memset(mgr->incompatible_vb, 0, sizeof(mgr->incompatible_vb));
691
692   if (!mgr->caps.fetch_dword_unaligned) {
693      /* Check if the strides and offsets are aligned to the size of DWORD. */
694      for (i = 0; i < count; i++) {
695         if (bufs[i].buffer) {
696            if (bufs[i].stride % 4 != 0 ||
697                bufs[i].buffer_offset % 4 != 0) {
698               mgr->incompatible_vb_layout = TRUE;
699               mgr->incompatible_vb[i] = TRUE;
700            }
701         }
702      }
703   }
704
705   for (i = 0; i < count; i++) {
706      const struct pipe_vertex_buffer *vb = &bufs[i];
707
708      pipe_resource_reference(&mgr->vertex_buffer[i].buffer, vb->buffer);
709
710      mgr->real_vertex_buffer[i].buffer_offset =
711      mgr->vertex_buffer[i].buffer_offset = vb->buffer_offset;
712
713      mgr->real_vertex_buffer[i].stride =
714      mgr->vertex_buffer[i].stride = vb->stride;
715
716      if (!vb->buffer ||
717          mgr->incompatible_vb[i]) {
718         pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL);
719         continue;
720      }
721
722      if (vb->buffer->user_ptr) {
723         pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL);
724         mgr->any_user_vbs = TRUE;
725         continue;
726      }
727
728      pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, vb->buffer);
729   }
730
731   for (i = count; i < mgr->nr_vertex_buffers; i++) {
732      pipe_resource_reference(&mgr->vertex_buffer[i].buffer, NULL);
733   }
734   for (i = count; i < mgr->nr_real_vertex_buffers; i++) {
735      pipe_resource_reference(&mgr->real_vertex_buffer[i].buffer, NULL);
736   }
737
738   mgr->nr_vertex_buffers = count;
739   mgr->nr_real_vertex_buffers = count;
740   mgr->vertex_buffers_dirty = TRUE;
741}
742
743void u_vbuf_set_index_buffer(struct u_vbuf *mgr,
744                             const struct pipe_index_buffer *ib)
745{
746   struct pipe_context *pipe = mgr->pipe;
747
748   if (ib && ib->buffer) {
749      assert(ib->offset % ib->index_size == 0);
750      pipe_resource_reference(&mgr->index_buffer.buffer, ib->buffer);
751      mgr->index_buffer.offset = ib->offset;
752      mgr->index_buffer.index_size = ib->index_size;
753   } else {
754      pipe_resource_reference(&mgr->index_buffer.buffer, NULL);
755   }
756
757   pipe->set_index_buffer(pipe, ib);
758}
759
760static void
761u_vbuf_upload_buffers(struct u_vbuf *mgr,
762                      int start_vertex, unsigned num_vertices,
763                      int start_instance, unsigned num_instances)
764{
765   unsigned i;
766   unsigned nr_velems = mgr->ve->count;
767   unsigned nr_vbufs = mgr->nr_vertex_buffers;
768   struct pipe_vertex_element *velems =
769         mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve;
770   unsigned start_offset[PIPE_MAX_ATTRIBS];
771   unsigned end_offset[PIPE_MAX_ATTRIBS] = {0};
772
773   /* Determine how much data needs to be uploaded. */
774   for (i = 0; i < nr_velems; i++) {
775      struct pipe_vertex_element *velem = &velems[i];
776      unsigned index = velem->vertex_buffer_index;
777      struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index];
778      unsigned instance_div, first, size;
779
780      /* Skip the buffers generated by translate. */
781      if (index == mgr->fallback_vbs[VB_VERTEX] ||
782          index == mgr->fallback_vbs[VB_INSTANCE] ||
783          index == mgr->fallback_vbs[VB_CONST]) {
784         continue;
785      }
786
787      assert(vb->buffer);
788
789      if (!vb->buffer->user_ptr) {
790         continue;
791      }
792
793      instance_div = velem->instance_divisor;
794      first = vb->buffer_offset + velem->src_offset;
795
796      if (!vb->stride) {
797         /* Constant attrib. */
798         size = mgr->ve->src_format_size[i];
799      } else if (instance_div) {
800         /* Per-instance attrib. */
801         unsigned count = (num_instances + instance_div - 1) / instance_div;
802         first += vb->stride * start_instance;
803         size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
804      } else {
805         /* Per-vertex attrib. */
806         first += vb->stride * start_vertex;
807         size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i];
808      }
809
810      /* Update offsets. */
811      if (!end_offset[index]) {
812         start_offset[index] = first;
813         end_offset[index] = first + size;
814      } else {
815         if (first < start_offset[index])
816            start_offset[index] = first;
817         if (first + size > end_offset[index])
818            end_offset[index] = first + size;
819      }
820   }
821
822   /* Upload buffers. */
823   for (i = 0; i < nr_vbufs; i++) {
824      unsigned start, end = end_offset[i];
825      struct pipe_vertex_buffer *real_vb;
826      uint8_t *ptr;
827
828      if (!end) {
829         continue;
830      }
831
832      start = start_offset[i];
833      assert(start < end);
834
835      real_vb = &mgr->real_vertex_buffer[i];
836      ptr = mgr->vertex_buffer[i].buffer->user_ptr;
837
838      u_upload_data(mgr->uploader, start, end - start, ptr + start,
839                    &real_vb->buffer_offset, &real_vb->buffer);
840
841      real_vb->buffer_offset -= start;
842   }
843}
844
845static boolean u_vbuf_need_minmax_index(struct u_vbuf *mgr)
846{
847   unsigned i, nr = mgr->ve->count;
848
849   for (i = 0; i < nr; i++) {
850      struct pipe_vertex_buffer *vb;
851      unsigned index;
852
853      /* Per-instance attribs don't need min/max_index. */
854      if (mgr->ve->ve[i].instance_divisor) {
855         continue;
856      }
857
858      index = mgr->ve->ve[i].vertex_buffer_index;
859      vb = &mgr->vertex_buffer[index];
860
861      /* Constant attribs don't need min/max_index. */
862      if (!vb->stride) {
863         continue;
864      }
865
866      /* Per-vertex attribs need min/max_index. */
867      if (vb->buffer->user_ptr ||
868          mgr->ve->incompatible_layout_elem[i] ||
869          mgr->incompatible_vb[index]) {
870         return TRUE;
871      }
872   }
873
874   return FALSE;
875}
876
877static boolean u_vbuf_mapping_vertex_buffer_blocks(struct u_vbuf *mgr)
878{
879   unsigned i, nr = mgr->ve->count;
880
881   for (i = 0; i < nr; i++) {
882      struct pipe_vertex_buffer *vb;
883      unsigned index;
884
885      /* Per-instance attribs are not per-vertex data. */
886      if (mgr->ve->ve[i].instance_divisor) {
887         continue;
888      }
889
890      index = mgr->ve->ve[i].vertex_buffer_index;
891      vb = &mgr->vertex_buffer[index];
892
893      /* Constant attribs are not per-vertex data. */
894      if (!vb->stride) {
895         continue;
896      }
897
898      /* Return true for the hw buffers which don't need to be translated. */
899      /* XXX we could use some kind of a is-busy query. */
900      if (!vb->buffer->user_ptr &&
901          !mgr->ve->incompatible_layout_elem[i] &&
902          !mgr->incompatible_vb[index]) {
903         return TRUE;
904      }
905   }
906
907   return FALSE;
908}
909
910static void u_vbuf_get_minmax_index(struct pipe_context *pipe,
911                                    struct pipe_index_buffer *ib,
912                                    const struct pipe_draw_info *info,
913                                    int *out_min_index,
914                                    int *out_max_index)
915{
916   struct pipe_transfer *transfer = NULL;
917   const void *indices;
918   unsigned i;
919   unsigned restart_index = info->restart_index;
920
921   if (ib->buffer->user_ptr) {
922      indices = ib->buffer->user_ptr +
923                ib->offset + info->start * ib->index_size;
924   } else {
925      indices = pipe_buffer_map_range(pipe, ib->buffer,
926                                      ib->offset + info->start * ib->index_size,
927                                      info->count * ib->index_size,
928                                      PIPE_TRANSFER_READ, &transfer);
929   }
930
931   switch (ib->index_size) {
932   case 4: {
933      const unsigned *ui_indices = (const unsigned*)indices;
934      unsigned max_ui = 0;
935      unsigned min_ui = ~0U;
936      if (info->primitive_restart) {
937         for (i = 0; i < info->count; i++) {
938            if (ui_indices[i] != restart_index) {
939               if (ui_indices[i] > max_ui) max_ui = ui_indices[i];
940               if (ui_indices[i] < min_ui) min_ui = ui_indices[i];
941            }
942         }
943      }
944      else {
945         for (i = 0; i < info->count; i++) {
946            if (ui_indices[i] > max_ui) max_ui = ui_indices[i];
947            if (ui_indices[i] < min_ui) min_ui = ui_indices[i];
948         }
949      }
950      *out_min_index = min_ui;
951      *out_max_index = max_ui;
952      break;
953   }
954   case 2: {
955      const unsigned short *us_indices = (const unsigned short*)indices;
956      unsigned max_us = 0;
957      unsigned min_us = ~0U;
958      if (info->primitive_restart) {
959         for (i = 0; i < info->count; i++) {
960            if (us_indices[i] != restart_index) {
961               if (us_indices[i] > max_us) max_us = us_indices[i];
962               if (us_indices[i] < min_us) min_us = us_indices[i];
963            }
964         }
965      }
966      else {
967         for (i = 0; i < info->count; i++) {
968            if (us_indices[i] > max_us) max_us = us_indices[i];
969            if (us_indices[i] < min_us) min_us = us_indices[i];
970         }
971      }
972      *out_min_index = min_us;
973      *out_max_index = max_us;
974      break;
975   }
976   case 1: {
977      const unsigned char *ub_indices = (const unsigned char*)indices;
978      unsigned max_ub = 0;
979      unsigned min_ub = ~0U;
980      if (info->primitive_restart) {
981         for (i = 0; i < info->count; i++) {
982            if (ub_indices[i] != restart_index) {
983               if (ub_indices[i] > max_ub) max_ub = ub_indices[i];
984               if (ub_indices[i] < min_ub) min_ub = ub_indices[i];
985            }
986         }
987      }
988      else {
989         for (i = 0; i < info->count; i++) {
990            if (ub_indices[i] > max_ub) max_ub = ub_indices[i];
991            if (ub_indices[i] < min_ub) min_ub = ub_indices[i];
992         }
993      }
994      *out_min_index = min_ub;
995      *out_max_index = max_ub;
996      break;
997   }
998   default:
999      assert(0);
1000      *out_min_index = 0;
1001      *out_max_index = 0;
1002   }
1003
1004   if (transfer) {
1005      pipe_buffer_unmap(pipe, transfer);
1006   }
1007}
1008
1009void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info)
1010{
1011   struct pipe_context *pipe = mgr->pipe;
1012   int start_vertex, min_index;
1013   unsigned num_vertices;
1014   boolean unroll_indices = FALSE;
1015
1016   /* Normal draw. No fallback and no user buffers. */
1017   if (!mgr->incompatible_vb_layout &&
1018       !mgr->ve->incompatible_layout &&
1019       !mgr->any_user_vbs) {
1020      /* Set vertex buffers if needed. */
1021      if (mgr->vertex_buffers_dirty) {
1022         pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers,
1023                                  mgr->real_vertex_buffer);
1024         mgr->vertex_buffers_dirty = FALSE;
1025      }
1026
1027      pipe->draw_vbo(pipe, info);
1028      return;
1029   }
1030
1031   if (info->indexed) {
1032      int max_index;
1033      boolean index_bounds_valid = FALSE;
1034
1035      if (info->max_index != ~0) {
1036         min_index = info->min_index;
1037         max_index = info->max_index;
1038         index_bounds_valid = TRUE;
1039      } else if (u_vbuf_need_minmax_index(mgr)) {
1040         u_vbuf_get_minmax_index(mgr->pipe, &mgr->index_buffer, info,
1041                                 &min_index, &max_index);
1042         index_bounds_valid = TRUE;
1043      }
1044
1045      /* If the index bounds are valid, it means some upload or translation
1046       * of per-vertex attribs will be performed. */
1047      if (index_bounds_valid) {
1048         assert(min_index <= max_index);
1049
1050         start_vertex = min_index + info->index_bias;
1051         num_vertices = max_index + 1 - min_index;
1052
1053         /* Primitive restart doesn't work when unrolling indices.
1054          * We would have to break this drawing operation into several ones. */
1055         /* Use some heuristic to see if unrolling indices improves
1056          * performance. */
1057         if (!info->primitive_restart &&
1058             num_vertices > info->count*2 &&
1059             num_vertices-info->count > 32 &&
1060             !u_vbuf_mapping_vertex_buffer_blocks(mgr)) {
1061            /*printf("num_vertices=%i count=%i\n", num_vertices, info->count);*/
1062            unroll_indices = TRUE;
1063         }
1064      } else {
1065         /* Nothing to do for per-vertex attribs. */
1066         start_vertex = 0;
1067         num_vertices = 0;
1068         min_index = 0;
1069      }
1070   } else {
1071      start_vertex = info->start;
1072      num_vertices = info->count;
1073      min_index = 0;
1074   }
1075
1076   /* Translate vertices with non-native layouts or formats. */
1077   if (unroll_indices ||
1078       mgr->incompatible_vb_layout ||
1079       mgr->ve->incompatible_layout) {
1080      /* XXX check the return value */
1081      u_vbuf_translate_begin(mgr, start_vertex, num_vertices,
1082                             info->start_instance, info->instance_count,
1083                             info->start, info->count, min_index,
1084                             unroll_indices);
1085   }
1086
1087   /* Upload user buffers. */
1088   if (mgr->any_user_vbs) {
1089      u_vbuf_upload_buffers(mgr, start_vertex, num_vertices,
1090                            info->start_instance, info->instance_count);
1091   }
1092
1093   /*
1094   if (unroll_indices) {
1095      printf("unrolling indices: start_vertex = %i, num_vertices = %i\n",
1096             start_vertex, num_vertices);
1097      util_dump_draw_info(stdout, info);
1098      printf("\n");
1099   }
1100
1101   unsigned i;
1102   for (i = 0; i < mgr->nr_vertex_buffers; i++) {
1103      printf("input %i: ", i);
1104      util_dump_vertex_buffer(stdout, mgr->vertex_buffer+i);
1105      printf("\n");
1106   }
1107   for (i = 0; i < mgr->nr_real_vertex_buffers; i++) {
1108      printf("real %i: ", i);
1109      util_dump_vertex_buffer(stdout, mgr->real_vertex_buffer+i);
1110      printf("\n");
1111   }
1112   */
1113
1114   u_upload_unmap(mgr->uploader);
1115   pipe->set_vertex_buffers(pipe, mgr->nr_real_vertex_buffers,
1116                            mgr->real_vertex_buffer);
1117
1118   if (unlikely(unroll_indices)) {
1119      struct pipe_draw_info new_info = *info;
1120      new_info.indexed = FALSE;
1121      new_info.index_bias = 0;
1122      new_info.min_index = 0;
1123      new_info.max_index = info->count - 1;
1124      new_info.start = 0;
1125
1126      pipe->draw_vbo(pipe, &new_info);
1127   } else {
1128      pipe->draw_vbo(pipe, info);
1129   }
1130
1131   if (mgr->using_translate) {
1132      u_vbuf_translate_end(mgr);
1133   }
1134   mgr->vertex_buffers_dirty = TRUE;
1135}
1136
1137void u_vbuf_save_vertex_elements(struct u_vbuf *mgr)
1138{
1139   assert(!mgr->ve_saved);
1140   mgr->ve_saved = mgr->ve;
1141}
1142
1143void u_vbuf_restore_vertex_elements(struct u_vbuf *mgr)
1144{
1145   if (mgr->ve != mgr->ve_saved) {
1146      struct pipe_context *pipe = mgr->pipe;
1147
1148      mgr->ve = mgr->ve_saved;
1149      pipe->bind_vertex_elements_state(pipe,
1150                                       mgr->ve ? mgr->ve->driver_cso : NULL);
1151   }
1152   mgr->ve_saved = NULL;
1153}
1154
1155void u_vbuf_save_vertex_buffers(struct u_vbuf *mgr)
1156{
1157   util_copy_vertex_buffers(mgr->vertex_buffer_saved,
1158                            &mgr->nr_vertex_buffers_saved,
1159                            mgr->vertex_buffer,
1160                            mgr->nr_vertex_buffers);
1161}
1162
1163void u_vbuf_restore_vertex_buffers(struct u_vbuf *mgr)
1164{
1165   unsigned i;
1166
1167   u_vbuf_set_vertex_buffers(mgr, mgr->nr_vertex_buffers_saved,
1168                             mgr->vertex_buffer_saved);
1169   for (i = 0; i < mgr->nr_vertex_buffers_saved; i++) {
1170      pipe_resource_reference(&mgr->vertex_buffer_saved[i].buffer, NULL);
1171   }
1172   mgr->nr_vertex_buffers_saved = 0;
1173}
1174