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/**
29 * This module uploads user buffers and translates the vertex buffers which
30 * contain incompatible vertices (i.e. not supported by the driver/hardware)
31 * into compatible ones, based on the Gallium CAPs.
32 *
33 * It does not upload index buffers.
34 *
35 * The module heavily uses bitmasks to represent per-buffer and
36 * per-vertex-element flags to avoid looping over the list of buffers just
37 * to see if there's a non-zero stride, or user buffer, or unsupported format,
38 * etc.
39 *
40 * There are 3 categories of vertex elements, which are processed separately:
41 * - per-vertex attribs (stride != 0, instance_divisor == 0)
42 * - instanced attribs (stride != 0, instance_divisor > 0)
43 * - constant attribs (stride == 0)
44 *
45 * All needed uploads and translations are performed every draw command, but
46 * only the subset of vertices needed for that draw command is uploaded or
47 * translated. (the module never translates whole buffers)
48 *
49 *
50 * The module consists of two main parts:
51 *
52 *
53 * 1) Translate (u_vbuf_translate_begin/end)
54 *
55 * This is pretty much a vertex fetch fallback. It translates vertices from
56 * one vertex buffer to another in an unused vertex buffer slot. It does
57 * whatever is needed to make the vertices readable by the hardware (changes
58 * vertex formats and aligns offsets and strides). The translate module is
59 * used here.
60 *
61 * Each of the 3 categories is translated to a separate buffer.
62 * Only the [min_index, max_index] range is translated. For instanced attribs,
63 * the range is [start_instance, start_instance+instance_count]. For constant
64 * attribs, the range is [0, 1].
65 *
66 *
67 * 2) User buffer uploading (u_vbuf_upload_buffers)
68 *
69 * Only the [min_index, max_index] range is uploaded (just like Translate)
70 * with a single memcpy.
71 *
72 * This method works best for non-indexed draw operations or indexed draw
73 * operations where the [min_index, max_index] range is not being way bigger
74 * than the vertex count.
75 *
76 * If the range is too big (e.g. one triangle with indices {0, 1, 10000}),
77 * the per-vertex attribs are uploaded via the translate module, all packed
78 * into one vertex buffer, and the indexed draw call is turned into
79 * a non-indexed one in the process. This adds additional complexity
80 * to the translate part, but it prevents bad apps from bringing your frame
81 * rate down.
82 *
83 *
84 * If there is nothing to do, it forwards every command to the driver.
85 * The module also has its own CSO cache of vertex element states.
86 */
87
88#include "util/u_vbuf.h"
89
90#include "util/u_dump.h"
91#include "util/u_format.h"
92#include "util/u_inlines.h"
93#include "util/u_memory.h"
94#include "util/u_upload_mgr.h"
95#include "translate/translate.h"
96#include "translate/translate_cache.h"
97#include "cso_cache/cso_cache.h"
98#include "cso_cache/cso_hash.h"
99
100struct u_vbuf_elements {
101   unsigned count;
102   struct pipe_vertex_element ve[PIPE_MAX_ATTRIBS];
103
104   unsigned src_format_size[PIPE_MAX_ATTRIBS];
105
106   /* If (velem[i].src_format != native_format[i]), the vertex buffer
107    * referenced by the vertex element cannot be used for rendering and
108    * its vertex data must be translated to native_format[i]. */
109   enum pipe_format native_format[PIPE_MAX_ATTRIBS];
110   unsigned native_format_size[PIPE_MAX_ATTRIBS];
111
112   /* Which buffers are used by the vertex element state. */
113   uint32_t used_vb_mask;
114   /* This might mean two things:
115    * - src_format != native_format, as discussed above.
116    * - src_offset % 4 != 0 (if the caps don't allow such an offset). */
117   uint32_t incompatible_elem_mask; /* each bit describes a corresp. attrib  */
118   /* Which buffer has at least one vertex element referencing it
119    * incompatible. */
120   uint32_t incompatible_vb_mask_any;
121   /* Which buffer has all vertex elements referencing it incompatible. */
122   uint32_t incompatible_vb_mask_all;
123   /* Which buffer has at least one vertex element referencing it
124    * compatible. */
125   uint32_t compatible_vb_mask_any;
126   /* Which buffer has all vertex elements referencing it compatible. */
127   uint32_t compatible_vb_mask_all;
128
129   /* Which buffer has at least one vertex element referencing it
130    * non-instanced. */
131   uint32_t noninstance_vb_mask_any;
132
133   void *driver_cso;
134};
135
136enum {
137   VB_VERTEX = 0,
138   VB_INSTANCE = 1,
139   VB_CONST = 2,
140   VB_NUM = 3
141};
142
143struct u_vbuf {
144   struct u_vbuf_caps caps;
145   bool has_signed_vb_offset;
146
147   struct pipe_context *pipe;
148   struct translate_cache *translate_cache;
149   struct cso_cache *cso_cache;
150
151   /* This is what was set in set_vertex_buffers.
152    * May contain user buffers. */
153   struct pipe_vertex_buffer vertex_buffer[PIPE_MAX_ATTRIBS];
154   uint32_t enabled_vb_mask;
155
156   /* Saved vertex buffer. */
157   struct pipe_vertex_buffer vertex_buffer0_saved;
158
159   /* Vertex buffers for the driver.
160    * There are usually no user buffers. */
161   struct pipe_vertex_buffer real_vertex_buffer[PIPE_MAX_ATTRIBS];
162   uint32_t dirty_real_vb_mask; /* which buffers are dirty since the last
163                                   call of set_vertex_buffers */
164
165   /* Vertex elements. */
166   struct u_vbuf_elements *ve, *ve_saved;
167
168   /* Vertex elements used for the translate fallback. */
169   struct pipe_vertex_element fallback_velems[PIPE_MAX_ATTRIBS];
170   /* If non-NULL, this is a vertex element state used for the translate
171    * fallback and therefore used for rendering too. */
172   boolean using_translate;
173   /* The vertex buffer slot index where translated vertices have been
174    * stored in. */
175   unsigned fallback_vbs[VB_NUM];
176
177   /* Which buffer is a user buffer. */
178   uint32_t user_vb_mask; /* each bit describes a corresp. buffer */
179   /* Which buffer is incompatible (unaligned). */
180   uint32_t incompatible_vb_mask; /* each bit describes a corresp. buffer */
181   /* Which buffer has a non-zero stride. */
182   uint32_t nonzero_stride_vb_mask; /* each bit describes a corresp. buffer */
183};
184
185static void *
186u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count,
187                              const struct pipe_vertex_element *attribs);
188static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso);
189
190static const struct {
191   enum pipe_format from, to;
192} vbuf_format_fallbacks[] = {
193   { PIPE_FORMAT_R32_FIXED,            PIPE_FORMAT_R32_FLOAT },
194   { PIPE_FORMAT_R32G32_FIXED,         PIPE_FORMAT_R32G32_FLOAT },
195   { PIPE_FORMAT_R32G32B32_FIXED,      PIPE_FORMAT_R32G32B32_FLOAT },
196   { PIPE_FORMAT_R32G32B32A32_FIXED,   PIPE_FORMAT_R32G32B32A32_FLOAT },
197   { PIPE_FORMAT_R16_FLOAT,            PIPE_FORMAT_R32_FLOAT },
198   { PIPE_FORMAT_R16G16_FLOAT,         PIPE_FORMAT_R32G32_FLOAT },
199   { PIPE_FORMAT_R16G16B16_FLOAT,      PIPE_FORMAT_R32G32B32_FLOAT },
200   { PIPE_FORMAT_R16G16B16A16_FLOAT,   PIPE_FORMAT_R32G32B32A32_FLOAT },
201   { PIPE_FORMAT_R64_FLOAT,            PIPE_FORMAT_R32_FLOAT },
202   { PIPE_FORMAT_R64G64_FLOAT,         PIPE_FORMAT_R32G32_FLOAT },
203   { PIPE_FORMAT_R64G64B64_FLOAT,      PIPE_FORMAT_R32G32B32_FLOAT },
204   { PIPE_FORMAT_R64G64B64A64_FLOAT,   PIPE_FORMAT_R32G32B32A32_FLOAT },
205   { PIPE_FORMAT_R32_UNORM,            PIPE_FORMAT_R32_FLOAT },
206   { PIPE_FORMAT_R32G32_UNORM,         PIPE_FORMAT_R32G32_FLOAT },
207   { PIPE_FORMAT_R32G32B32_UNORM,      PIPE_FORMAT_R32G32B32_FLOAT },
208   { PIPE_FORMAT_R32G32B32A32_UNORM,   PIPE_FORMAT_R32G32B32A32_FLOAT },
209   { PIPE_FORMAT_R32_SNORM,            PIPE_FORMAT_R32_FLOAT },
210   { PIPE_FORMAT_R32G32_SNORM,         PIPE_FORMAT_R32G32_FLOAT },
211   { PIPE_FORMAT_R32G32B32_SNORM,      PIPE_FORMAT_R32G32B32_FLOAT },
212   { PIPE_FORMAT_R32G32B32A32_SNORM,   PIPE_FORMAT_R32G32B32A32_FLOAT },
213   { PIPE_FORMAT_R32_USCALED,          PIPE_FORMAT_R32_FLOAT },
214   { PIPE_FORMAT_R32G32_USCALED,       PIPE_FORMAT_R32G32_FLOAT },
215   { PIPE_FORMAT_R32G32B32_USCALED,    PIPE_FORMAT_R32G32B32_FLOAT },
216   { PIPE_FORMAT_R32G32B32A32_USCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
217   { PIPE_FORMAT_R32_SSCALED,          PIPE_FORMAT_R32_FLOAT },
218   { PIPE_FORMAT_R32G32_SSCALED,       PIPE_FORMAT_R32G32_FLOAT },
219   { PIPE_FORMAT_R32G32B32_SSCALED,    PIPE_FORMAT_R32G32B32_FLOAT },
220   { PIPE_FORMAT_R32G32B32A32_SSCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
221   { PIPE_FORMAT_R16_UNORM,            PIPE_FORMAT_R32_FLOAT },
222   { PIPE_FORMAT_R16G16_UNORM,         PIPE_FORMAT_R32G32_FLOAT },
223   { PIPE_FORMAT_R16G16B16_UNORM,      PIPE_FORMAT_R32G32B32_FLOAT },
224   { PIPE_FORMAT_R16G16B16A16_UNORM,   PIPE_FORMAT_R32G32B32A32_FLOAT },
225   { PIPE_FORMAT_R16_SNORM,            PIPE_FORMAT_R32_FLOAT },
226   { PIPE_FORMAT_R16G16_SNORM,         PIPE_FORMAT_R32G32_FLOAT },
227   { PIPE_FORMAT_R16G16B16_SNORM,      PIPE_FORMAT_R32G32B32_FLOAT },
228   { PIPE_FORMAT_R16G16B16A16_SNORM,   PIPE_FORMAT_R32G32B32A32_FLOAT },
229   { PIPE_FORMAT_R16_USCALED,          PIPE_FORMAT_R32_FLOAT },
230   { PIPE_FORMAT_R16G16_USCALED,       PIPE_FORMAT_R32G32_FLOAT },
231   { PIPE_FORMAT_R16G16B16_USCALED,    PIPE_FORMAT_R32G32B32_FLOAT },
232   { PIPE_FORMAT_R16G16B16A16_USCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
233   { PIPE_FORMAT_R16_SSCALED,          PIPE_FORMAT_R32_FLOAT },
234   { PIPE_FORMAT_R16G16_SSCALED,       PIPE_FORMAT_R32G32_FLOAT },
235   { PIPE_FORMAT_R16G16B16_SSCALED,    PIPE_FORMAT_R32G32B32_FLOAT },
236   { PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
237   { PIPE_FORMAT_R8_UNORM,             PIPE_FORMAT_R32_FLOAT },
238   { PIPE_FORMAT_R8G8_UNORM,           PIPE_FORMAT_R32G32_FLOAT },
239   { PIPE_FORMAT_R8G8B8_UNORM,         PIPE_FORMAT_R32G32B32_FLOAT },
240   { PIPE_FORMAT_R8G8B8A8_UNORM,       PIPE_FORMAT_R32G32B32A32_FLOAT },
241   { PIPE_FORMAT_R8_SNORM,             PIPE_FORMAT_R32_FLOAT },
242   { PIPE_FORMAT_R8G8_SNORM,           PIPE_FORMAT_R32G32_FLOAT },
243   { PIPE_FORMAT_R8G8B8_SNORM,         PIPE_FORMAT_R32G32B32_FLOAT },
244   { PIPE_FORMAT_R8G8B8A8_SNORM,       PIPE_FORMAT_R32G32B32A32_FLOAT },
245   { PIPE_FORMAT_R8_USCALED,           PIPE_FORMAT_R32_FLOAT },
246   { PIPE_FORMAT_R8G8_USCALED,         PIPE_FORMAT_R32G32_FLOAT },
247   { PIPE_FORMAT_R8G8B8_USCALED,       PIPE_FORMAT_R32G32B32_FLOAT },
248   { PIPE_FORMAT_R8G8B8A8_USCALED,     PIPE_FORMAT_R32G32B32A32_FLOAT },
249   { PIPE_FORMAT_R8_SSCALED,           PIPE_FORMAT_R32_FLOAT },
250   { PIPE_FORMAT_R8G8_SSCALED,         PIPE_FORMAT_R32G32_FLOAT },
251   { PIPE_FORMAT_R8G8B8_SSCALED,       PIPE_FORMAT_R32G32B32_FLOAT },
252   { PIPE_FORMAT_R8G8B8A8_SSCALED,     PIPE_FORMAT_R32G32B32A32_FLOAT },
253};
254
255boolean u_vbuf_get_caps(struct pipe_screen *screen, struct u_vbuf_caps *caps,
256                        unsigned flags)
257{
258   unsigned i;
259   boolean fallback = FALSE;
260
261   /* I'd rather have a bitfield of which formats are supported and a static
262    * table of the translations indexed by format, but since we don't have C99
263    * we can't easily make a sparsely-populated table indexed by format.  So,
264    * we construct the sparse table here.
265    */
266   for (i = 0; i < PIPE_FORMAT_COUNT; i++)
267      caps->format_translation[i] = i;
268
269   for (i = 0; i < ARRAY_SIZE(vbuf_format_fallbacks); i++) {
270      enum pipe_format format = vbuf_format_fallbacks[i].from;
271
272      if (!screen->is_format_supported(screen, format, PIPE_BUFFER, 0, 0,
273                                       PIPE_BIND_VERTEX_BUFFER)) {
274         caps->format_translation[format] = vbuf_format_fallbacks[i].to;
275         fallback = TRUE;
276      }
277   }
278
279   caps->buffer_offset_unaligned =
280      !screen->get_param(screen,
281                         PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY);
282   caps->buffer_stride_unaligned =
283     !screen->get_param(screen,
284                        PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY);
285   caps->velem_src_offset_unaligned =
286      !screen->get_param(screen,
287                         PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY);
288   caps->user_vertex_buffers =
289      screen->get_param(screen, PIPE_CAP_USER_VERTEX_BUFFERS);
290
291   if (!caps->buffer_offset_unaligned ||
292       !caps->buffer_stride_unaligned ||
293       !caps->velem_src_offset_unaligned ||
294       (!(flags & U_VBUF_FLAG_NO_USER_VBOS) && !caps->user_vertex_buffers)) {
295      fallback = TRUE;
296   }
297
298   return fallback;
299}
300
301struct u_vbuf *
302u_vbuf_create(struct pipe_context *pipe, struct u_vbuf_caps *caps)
303{
304   struct u_vbuf *mgr = CALLOC_STRUCT(u_vbuf);
305
306   mgr->caps = *caps;
307   mgr->pipe = pipe;
308   mgr->cso_cache = cso_cache_create();
309   mgr->translate_cache = translate_cache_create();
310   memset(mgr->fallback_vbs, ~0, sizeof(mgr->fallback_vbs));
311
312   mgr->has_signed_vb_offset =
313      pipe->screen->get_param(pipe->screen,
314                              PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET);
315
316   return mgr;
317}
318
319/* u_vbuf uses its own caching for vertex elements, because it needs to keep
320 * its own preprocessed state per vertex element CSO. */
321static struct u_vbuf_elements *
322u_vbuf_set_vertex_elements_internal(struct u_vbuf *mgr, unsigned count,
323                                    const struct pipe_vertex_element *states)
324{
325   struct pipe_context *pipe = mgr->pipe;
326   unsigned key_size, hash_key;
327   struct cso_hash_iter iter;
328   struct u_vbuf_elements *ve;
329   struct cso_velems_state velems_state;
330
331   /* need to include the count into the stored state data too. */
332   key_size = sizeof(struct pipe_vertex_element) * count + sizeof(unsigned);
333   velems_state.count = count;
334   memcpy(velems_state.velems, states,
335          sizeof(struct pipe_vertex_element) * count);
336   hash_key = cso_construct_key((void*)&velems_state, key_size);
337   iter = cso_find_state_template(mgr->cso_cache, hash_key, CSO_VELEMENTS,
338                                  (void*)&velems_state, key_size);
339
340   if (cso_hash_iter_is_null(iter)) {
341      struct cso_velements *cso = MALLOC_STRUCT(cso_velements);
342      memcpy(&cso->state, &velems_state, key_size);
343      cso->data = u_vbuf_create_vertex_elements(mgr, count, states);
344      cso->delete_state = (cso_state_callback)u_vbuf_delete_vertex_elements;
345      cso->context = (void*)mgr;
346
347      iter = cso_insert_state(mgr->cso_cache, hash_key, CSO_VELEMENTS, cso);
348      ve = cso->data;
349   } else {
350      ve = ((struct cso_velements *)cso_hash_iter_data(iter))->data;
351   }
352
353   assert(ve);
354
355   if (ve != mgr->ve)
356      pipe->bind_vertex_elements_state(pipe, ve->driver_cso);
357
358   return ve;
359}
360
361void u_vbuf_set_vertex_elements(struct u_vbuf *mgr, unsigned count,
362                               const struct pipe_vertex_element *states)
363{
364   mgr->ve = u_vbuf_set_vertex_elements_internal(mgr, count, states);
365}
366
367void u_vbuf_destroy(struct u_vbuf *mgr)
368{
369   struct pipe_screen *screen = mgr->pipe->screen;
370   unsigned i;
371   const unsigned num_vb = screen->get_shader_param(screen, PIPE_SHADER_VERTEX,
372                                                    PIPE_SHADER_CAP_MAX_INPUTS);
373
374   mgr->pipe->set_vertex_buffers(mgr->pipe, 0, num_vb, NULL);
375
376   for (i = 0; i < PIPE_MAX_ATTRIBS; i++)
377      pipe_vertex_buffer_unreference(&mgr->vertex_buffer[i]);
378   for (i = 0; i < PIPE_MAX_ATTRIBS; i++)
379      pipe_vertex_buffer_unreference(&mgr->real_vertex_buffer[i]);
380
381   pipe_vertex_buffer_unreference(&mgr->vertex_buffer0_saved);
382
383   translate_cache_destroy(mgr->translate_cache);
384   cso_cache_delete(mgr->cso_cache);
385   FREE(mgr);
386}
387
388static enum pipe_error
389u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key,
390                         const struct pipe_draw_info *info,
391                         unsigned vb_mask, unsigned out_vb,
392                         int start_vertex, unsigned num_vertices,
393                         int min_index, boolean unroll_indices)
394{
395   struct translate *tr;
396   struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0};
397   struct pipe_resource *out_buffer = NULL;
398   uint8_t *out_map;
399   unsigned out_offset, mask;
400
401   /* Get a translate object. */
402   tr = translate_cache_find(mgr->translate_cache, key);
403
404   /* Map buffers we want to translate. */
405   mask = vb_mask;
406   while (mask) {
407      struct pipe_vertex_buffer *vb;
408      unsigned offset;
409      uint8_t *map;
410      unsigned i = u_bit_scan(&mask);
411
412      vb = &mgr->vertex_buffer[i];
413      offset = vb->buffer_offset + vb->stride * start_vertex;
414
415      if (vb->is_user_buffer) {
416         map = (uint8_t*)vb->buffer.user + offset;
417      } else {
418         unsigned size = vb->stride ? num_vertices * vb->stride
419                                    : sizeof(double)*4;
420
421         if (!vb->buffer.resource)
422            continue;
423
424         if (offset + size > vb->buffer.resource->width0) {
425            /* Don't try to map past end of buffer.  This often happens when
426             * we're translating an attribute that's at offset > 0 from the
427             * start of the vertex.  If we'd subtract attrib's offset from
428             * the size, this probably wouldn't happen.
429             */
430            size = vb->buffer.resource->width0 - offset;
431
432            /* Also adjust num_vertices.  A common user error is to call
433             * glDrawRangeElements() with incorrect 'end' argument.  The 'end
434             * value should be the max index value, but people often
435             * accidentally add one to this value.  This adjustment avoids
436             * crashing (by reading past the end of a hardware buffer mapping)
437             * when people do that.
438             */
439            num_vertices = (size + vb->stride - 1) / vb->stride;
440         }
441
442         map = pipe_buffer_map_range(mgr->pipe, vb->buffer.resource, offset, size,
443                                     PIPE_TRANSFER_READ, &vb_transfer[i]);
444      }
445
446      /* Subtract min_index so that indexing with the index buffer works. */
447      if (unroll_indices) {
448         map -= (ptrdiff_t)vb->stride * min_index;
449      }
450
451      tr->set_buffer(tr, i, map, vb->stride, info->max_index);
452   }
453
454   /* Translate. */
455   if (unroll_indices) {
456      struct pipe_transfer *transfer = NULL;
457      const unsigned offset = info->start * info->index_size;
458      uint8_t *map;
459
460      /* Create and map the output buffer. */
461      u_upload_alloc(mgr->pipe->stream_uploader, 0,
462                     key->output_stride * info->count, 4,
463                     &out_offset, &out_buffer,
464                     (void**)&out_map);
465      if (!out_buffer)
466         return PIPE_ERROR_OUT_OF_MEMORY;
467
468      if (info->has_user_indices) {
469         map = (uint8_t*)info->index.user + offset;
470      } else {
471         map = pipe_buffer_map_range(mgr->pipe, info->index.resource, offset,
472                                     info->count * info->index_size,
473                                     PIPE_TRANSFER_READ, &transfer);
474      }
475
476      switch (info->index_size) {
477      case 4:
478         tr->run_elts(tr, (unsigned*)map, info->count, 0, 0, out_map);
479         break;
480      case 2:
481         tr->run_elts16(tr, (uint16_t*)map, info->count, 0, 0, out_map);
482         break;
483      case 1:
484         tr->run_elts8(tr, map, info->count, 0, 0, out_map);
485         break;
486      }
487
488      if (transfer) {
489         pipe_buffer_unmap(mgr->pipe, transfer);
490      }
491   } else {
492      /* Create and map the output buffer. */
493      u_upload_alloc(mgr->pipe->stream_uploader,
494                     mgr->has_signed_vb_offset ?
495                        0 : key->output_stride * start_vertex,
496                     key->output_stride * num_vertices, 4,
497                     &out_offset, &out_buffer,
498                     (void**)&out_map);
499      if (!out_buffer)
500         return PIPE_ERROR_OUT_OF_MEMORY;
501
502      out_offset -= key->output_stride * start_vertex;
503
504      tr->run(tr, 0, num_vertices, 0, 0, out_map);
505   }
506
507   /* Unmap all buffers. */
508   mask = vb_mask;
509   while (mask) {
510      unsigned i = u_bit_scan(&mask);
511
512      if (vb_transfer[i]) {
513         pipe_buffer_unmap(mgr->pipe, vb_transfer[i]);
514      }
515   }
516
517   /* Setup the new vertex buffer. */
518   mgr->real_vertex_buffer[out_vb].buffer_offset = out_offset;
519   mgr->real_vertex_buffer[out_vb].stride = key->output_stride;
520
521   /* Move the buffer reference. */
522   pipe_vertex_buffer_unreference(&mgr->real_vertex_buffer[out_vb]);
523   mgr->real_vertex_buffer[out_vb].buffer.resource = out_buffer;
524   mgr->real_vertex_buffer[out_vb].is_user_buffer = false;
525
526   return PIPE_OK;
527}
528
529static boolean
530u_vbuf_translate_find_free_vb_slots(struct u_vbuf *mgr,
531                                    unsigned mask[VB_NUM])
532{
533   unsigned type;
534   unsigned fallback_vbs[VB_NUM];
535   /* Set the bit for each buffer which is incompatible, or isn't set. */
536   uint32_t unused_vb_mask =
537      mgr->ve->incompatible_vb_mask_all | mgr->incompatible_vb_mask |
538      ~mgr->enabled_vb_mask;
539
540   memset(fallback_vbs, ~0, sizeof(fallback_vbs));
541
542   /* Find free slots for each type if needed. */
543   for (type = 0; type < VB_NUM; type++) {
544      if (mask[type]) {
545         uint32_t index;
546
547         if (!unused_vb_mask) {
548            return FALSE;
549         }
550
551         index = ffs(unused_vb_mask) - 1;
552         fallback_vbs[type] = index;
553         unused_vb_mask &= ~(1 << index);
554         /*printf("found slot=%i for type=%i\n", index, type);*/
555      }
556   }
557
558   for (type = 0; type < VB_NUM; type++) {
559      if (mask[type]) {
560         mgr->dirty_real_vb_mask |= 1 << fallback_vbs[type];
561      }
562   }
563
564   memcpy(mgr->fallback_vbs, fallback_vbs, sizeof(fallback_vbs));
565   return TRUE;
566}
567
568static boolean
569u_vbuf_translate_begin(struct u_vbuf *mgr,
570                       const struct pipe_draw_info *info,
571                       int start_vertex, unsigned num_vertices,
572                       int min_index, boolean unroll_indices)
573{
574   unsigned mask[VB_NUM] = {0};
575   struct translate_key key[VB_NUM];
576   unsigned elem_index[VB_NUM][PIPE_MAX_ATTRIBS]; /* ... into key.elements */
577   unsigned i, type;
578   const unsigned incompatible_vb_mask = mgr->incompatible_vb_mask &
579                                         mgr->ve->used_vb_mask;
580
581   const int start[VB_NUM] = {
582      start_vertex,           /* VERTEX */
583      info->start_instance,   /* INSTANCE */
584      0                       /* CONST */
585   };
586
587   const unsigned num[VB_NUM] = {
588      num_vertices,           /* VERTEX */
589      info->instance_count,   /* INSTANCE */
590      1                       /* CONST */
591   };
592
593   memset(key, 0, sizeof(key));
594   memset(elem_index, ~0, sizeof(elem_index));
595
596   /* See if there are vertex attribs of each type to translate and
597    * which ones. */
598   for (i = 0; i < mgr->ve->count; i++) {
599      unsigned vb_index = mgr->ve->ve[i].vertex_buffer_index;
600
601      if (!mgr->vertex_buffer[vb_index].stride) {
602         if (!(mgr->ve->incompatible_elem_mask & (1 << i)) &&
603             !(incompatible_vb_mask & (1 << vb_index))) {
604            continue;
605         }
606         mask[VB_CONST] |= 1 << vb_index;
607      } else if (mgr->ve->ve[i].instance_divisor) {
608         if (!(mgr->ve->incompatible_elem_mask & (1 << i)) &&
609             !(incompatible_vb_mask & (1 << vb_index))) {
610            continue;
611         }
612         mask[VB_INSTANCE] |= 1 << vb_index;
613      } else {
614         if (!unroll_indices &&
615             !(mgr->ve->incompatible_elem_mask & (1 << i)) &&
616             !(incompatible_vb_mask & (1 << vb_index))) {
617            continue;
618         }
619         mask[VB_VERTEX] |= 1 << vb_index;
620      }
621   }
622
623   assert(mask[VB_VERTEX] || mask[VB_INSTANCE] || mask[VB_CONST]);
624
625   /* Find free vertex buffer slots. */
626   if (!u_vbuf_translate_find_free_vb_slots(mgr, mask)) {
627      return FALSE;
628   }
629
630   /* Initialize the translate keys. */
631   for (i = 0; i < mgr->ve->count; i++) {
632      struct translate_key *k;
633      struct translate_element *te;
634      enum pipe_format output_format = mgr->ve->native_format[i];
635      unsigned bit, vb_index = mgr->ve->ve[i].vertex_buffer_index;
636      bit = 1 << vb_index;
637
638      if (!(mgr->ve->incompatible_elem_mask & (1 << i)) &&
639          !(incompatible_vb_mask & (1 << vb_index)) &&
640          (!unroll_indices || !(mask[VB_VERTEX] & bit))) {
641         continue;
642      }
643
644      /* Set type to what we will translate.
645       * Whether vertex, instance, or constant attribs. */
646      for (type = 0; type < VB_NUM; type++) {
647         if (mask[type] & bit) {
648            break;
649         }
650      }
651      assert(type < VB_NUM);
652      if (mgr->ve->ve[i].src_format != output_format)
653         assert(translate_is_output_format_supported(output_format));
654      /*printf("velem=%i type=%i\n", i, type);*/
655
656      /* Add the vertex element. */
657      k = &key[type];
658      elem_index[type][i] = k->nr_elements;
659
660      te = &k->element[k->nr_elements];
661      te->type = TRANSLATE_ELEMENT_NORMAL;
662      te->instance_divisor = 0;
663      te->input_buffer = vb_index;
664      te->input_format = mgr->ve->ve[i].src_format;
665      te->input_offset = mgr->ve->ve[i].src_offset;
666      te->output_format = output_format;
667      te->output_offset = k->output_stride;
668
669      k->output_stride += mgr->ve->native_format_size[i];
670      k->nr_elements++;
671   }
672
673   /* Translate buffers. */
674   for (type = 0; type < VB_NUM; type++) {
675      if (key[type].nr_elements) {
676         enum pipe_error err;
677         err = u_vbuf_translate_buffers(mgr, &key[type], info, mask[type],
678                                        mgr->fallback_vbs[type],
679                                        start[type], num[type], min_index,
680                                        unroll_indices && type == VB_VERTEX);
681         if (err != PIPE_OK)
682            return FALSE;
683
684         /* Fixup the stride for constant attribs. */
685         if (type == VB_CONST) {
686            mgr->real_vertex_buffer[mgr->fallback_vbs[VB_CONST]].stride = 0;
687         }
688      }
689   }
690
691   /* Setup new vertex elements. */
692   for (i = 0; i < mgr->ve->count; i++) {
693      for (type = 0; type < VB_NUM; type++) {
694         if (elem_index[type][i] < key[type].nr_elements) {
695            struct translate_element *te = &key[type].element[elem_index[type][i]];
696            mgr->fallback_velems[i].instance_divisor = mgr->ve->ve[i].instance_divisor;
697            mgr->fallback_velems[i].src_format = te->output_format;
698            mgr->fallback_velems[i].src_offset = te->output_offset;
699            mgr->fallback_velems[i].vertex_buffer_index = mgr->fallback_vbs[type];
700
701            /* elem_index[type][i] can only be set for one type. */
702            assert(type > VB_INSTANCE || elem_index[type+1][i] == ~0u);
703            assert(type > VB_VERTEX   || elem_index[type+2][i] == ~0u);
704            break;
705         }
706      }
707      /* No translating, just copy the original vertex element over. */
708      if (type == VB_NUM) {
709         memcpy(&mgr->fallback_velems[i], &mgr->ve->ve[i],
710                sizeof(struct pipe_vertex_element));
711      }
712   }
713
714   u_vbuf_set_vertex_elements_internal(mgr, mgr->ve->count,
715                                       mgr->fallback_velems);
716   mgr->using_translate = TRUE;
717   return TRUE;
718}
719
720static void u_vbuf_translate_end(struct u_vbuf *mgr)
721{
722   unsigned i;
723
724   /* Restore vertex elements. */
725   mgr->pipe->bind_vertex_elements_state(mgr->pipe, mgr->ve->driver_cso);
726   mgr->using_translate = FALSE;
727
728   /* Unreference the now-unused VBOs. */
729   for (i = 0; i < VB_NUM; i++) {
730      unsigned vb = mgr->fallback_vbs[i];
731      if (vb != ~0u) {
732         pipe_resource_reference(&mgr->real_vertex_buffer[vb].buffer.resource, NULL);
733         mgr->fallback_vbs[i] = ~0;
734
735         /* This will cause the buffer to be unbound in the driver later. */
736         mgr->dirty_real_vb_mask |= 1 << vb;
737      }
738   }
739}
740
741static void *
742u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count,
743                              const struct pipe_vertex_element *attribs)
744{
745   struct pipe_context *pipe = mgr->pipe;
746   unsigned i;
747   struct pipe_vertex_element driver_attribs[PIPE_MAX_ATTRIBS];
748   struct u_vbuf_elements *ve = CALLOC_STRUCT(u_vbuf_elements);
749   uint32_t used_buffers = 0;
750
751   ve->count = count;
752
753   memcpy(ve->ve, attribs, sizeof(struct pipe_vertex_element) * count);
754   memcpy(driver_attribs, attribs, sizeof(struct pipe_vertex_element) * count);
755
756   /* Set the best native format in case the original format is not
757    * supported. */
758   for (i = 0; i < count; i++) {
759      enum pipe_format format = ve->ve[i].src_format;
760
761      ve->src_format_size[i] = util_format_get_blocksize(format);
762
763      used_buffers |= 1 << ve->ve[i].vertex_buffer_index;
764
765      if (!ve->ve[i].instance_divisor) {
766         ve->noninstance_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index;
767      }
768
769      format = mgr->caps.format_translation[format];
770
771      driver_attribs[i].src_format = format;
772      ve->native_format[i] = format;
773      ve->native_format_size[i] =
774            util_format_get_blocksize(ve->native_format[i]);
775
776      if (ve->ve[i].src_format != format ||
777          (!mgr->caps.velem_src_offset_unaligned &&
778           ve->ve[i].src_offset % 4 != 0)) {
779         ve->incompatible_elem_mask |= 1 << i;
780         ve->incompatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index;
781      } else {
782         ve->compatible_vb_mask_any |= 1 << ve->ve[i].vertex_buffer_index;
783      }
784   }
785
786   ve->used_vb_mask = used_buffers;
787   ve->compatible_vb_mask_all = ~ve->incompatible_vb_mask_any & used_buffers;
788   ve->incompatible_vb_mask_all = ~ve->compatible_vb_mask_any & used_buffers;
789
790   /* Align the formats and offsets to the size of DWORD if needed. */
791   if (!mgr->caps.velem_src_offset_unaligned) {
792      for (i = 0; i < count; i++) {
793         ve->native_format_size[i] = align(ve->native_format_size[i], 4);
794         driver_attribs[i].src_offset = align(ve->ve[i].src_offset, 4);
795      }
796   }
797
798   ve->driver_cso =
799      pipe->create_vertex_elements_state(pipe, count, driver_attribs);
800   return ve;
801}
802
803static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso)
804{
805   struct pipe_context *pipe = mgr->pipe;
806   struct u_vbuf_elements *ve = cso;
807
808   pipe->delete_vertex_elements_state(pipe, ve->driver_cso);
809   FREE(ve);
810}
811
812void u_vbuf_set_vertex_buffers(struct u_vbuf *mgr,
813                               unsigned start_slot, unsigned count,
814                               const struct pipe_vertex_buffer *bufs)
815{
816   unsigned i;
817   /* which buffers are enabled */
818   uint32_t enabled_vb_mask = 0;
819   /* which buffers are in user memory */
820   uint32_t user_vb_mask = 0;
821   /* which buffers are incompatible with the driver */
822   uint32_t incompatible_vb_mask = 0;
823   /* which buffers have a non-zero stride */
824   uint32_t nonzero_stride_vb_mask = 0;
825   const uint32_t mask = ~(((1ull << count) - 1) << start_slot);
826
827   /* Zero out the bits we are going to rewrite completely. */
828   mgr->user_vb_mask &= mask;
829   mgr->incompatible_vb_mask &= mask;
830   mgr->nonzero_stride_vb_mask &= mask;
831   mgr->enabled_vb_mask &= mask;
832
833   if (!bufs) {
834      struct pipe_context *pipe = mgr->pipe;
835      /* Unbind. */
836      mgr->dirty_real_vb_mask &= mask;
837
838      for (i = 0; i < count; i++) {
839         unsigned dst_index = start_slot + i;
840
841         pipe_vertex_buffer_unreference(&mgr->vertex_buffer[dst_index]);
842         pipe_vertex_buffer_unreference(&mgr->real_vertex_buffer[dst_index]);
843      }
844
845      pipe->set_vertex_buffers(pipe, start_slot, count, NULL);
846      return;
847   }
848
849   for (i = 0; i < count; i++) {
850      unsigned dst_index = start_slot + i;
851      const struct pipe_vertex_buffer *vb = &bufs[i];
852      struct pipe_vertex_buffer *orig_vb = &mgr->vertex_buffer[dst_index];
853      struct pipe_vertex_buffer *real_vb = &mgr->real_vertex_buffer[dst_index];
854
855      if (!vb->buffer.resource) {
856         pipe_vertex_buffer_unreference(orig_vb);
857         pipe_vertex_buffer_unreference(real_vb);
858         continue;
859      }
860
861      pipe_vertex_buffer_reference(orig_vb, vb);
862
863      if (vb->stride) {
864         nonzero_stride_vb_mask |= 1 << dst_index;
865      }
866      enabled_vb_mask |= 1 << dst_index;
867
868      if ((!mgr->caps.buffer_offset_unaligned && vb->buffer_offset % 4 != 0) ||
869          (!mgr->caps.buffer_stride_unaligned && vb->stride % 4 != 0)) {
870         incompatible_vb_mask |= 1 << dst_index;
871         real_vb->buffer_offset = vb->buffer_offset;
872         real_vb->stride = vb->stride;
873         pipe_vertex_buffer_unreference(real_vb);
874         real_vb->is_user_buffer = false;
875         continue;
876      }
877
878      if (!mgr->caps.user_vertex_buffers && vb->is_user_buffer) {
879         user_vb_mask |= 1 << dst_index;
880         real_vb->buffer_offset = vb->buffer_offset;
881         real_vb->stride = vb->stride;
882         pipe_vertex_buffer_unreference(real_vb);
883         real_vb->is_user_buffer = false;
884         continue;
885      }
886
887      pipe_vertex_buffer_reference(real_vb, vb);
888   }
889
890   mgr->user_vb_mask |= user_vb_mask;
891   mgr->incompatible_vb_mask |= incompatible_vb_mask;
892   mgr->nonzero_stride_vb_mask |= nonzero_stride_vb_mask;
893   mgr->enabled_vb_mask |= enabled_vb_mask;
894
895   /* All changed buffers are marked as dirty, even the NULL ones,
896    * which will cause the NULL buffers to be unbound in the driver later. */
897   mgr->dirty_real_vb_mask |= ~mask;
898}
899
900static enum pipe_error
901u_vbuf_upload_buffers(struct u_vbuf *mgr,
902                      int start_vertex, unsigned num_vertices,
903                      int start_instance, unsigned num_instances)
904{
905   unsigned i;
906   unsigned nr_velems = mgr->ve->count;
907   const struct pipe_vertex_element *velems =
908         mgr->using_translate ? mgr->fallback_velems : mgr->ve->ve;
909   unsigned start_offset[PIPE_MAX_ATTRIBS];
910   unsigned end_offset[PIPE_MAX_ATTRIBS];
911   uint32_t buffer_mask = 0;
912
913   /* Determine how much data needs to be uploaded. */
914   for (i = 0; i < nr_velems; i++) {
915      const struct pipe_vertex_element *velem = &velems[i];
916      unsigned index = velem->vertex_buffer_index;
917      struct pipe_vertex_buffer *vb = &mgr->vertex_buffer[index];
918      unsigned instance_div, first, size, index_bit;
919
920      /* Skip the buffers generated by translate. */
921      if (index == mgr->fallback_vbs[VB_VERTEX] ||
922          index == mgr->fallback_vbs[VB_INSTANCE] ||
923          index == mgr->fallback_vbs[VB_CONST]) {
924         continue;
925      }
926
927      if (!vb->is_user_buffer) {
928         continue;
929      }
930
931      instance_div = velem->instance_divisor;
932      first = vb->buffer_offset + velem->src_offset;
933
934      if (!vb->stride) {
935         /* Constant attrib. */
936         size = mgr->ve->src_format_size[i];
937      } else if (instance_div) {
938         /* Per-instance attrib. */
939
940         /* Figure out how many instances we'll render given instance_div.  We
941          * can't use the typical div_round_up() pattern because the CTS uses
942          * instance_div = ~0 for a test, which overflows div_round_up()'s
943          * addition.
944          */
945         unsigned count = num_instances / instance_div;
946         if (count * instance_div != num_instances)
947            count++;
948
949         first += vb->stride * start_instance;
950         size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
951      } else {
952         /* Per-vertex attrib. */
953         first += vb->stride * start_vertex;
954         size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i];
955      }
956
957      index_bit = 1 << index;
958
959      /* Update offsets. */
960      if (!(buffer_mask & index_bit)) {
961         start_offset[index] = first;
962         end_offset[index] = first + size;
963      } else {
964         if (first < start_offset[index])
965            start_offset[index] = first;
966         if (first + size > end_offset[index])
967            end_offset[index] = first + size;
968      }
969
970      buffer_mask |= index_bit;
971   }
972
973   /* Upload buffers. */
974   while (buffer_mask) {
975      unsigned start, end;
976      struct pipe_vertex_buffer *real_vb;
977      const uint8_t *ptr;
978
979      i = u_bit_scan(&buffer_mask);
980
981      start = start_offset[i];
982      end = end_offset[i];
983      assert(start < end);
984
985      real_vb = &mgr->real_vertex_buffer[i];
986      ptr = mgr->vertex_buffer[i].buffer.user;
987
988      u_upload_data(mgr->pipe->stream_uploader,
989                    mgr->has_signed_vb_offset ? 0 : start,
990                    end - start, 4,
991                    ptr + start, &real_vb->buffer_offset, &real_vb->buffer.resource);
992      if (!real_vb->buffer.resource)
993         return PIPE_ERROR_OUT_OF_MEMORY;
994
995      real_vb->buffer_offset -= start;
996   }
997
998   return PIPE_OK;
999}
1000
1001static boolean u_vbuf_need_minmax_index(const struct u_vbuf *mgr)
1002{
1003   /* See if there are any per-vertex attribs which will be uploaded or
1004    * translated. Use bitmasks to get the info instead of looping over vertex
1005    * elements. */
1006   return (mgr->ve->used_vb_mask &
1007           ((mgr->user_vb_mask |
1008             mgr->incompatible_vb_mask |
1009             mgr->ve->incompatible_vb_mask_any) &
1010            mgr->ve->noninstance_vb_mask_any &
1011            mgr->nonzero_stride_vb_mask)) != 0;
1012}
1013
1014static boolean u_vbuf_mapping_vertex_buffer_blocks(const struct u_vbuf *mgr)
1015{
1016   /* Return true if there are hw buffers which don't need to be translated.
1017    *
1018    * We could query whether each buffer is busy, but that would
1019    * be way more costly than this. */
1020   return (mgr->ve->used_vb_mask &
1021           (~mgr->user_vb_mask &
1022            ~mgr->incompatible_vb_mask &
1023            mgr->ve->compatible_vb_mask_all &
1024            mgr->ve->noninstance_vb_mask_any &
1025            mgr->nonzero_stride_vb_mask)) != 0;
1026}
1027
1028static void
1029u_vbuf_get_minmax_index_mapped(const struct pipe_draw_info *info,
1030                               const void *indices, unsigned *out_min_index,
1031                               unsigned *out_max_index)
1032{
1033   unsigned max = 0;
1034   unsigned min = ~0u;
1035
1036   switch (info->index_size) {
1037   case 4: {
1038      const unsigned *ui_indices = (const unsigned*)indices;
1039      if (info->primitive_restart) {
1040         for (unsigned i = 0; i < info->count; i++) {
1041            if (ui_indices[i] != info->restart_index) {
1042               if (ui_indices[i] > max) max = ui_indices[i];
1043               if (ui_indices[i] < min) min = ui_indices[i];
1044            }
1045         }
1046      }
1047      else {
1048         for (unsigned i = 0; i < info->count; i++) {
1049            if (ui_indices[i] > max) max = ui_indices[i];
1050            if (ui_indices[i] < min) min = ui_indices[i];
1051         }
1052      }
1053      break;
1054   }
1055   case 2: {
1056      const unsigned short *us_indices = (const unsigned short*)indices;
1057      if (info->primitive_restart) {
1058         for (unsigned i = 0; i < info->count; i++) {
1059            if (us_indices[i] != info->restart_index) {
1060               if (us_indices[i] > max) max = us_indices[i];
1061               if (us_indices[i] < min) min = us_indices[i];
1062            }
1063         }
1064      }
1065      else {
1066         for (unsigned i = 0; i < info->count; i++) {
1067            if (us_indices[i] > max) max = us_indices[i];
1068            if (us_indices[i] < min) min = us_indices[i];
1069         }
1070      }
1071      break;
1072   }
1073   case 1: {
1074      const unsigned char *ub_indices = (const unsigned char*)indices;
1075      if (info->primitive_restart) {
1076         for (unsigned i = 0; i < info->count; i++) {
1077            if (ub_indices[i] != info->restart_index) {
1078               if (ub_indices[i] > max) max = ub_indices[i];
1079               if (ub_indices[i] < min) min = ub_indices[i];
1080            }
1081         }
1082      }
1083      else {
1084         for (unsigned i = 0; i < info->count; i++) {
1085            if (ub_indices[i] > max) max = ub_indices[i];
1086            if (ub_indices[i] < min) min = ub_indices[i];
1087         }
1088      }
1089      break;
1090   }
1091   default:
1092      assert(0);
1093   }
1094
1095   *out_min_index = min;
1096   *out_max_index = max;
1097}
1098
1099void u_vbuf_get_minmax_index(struct pipe_context *pipe,
1100                             const struct pipe_draw_info *info,
1101                             unsigned *out_min_index, unsigned *out_max_index)
1102{
1103   struct pipe_transfer *transfer = NULL;
1104   const void *indices;
1105
1106   if (info->has_user_indices) {
1107      indices = (uint8_t*)info->index.user +
1108                info->start * info->index_size;
1109   } else {
1110      indices = pipe_buffer_map_range(pipe, info->index.resource,
1111                                      info->start * info->index_size,
1112                                      info->count * info->index_size,
1113                                      PIPE_TRANSFER_READ, &transfer);
1114   }
1115
1116   u_vbuf_get_minmax_index_mapped(info, indices, out_min_index, out_max_index);
1117
1118   if (transfer) {
1119      pipe_buffer_unmap(pipe, transfer);
1120   }
1121}
1122
1123static void u_vbuf_set_driver_vertex_buffers(struct u_vbuf *mgr)
1124{
1125   struct pipe_context *pipe = mgr->pipe;
1126   unsigned start_slot, count;
1127
1128   start_slot = ffs(mgr->dirty_real_vb_mask) - 1;
1129   count = util_last_bit(mgr->dirty_real_vb_mask >> start_slot);
1130
1131   pipe->set_vertex_buffers(pipe, start_slot, count,
1132                            mgr->real_vertex_buffer + start_slot);
1133   mgr->dirty_real_vb_mask = 0;
1134}
1135
1136static void
1137u_vbuf_split_indexed_multidraw(struct u_vbuf *mgr, struct pipe_draw_info *info,
1138                               unsigned *indirect_data, unsigned stride,
1139                               unsigned draw_count)
1140{
1141   assert(info->index_size);
1142   info->indirect = NULL;
1143
1144   for (unsigned i = 0; i < draw_count; i++) {
1145      unsigned offset = i * stride / 4;
1146
1147      info->count = indirect_data[offset + 0];
1148      info->instance_count = indirect_data[offset + 1];
1149
1150      if (!info->count || !info->instance_count)
1151         continue;
1152
1153      info->start = indirect_data[offset + 2];
1154      info->index_bias = indirect_data[offset + 3];
1155      info->start_instance = indirect_data[offset + 4];
1156
1157      u_vbuf_draw_vbo(mgr, info);
1158   }
1159}
1160
1161void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info)
1162{
1163   struct pipe_context *pipe = mgr->pipe;
1164   int start_vertex;
1165   unsigned min_index;
1166   unsigned num_vertices;
1167   boolean unroll_indices = FALSE;
1168   const uint32_t used_vb_mask = mgr->ve->used_vb_mask;
1169   uint32_t user_vb_mask = mgr->user_vb_mask & used_vb_mask;
1170   const uint32_t incompatible_vb_mask =
1171      mgr->incompatible_vb_mask & used_vb_mask;
1172   struct pipe_draw_info new_info;
1173
1174   /* Normal draw. No fallback and no user buffers. */
1175   if (!incompatible_vb_mask &&
1176       !mgr->ve->incompatible_elem_mask &&
1177       !user_vb_mask) {
1178
1179      /* Set vertex buffers if needed. */
1180      if (mgr->dirty_real_vb_mask & used_vb_mask) {
1181         u_vbuf_set_driver_vertex_buffers(mgr);
1182      }
1183
1184      pipe->draw_vbo(pipe, info);
1185      return;
1186   }
1187
1188   new_info = *info;
1189
1190   /* Handle indirect (multi)draws. */
1191   if (new_info.indirect) {
1192      const struct pipe_draw_indirect_info *indirect = new_info.indirect;
1193      unsigned draw_count = 0;
1194
1195      /* Get the number of draws. */
1196      if (indirect->indirect_draw_count) {
1197         pipe_buffer_read(pipe, indirect->indirect_draw_count,
1198                          indirect->indirect_draw_count_offset,
1199                          4, &draw_count);
1200      } else {
1201         draw_count = indirect->draw_count;
1202      }
1203
1204      if (!draw_count)
1205         return;
1206
1207      unsigned data_size = (draw_count - 1) * indirect->stride +
1208                           (new_info.index_size ? 20 : 16);
1209      unsigned *data = malloc(data_size);
1210      if (!data)
1211         return; /* report an error? */
1212
1213      /* Read the used buffer range only once, because the read can be
1214       * uncached.
1215       */
1216      pipe_buffer_read(pipe, indirect->buffer, indirect->offset, data_size,
1217                       data);
1218
1219      if (info->index_size) {
1220         /* Indexed multidraw. */
1221         unsigned index_bias0 = data[3];
1222         bool index_bias_same = true;
1223
1224         /* If we invoke the translate path, we have to split the multidraw. */
1225         if (incompatible_vb_mask ||
1226             mgr->ve->incompatible_elem_mask) {
1227            u_vbuf_split_indexed_multidraw(mgr, &new_info, data,
1228                                           indirect->stride, draw_count);
1229            free(data);
1230            return;
1231         }
1232
1233         /* See if index_bias is the same for all draws. */
1234         for (unsigned i = 1; i < draw_count; i++) {
1235            if (data[i * indirect->stride / 4 + 3] != index_bias0) {
1236               index_bias_same = false;
1237               break;
1238            }
1239         }
1240
1241         /* Split the multidraw if index_bias is different. */
1242         if (!index_bias_same) {
1243            u_vbuf_split_indexed_multidraw(mgr, &new_info, data,
1244                                           indirect->stride, draw_count);
1245            free(data);
1246            return;
1247         }
1248
1249         /* If we don't need to use the translate path and index_bias is
1250          * the same, we can process the multidraw with the time complexity
1251          * equal to 1 draw call (except for the index range computation).
1252          * We only need to compute the index range covering all draw calls
1253          * of the multidraw.
1254          *
1255          * The driver will not look at these values because indirect != NULL.
1256          * These values determine the user buffer bounds to upload.
1257          */
1258         new_info.index_bias = index_bias0;
1259         new_info.min_index = ~0u;
1260         new_info.max_index = 0;
1261         new_info.start_instance = ~0u;
1262         unsigned end_instance = 0;
1263
1264         struct pipe_transfer *transfer = NULL;
1265         const uint8_t *indices;
1266
1267         if (info->has_user_indices) {
1268            indices = (uint8_t*)info->index.user;
1269         } else {
1270            indices = (uint8_t*)pipe_buffer_map(pipe, info->index.resource,
1271                                                PIPE_TRANSFER_READ, &transfer);
1272         }
1273
1274         for (unsigned i = 0; i < draw_count; i++) {
1275            unsigned offset = i * indirect->stride / 4;
1276            unsigned start = data[offset + 2];
1277            unsigned count = data[offset + 0];
1278            unsigned start_instance = data[offset + 4];
1279            unsigned instance_count = data[offset + 1];
1280
1281            if (!count || !instance_count)
1282               continue;
1283
1284            /* Update the ranges of instances. */
1285            new_info.start_instance = MIN2(new_info.start_instance,
1286                                           start_instance);
1287            end_instance = MAX2(end_instance, start_instance + instance_count);
1288
1289            /* Update the index range. */
1290            unsigned min, max;
1291            new_info.count = count; /* only used by get_minmax_index */
1292            u_vbuf_get_minmax_index_mapped(&new_info,
1293                                           indices +
1294                                           new_info.index_size * start,
1295                                           &min, &max);
1296
1297            new_info.min_index = MIN2(new_info.min_index, min);
1298            new_info.max_index = MAX2(new_info.max_index, max);
1299         }
1300         free(data);
1301
1302         if (transfer)
1303            pipe_buffer_unmap(pipe, transfer);
1304
1305         /* Set the final instance count. */
1306         new_info.instance_count = end_instance - new_info.start_instance;
1307
1308         if (new_info.start_instance == ~0u || !new_info.instance_count)
1309            return;
1310      } else {
1311         /* Non-indexed multidraw.
1312          *
1313          * Keep the draw call indirect and compute minimums & maximums,
1314          * which will determine the user buffer bounds to upload, but
1315          * the driver will not look at these values because indirect != NULL.
1316          *
1317          * This efficiently processes the multidraw with the time complexity
1318          * equal to 1 draw call.
1319          */
1320         new_info.start = ~0u;
1321         new_info.start_instance = ~0u;
1322         unsigned end_vertex = 0;
1323         unsigned end_instance = 0;
1324
1325         for (unsigned i = 0; i < draw_count; i++) {
1326            unsigned offset = i * indirect->stride / 4;
1327            unsigned start = data[offset + 2];
1328            unsigned count = data[offset + 0];
1329            unsigned start_instance = data[offset + 3];
1330            unsigned instance_count = data[offset + 1];
1331
1332            new_info.start = MIN2(new_info.start, start);
1333            new_info.start_instance = MIN2(new_info.start_instance,
1334                                           start_instance);
1335
1336            end_vertex = MAX2(end_vertex, start + count);
1337            end_instance = MAX2(end_instance, start_instance + instance_count);
1338         }
1339         free(data);
1340
1341         /* Set the final counts. */
1342         new_info.count = end_vertex - new_info.start;
1343         new_info.instance_count = end_instance - new_info.start_instance;
1344
1345         if (new_info.start == ~0u || !new_info.count || !new_info.instance_count)
1346            return;
1347      }
1348   }
1349
1350   if (new_info.index_size) {
1351      /* See if anything needs to be done for per-vertex attribs. */
1352      if (u_vbuf_need_minmax_index(mgr)) {
1353         unsigned max_index;
1354
1355         if (new_info.max_index != ~0u) {
1356            min_index = new_info.min_index;
1357            max_index = new_info.max_index;
1358         } else {
1359            u_vbuf_get_minmax_index(mgr->pipe, &new_info,
1360                                    &min_index, &max_index);
1361         }
1362
1363         assert(min_index <= max_index);
1364
1365         start_vertex = min_index + new_info.index_bias;
1366         num_vertices = max_index + 1 - min_index;
1367
1368         /* Primitive restart doesn't work when unrolling indices.
1369          * We would have to break this drawing operation into several ones. */
1370         /* Use some heuristic to see if unrolling indices improves
1371          * performance. */
1372         if (!info->indirect &&
1373             !new_info.primitive_restart &&
1374             num_vertices > new_info.count*2 &&
1375             num_vertices - new_info.count > 32 &&
1376             !u_vbuf_mapping_vertex_buffer_blocks(mgr)) {
1377            unroll_indices = TRUE;
1378            user_vb_mask &= ~(mgr->nonzero_stride_vb_mask &
1379                              mgr->ve->noninstance_vb_mask_any);
1380         }
1381      } else {
1382         /* Nothing to do for per-vertex attribs. */
1383         start_vertex = 0;
1384         num_vertices = 0;
1385         min_index = 0;
1386      }
1387   } else {
1388      start_vertex = new_info.start;
1389      num_vertices = new_info.count;
1390      min_index = 0;
1391   }
1392
1393   /* Translate vertices with non-native layouts or formats. */
1394   if (unroll_indices ||
1395       incompatible_vb_mask ||
1396       mgr->ve->incompatible_elem_mask) {
1397      if (!u_vbuf_translate_begin(mgr, &new_info, start_vertex, num_vertices,
1398                                  min_index, unroll_indices)) {
1399         debug_warn_once("u_vbuf_translate_begin() failed");
1400         return;
1401      }
1402
1403      if (unroll_indices) {
1404         new_info.index_size = 0;
1405         new_info.index_bias = 0;
1406         new_info.min_index = 0;
1407         new_info.max_index = new_info.count - 1;
1408         new_info.start = 0;
1409      }
1410
1411      user_vb_mask &= ~(incompatible_vb_mask |
1412                        mgr->ve->incompatible_vb_mask_all);
1413   }
1414
1415   /* Upload user buffers. */
1416   if (user_vb_mask) {
1417      if (u_vbuf_upload_buffers(mgr, start_vertex, num_vertices,
1418                                new_info.start_instance,
1419                                new_info.instance_count) != PIPE_OK) {
1420         debug_warn_once("u_vbuf_upload_buffers() failed");
1421         return;
1422      }
1423
1424      mgr->dirty_real_vb_mask |= user_vb_mask;
1425   }
1426
1427   /*
1428   if (unroll_indices) {
1429      printf("unrolling indices: start_vertex = %i, num_vertices = %i\n",
1430             start_vertex, num_vertices);
1431      util_dump_draw_info(stdout, info);
1432      printf("\n");
1433   }
1434
1435   unsigned i;
1436   for (i = 0; i < mgr->nr_vertex_buffers; i++) {
1437      printf("input %i: ", i);
1438      util_dump_vertex_buffer(stdout, mgr->vertex_buffer+i);
1439      printf("\n");
1440   }
1441   for (i = 0; i < mgr->nr_real_vertex_buffers; i++) {
1442      printf("real %i: ", i);
1443      util_dump_vertex_buffer(stdout, mgr->real_vertex_buffer+i);
1444      printf("\n");
1445   }
1446   */
1447
1448   u_upload_unmap(pipe->stream_uploader);
1449   u_vbuf_set_driver_vertex_buffers(mgr);
1450
1451   pipe->draw_vbo(pipe, &new_info);
1452
1453   if (mgr->using_translate) {
1454      u_vbuf_translate_end(mgr);
1455   }
1456}
1457
1458void u_vbuf_save_vertex_elements(struct u_vbuf *mgr)
1459{
1460   assert(!mgr->ve_saved);
1461   mgr->ve_saved = mgr->ve;
1462}
1463
1464void u_vbuf_restore_vertex_elements(struct u_vbuf *mgr)
1465{
1466   if (mgr->ve != mgr->ve_saved) {
1467      struct pipe_context *pipe = mgr->pipe;
1468
1469      mgr->ve = mgr->ve_saved;
1470      pipe->bind_vertex_elements_state(pipe,
1471                                       mgr->ve ? mgr->ve->driver_cso : NULL);
1472   }
1473   mgr->ve_saved = NULL;
1474}
1475
1476void u_vbuf_save_vertex_buffer0(struct u_vbuf *mgr)
1477{
1478   pipe_vertex_buffer_reference(&mgr->vertex_buffer0_saved,
1479                                &mgr->vertex_buffer[0]);
1480}
1481
1482void u_vbuf_restore_vertex_buffer0(struct u_vbuf *mgr)
1483{
1484   u_vbuf_set_vertex_buffers(mgr, 0, 1, &mgr->vertex_buffer0_saved);
1485   pipe_vertex_buffer_unreference(&mgr->vertex_buffer0_saved);
1486}
1487