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
100 struct 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
136 enum {
137 VB_VERTEX = 0,
138 VB_INSTANCE = 1,
139 VB_CONST = 2,
140 VB_NUM = 3
141 };
142
143 struct 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 unsigned aux_vertex_buffer_slot;
158 struct pipe_vertex_buffer aux_vertex_buffer_saved;
159
160 /* Vertex buffers for the driver.
161 * There are usually no user buffers. */
162 struct pipe_vertex_buffer real_vertex_buffer[PIPE_MAX_ATTRIBS];
163 uint32_t dirty_real_vb_mask; /* which buffers are dirty since the last
164 call of set_vertex_buffers */
165
166 /* Vertex elements. */
167 struct u_vbuf_elements *ve, *ve_saved;
168
169 /* Vertex elements used for the translate fallback. */
170 struct pipe_vertex_element fallback_velems[PIPE_MAX_ATTRIBS];
171 /* If non-NULL, this is a vertex element state used for the translate
172 * fallback and therefore used for rendering too. */
173 boolean using_translate;
174 /* The vertex buffer slot index where translated vertices have been
175 * stored in. */
176 unsigned fallback_vbs[VB_NUM];
177
178 /* Which buffer is a user buffer. */
179 uint32_t user_vb_mask; /* each bit describes a corresp. buffer */
180 /* Which buffer is incompatible (unaligned). */
181 uint32_t incompatible_vb_mask; /* each bit describes a corresp. buffer */
182 /* Which buffer has a non-zero stride. */
183 uint32_t nonzero_stride_vb_mask; /* each bit describes a corresp. buffer */
184 };
185
186 static void *
187 u_vbuf_create_vertex_elements(struct u_vbuf *mgr, unsigned count,
188 const struct pipe_vertex_element *attribs);
189 static void u_vbuf_delete_vertex_elements(struct u_vbuf *mgr, void *cso);
190
191 static const struct {
192 enum pipe_format from, to;
193 } vbuf_format_fallbacks[] = {
194 { PIPE_FORMAT_R32_FIXED, PIPE_FORMAT_R32_FLOAT },
195 { PIPE_FORMAT_R32G32_FIXED, PIPE_FORMAT_R32G32_FLOAT },
196 { PIPE_FORMAT_R32G32B32_FIXED, PIPE_FORMAT_R32G32B32_FLOAT },
197 { PIPE_FORMAT_R32G32B32A32_FIXED, PIPE_FORMAT_R32G32B32A32_FLOAT },
198 { PIPE_FORMAT_R16_FLOAT, PIPE_FORMAT_R32_FLOAT },
199 { PIPE_FORMAT_R16G16_FLOAT, PIPE_FORMAT_R32G32_FLOAT },
200 { PIPE_FORMAT_R16G16B16_FLOAT, PIPE_FORMAT_R32G32B32_FLOAT },
201 { PIPE_FORMAT_R16G16B16A16_FLOAT, PIPE_FORMAT_R32G32B32A32_FLOAT },
202 { PIPE_FORMAT_R64_FLOAT, PIPE_FORMAT_R32_FLOAT },
203 { PIPE_FORMAT_R64G64_FLOAT, PIPE_FORMAT_R32G32_FLOAT },
204 { PIPE_FORMAT_R64G64B64_FLOAT, PIPE_FORMAT_R32G32B32_FLOAT },
205 { PIPE_FORMAT_R64G64B64A64_FLOAT, PIPE_FORMAT_R32G32B32A32_FLOAT },
206 { PIPE_FORMAT_R32_UNORM, PIPE_FORMAT_R32_FLOAT },
207 { PIPE_FORMAT_R32G32_UNORM, PIPE_FORMAT_R32G32_FLOAT },
208 { PIPE_FORMAT_R32G32B32_UNORM, PIPE_FORMAT_R32G32B32_FLOAT },
209 { PIPE_FORMAT_R32G32B32A32_UNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
210 { PIPE_FORMAT_R32_SNORM, PIPE_FORMAT_R32_FLOAT },
211 { PIPE_FORMAT_R32G32_SNORM, PIPE_FORMAT_R32G32_FLOAT },
212 { PIPE_FORMAT_R32G32B32_SNORM, PIPE_FORMAT_R32G32B32_FLOAT },
213 { PIPE_FORMAT_R32G32B32A32_SNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
214 { PIPE_FORMAT_R32_USCALED, PIPE_FORMAT_R32_FLOAT },
215 { PIPE_FORMAT_R32G32_USCALED, PIPE_FORMAT_R32G32_FLOAT },
216 { PIPE_FORMAT_R32G32B32_USCALED, PIPE_FORMAT_R32G32B32_FLOAT },
217 { PIPE_FORMAT_R32G32B32A32_USCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
218 { PIPE_FORMAT_R32_SSCALED, PIPE_FORMAT_R32_FLOAT },
219 { PIPE_FORMAT_R32G32_SSCALED, PIPE_FORMAT_R32G32_FLOAT },
220 { PIPE_FORMAT_R32G32B32_SSCALED, PIPE_FORMAT_R32G32B32_FLOAT },
221 { PIPE_FORMAT_R32G32B32A32_SSCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
222 { PIPE_FORMAT_R16_UNORM, PIPE_FORMAT_R32_FLOAT },
223 { PIPE_FORMAT_R16G16_UNORM, PIPE_FORMAT_R32G32_FLOAT },
224 { PIPE_FORMAT_R16G16B16_UNORM, PIPE_FORMAT_R32G32B32_FLOAT },
225 { PIPE_FORMAT_R16G16B16A16_UNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
226 { PIPE_FORMAT_R16_SNORM, PIPE_FORMAT_R32_FLOAT },
227 { PIPE_FORMAT_R16G16_SNORM, PIPE_FORMAT_R32G32_FLOAT },
228 { PIPE_FORMAT_R16G16B16_SNORM, PIPE_FORMAT_R32G32B32_FLOAT },
229 { PIPE_FORMAT_R16G16B16A16_SNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
230 { PIPE_FORMAT_R16_USCALED, PIPE_FORMAT_R32_FLOAT },
231 { PIPE_FORMAT_R16G16_USCALED, PIPE_FORMAT_R32G32_FLOAT },
232 { PIPE_FORMAT_R16G16B16_USCALED, PIPE_FORMAT_R32G32B32_FLOAT },
233 { PIPE_FORMAT_R16G16B16A16_USCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
234 { PIPE_FORMAT_R16_SSCALED, PIPE_FORMAT_R32_FLOAT },
235 { PIPE_FORMAT_R16G16_SSCALED, PIPE_FORMAT_R32G32_FLOAT },
236 { PIPE_FORMAT_R16G16B16_SSCALED, PIPE_FORMAT_R32G32B32_FLOAT },
237 { PIPE_FORMAT_R16G16B16A16_SSCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
238 { PIPE_FORMAT_R8_UNORM, PIPE_FORMAT_R32_FLOAT },
239 { PIPE_FORMAT_R8G8_UNORM, PIPE_FORMAT_R32G32_FLOAT },
240 { PIPE_FORMAT_R8G8B8_UNORM, PIPE_FORMAT_R32G32B32_FLOAT },
241 { PIPE_FORMAT_R8G8B8A8_UNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
242 { PIPE_FORMAT_R8_SNORM, PIPE_FORMAT_R32_FLOAT },
243 { PIPE_FORMAT_R8G8_SNORM, PIPE_FORMAT_R32G32_FLOAT },
244 { PIPE_FORMAT_R8G8B8_SNORM, PIPE_FORMAT_R32G32B32_FLOAT },
245 { PIPE_FORMAT_R8G8B8A8_SNORM, PIPE_FORMAT_R32G32B32A32_FLOAT },
246 { PIPE_FORMAT_R8_USCALED, PIPE_FORMAT_R32_FLOAT },
247 { PIPE_FORMAT_R8G8_USCALED, PIPE_FORMAT_R32G32_FLOAT },
248 { PIPE_FORMAT_R8G8B8_USCALED, PIPE_FORMAT_R32G32B32_FLOAT },
249 { PIPE_FORMAT_R8G8B8A8_USCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
250 { PIPE_FORMAT_R8_SSCALED, PIPE_FORMAT_R32_FLOAT },
251 { PIPE_FORMAT_R8G8_SSCALED, PIPE_FORMAT_R32G32_FLOAT },
252 { PIPE_FORMAT_R8G8B8_SSCALED, PIPE_FORMAT_R32G32B32_FLOAT },
253 { PIPE_FORMAT_R8G8B8A8_SSCALED, PIPE_FORMAT_R32G32B32A32_FLOAT },
254 };
255
u_vbuf_get_caps(struct pipe_screen * screen,struct u_vbuf_caps * caps,unsigned flags)256 boolean u_vbuf_get_caps(struct pipe_screen *screen, struct u_vbuf_caps *caps,
257 unsigned flags)
258 {
259 unsigned i;
260 boolean fallback = FALSE;
261
262 /* I'd rather have a bitfield of which formats are supported and a static
263 * table of the translations indexed by format, but since we don't have C99
264 * we can't easily make a sparsely-populated table indexed by format. So,
265 * we construct the sparse table here.
266 */
267 for (i = 0; i < PIPE_FORMAT_COUNT; i++)
268 caps->format_translation[i] = i;
269
270 for (i = 0; i < ARRAY_SIZE(vbuf_format_fallbacks); i++) {
271 enum pipe_format format = vbuf_format_fallbacks[i].from;
272
273 if (!screen->is_format_supported(screen, format, PIPE_BUFFER, 0,
274 PIPE_BIND_VERTEX_BUFFER)) {
275 caps->format_translation[format] = vbuf_format_fallbacks[i].to;
276 fallback = TRUE;
277 }
278 }
279
280 caps->buffer_offset_unaligned =
281 !screen->get_param(screen,
282 PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY);
283 caps->buffer_stride_unaligned =
284 !screen->get_param(screen,
285 PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY);
286 caps->velem_src_offset_unaligned =
287 !screen->get_param(screen,
288 PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY);
289 caps->user_vertex_buffers =
290 screen->get_param(screen, PIPE_CAP_USER_VERTEX_BUFFERS);
291
292 if (!caps->buffer_offset_unaligned ||
293 !caps->buffer_stride_unaligned ||
294 !caps->velem_src_offset_unaligned ||
295 (!(flags & U_VBUF_FLAG_NO_USER_VBOS) && !caps->user_vertex_buffers)) {
296 fallback = TRUE;
297 }
298
299 return fallback;
300 }
301
302 struct u_vbuf *
u_vbuf_create(struct pipe_context * pipe,struct u_vbuf_caps * caps,unsigned aux_vertex_buffer_index)303 u_vbuf_create(struct pipe_context *pipe,
304 struct u_vbuf_caps *caps, unsigned aux_vertex_buffer_index)
305 {
306 struct u_vbuf *mgr = CALLOC_STRUCT(u_vbuf);
307
308 mgr->caps = *caps;
309 mgr->aux_vertex_buffer_slot = aux_vertex_buffer_index;
310 mgr->pipe = pipe;
311 mgr->cso_cache = cso_cache_create();
312 mgr->translate_cache = translate_cache_create();
313 memset(mgr->fallback_vbs, ~0, sizeof(mgr->fallback_vbs));
314
315 mgr->has_signed_vb_offset =
316 pipe->screen->get_param(pipe->screen,
317 PIPE_CAP_SIGNED_VERTEX_BUFFER_OFFSET);
318
319 return mgr;
320 }
321
322 /* u_vbuf uses its own caching for vertex elements, because it needs to keep
323 * its own preprocessed state per vertex element CSO. */
324 static struct u_vbuf_elements *
u_vbuf_set_vertex_elements_internal(struct u_vbuf * mgr,unsigned count,const struct pipe_vertex_element * states)325 u_vbuf_set_vertex_elements_internal(struct u_vbuf *mgr, unsigned count,
326 const struct pipe_vertex_element *states)
327 {
328 struct pipe_context *pipe = mgr->pipe;
329 unsigned key_size, hash_key;
330 struct cso_hash_iter iter;
331 struct u_vbuf_elements *ve;
332 struct cso_velems_state velems_state;
333
334 /* need to include the count into the stored state data too. */
335 key_size = sizeof(struct pipe_vertex_element) * count + sizeof(unsigned);
336 velems_state.count = count;
337 memcpy(velems_state.velems, states,
338 sizeof(struct pipe_vertex_element) * count);
339 hash_key = cso_construct_key((void*)&velems_state, key_size);
340 iter = cso_find_state_template(mgr->cso_cache, hash_key, CSO_VELEMENTS,
341 (void*)&velems_state, key_size);
342
343 if (cso_hash_iter_is_null(iter)) {
344 struct cso_velements *cso = MALLOC_STRUCT(cso_velements);
345 memcpy(&cso->state, &velems_state, key_size);
346 cso->data = u_vbuf_create_vertex_elements(mgr, count, states);
347 cso->delete_state = (cso_state_callback)u_vbuf_delete_vertex_elements;
348 cso->context = (void*)mgr;
349
350 iter = cso_insert_state(mgr->cso_cache, hash_key, CSO_VELEMENTS, cso);
351 ve = cso->data;
352 } else {
353 ve = ((struct cso_velements *)cso_hash_iter_data(iter))->data;
354 }
355
356 assert(ve);
357
358 if (ve != mgr->ve)
359 pipe->bind_vertex_elements_state(pipe, ve->driver_cso);
360
361 return ve;
362 }
363
u_vbuf_set_vertex_elements(struct u_vbuf * mgr,unsigned count,const struct pipe_vertex_element * states)364 void u_vbuf_set_vertex_elements(struct u_vbuf *mgr, unsigned count,
365 const struct pipe_vertex_element *states)
366 {
367 mgr->ve = u_vbuf_set_vertex_elements_internal(mgr, count, states);
368 }
369
u_vbuf_destroy(struct u_vbuf * mgr)370 void u_vbuf_destroy(struct u_vbuf *mgr)
371 {
372 struct pipe_screen *screen = mgr->pipe->screen;
373 unsigned i;
374 const unsigned num_vb = screen->get_shader_param(screen, PIPE_SHADER_VERTEX,
375 PIPE_SHADER_CAP_MAX_INPUTS);
376
377 mgr->pipe->set_vertex_buffers(mgr->pipe, 0, num_vb, NULL);
378
379 for (i = 0; i < PIPE_MAX_ATTRIBS; i++)
380 pipe_vertex_buffer_unreference(&mgr->vertex_buffer[i]);
381 for (i = 0; i < PIPE_MAX_ATTRIBS; i++)
382 pipe_vertex_buffer_unreference(&mgr->real_vertex_buffer[i]);
383
384 pipe_vertex_buffer_unreference(&mgr->aux_vertex_buffer_saved);
385
386 translate_cache_destroy(mgr->translate_cache);
387 cso_cache_delete(mgr->cso_cache);
388 FREE(mgr);
389 }
390
391 static enum pipe_error
u_vbuf_translate_buffers(struct u_vbuf * mgr,struct translate_key * key,const struct pipe_draw_info * info,unsigned vb_mask,unsigned out_vb,int start_vertex,unsigned num_vertices,int min_index,boolean unroll_indices)392 u_vbuf_translate_buffers(struct u_vbuf *mgr, struct translate_key *key,
393 const struct pipe_draw_info *info,
394 unsigned vb_mask, unsigned out_vb,
395 int start_vertex, unsigned num_vertices,
396 int min_index, boolean unroll_indices)
397 {
398 struct translate *tr;
399 struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS] = {0};
400 struct pipe_resource *out_buffer = NULL;
401 uint8_t *out_map;
402 unsigned out_offset, mask;
403
404 /* Get a translate object. */
405 tr = translate_cache_find(mgr->translate_cache, key);
406
407 /* Map buffers we want to translate. */
408 mask = vb_mask;
409 while (mask) {
410 struct pipe_vertex_buffer *vb;
411 unsigned offset;
412 uint8_t *map;
413 unsigned i = u_bit_scan(&mask);
414
415 vb = &mgr->vertex_buffer[i];
416 offset = vb->buffer_offset + vb->stride * start_vertex;
417
418 if (vb->is_user_buffer) {
419 map = (uint8_t*)vb->buffer.user + offset;
420 } else {
421 unsigned size = vb->stride ? num_vertices * vb->stride
422 : sizeof(double)*4;
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, ~0);
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
529 static boolean
u_vbuf_translate_find_free_vb_slots(struct u_vbuf * mgr,unsigned mask[VB_NUM])530 u_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
568 static boolean
u_vbuf_translate_begin(struct u_vbuf * mgr,const struct pipe_draw_info * info,int start_vertex,unsigned num_vertices,int min_index,boolean unroll_indices)569 u_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
u_vbuf_translate_end(struct u_vbuf * mgr)720 static 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
741 static void *
u_vbuf_create_vertex_elements(struct u_vbuf * mgr,unsigned count,const struct pipe_vertex_element * attribs)742 u_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
u_vbuf_delete_vertex_elements(struct u_vbuf * mgr,void * cso)803 static 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
u_vbuf_set_vertex_buffers(struct u_vbuf * mgr,unsigned start_slot,unsigned count,const struct pipe_vertex_buffer * bufs)812 void 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
900 static enum pipe_error
u_vbuf_upload_buffers(struct u_vbuf * mgr,int start_vertex,unsigned num_vertices,int start_instance,unsigned num_instances)901 u_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 unsigned count = (num_instances + instance_div - 1) / instance_div;
940 first += vb->stride * start_instance;
941 size = vb->stride * (count - 1) + mgr->ve->src_format_size[i];
942 } else {
943 /* Per-vertex attrib. */
944 first += vb->stride * start_vertex;
945 size = vb->stride * (num_vertices - 1) + mgr->ve->src_format_size[i];
946 }
947
948 index_bit = 1 << index;
949
950 /* Update offsets. */
951 if (!(buffer_mask & index_bit)) {
952 start_offset[index] = first;
953 end_offset[index] = first + size;
954 } else {
955 if (first < start_offset[index])
956 start_offset[index] = first;
957 if (first + size > end_offset[index])
958 end_offset[index] = first + size;
959 }
960
961 buffer_mask |= index_bit;
962 }
963
964 /* Upload buffers. */
965 while (buffer_mask) {
966 unsigned start, end;
967 struct pipe_vertex_buffer *real_vb;
968 const uint8_t *ptr;
969
970 i = u_bit_scan(&buffer_mask);
971
972 start = start_offset[i];
973 end = end_offset[i];
974 assert(start < end);
975
976 real_vb = &mgr->real_vertex_buffer[i];
977 ptr = mgr->vertex_buffer[i].buffer.user;
978
979 u_upload_data(mgr->pipe->stream_uploader,
980 mgr->has_signed_vb_offset ? 0 : start,
981 end - start, 4,
982 ptr + start, &real_vb->buffer_offset, &real_vb->buffer.resource);
983 if (!real_vb->buffer.resource)
984 return PIPE_ERROR_OUT_OF_MEMORY;
985
986 real_vb->buffer_offset -= start;
987 }
988
989 return PIPE_OK;
990 }
991
u_vbuf_need_minmax_index(const struct u_vbuf * mgr)992 static boolean u_vbuf_need_minmax_index(const struct u_vbuf *mgr)
993 {
994 /* See if there are any per-vertex attribs which will be uploaded or
995 * translated. Use bitmasks to get the info instead of looping over vertex
996 * elements. */
997 return (mgr->ve->used_vb_mask &
998 ((mgr->user_vb_mask |
999 mgr->incompatible_vb_mask |
1000 mgr->ve->incompatible_vb_mask_any) &
1001 mgr->ve->noninstance_vb_mask_any &
1002 mgr->nonzero_stride_vb_mask)) != 0;
1003 }
1004
u_vbuf_mapping_vertex_buffer_blocks(const struct u_vbuf * mgr)1005 static boolean u_vbuf_mapping_vertex_buffer_blocks(const struct u_vbuf *mgr)
1006 {
1007 /* Return true if there are hw buffers which don't need to be translated.
1008 *
1009 * We could query whether each buffer is busy, but that would
1010 * be way more costly than this. */
1011 return (mgr->ve->used_vb_mask &
1012 (~mgr->user_vb_mask &
1013 ~mgr->incompatible_vb_mask &
1014 mgr->ve->compatible_vb_mask_all &
1015 mgr->ve->noninstance_vb_mask_any &
1016 mgr->nonzero_stride_vb_mask)) != 0;
1017 }
1018
u_vbuf_get_minmax_index(struct pipe_context * pipe,const struct pipe_draw_info * info,int * out_min_index,int * out_max_index)1019 static void u_vbuf_get_minmax_index(struct pipe_context *pipe,
1020 const struct pipe_draw_info *info,
1021 int *out_min_index, int *out_max_index)
1022 {
1023 struct pipe_transfer *transfer = NULL;
1024 const void *indices;
1025 unsigned i;
1026
1027 if (info->has_user_indices) {
1028 indices = (uint8_t*)info->index.user +
1029 info->start * info->index_size;
1030 } else {
1031 indices = pipe_buffer_map_range(pipe, info->index.resource,
1032 info->start * info->index_size,
1033 info->count * info->index_size,
1034 PIPE_TRANSFER_READ, &transfer);
1035 }
1036
1037 switch (info->index_size) {
1038 case 4: {
1039 const unsigned *ui_indices = (const unsigned*)indices;
1040 unsigned max_ui = 0;
1041 unsigned min_ui = ~0U;
1042 if (info->primitive_restart) {
1043 for (i = 0; i < info->count; i++) {
1044 if (ui_indices[i] != info->restart_index) {
1045 if (ui_indices[i] > max_ui) max_ui = ui_indices[i];
1046 if (ui_indices[i] < min_ui) min_ui = ui_indices[i];
1047 }
1048 }
1049 }
1050 else {
1051 for (i = 0; i < info->count; i++) {
1052 if (ui_indices[i] > max_ui) max_ui = ui_indices[i];
1053 if (ui_indices[i] < min_ui) min_ui = ui_indices[i];
1054 }
1055 }
1056 *out_min_index = min_ui;
1057 *out_max_index = max_ui;
1058 break;
1059 }
1060 case 2: {
1061 const unsigned short *us_indices = (const unsigned short*)indices;
1062 unsigned max_us = 0;
1063 unsigned min_us = ~0U;
1064 if (info->primitive_restart) {
1065 for (i = 0; i < info->count; i++) {
1066 if (us_indices[i] != info->restart_index) {
1067 if (us_indices[i] > max_us) max_us = us_indices[i];
1068 if (us_indices[i] < min_us) min_us = us_indices[i];
1069 }
1070 }
1071 }
1072 else {
1073 for (i = 0; i < info->count; i++) {
1074 if (us_indices[i] > max_us) max_us = us_indices[i];
1075 if (us_indices[i] < min_us) min_us = us_indices[i];
1076 }
1077 }
1078 *out_min_index = min_us;
1079 *out_max_index = max_us;
1080 break;
1081 }
1082 case 1: {
1083 const unsigned char *ub_indices = (const unsigned char*)indices;
1084 unsigned max_ub = 0;
1085 unsigned min_ub = ~0U;
1086 if (info->primitive_restart) {
1087 for (i = 0; i < info->count; i++) {
1088 if (ub_indices[i] != info->restart_index) {
1089 if (ub_indices[i] > max_ub) max_ub = ub_indices[i];
1090 if (ub_indices[i] < min_ub) min_ub = ub_indices[i];
1091 }
1092 }
1093 }
1094 else {
1095 for (i = 0; i < info->count; i++) {
1096 if (ub_indices[i] > max_ub) max_ub = ub_indices[i];
1097 if (ub_indices[i] < min_ub) min_ub = ub_indices[i];
1098 }
1099 }
1100 *out_min_index = min_ub;
1101 *out_max_index = max_ub;
1102 break;
1103 }
1104 default:
1105 assert(0);
1106 *out_min_index = 0;
1107 *out_max_index = 0;
1108 }
1109
1110 if (transfer) {
1111 pipe_buffer_unmap(pipe, transfer);
1112 }
1113 }
1114
u_vbuf_set_driver_vertex_buffers(struct u_vbuf * mgr)1115 static void u_vbuf_set_driver_vertex_buffers(struct u_vbuf *mgr)
1116 {
1117 struct pipe_context *pipe = mgr->pipe;
1118 unsigned start_slot, count;
1119
1120 start_slot = ffs(mgr->dirty_real_vb_mask) - 1;
1121 count = util_last_bit(mgr->dirty_real_vb_mask >> start_slot);
1122
1123 pipe->set_vertex_buffers(pipe, start_slot, count,
1124 mgr->real_vertex_buffer + start_slot);
1125 mgr->dirty_real_vb_mask = 0;
1126 }
1127
u_vbuf_draw_vbo(struct u_vbuf * mgr,const struct pipe_draw_info * info)1128 void u_vbuf_draw_vbo(struct u_vbuf *mgr, const struct pipe_draw_info *info)
1129 {
1130 struct pipe_context *pipe = mgr->pipe;
1131 int start_vertex, min_index;
1132 unsigned num_vertices;
1133 boolean unroll_indices = FALSE;
1134 const uint32_t used_vb_mask = mgr->ve->used_vb_mask;
1135 uint32_t user_vb_mask = mgr->user_vb_mask & used_vb_mask;
1136 const uint32_t incompatible_vb_mask =
1137 mgr->incompatible_vb_mask & used_vb_mask;
1138 struct pipe_draw_info new_info;
1139
1140 /* Normal draw. No fallback and no user buffers. */
1141 if (!incompatible_vb_mask &&
1142 !mgr->ve->incompatible_elem_mask &&
1143 !user_vb_mask) {
1144
1145 /* Set vertex buffers if needed. */
1146 if (mgr->dirty_real_vb_mask & used_vb_mask) {
1147 u_vbuf_set_driver_vertex_buffers(mgr);
1148 }
1149
1150 pipe->draw_vbo(pipe, info);
1151 return;
1152 }
1153
1154 new_info = *info;
1155
1156 /* Fallback. We need to know all the parameters. */
1157 if (new_info.indirect) {
1158 struct pipe_transfer *transfer = NULL;
1159 int *data;
1160
1161 if (new_info.index_size) {
1162 data = pipe_buffer_map_range(pipe, new_info.indirect->buffer,
1163 new_info.indirect->offset, 20,
1164 PIPE_TRANSFER_READ, &transfer);
1165 new_info.index_bias = data[3];
1166 new_info.start_instance = data[4];
1167 }
1168 else {
1169 data = pipe_buffer_map_range(pipe, new_info.indirect->buffer,
1170 new_info.indirect->offset, 16,
1171 PIPE_TRANSFER_READ, &transfer);
1172 new_info.start_instance = data[3];
1173 }
1174
1175 new_info.count = data[0];
1176 new_info.instance_count = data[1];
1177 new_info.start = data[2];
1178 pipe_buffer_unmap(pipe, transfer);
1179 new_info.indirect = NULL;
1180 }
1181
1182 if (new_info.index_size) {
1183 /* See if anything needs to be done for per-vertex attribs. */
1184 if (u_vbuf_need_minmax_index(mgr)) {
1185 int max_index;
1186
1187 if (new_info.max_index != ~0u) {
1188 min_index = new_info.min_index;
1189 max_index = new_info.max_index;
1190 } else {
1191 u_vbuf_get_minmax_index(mgr->pipe, &new_info,
1192 &min_index, &max_index);
1193 }
1194
1195 assert(min_index <= max_index);
1196
1197 start_vertex = min_index + new_info.index_bias;
1198 num_vertices = max_index + 1 - min_index;
1199
1200 /* Primitive restart doesn't work when unrolling indices.
1201 * We would have to break this drawing operation into several ones. */
1202 /* Use some heuristic to see if unrolling indices improves
1203 * performance. */
1204 if (!new_info.primitive_restart &&
1205 num_vertices > new_info.count*2 &&
1206 num_vertices - new_info.count > 32 &&
1207 !u_vbuf_mapping_vertex_buffer_blocks(mgr)) {
1208 unroll_indices = TRUE;
1209 user_vb_mask &= ~(mgr->nonzero_stride_vb_mask &
1210 mgr->ve->noninstance_vb_mask_any);
1211 }
1212 } else {
1213 /* Nothing to do for per-vertex attribs. */
1214 start_vertex = 0;
1215 num_vertices = 0;
1216 min_index = 0;
1217 }
1218 } else {
1219 start_vertex = new_info.start;
1220 num_vertices = new_info.count;
1221 min_index = 0;
1222 }
1223
1224 /* Translate vertices with non-native layouts or formats. */
1225 if (unroll_indices ||
1226 incompatible_vb_mask ||
1227 mgr->ve->incompatible_elem_mask) {
1228 if (!u_vbuf_translate_begin(mgr, &new_info, start_vertex, num_vertices,
1229 min_index, unroll_indices)) {
1230 debug_warn_once("u_vbuf_translate_begin() failed");
1231 return;
1232 }
1233
1234 if (unroll_indices) {
1235 new_info.index_size = 0;
1236 new_info.index_bias = 0;
1237 new_info.min_index = 0;
1238 new_info.max_index = new_info.count - 1;
1239 new_info.start = 0;
1240 }
1241
1242 user_vb_mask &= ~(incompatible_vb_mask |
1243 mgr->ve->incompatible_vb_mask_all);
1244 }
1245
1246 /* Upload user buffers. */
1247 if (user_vb_mask) {
1248 if (u_vbuf_upload_buffers(mgr, start_vertex, num_vertices,
1249 new_info.start_instance,
1250 new_info.instance_count) != PIPE_OK) {
1251 debug_warn_once("u_vbuf_upload_buffers() failed");
1252 return;
1253 }
1254
1255 mgr->dirty_real_vb_mask |= user_vb_mask;
1256 }
1257
1258 /*
1259 if (unroll_indices) {
1260 printf("unrolling indices: start_vertex = %i, num_vertices = %i\n",
1261 start_vertex, num_vertices);
1262 util_dump_draw_info(stdout, info);
1263 printf("\n");
1264 }
1265
1266 unsigned i;
1267 for (i = 0; i < mgr->nr_vertex_buffers; i++) {
1268 printf("input %i: ", i);
1269 util_dump_vertex_buffer(stdout, mgr->vertex_buffer+i);
1270 printf("\n");
1271 }
1272 for (i = 0; i < mgr->nr_real_vertex_buffers; i++) {
1273 printf("real %i: ", i);
1274 util_dump_vertex_buffer(stdout, mgr->real_vertex_buffer+i);
1275 printf("\n");
1276 }
1277 */
1278
1279 u_upload_unmap(pipe->stream_uploader);
1280 u_vbuf_set_driver_vertex_buffers(mgr);
1281
1282 pipe->draw_vbo(pipe, &new_info);
1283
1284 if (mgr->using_translate) {
1285 u_vbuf_translate_end(mgr);
1286 }
1287 }
1288
u_vbuf_save_vertex_elements(struct u_vbuf * mgr)1289 void u_vbuf_save_vertex_elements(struct u_vbuf *mgr)
1290 {
1291 assert(!mgr->ve_saved);
1292 mgr->ve_saved = mgr->ve;
1293 }
1294
u_vbuf_restore_vertex_elements(struct u_vbuf * mgr)1295 void u_vbuf_restore_vertex_elements(struct u_vbuf *mgr)
1296 {
1297 if (mgr->ve != mgr->ve_saved) {
1298 struct pipe_context *pipe = mgr->pipe;
1299
1300 mgr->ve = mgr->ve_saved;
1301 pipe->bind_vertex_elements_state(pipe,
1302 mgr->ve ? mgr->ve->driver_cso : NULL);
1303 }
1304 mgr->ve_saved = NULL;
1305 }
1306
u_vbuf_save_aux_vertex_buffer_slot(struct u_vbuf * mgr)1307 void u_vbuf_save_aux_vertex_buffer_slot(struct u_vbuf *mgr)
1308 {
1309 pipe_vertex_buffer_reference(&mgr->aux_vertex_buffer_saved,
1310 &mgr->vertex_buffer[mgr->aux_vertex_buffer_slot]);
1311 }
1312
u_vbuf_restore_aux_vertex_buffer_slot(struct u_vbuf * mgr)1313 void u_vbuf_restore_aux_vertex_buffer_slot(struct u_vbuf *mgr)
1314 {
1315 u_vbuf_set_vertex_buffers(mgr, mgr->aux_vertex_buffer_slot, 1,
1316 &mgr->aux_vertex_buffer_saved);
1317 pipe_vertex_buffer_unreference(&mgr->aux_vertex_buffer_saved);
1318 }
1319