1/*
2 * Copyright 2010 Christoph Bumiller
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23#include "pipe/p_context.h"
24#include "pipe/p_state.h"
25#include "util/u_draw.h"
26#include "util/u_inlines.h"
27#include "util/u_prim.h"
28#include "util/format/u_format.h"
29#include "translate/translate.h"
30
31#include "nv50/nv50_context.h"
32#include "nv50/nv50_query_hw.h"
33#include "nv50/nv50_resource.h"
34
35#include "nv50/nv50_3d.xml.h"
36
37void
38nv50_vertex_state_delete(struct pipe_context *pipe,
39                         void *hwcso)
40{
41   struct nv50_vertex_stateobj *so = hwcso;
42
43   if (so->translate)
44      so->translate->release(so->translate);
45   FREE(hwcso);
46}
47
48void *
49nv50_vertex_state_create(struct pipe_context *pipe,
50                         unsigned num_elements,
51                         const struct pipe_vertex_element *elements)
52{
53    struct nv50_vertex_stateobj *so;
54    struct translate_key transkey;
55    unsigned i;
56
57    so = MALLOC(sizeof(*so) +
58                num_elements * sizeof(struct nv50_vertex_element));
59    if (!so)
60        return NULL;
61    so->num_elements = num_elements;
62    so->instance_elts = 0;
63    so->instance_bufs = 0;
64    so->need_conversion = false;
65
66    memset(so->vb_access_size, 0, sizeof(so->vb_access_size));
67
68    for (i = 0; i < PIPE_MAX_ATTRIBS; ++i)
69       so->min_instance_div[i] = 0xffffffff;
70
71    transkey.nr_elements = 0;
72    transkey.output_stride = 0;
73
74    for (i = 0; i < num_elements; ++i) {
75        const struct pipe_vertex_element *ve = &elements[i];
76        const unsigned vbi = ve->vertex_buffer_index;
77        unsigned size;
78        enum pipe_format fmt = ve->src_format;
79
80        so->element[i].pipe = elements[i];
81        so->element[i].state = nv50_vertex_format[fmt].vtx;
82
83        if (!so->element[i].state) {
84            switch (util_format_get_nr_components(fmt)) {
85            case 1: fmt = PIPE_FORMAT_R32_FLOAT; break;
86            case 2: fmt = PIPE_FORMAT_R32G32_FLOAT; break;
87            case 3: fmt = PIPE_FORMAT_R32G32B32_FLOAT; break;
88            case 4: fmt = PIPE_FORMAT_R32G32B32A32_FLOAT; break;
89            default:
90                assert(0);
91                FREE(so);
92                return NULL;
93            }
94            so->element[i].state = nv50_vertex_format[fmt].vtx;
95            so->need_conversion = true;
96            util_debug_message(&nouveau_context(pipe)->debug, FALLBACK,
97                               "Converting vertex element %d, no hw format %s",
98                               i, util_format_name(ve->src_format));
99        }
100        so->element[i].state |= i;
101
102        size = util_format_get_blocksize(fmt);
103        if (so->vb_access_size[vbi] < (ve->src_offset + size))
104           so->vb_access_size[vbi] = ve->src_offset + size;
105
106        if (1) {
107            unsigned j = transkey.nr_elements++;
108
109            transkey.element[j].type = TRANSLATE_ELEMENT_NORMAL;
110            transkey.element[j].input_format = ve->src_format;
111            transkey.element[j].input_buffer = vbi;
112            transkey.element[j].input_offset = ve->src_offset;
113            transkey.element[j].instance_divisor = ve->instance_divisor;
114
115            transkey.element[j].output_format = fmt;
116            transkey.element[j].output_offset = transkey.output_stride;
117            transkey.output_stride += (util_format_get_stride(fmt, 1) + 3) & ~3;
118
119            if (unlikely(ve->instance_divisor)) {
120               so->instance_elts |= 1 << i;
121               so->instance_bufs |= 1 << vbi;
122               if (ve->instance_divisor < so->min_instance_div[vbi])
123                  so->min_instance_div[vbi] = ve->instance_divisor;
124            }
125        }
126    }
127
128    so->translate = translate_create(&transkey);
129    so->vertex_size = transkey.output_stride / 4;
130    so->packet_vertex_limit = NV04_PFIFO_MAX_PACKET_LEN /
131       MAX2(so->vertex_size, 1);
132
133    return so;
134}
135
136#define NV50_3D_VERTEX_ATTRIB_INACTIVE              \
137   NV50_3D_VERTEX_ARRAY_ATTRIB_TYPE_FLOAT |         \
138   NV50_3D_VERTEX_ARRAY_ATTRIB_FORMAT_32_32_32_32 | \
139   NV50_3D_VERTEX_ARRAY_ATTRIB_CONST
140
141static void
142nv50_emit_vtxattr(struct nv50_context *nv50, struct pipe_vertex_buffer *vb,
143                  struct pipe_vertex_element *ve, unsigned attr)
144{
145   struct nouveau_pushbuf *push = nv50->base.pushbuf;
146   const void *data = (const uint8_t *)vb->buffer.user + ve->src_offset;
147   float v[4];
148   const unsigned nc = util_format_get_nr_components(ve->src_format);
149
150   assert(vb->is_user_buffer);
151
152   util_format_unpack_rgba(ve->src_format, v, data, 1);
153
154   switch (nc) {
155   case 4:
156      BEGIN_NV04(push, NV50_3D(VTX_ATTR_4F_X(attr)), 4);
157      PUSH_DATAf(push, v[0]);
158      PUSH_DATAf(push, v[1]);
159      PUSH_DATAf(push, v[2]);
160      PUSH_DATAf(push, v[3]);
161      break;
162   case 3:
163      BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(attr)), 3);
164      PUSH_DATAf(push, v[0]);
165      PUSH_DATAf(push, v[1]);
166      PUSH_DATAf(push, v[2]);
167      break;
168   case 2:
169      BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(attr)), 2);
170      PUSH_DATAf(push, v[0]);
171      PUSH_DATAf(push, v[1]);
172      break;
173   case 1:
174      if (attr == nv50->vertprog->vp.edgeflag) {
175         BEGIN_NV04(push, NV50_3D(EDGEFLAG), 1);
176         PUSH_DATA (push, v[0] ? 1 : 0);
177      }
178      BEGIN_NV04(push, NV50_3D(VTX_ATTR_1F(attr)), 1);
179      PUSH_DATAf(push, v[0]);
180      break;
181   default:
182      assert(0);
183      break;
184   }
185}
186
187static inline void
188nv50_user_vbuf_range(struct nv50_context *nv50, unsigned vbi,
189                     uint32_t *base, uint32_t *size)
190{
191   assert(vbi < PIPE_MAX_ATTRIBS);
192   if (unlikely(nv50->vertex->instance_bufs & (1 << vbi))) {
193      const uint32_t div = nv50->vertex->min_instance_div[vbi];
194      *base = nv50->instance_off * nv50->vtxbuf[vbi].stride;
195      *size = (nv50->instance_max / div) * nv50->vtxbuf[vbi].stride +
196         nv50->vertex->vb_access_size[vbi];
197   } else {
198      /* NOTE: if there are user buffers, we *must* have index bounds */
199      assert(nv50->vb_elt_limit != ~0);
200      *base = nv50->vb_elt_first * nv50->vtxbuf[vbi].stride;
201      *size = nv50->vb_elt_limit * nv50->vtxbuf[vbi].stride +
202         nv50->vertex->vb_access_size[vbi];
203   }
204}
205
206static void
207nv50_upload_user_buffers(struct nv50_context *nv50,
208                         uint64_t addrs[], uint32_t limits[])
209{
210   unsigned b;
211
212   assert(nv50->num_vtxbufs <= PIPE_MAX_ATTRIBS);
213   for (b = 0; b < nv50->num_vtxbufs; ++b) {
214      struct nouveau_bo *bo;
215      const struct pipe_vertex_buffer *vb = &nv50->vtxbuf[b];
216      uint32_t base, size;
217
218      if (!(nv50->vbo_user & (1 << b)) || !vb->stride)
219         continue;
220      nv50_user_vbuf_range(nv50, b, &base, &size);
221
222      limits[b] = base + size - 1;
223      addrs[b] = nouveau_scratch_data(&nv50->base, vb->buffer.user, base, size,
224                                      &bo);
225      if (addrs[b])
226         BCTX_REFN_bo(nv50->bufctx_3d, 3D_VERTEX_TMP, NOUVEAU_BO_GART |
227                      NOUVEAU_BO_RD, bo);
228   }
229   nv50->base.vbo_dirty = true;
230}
231
232static void
233nv50_update_user_vbufs(struct nv50_context *nv50)
234{
235   uint64_t address[PIPE_MAX_ATTRIBS];
236   struct nouveau_pushbuf *push = nv50->base.pushbuf;
237   unsigned i;
238   uint32_t written = 0;
239
240   for (i = 0; i < nv50->vertex->num_elements; ++i) {
241      struct pipe_vertex_element *ve = &nv50->vertex->element[i].pipe;
242      const unsigned b = ve->vertex_buffer_index;
243      struct pipe_vertex_buffer *vb;
244      uint32_t base, size;
245
246      assert(b < PIPE_MAX_ATTRIBS);
247      vb = &nv50->vtxbuf[b];
248
249      if (!(nv50->vbo_user & (1 << b)))
250         continue;
251
252      if (!vb->stride) {
253         nv50_emit_vtxattr(nv50, vb, ve, i);
254         continue;
255      }
256      nv50_user_vbuf_range(nv50, b, &base, &size);
257
258      if (!(written & (1 << b))) {
259         struct nouveau_bo *bo;
260         const uint32_t bo_flags = NOUVEAU_BO_GART | NOUVEAU_BO_RD;
261         written |= 1 << b;
262         address[b] = nouveau_scratch_data(&nv50->base, vb->buffer.user,
263                                           base, size, &bo);
264         if (address[b])
265            BCTX_REFN_bo(nv50->bufctx_3d, 3D_VERTEX_TMP, bo_flags, bo);
266      }
267
268      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
269      PUSH_DATAh(push, address[b] + base + size - 1);
270      PUSH_DATA (push, address[b] + base + size - 1);
271      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_START_HIGH(i)), 2);
272      PUSH_DATAh(push, address[b] + ve->src_offset);
273      PUSH_DATA (push, address[b] + ve->src_offset);
274   }
275   nv50->base.vbo_dirty = true;
276}
277
278static inline void
279nv50_release_user_vbufs(struct nv50_context *nv50)
280{
281   if (nv50->vbo_user) {
282      nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_3D_VERTEX_TMP);
283      nouveau_scratch_done(&nv50->base);
284   }
285}
286
287void
288nv50_vertex_arrays_validate(struct nv50_context *nv50)
289{
290   uint64_t addrs[PIPE_MAX_ATTRIBS];
291   uint32_t limits[PIPE_MAX_ATTRIBS];
292   struct nouveau_pushbuf *push = nv50->base.pushbuf;
293   struct nv50_vertex_stateobj *vertex = nv50->vertex;
294   struct pipe_vertex_buffer *vb;
295   struct nv50_vertex_element *ve;
296   uint32_t mask;
297   uint32_t refd = 0;
298   unsigned i;
299   const unsigned n = MAX2(vertex->num_elements, nv50->state.num_vtxelts);
300
301   if (unlikely(vertex->need_conversion))
302      nv50->vbo_fifo = ~0;
303   else
304   if (nv50->vbo_user & ~nv50->vbo_constant)
305      nv50->vbo_fifo = nv50->vbo_push_hint ? ~0 : 0;
306   else
307      nv50->vbo_fifo = 0;
308
309   if (!nv50->vbo_fifo) {
310      /* if vertex buffer was written by GPU - flush VBO cache */
311      assert(nv50->num_vtxbufs <= PIPE_MAX_ATTRIBS);
312      for (i = 0; i < nv50->num_vtxbufs; ++i) {
313         struct nv04_resource *buf = nv04_resource(nv50->vtxbuf[i].buffer.resource);
314         if (!nv50->vtxbuf[i].is_user_buffer &&
315             buf && buf->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
316            buf->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
317            nv50->base.vbo_dirty = true;
318         }
319      }
320   }
321
322   /* update vertex format state */
323   BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_ATTRIB(0)), n);
324   if (nv50->vbo_fifo) {
325      nv50->state.num_vtxelts = vertex->num_elements;
326      for (i = 0; i < vertex->num_elements; ++i)
327         PUSH_DATA (push, vertex->element[i].state);
328      for (; i < n; ++i)
329         PUSH_DATA (push, NV50_3D_VERTEX_ATTRIB_INACTIVE);
330      for (i = 0; i < n; ++i) {
331         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
332         PUSH_DATA (push, 0);
333      }
334      return;
335   }
336   for (i = 0; i < vertex->num_elements; ++i) {
337      const unsigned b = vertex->element[i].pipe.vertex_buffer_index;
338
339      assert(b < PIPE_MAX_ATTRIBS);
340      ve = &vertex->element[i];
341      vb = &nv50->vtxbuf[b];
342
343      if (likely(vb->stride) || !(nv50->vbo_user & (1 << b)))
344         PUSH_DATA(push, ve->state);
345      else
346         PUSH_DATA(push, ve->state | NV50_3D_VERTEX_ARRAY_ATTRIB_CONST);
347   }
348   for (; i < n; ++i)
349      PUSH_DATA(push, NV50_3D_VERTEX_ATTRIB_INACTIVE);
350
351   /* update per-instance enables */
352   mask = vertex->instance_elts ^ nv50->state.instance_elts;
353   while (mask) {
354      const int i = ffs(mask) - 1;
355      mask &= ~(1 << i);
356      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_PER_INSTANCE(i)), 1);
357      PUSH_DATA (push, (vertex->instance_elts >> i) & 1);
358   }
359   nv50->state.instance_elts = vertex->instance_elts;
360
361   if (nv50->vbo_user & ~nv50->vbo_constant)
362      nv50_upload_user_buffers(nv50, addrs, limits);
363
364   /* update buffers and set constant attributes */
365   for (i = 0; i < vertex->num_elements; ++i) {
366      uint64_t address, limit;
367      const unsigned b = vertex->element[i].pipe.vertex_buffer_index;
368
369      assert(b < PIPE_MAX_ATTRIBS);
370      ve = &vertex->element[i];
371      vb = &nv50->vtxbuf[b];
372
373      if (unlikely(nv50->vbo_constant & (1 << b))) {
374         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
375         PUSH_DATA (push, 0);
376         nv50_emit_vtxattr(nv50, vb, &ve->pipe, i);
377         continue;
378      } else
379      if (nv50->vbo_user & (1 << b)) {
380         address = addrs[b] + ve->pipe.src_offset;
381         limit = addrs[b] + limits[b];
382      } else
383      if (!vb->buffer.resource) {
384         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
385         PUSH_DATA (push, 0);
386         continue;
387      } else {
388         struct nv04_resource *buf = nv04_resource(vb->buffer.resource);
389         if (!(refd & (1 << b))) {
390            refd |= 1 << b;
391            BCTX_REFN(nv50->bufctx_3d, 3D_VERTEX, buf, RD);
392         }
393         address = buf->address + vb->buffer_offset + ve->pipe.src_offset;
394         limit = buf->address + buf->base.width0 - 1;
395      }
396
397      if (unlikely(ve->pipe.instance_divisor)) {
398         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 4);
399         PUSH_DATA (push, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
400         PUSH_DATAh(push, address);
401         PUSH_DATA (push, address);
402         PUSH_DATA (push, ve->pipe.instance_divisor);
403      } else {
404         BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 3);
405         PUSH_DATA (push, NV50_3D_VERTEX_ARRAY_FETCH_ENABLE | vb->stride);
406         PUSH_DATAh(push, address);
407         PUSH_DATA (push, address);
408      }
409      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_LIMIT_HIGH(i)), 2);
410      PUSH_DATAh(push, limit);
411      PUSH_DATA (push, limit);
412   }
413   for (; i < nv50->state.num_vtxelts; ++i) {
414      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FETCH(i)), 1);
415      PUSH_DATA (push, 0);
416   }
417   nv50->state.num_vtxelts = vertex->num_elements;
418}
419
420#define NV50_PRIM_GL_CASE(n) \
421   case PIPE_PRIM_##n: return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_##n
422
423static inline unsigned
424nv50_prim_gl(unsigned prim)
425{
426   switch (prim) {
427   NV50_PRIM_GL_CASE(POINTS);
428   NV50_PRIM_GL_CASE(LINES);
429   NV50_PRIM_GL_CASE(LINE_LOOP);
430   NV50_PRIM_GL_CASE(LINE_STRIP);
431   NV50_PRIM_GL_CASE(TRIANGLES);
432   NV50_PRIM_GL_CASE(TRIANGLE_STRIP);
433   NV50_PRIM_GL_CASE(TRIANGLE_FAN);
434   NV50_PRIM_GL_CASE(QUADS);
435   NV50_PRIM_GL_CASE(QUAD_STRIP);
436   NV50_PRIM_GL_CASE(POLYGON);
437   NV50_PRIM_GL_CASE(LINES_ADJACENCY);
438   NV50_PRIM_GL_CASE(LINE_STRIP_ADJACENCY);
439   NV50_PRIM_GL_CASE(TRIANGLES_ADJACENCY);
440   NV50_PRIM_GL_CASE(TRIANGLE_STRIP_ADJACENCY);
441   default:
442      return NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_POINTS;
443      break;
444   }
445}
446
447/* For pre-nva0 transform feedback. */
448static const uint8_t nv50_pipe_prim_to_prim_size[PIPE_PRIM_MAX + 1] =
449{
450   [PIPE_PRIM_POINTS] = 1,
451   [PIPE_PRIM_LINES] = 2,
452   [PIPE_PRIM_LINE_LOOP] = 2,
453   [PIPE_PRIM_LINE_STRIP] = 2,
454   [PIPE_PRIM_TRIANGLES] = 3,
455   [PIPE_PRIM_TRIANGLE_STRIP] = 3,
456   [PIPE_PRIM_TRIANGLE_FAN] = 3,
457   [PIPE_PRIM_QUADS] = 3,
458   [PIPE_PRIM_QUAD_STRIP] = 3,
459   [PIPE_PRIM_POLYGON] = 3,
460   [PIPE_PRIM_LINES_ADJACENCY] = 2,
461   [PIPE_PRIM_LINE_STRIP_ADJACENCY] = 2,
462   [PIPE_PRIM_TRIANGLES_ADJACENCY] = 3,
463   [PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY] = 3
464};
465
466static void
467nv50_draw_arrays(struct nv50_context *nv50,
468                 unsigned mode, unsigned start, unsigned count,
469                 unsigned instance_count)
470{
471   struct nouveau_pushbuf *push = nv50->base.pushbuf;
472   unsigned prim;
473
474   if (nv50->state.index_bias) {
475      BEGIN_NV04(push, NV50_3D(VB_ELEMENT_BASE), 1);
476      PUSH_DATA (push, 0);
477      if (nv50->screen->base.class_3d >= NV84_3D_CLASS) {
478         BEGIN_NV04(push, NV84_3D(VERTEX_ID_BASE), 1);
479         PUSH_DATA (push, 0);
480      }
481      nv50->state.index_bias = 0;
482   }
483
484   prim = nv50_prim_gl(mode);
485
486   while (instance_count--) {
487      BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
488      PUSH_DATA (push, prim);
489      BEGIN_NV04(push, NV50_3D(VERTEX_BUFFER_FIRST), 2);
490      PUSH_DATA (push, start);
491      PUSH_DATA (push, count);
492      BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
493      PUSH_DATA (push, 0);
494
495      prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
496   }
497}
498
499static void
500nv50_draw_elements_inline_u08(struct nouveau_pushbuf *push, const uint8_t *map,
501                              unsigned start, unsigned count)
502{
503   map += start;
504
505   if (count & 3) {
506      unsigned i;
507      BEGIN_NI04(push, NV50_3D(VB_ELEMENT_U32), count & 3);
508      for (i = 0; i < (count & 3); ++i)
509         PUSH_DATA(push, *map++);
510      count &= ~3;
511   }
512   while (count) {
513      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 4) / 4;
514
515      BEGIN_NI04(push, NV50_3D(VB_ELEMENT_U8), nr);
516      for (i = 0; i < nr; ++i) {
517         PUSH_DATA(push,
518                   (map[3] << 24) | (map[2] << 16) | (map[1] << 8) | map[0]);
519         map += 4;
520      }
521      count -= nr * 4;
522   }
523}
524
525static void
526nv50_draw_elements_inline_u16(struct nouveau_pushbuf *push, const uint16_t *map,
527                              unsigned start, unsigned count)
528{
529   map += start;
530
531   if (count & 1) {
532      count &= ~1;
533      BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U32), 1);
534      PUSH_DATA (push, *map++);
535   }
536   while (count) {
537      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
538
539      BEGIN_NI04(push, NV50_3D(VB_ELEMENT_U16), nr);
540      for (i = 0; i < nr; ++i) {
541         PUSH_DATA(push, (map[1] << 16) | map[0]);
542         map += 2;
543      }
544      count -= nr * 2;
545   }
546}
547
548static void
549nv50_draw_elements_inline_u32(struct nouveau_pushbuf *push, const uint32_t *map,
550                              unsigned start, unsigned count)
551{
552   map += start;
553
554   while (count) {
555      const unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
556
557      BEGIN_NI04(push, NV50_3D(VB_ELEMENT_U32), nr);
558      PUSH_DATAp(push, map, nr);
559
560      map += nr;
561      count -= nr;
562   }
563}
564
565static void
566nv50_draw_elements_inline_u32_short(struct nouveau_pushbuf *push,
567                                    const uint32_t *map,
568                                    unsigned start, unsigned count)
569{
570   map += start;
571
572   if (count & 1) {
573      count--;
574      BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U32), 1);
575      PUSH_DATA (push, *map++);
576   }
577   while (count) {
578      unsigned i, nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN * 2) / 2;
579
580      BEGIN_NI04(push, NV50_3D(VB_ELEMENT_U16), nr);
581      for (i = 0; i < nr; ++i) {
582         PUSH_DATA(push, (map[1] << 16) | map[0]);
583         map += 2;
584      }
585      count -= nr * 2;
586   }
587}
588
589static void
590nv50_draw_elements(struct nv50_context *nv50, bool shorten,
591                   const struct pipe_draw_info *info,
592                   unsigned mode, unsigned start, unsigned count,
593                   unsigned instance_count, int32_t index_bias,
594		   unsigned index_size)
595{
596   struct nouveau_pushbuf *push = nv50->base.pushbuf;
597   unsigned prim;
598
599   prim = nv50_prim_gl(mode);
600
601   if (index_bias != nv50->state.index_bias) {
602      BEGIN_NV04(push, NV50_3D(VB_ELEMENT_BASE), 1);
603      PUSH_DATA (push, index_bias);
604      if (nv50->screen->base.class_3d >= NV84_3D_CLASS) {
605         BEGIN_NV04(push, NV84_3D(VERTEX_ID_BASE), 1);
606         PUSH_DATA (push, index_bias);
607      }
608      nv50->state.index_bias = index_bias;
609   }
610
611   if (!info->has_user_indices) {
612      struct nv04_resource *buf = nv04_resource(info->index.resource);
613      unsigned pb_start;
614      unsigned pb_bytes;
615      const unsigned base = buf->offset & ~3;
616
617      start += (buf->offset & 3) >> (index_size >> 1);
618
619      assert(nouveau_resource_mapped_by_gpu(info->index.resource));
620
621      /* This shouldn't have to be here. The going theory is that the buffer
622       * is being filled in by PGRAPH, and it's not done yet by the time it
623       * gets submitted to PFIFO, which in turn starts immediately prefetching
624       * the not-yet-written data. Ideally this wait would only happen on
625       * pushbuf submit, but it's probably not a big performance difference.
626       */
627      if (buf->fence_wr && !nouveau_fence_signalled(buf->fence_wr))
628         nouveau_fence_wait(buf->fence_wr, &nv50->base.debug);
629
630      while (instance_count--) {
631         BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
632         PUSH_DATA (push, prim);
633
634         nouveau_pushbuf_space(push, 16, 0, 1);
635         PUSH_REFN(push, buf->bo, NOUVEAU_BO_RD | buf->domain);
636
637         switch (index_size) {
638         case 4:
639            BEGIN_NL50(push, NV50_3D(VB_ELEMENT_U32), count);
640            nouveau_pushbuf_data(push, buf->bo, base + start * 4, count * 4);
641            break;
642         case 2:
643            pb_start = (start & ~1) * 2;
644            pb_bytes = ((start + count + 1) & ~1) * 2 - pb_start;
645
646            BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U16_SETUP), 1);
647            PUSH_DATA (push, (start << 31) | count);
648            BEGIN_NL50(push, NV50_3D(VB_ELEMENT_U16), pb_bytes / 4);
649            nouveau_pushbuf_data(push, buf->bo, base + pb_start, pb_bytes);
650            BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U16_SETUP), 1);
651            PUSH_DATA (push, 0);
652            break;
653         default:
654            assert(index_size == 1);
655            pb_start = start & ~3;
656            pb_bytes = ((start + count + 3) & ~3) - pb_start;
657
658            BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U8_SETUP), 1);
659            PUSH_DATA (push, (start << 30) | count);
660            BEGIN_NL50(push, NV50_3D(VB_ELEMENT_U8), pb_bytes / 4);
661            nouveau_pushbuf_data(push, buf->bo, base + pb_start, pb_bytes);
662            BEGIN_NV04(push, NV50_3D(VB_ELEMENT_U8_SETUP), 1);
663            PUSH_DATA (push, 0);
664            break;
665         }
666         BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
667         PUSH_DATA (push, 0);
668
669         prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
670      }
671   } else {
672      const void *data = info->index.user;
673
674      while (instance_count--) {
675         BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
676         PUSH_DATA (push, prim);
677         switch (index_size) {
678         case 1:
679            nv50_draw_elements_inline_u08(push, data, start, count);
680            break;
681         case 2:
682            nv50_draw_elements_inline_u16(push, data, start, count);
683            break;
684         case 4:
685            if (shorten)
686               nv50_draw_elements_inline_u32_short(push, data, start, count);
687            else
688               nv50_draw_elements_inline_u32(push, data, start, count);
689            break;
690         default:
691            assert(0);
692            return;
693         }
694         BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
695         PUSH_DATA (push, 0);
696
697         prim |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
698      }
699   }
700   NOUVEAU_DRV_STAT(&nv50->screen->base, draw_calls_indexed, 1);
701}
702
703static void
704nva0_draw_stream_output(struct nv50_context *nv50,
705                        const struct pipe_draw_info *info,
706                        const struct pipe_draw_indirect_info *indirect)
707{
708   struct nouveau_pushbuf *push = nv50->base.pushbuf;
709   struct nv50_so_target *so = nv50_so_target(indirect->count_from_stream_output);
710   struct nv04_resource *res = nv04_resource(so->pipe.buffer);
711   unsigned num_instances = info->instance_count;
712   unsigned mode = nv50_prim_gl(info->mode);
713
714   if (unlikely(nv50->screen->base.class_3d < NVA0_3D_CLASS)) {
715      /* A proper implementation without waiting doesn't seem possible,
716       * so don't bother.
717       */
718      NOUVEAU_ERR("draw_stream_output not supported on pre-NVA0 cards\n");
719      return;
720   }
721
722   if (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
723      res->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
724      PUSH_SPACE(push, 4);
725      BEGIN_NV04(push, SUBC_3D(NV50_GRAPH_SERIALIZE), 1);
726      PUSH_DATA (push, 0);
727      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FLUSH), 1);
728      PUSH_DATA (push, 0);
729   }
730
731   assert(num_instances);
732   do {
733      PUSH_SPACE(push, 8);
734      BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
735      PUSH_DATA (push, mode);
736      BEGIN_NV04(push, NVA0_3D(DRAW_TFB_BASE), 1);
737      PUSH_DATA (push, 0);
738      BEGIN_NV04(push, NVA0_3D(DRAW_TFB_STRIDE), 1);
739      PUSH_DATA (push, so->stride);
740      nv50_hw_query_pushbuf_submit(push, NVA0_3D_DRAW_TFB_BYTES,
741                                   nv50_query(so->pq), 0x4);
742      BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
743      PUSH_DATA (push, 0);
744
745      mode |= NV50_3D_VERTEX_BEGIN_GL_INSTANCE_NEXT;
746   } while (--num_instances);
747}
748
749static void
750nv50_draw_vbo_kick_notify(struct nouveau_pushbuf *chan)
751{
752   struct nv50_screen *screen = chan->user_priv;
753
754   nouveau_fence_update(&screen->base, true);
755
756   nv50_bufctx_fence(screen->cur_ctx->bufctx_3d, true);
757}
758
759void
760nv50_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info,
761              unsigned drawid_offset,
762              const struct pipe_draw_indirect_info *indirect,
763              const struct pipe_draw_start_count_bias *draws,
764              unsigned num_draws)
765{
766   if (num_draws > 1) {
767      util_draw_multi(pipe, info, drawid_offset, indirect, draws, num_draws);
768      return;
769   }
770
771   if (!indirect && (!draws[0].count || !info->instance_count))
772      return;
773
774   /* We don't actually support indirect draws, so add a fallback for ES 3.1's
775    * benefit.
776    */
777   if (indirect && indirect->buffer) {
778      util_draw_indirect(pipe, info, indirect);
779      return;
780   }
781
782   struct nv50_context *nv50 = nv50_context(pipe);
783   struct nouveau_pushbuf *push = nv50->base.pushbuf;
784   bool tex_dirty = false;
785   int s;
786
787   if (info->index_size && !info->has_user_indices)
788      BCTX_REFN(nv50->bufctx_3d, 3D_INDEX, nv04_resource(info->index.resource), RD);
789
790   /* NOTE: caller must ensure that (min_index + index_bias) is >= 0 */
791   if (info->index_bounds_valid) {
792      nv50->vb_elt_first = info->min_index + (info->index_size ? draws->index_bias : 0);
793      nv50->vb_elt_limit = info->max_index - info->min_index;
794   } else {
795      nv50->vb_elt_first = 0;
796      nv50->vb_elt_limit = ~0;
797   }
798   nv50->instance_off = info->start_instance;
799   nv50->instance_max = info->instance_count - 1;
800
801   /* For picking only a few vertices from a large user buffer, push is better,
802    * if index count is larger and we expect repeated vertices, suggest upload.
803    */
804   nv50->vbo_push_hint = /* the 64 is heuristic */
805      !(info->index_size && ((nv50->vb_elt_limit + 64) < draws[0].count));
806
807   if (nv50->vbo_user && !(nv50->dirty_3d & (NV50_NEW_3D_ARRAYS | NV50_NEW_3D_VERTEX))) {
808      if (!!nv50->vbo_fifo != nv50->vbo_push_hint)
809         nv50->dirty_3d |= NV50_NEW_3D_ARRAYS;
810      else
811      if (!nv50->vbo_fifo)
812         nv50_update_user_vbufs(nv50);
813   }
814
815   if (unlikely(nv50->num_so_targets && !nv50->gmtyprog))
816      nv50->state.prim_size = nv50_pipe_prim_to_prim_size[info->mode];
817
818   nv50_state_validate_3d(nv50, ~0);
819
820   push->kick_notify = nv50_draw_vbo_kick_notify;
821
822   for (s = 0; s < NV50_MAX_3D_SHADER_STAGES && !nv50->cb_dirty; ++s) {
823      if (nv50->constbuf_coherent[s])
824         nv50->cb_dirty = true;
825   }
826
827   /* If there are any coherent constbufs, flush the cache */
828   if (nv50->cb_dirty) {
829      BEGIN_NV04(push, NV50_3D(CODE_CB_FLUSH), 1);
830      PUSH_DATA (push, 0);
831      nv50->cb_dirty = false;
832   }
833
834   for (s = 0; s < NV50_MAX_3D_SHADER_STAGES && !tex_dirty; ++s) {
835      if (nv50->textures_coherent[s])
836         tex_dirty = true;
837   }
838
839   if (tex_dirty) {
840      BEGIN_NV04(push, NV50_3D(TEX_CACHE_CTL), 1);
841      PUSH_DATA (push, 0x20);
842   }
843
844   if (nv50->screen->base.class_3d >= NVA0_3D_CLASS &&
845       nv50->seamless_cube_map != nv50->state.seamless_cube_map) {
846      nv50->state.seamless_cube_map = nv50->seamless_cube_map;
847      BEGIN_NV04(push, SUBC_3D(NVA0_3D_TEX_MISC), 1);
848      PUSH_DATA (push, nv50->seamless_cube_map ? NVA0_3D_TEX_MISC_SEAMLESS_CUBE_MAP : 0);
849   }
850
851   if (nv50->vertprog->mul_zero_wins != nv50->state.mul_zero_wins) {
852      nv50->state.mul_zero_wins = nv50->vertprog->mul_zero_wins;
853      BEGIN_NV04(push, NV50_3D(UNK1690), 1);
854      PUSH_DATA (push, 0x00010000 * !!nv50->state.mul_zero_wins);
855   }
856
857   /* Make starting/pausing streamout work pre-NVA0 enough for ES3.0. This
858    * means counting vertices in a vertex shader when it has so outputs.
859    */
860   if (nv50->screen->base.class_3d < NVA0_3D_CLASS &&
861       nv50->vertprog->pipe.stream_output.num_outputs) {
862      for (int i = 0; i < nv50->num_so_targets; i++) {
863         nv50->so_used[i] += info->instance_count *
864            u_stream_outputs_for_vertices(info->mode, draws[0].count) *
865            nv50->vertprog->pipe.stream_output.stride[i] * 4;
866      }
867   }
868
869   if (nv50->vbo_fifo) {
870      nv50_push_vbo(nv50, info, indirect, &draws[0]);
871      goto cleanup;
872   }
873
874   if (nv50->state.instance_base != info->start_instance) {
875      nv50->state.instance_base = info->start_instance;
876      /* NOTE: this does not affect the shader input, should it ? */
877      BEGIN_NV04(push, NV50_3D(VB_INSTANCE_BASE), 1);
878      PUSH_DATA (push, info->start_instance);
879   }
880
881   nv50->base.vbo_dirty |= !!nv50->vtxbufs_coherent;
882
883   if (nv50->base.vbo_dirty) {
884      BEGIN_NV04(push, NV50_3D(VERTEX_ARRAY_FLUSH), 1);
885      PUSH_DATA (push, 0);
886      nv50->base.vbo_dirty = false;
887   }
888
889   if (info->index_size) {
890      bool shorten = info->index_bounds_valid && info->max_index <= 65535;
891
892      if (info->primitive_restart != nv50->state.prim_restart) {
893         if (info->primitive_restart) {
894            BEGIN_NV04(push, NV50_3D(PRIM_RESTART_ENABLE), 2);
895            PUSH_DATA (push, 1);
896            PUSH_DATA (push, info->restart_index);
897
898            if (info->restart_index > 65535)
899               shorten = false;
900         } else {
901            BEGIN_NV04(push, NV50_3D(PRIM_RESTART_ENABLE), 1);
902            PUSH_DATA (push, 0);
903         }
904         nv50->state.prim_restart = info->primitive_restart;
905      } else
906      if (info->primitive_restart) {
907         BEGIN_NV04(push, NV50_3D(PRIM_RESTART_INDEX), 1);
908         PUSH_DATA (push, info->restart_index);
909
910         if (info->restart_index > 65535)
911            shorten = false;
912      }
913
914      nv50_draw_elements(nv50, shorten, info,
915                         info->mode, draws[0].start, draws[0].count,
916                         info->instance_count, draws->index_bias, info->index_size);
917   } else
918   if (unlikely(indirect && indirect->count_from_stream_output)) {
919      nva0_draw_stream_output(nv50, info, indirect);
920   } else {
921      nv50_draw_arrays(nv50,
922                       info->mode, draws[0].start, draws[0].count,
923                       info->instance_count);
924   }
925
926cleanup:
927   push->kick_notify = nv50_default_kick_notify;
928
929   nv50_release_user_vbufs(nv50);
930
931   nouveau_pushbuf_bufctx(push, NULL);
932
933   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_3D_INDEX);
934}
935