1/*
2 * Copyright 2012 Nouveau Project
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 * Authors: Christoph Bumiller
23 */
24
25#include "nvc0/nvc0_context.h"
26#include "nvc0/nve4_compute.h"
27
28#include "codegen/nv50_ir_driver.h"
29
30#ifdef DEBUG
31static void nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc *);
32static void gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc *);
33#endif
34
35
36int
37nve4_screen_compute_setup(struct nvc0_screen *screen,
38                          struct nouveau_pushbuf *push)
39{
40   struct nouveau_device *dev = screen->base.device;
41   struct nouveau_object *chan = screen->base.channel;
42   int i;
43   int ret;
44   uint32_t obj_class;
45   uint64_t address;
46
47   switch (dev->chipset & ~0xf) {
48   case 0x100:
49   case 0xf0:
50      obj_class = NVF0_COMPUTE_CLASS; /* GK110 */
51      break;
52   case 0xe0:
53      obj_class = NVE4_COMPUTE_CLASS; /* GK104 */
54      break;
55   case 0x110:
56      obj_class = GM107_COMPUTE_CLASS;
57      break;
58   case 0x120:
59      obj_class = GM200_COMPUTE_CLASS;
60      break;
61   case 0x130:
62      obj_class = (dev->chipset == 0x130 || dev->chipset == 0x13b) ?
63                      GP100_COMPUTE_CLASS : GP104_COMPUTE_CLASS;
64      break;
65   default:
66      NOUVEAU_ERR("unsupported chipset: NV%02x\n", dev->chipset);
67      return -1;
68   }
69
70   ret = nouveau_object_new(chan, 0xbeef00c0, obj_class, NULL, 0,
71                            &screen->compute);
72   if (ret) {
73      NOUVEAU_ERR("Failed to allocate compute object: %d\n", ret);
74      return ret;
75   }
76
77   BEGIN_NVC0(push, SUBC_CP(NV01_SUBCHAN_OBJECT), 1);
78   PUSH_DATA (push, screen->compute->oclass);
79
80   BEGIN_NVC0(push, NVE4_CP(TEMP_ADDRESS_HIGH), 2);
81   PUSH_DATAh(push, screen->tls->offset);
82   PUSH_DATA (push, screen->tls->offset);
83   /* No idea why there are 2. Divide size by 2 to be safe.
84    * Actually this might be per-MP TEMP size and looks like I'm only using
85    * 2 MPs instead of all 8.
86    */
87   BEGIN_NVC0(push, NVE4_CP(MP_TEMP_SIZE_HIGH(0)), 3);
88   PUSH_DATAh(push, screen->tls->size / screen->mp_count);
89   PUSH_DATA (push, (screen->tls->size / screen->mp_count) & ~0x7fff);
90   PUSH_DATA (push, 0xff);
91   BEGIN_NVC0(push, NVE4_CP(MP_TEMP_SIZE_HIGH(1)), 3);
92   PUSH_DATAh(push, screen->tls->size / screen->mp_count);
93   PUSH_DATA (push, (screen->tls->size / screen->mp_count) & ~0x7fff);
94   PUSH_DATA (push, 0xff);
95
96   /* Unified address space ? Who needs that ? Certainly not OpenCL.
97    *
98    * FATAL: Buffers with addresses inside [0x1000000, 0x3000000] will NOT be
99    *  accessible. We cannot prevent that at the moment, so expect failure.
100    */
101   BEGIN_NVC0(push, NVE4_CP(LOCAL_BASE), 1);
102   PUSH_DATA (push, 0xff << 24);
103   BEGIN_NVC0(push, NVE4_CP(SHARED_BASE), 1);
104   PUSH_DATA (push, 0xfe << 24);
105
106   BEGIN_NVC0(push, NVE4_CP(CODE_ADDRESS_HIGH), 2);
107   PUSH_DATAh(push, screen->text->offset);
108   PUSH_DATA (push, screen->text->offset);
109
110   BEGIN_NVC0(push, SUBC_CP(0x0310), 1);
111   PUSH_DATA (push, (obj_class >= NVF0_COMPUTE_CLASS) ? 0x400 : 0x300);
112
113   /* NOTE: these do not affect the state used by the 3D object */
114   BEGIN_NVC0(push, NVE4_CP(TIC_ADDRESS_HIGH), 3);
115   PUSH_DATAh(push, screen->txc->offset);
116   PUSH_DATA (push, screen->txc->offset);
117   PUSH_DATA (push, NVC0_TIC_MAX_ENTRIES - 1);
118   BEGIN_NVC0(push, NVE4_CP(TSC_ADDRESS_HIGH), 3);
119   PUSH_DATAh(push, screen->txc->offset + 65536);
120   PUSH_DATA (push, screen->txc->offset + 65536);
121   PUSH_DATA (push, NVC0_TSC_MAX_ENTRIES - 1);
122
123   if (obj_class >= NVF0_COMPUTE_CLASS) {
124      /* The blob calls GK110_COMPUTE.FIRMWARE[0x6], along with the args (0x1)
125       * passed with GK110_COMPUTE.GRAPH.SCRATCH[0x2]. This is currently
126       * disabled because our firmware doesn't support these commands and the
127       * GPU hangs if they are used. */
128      BEGIN_NIC0(push, SUBC_CP(0x0248), 64);
129      for (i = 63; i >= 0; i--)
130         PUSH_DATA(push, 0x38000 | i);
131      IMMED_NVC0(push, SUBC_CP(NV50_GRAPH_SERIALIZE), 0);
132   }
133
134   BEGIN_NVC0(push, NVE4_CP(TEX_CB_INDEX), 1);
135   PUSH_DATA (push, 7); /* does not interfere with 3D */
136
137   /* Disabling this UNK command avoid a read fault when using texelFetch()
138    * from a compute shader for weird reasons.
139   if (obj_class == NVF0_COMPUTE_CLASS)
140      IMMED_NVC0(push, SUBC_CP(0x02c4), 1);
141   */
142
143   address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5);
144
145   /* MS sample coordinate offsets: these do not work with _ALT modes ! */
146   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
147   PUSH_DATAh(push, address + NVC0_CB_AUX_MS_INFO);
148   PUSH_DATA (push, address + NVC0_CB_AUX_MS_INFO);
149   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
150   PUSH_DATA (push, 64);
151   PUSH_DATA (push, 1);
152   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 17);
153   PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
154   PUSH_DATA (push, 0); /* 0 */
155   PUSH_DATA (push, 0);
156   PUSH_DATA (push, 1); /* 1 */
157   PUSH_DATA (push, 0);
158   PUSH_DATA (push, 0); /* 2 */
159   PUSH_DATA (push, 1);
160   PUSH_DATA (push, 1); /* 3 */
161   PUSH_DATA (push, 1);
162   PUSH_DATA (push, 2); /* 4 */
163   PUSH_DATA (push, 0);
164   PUSH_DATA (push, 3); /* 5 */
165   PUSH_DATA (push, 0);
166   PUSH_DATA (push, 2); /* 6 */
167   PUSH_DATA (push, 1);
168   PUSH_DATA (push, 3); /* 7 */
169   PUSH_DATA (push, 1);
170
171#ifdef NOUVEAU_NVE4_MP_TRAP_HANDLER
172   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
173   PUSH_DATAh(push, screen->parm->offset + NVE4_CP_INPUT_TRAP_INFO_PTR);
174   PUSH_DATA (push, screen->parm->offset + NVE4_CP_INPUT_TRAP_INFO_PTR);
175   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
176   PUSH_DATA (push, 28);
177   PUSH_DATA (push, 1);
178   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 8);
179   PUSH_DATA (push, 1);
180   PUSH_DATA (push, screen->parm->offset + NVE4_CP_PARAM_TRAP_INFO);
181   PUSH_DATAh(push, screen->parm->offset + NVE4_CP_PARAM_TRAP_INFO);
182   PUSH_DATA (push, screen->tls->offset);
183   PUSH_DATAh(push, screen->tls->offset);
184   PUSH_DATA (push, screen->tls->size / 2); /* MP TEMP block size */
185   PUSH_DATA (push, screen->tls->size / 2 / 64); /* warp TEMP block size */
186   PUSH_DATA (push, 0); /* warp cfstack size */
187#endif
188
189   BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
190   PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
191
192   return 0;
193}
194
195static void
196gm107_compute_validate_surfaces(struct nvc0_context *nvc0,
197                                struct pipe_image_view *view, int slot)
198{
199   struct nv04_resource *res = nv04_resource(view->resource);
200   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
201   struct nvc0_screen *screen = nvc0->screen;
202   struct nouveau_bo *txc = nvc0->screen->txc;
203   struct nv50_tic_entry *tic;
204   uint64_t address;
205   const int s = 5;
206
207   tic = nv50_tic_entry(nvc0->images_tic[s][slot]);
208
209   res = nv04_resource(tic->pipe.texture);
210   nvc0_update_tic(nvc0, tic, res);
211
212   if (tic->id < 0) {
213      tic->id = nvc0_screen_tic_alloc(nvc0->screen, tic);
214
215      /* upload the texture view */
216      PUSH_SPACE(push, 16);
217      BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
218      PUSH_DATAh(push, txc->offset + (tic->id * 32));
219      PUSH_DATA (push, txc->offset + (tic->id * 32));
220      BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
221      PUSH_DATA (push, 32);
222      PUSH_DATA (push, 1);
223      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 9);
224      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
225      PUSH_DATAp(push, &tic->tic[0], 8);
226
227      BEGIN_NIC0(push, NVE4_CP(TIC_FLUSH), 1);
228      PUSH_DATA (push, (tic->id << 4) | 1);
229   } else
230   if (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
231      BEGIN_NIC0(push, NVE4_CP(TEX_CACHE_CTL), 1);
232      PUSH_DATA (push, (tic->id << 4) | 1);
233   }
234   nvc0->screen->tic.lock[tic->id / 32] |= 1 << (tic->id % 32);
235
236   res->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
237   res->status |=  NOUVEAU_BUFFER_STATUS_GPU_READING;
238
239   BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RD);
240
241   address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
242
243   /* upload the texture handle */
244   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
245   PUSH_DATAh(push, address + NVC0_CB_AUX_TEX_INFO(slot + 32));
246   PUSH_DATA (push, address + NVC0_CB_AUX_TEX_INFO(slot + 32));
247   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
248   PUSH_DATA (push, 4);
249   PUSH_DATA (push, 0x1);
250   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 2);
251   PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
252   PUSH_DATA (push, tic->id);
253
254   BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
255   PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
256}
257
258static void
259nve4_compute_validate_surfaces(struct nvc0_context *nvc0)
260{
261   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
262   uint64_t address;
263   const int s = 5;
264   int i, j;
265
266   if (!nvc0->images_dirty[s])
267      return;
268
269   address = nvc0->screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
270
271   for (i = 0; i < NVC0_MAX_IMAGES; ++i) {
272      struct pipe_image_view *view = &nvc0->images[s][i];
273
274      BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
275      PUSH_DATAh(push, address + NVC0_CB_AUX_SU_INFO(i));
276      PUSH_DATA (push, address + NVC0_CB_AUX_SU_INFO(i));
277      BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
278      PUSH_DATA (push, 16 * 4);
279      PUSH_DATA (push, 0x1);
280      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 16);
281      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
282
283      if (view->resource) {
284         struct nv04_resource *res = nv04_resource(view->resource);
285
286         if (res->base.target == PIPE_BUFFER) {
287            if (view->access & PIPE_IMAGE_ACCESS_WRITE)
288               nvc0_mark_image_range_valid(view);
289         }
290
291         nve4_set_surface_info(push, view, nvc0);
292         BCTX_REFN(nvc0->bufctx_cp, CP_SUF, res, RDWR);
293
294         if (nvc0->screen->base.class_3d >= GM107_3D_CLASS)
295            gm107_compute_validate_surfaces(nvc0, view, i);
296      } else {
297         for (j = 0; j < 16; j++)
298            PUSH_DATA(push, 0);
299      }
300   }
301}
302
303/* Thankfully, textures with samplers follow the normal rules. */
304static void
305nve4_compute_validate_samplers(struct nvc0_context *nvc0)
306{
307   bool need_flush = nve4_validate_tsc(nvc0, 5);
308   if (need_flush) {
309      BEGIN_NVC0(nvc0->base.pushbuf, NVE4_CP(TSC_FLUSH), 1);
310      PUSH_DATA (nvc0->base.pushbuf, 0);
311   }
312
313   /* Invalidate all 3D samplers because they are aliased. */
314   for (int s = 0; s < 5; s++)
315      nvc0->samplers_dirty[s] = ~0;
316   nvc0->dirty_3d |= NVC0_NEW_3D_SAMPLERS;
317}
318
319/* (Code duplicated at bottom for various non-convincing reasons.
320 *  E.g. we might want to use the COMPUTE subchannel to upload TIC/TSC
321 *  entries to avoid a subchannel switch.
322 *  Same for texture cache flushes.
323 *  Also, the bufctx differs, and more IFs in the 3D version looks ugly.)
324 */
325static void nve4_compute_validate_textures(struct nvc0_context *);
326
327static void
328nve4_compute_set_tex_handles(struct nvc0_context *nvc0)
329{
330   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
331   struct nvc0_screen *screen = nvc0->screen;
332   uint64_t address;
333   const unsigned s = nvc0_shader_stage(PIPE_SHADER_COMPUTE);
334   unsigned i, n;
335   uint32_t dirty = nvc0->textures_dirty[s] | nvc0->samplers_dirty[s];
336
337   if (!dirty)
338      return;
339   i = ffs(dirty) - 1;
340   n = util_logbase2(dirty) + 1 - i;
341   assert(n);
342
343   address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
344
345   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
346   PUSH_DATAh(push, address + NVC0_CB_AUX_TEX_INFO(i));
347   PUSH_DATA (push, address + NVC0_CB_AUX_TEX_INFO(i));
348   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
349   PUSH_DATA (push, n * 4);
350   PUSH_DATA (push, 0x1);
351   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + n);
352   PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
353   PUSH_DATAp(push, &nvc0->tex_handles[s][i], n);
354
355   BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
356   PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
357
358   nvc0->textures_dirty[s] = 0;
359   nvc0->samplers_dirty[s] = 0;
360}
361
362static void
363nve4_compute_validate_constbufs(struct nvc0_context *nvc0)
364{
365   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
366   const int s = 5;
367
368   while (nvc0->constbuf_dirty[s]) {
369      int i = ffs(nvc0->constbuf_dirty[s]) - 1;
370      nvc0->constbuf_dirty[s] &= ~(1 << i);
371
372      if (nvc0->constbuf[s][i].user) {
373         struct nouveau_bo *bo = nvc0->screen->uniform_bo;
374         const unsigned base = NVC0_CB_USR_INFO(s);
375         const unsigned size = nvc0->constbuf[s][0].size;
376         assert(i == 0); /* we really only want OpenGL uniforms here */
377         assert(nvc0->constbuf[s][0].u.data);
378
379         BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
380         PUSH_DATAh(push, bo->offset + base);
381         PUSH_DATA (push, bo->offset + base);
382         BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
383         PUSH_DATA (push, size);
384         PUSH_DATA (push, 0x1);
385         BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (size / 4));
386         PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
387         PUSH_DATAp(push, nvc0->constbuf[s][0].u.data, size / 4);
388      }
389      else {
390         struct nv04_resource *res =
391            nv04_resource(nvc0->constbuf[s][i].u.buf);
392         if (res) {
393            uint64_t address
394               = nvc0->screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
395
396            /* constbufs above 0 will are fetched via ubo info in the shader */
397            if (i > 0) {
398               BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
399               PUSH_DATAh(push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
400               PUSH_DATA (push, address + NVC0_CB_AUX_UBO_INFO(i - 1));
401               BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
402               PUSH_DATA (push, 4 * 4);
403               PUSH_DATA (push, 0x1);
404               BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 4);
405               PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
406
407               PUSH_DATA (push, res->address + nvc0->constbuf[s][i].offset);
408               PUSH_DATAh(push, res->address + nvc0->constbuf[s][i].offset);
409               PUSH_DATA (push, nvc0->constbuf[s][i].size);
410               PUSH_DATA (push, 0);
411            }
412
413            BCTX_REFN(nvc0->bufctx_cp, CP_CB(i), res, RD);
414            res->cb_bindings[s] |= 1 << i;
415         }
416      }
417   }
418
419   BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
420   PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
421}
422
423static void
424nve4_compute_validate_buffers(struct nvc0_context *nvc0)
425{
426   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
427   uint64_t address;
428   const int s = 5;
429   int i;
430
431   address = nvc0->screen->uniform_bo->offset + NVC0_CB_AUX_INFO(s);
432
433   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
434   PUSH_DATAh(push, address + NVC0_CB_AUX_BUF_INFO(0));
435   PUSH_DATA (push, address + NVC0_CB_AUX_BUF_INFO(0));
436   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
437   PUSH_DATA (push, 4 * NVC0_MAX_BUFFERS * 4);
438   PUSH_DATA (push, 0x1);
439   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 4 * NVC0_MAX_BUFFERS);
440   PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
441
442   for (i = 0; i < NVC0_MAX_BUFFERS; i++) {
443      if (nvc0->buffers[s][i].buffer) {
444         struct nv04_resource *res =
445            nv04_resource(nvc0->buffers[s][i].buffer);
446         PUSH_DATA (push, res->address + nvc0->buffers[s][i].buffer_offset);
447         PUSH_DATAh(push, res->address + nvc0->buffers[s][i].buffer_offset);
448         PUSH_DATA (push, nvc0->buffers[s][i].buffer_size);
449         PUSH_DATA (push, 0);
450         BCTX_REFN(nvc0->bufctx_cp, CP_BUF, res, RDWR);
451         util_range_add(&res->valid_buffer_range,
452                        nvc0->buffers[s][i].buffer_offset,
453                        nvc0->buffers[s][i].buffer_offset +
454                        nvc0->buffers[s][i].buffer_size);
455      } else {
456         PUSH_DATA (push, 0);
457         PUSH_DATA (push, 0);
458         PUSH_DATA (push, 0);
459         PUSH_DATA (push, 0);
460      }
461   }
462}
463
464static struct nvc0_state_validate
465validate_list_cp[] = {
466   { nvc0_compprog_validate,              NVC0_NEW_CP_PROGRAM     },
467   { nve4_compute_validate_textures,      NVC0_NEW_CP_TEXTURES    },
468   { nve4_compute_validate_samplers,      NVC0_NEW_CP_SAMPLERS    },
469   { nve4_compute_set_tex_handles,        NVC0_NEW_CP_TEXTURES |
470                                          NVC0_NEW_CP_SAMPLERS    },
471   { nve4_compute_validate_surfaces,      NVC0_NEW_CP_SURFACES    },
472   { nvc0_compute_validate_globals,       NVC0_NEW_CP_GLOBALS     },
473   { nve4_compute_validate_buffers,       NVC0_NEW_CP_BUFFERS     },
474   { nve4_compute_validate_constbufs,     NVC0_NEW_CP_CONSTBUF    },
475};
476
477static bool
478nve4_state_validate_cp(struct nvc0_context *nvc0, uint32_t mask)
479{
480   bool ret;
481
482   ret = nvc0_state_validate(nvc0, mask, validate_list_cp,
483                             ARRAY_SIZE(validate_list_cp), &nvc0->dirty_cp,
484                             nvc0->bufctx_cp);
485
486   if (unlikely(nvc0->state.flushed))
487      nvc0_bufctx_fence(nvc0, nvc0->bufctx_cp, true);
488   return ret;
489}
490
491static void
492nve4_compute_upload_input(struct nvc0_context *nvc0,
493                          const struct pipe_grid_info *info)
494{
495   struct nvc0_screen *screen = nvc0->screen;
496   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
497   struct nvc0_program *cp = nvc0->compprog;
498   uint64_t address;
499
500   address = screen->uniform_bo->offset + NVC0_CB_AUX_INFO(5);
501
502   if (cp->parm_size) {
503      BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
504      PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_USR_INFO(5));
505      PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_USR_INFO(5));
506      BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
507      PUSH_DATA (push, cp->parm_size);
508      PUSH_DATA (push, 0x1);
509      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (cp->parm_size / 4));
510      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
511      PUSH_DATAp(push, info->input, cp->parm_size / 4);
512   }
513   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
514   PUSH_DATAh(push, address + NVC0_CB_AUX_GRID_INFO(0));
515   PUSH_DATA (push, address + NVC0_CB_AUX_GRID_INFO(0));
516   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
517   PUSH_DATA (push, 8 * 4);
518   PUSH_DATA (push, 0x1);
519
520   if (unlikely(info->indirect)) {
521      struct nv04_resource *res = nv04_resource(info->indirect);
522      uint32_t offset = res->offset + info->indirect_offset;
523
524      nouveau_pushbuf_space(push, 32, 0, 1);
525      PUSH_REFN(push, res->bo, NOUVEAU_BO_RD | res->domain);
526
527      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 8);
528      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
529      PUSH_DATAp(push, info->block, 3);
530      nouveau_pushbuf_data(push, res->bo, offset,
531                           NVC0_IB_ENTRY_1_NO_PREFETCH | 3 * 4);
532   } else {
533      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + 8);
534      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
535      PUSH_DATAp(push, info->block, 3);
536      PUSH_DATAp(push, info->grid, 3);
537   }
538   PUSH_DATA (push, 0);
539   PUSH_DATA (push, info->work_dim);
540
541   BEGIN_NVC0(push, NVE4_CP(FLUSH), 1);
542   PUSH_DATA (push, NVE4_COMPUTE_FLUSH_CB);
543}
544
545static inline uint8_t
546nve4_compute_derive_cache_split(struct nvc0_context *nvc0, uint32_t shared_size)
547{
548   if (shared_size > (32 << 10))
549      return NVC0_3D_CACHE_SPLIT_48K_SHARED_16K_L1;
550   if (shared_size > (16 << 10))
551      return NVE4_3D_CACHE_SPLIT_32K_SHARED_32K_L1;
552   return NVC1_3D_CACHE_SPLIT_16K_SHARED_48K_L1;
553}
554
555static void
556nve4_compute_setup_buf_cb(struct nvc0_context *nvc0, bool gp100, void *desc)
557{
558   // only user constant buffers 0-6 can be put in the descriptor, the rest are
559   // loaded through global memory
560   for (int i = 0; i <= 6; i++) {
561      if (nvc0->constbuf[5][i].user || !nvc0->constbuf[5][i].u.buf)
562         continue;
563
564      struct nv04_resource *res =
565         nv04_resource(nvc0->constbuf[5][i].u.buf);
566
567      uint32_t base = res->offset + nvc0->constbuf[5][i].offset;
568      uint32_t size = nvc0->constbuf[5][i].size;
569      if (gp100)
570         gp100_cp_launch_desc_set_cb(desc, i, res->bo, base, size);
571      else
572         nve4_cp_launch_desc_set_cb(desc, i, res->bo, base, size);
573   }
574
575   // there is no need to do FLUSH(NVE4_COMPUTE_FLUSH_CB) because
576   // nve4_compute_upload_input() does it later
577}
578
579static void
580nve4_compute_setup_launch_desc(struct nvc0_context *nvc0,
581                               struct nve4_cp_launch_desc *desc,
582                               const struct pipe_grid_info *info)
583{
584   const struct nvc0_screen *screen = nvc0->screen;
585   const struct nvc0_program *cp = nvc0->compprog;
586
587   nve4_cp_launch_desc_init_default(desc);
588
589   desc->entry = nvc0_program_symbol_offset(cp, info->pc);
590
591   desc->griddim_x = info->grid[0];
592   desc->griddim_y = info->grid[1];
593   desc->griddim_z = info->grid[2];
594   desc->blockdim_x = info->block[0];
595   desc->blockdim_y = info->block[1];
596   desc->blockdim_z = info->block[2];
597
598   desc->shared_size = align(cp->cp.smem_size, 0x100);
599   desc->local_size_p = (cp->hdr[1] & 0xfffff0) + align(cp->cp.lmem_size, 0x10);
600   desc->local_size_n = 0;
601   desc->cstack_size = 0x800;
602   desc->cache_split = nve4_compute_derive_cache_split(nvc0, cp->cp.smem_size);
603
604   desc->gpr_alloc = cp->num_gprs;
605   desc->bar_alloc = cp->num_barriers;
606
607   // Only bind user uniforms and the driver constant buffer through the
608   // launch descriptor because UBOs are sticked to the driver cb to avoid the
609   // limitation of 8 CBs.
610   if (nvc0->constbuf[5][0].user || cp->parm_size) {
611      nve4_cp_launch_desc_set_cb(desc, 0, screen->uniform_bo,
612                                 NVC0_CB_USR_INFO(5), 1 << 16);
613
614      // Later logic will attempt to bind a real buffer at position 0. That
615      // should not happen if we've bound a user buffer.
616      assert(nvc0->constbuf[5][0].user || !nvc0->constbuf[5][0].u.buf);
617   }
618   nve4_cp_launch_desc_set_cb(desc, 7, screen->uniform_bo,
619                              NVC0_CB_AUX_INFO(5), 1 << 11);
620
621   nve4_compute_setup_buf_cb(nvc0, false, desc);
622}
623
624static void
625gp100_compute_setup_launch_desc(struct nvc0_context *nvc0,
626                                struct gp100_cp_launch_desc *desc,
627                                const struct pipe_grid_info *info)
628{
629   const struct nvc0_screen *screen = nvc0->screen;
630   const struct nvc0_program *cp = nvc0->compprog;
631
632   gp100_cp_launch_desc_init_default(desc);
633
634   desc->entry = nvc0_program_symbol_offset(cp, info->pc);
635
636   desc->griddim_x = info->grid[0];
637   desc->griddim_y = info->grid[1];
638   desc->griddim_z = info->grid[2];
639   desc->blockdim_x = info->block[0];
640   desc->blockdim_y = info->block[1];
641   desc->blockdim_z = info->block[2];
642
643   desc->shared_size = align(cp->cp.smem_size, 0x100);
644   desc->local_size_p = (cp->hdr[1] & 0xfffff0) + align(cp->cp.lmem_size, 0x10);
645   desc->local_size_n = 0;
646   desc->cstack_size = 0x800;
647
648   desc->gpr_alloc = cp->num_gprs;
649   desc->bar_alloc = cp->num_barriers;
650
651   // Only bind user uniforms and the driver constant buffer through the
652   // launch descriptor because UBOs are sticked to the driver cb to avoid the
653   // limitation of 8 CBs.
654   if (nvc0->constbuf[5][0].user || cp->parm_size) {
655      gp100_cp_launch_desc_set_cb(desc, 0, screen->uniform_bo,
656                                  NVC0_CB_USR_INFO(5), 1 << 16);
657
658      // Later logic will attempt to bind a real buffer at position 0. That
659      // should not happen if we've bound a user buffer.
660      assert(nvc0->constbuf[5][0].user || !nvc0->constbuf[5][0].u.buf);
661   }
662   gp100_cp_launch_desc_set_cb(desc, 7, screen->uniform_bo,
663                               NVC0_CB_AUX_INFO(5), 1 << 11);
664
665   nve4_compute_setup_buf_cb(nvc0, true, desc);
666}
667
668static inline void *
669nve4_compute_alloc_launch_desc(struct nouveau_context *nv,
670                               struct nouveau_bo **pbo, uint64_t *pgpuaddr)
671{
672   uint8_t *ptr = nouveau_scratch_get(nv, 512, pgpuaddr, pbo);
673   if (!ptr)
674      return NULL;
675   if (*pgpuaddr & 255) {
676      unsigned adj = 256 - (*pgpuaddr & 255);
677      ptr += adj;
678      *pgpuaddr += adj;
679   }
680   return ptr;
681}
682
683static void
684nve4_upload_indirect_desc(struct nouveau_pushbuf *push,
685                          struct nv04_resource *res,  uint64_t gpuaddr,
686                          uint32_t length, uint32_t bo_offset)
687{
688   BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
689   PUSH_DATAh(push, gpuaddr);
690   PUSH_DATA (push, gpuaddr);
691   BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
692   PUSH_DATA (push, length);
693   PUSH_DATA (push, 1);
694
695   nouveau_pushbuf_space(push, 32, 0, 1);
696   PUSH_REFN(push, res->bo, NOUVEAU_BO_RD | res->domain);
697
698   BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (length / 4));
699   PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
700   nouveau_pushbuf_data(push, res->bo, bo_offset,
701                        NVC0_IB_ENTRY_1_NO_PREFETCH | length);
702}
703
704void
705nve4_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
706{
707   struct nvc0_context *nvc0 = nvc0_context(pipe);
708   struct nvc0_screen *screen = nvc0->screen;
709   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
710   void *desc;
711   uint64_t desc_gpuaddr;
712   struct nouveau_bo *desc_bo;
713   int ret;
714
715   desc = nve4_compute_alloc_launch_desc(&nvc0->base, &desc_bo, &desc_gpuaddr);
716   if (!desc) {
717      ret = -1;
718      goto out;
719   }
720   BCTX_REFN_bo(nvc0->bufctx_cp, CP_DESC, NOUVEAU_BO_GART | NOUVEAU_BO_RD,
721                desc_bo);
722
723   list_for_each_entry(struct nvc0_resident, resident, &nvc0->tex_head, list) {
724      nvc0_add_resident(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS, resident->buf,
725                        resident->flags);
726   }
727
728   list_for_each_entry(struct nvc0_resident, resident, &nvc0->img_head, list) {
729      nvc0_add_resident(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS, resident->buf,
730                        resident->flags);
731   }
732
733   ret = !nve4_state_validate_cp(nvc0, ~0);
734   if (ret)
735      goto out;
736
737   if (nvc0->screen->compute->oclass >= GP100_COMPUTE_CLASS)
738      gp100_compute_setup_launch_desc(nvc0, desc, info);
739   else
740      nve4_compute_setup_launch_desc(nvc0, desc, info);
741
742   nve4_compute_upload_input(nvc0, info);
743
744#ifdef DEBUG
745   if (debug_get_num_option("NV50_PROG_DEBUG", 0)) {
746      if (nvc0->screen->compute->oclass >= GP100_COMPUTE_CLASS)
747         gp100_compute_dump_launch_desc(desc);
748      else
749         nve4_compute_dump_launch_desc(desc);
750   }
751#endif
752
753   if (unlikely(info->indirect)) {
754      struct nv04_resource *res = nv04_resource(info->indirect);
755      uint32_t offset = res->offset + info->indirect_offset;
756
757      /* upload the descriptor */
758      BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
759      PUSH_DATAh(push, desc_gpuaddr);
760      PUSH_DATA (push, desc_gpuaddr);
761      BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
762      PUSH_DATA (push, 256);
763      PUSH_DATA (push, 1);
764      BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 1 + (256 / 4));
765      PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x08 << 1));
766      PUSH_DATAp(push, (const uint32_t *)desc, 256 / 4);
767
768      if (nvc0->screen->compute->oclass >= GP100_COMPUTE_CLASS) {
769         nve4_upload_indirect_desc(push, res, desc_gpuaddr + 48, 12, offset);
770      } else {
771         /* overwrite griddim_x and griddim_y as two 32-bits integers even
772          * if griddim_y must be a 16-bits integer */
773         nve4_upload_indirect_desc(push, res, desc_gpuaddr + 48, 8, offset);
774
775         /* overwrite the 16 high bits of griddim_y with griddim_z because
776          * we need (z << 16) | x */
777         nve4_upload_indirect_desc(push, res, desc_gpuaddr + 54, 4, offset + 8);
778      }
779   }
780
781   /* upload descriptor and flush */
782   nouveau_pushbuf_space(push, 32, 1, 0);
783   PUSH_REFN(push, screen->text, NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_RD);
784   BEGIN_NVC0(push, NVE4_CP(LAUNCH_DESC_ADDRESS), 1);
785   PUSH_DATA (push, desc_gpuaddr >> 8);
786   BEGIN_NVC0(push, NVE4_CP(LAUNCH), 1);
787   PUSH_DATA (push, 0x3);
788   BEGIN_NVC0(push, SUBC_CP(NV50_GRAPH_SERIALIZE), 1);
789   PUSH_DATA (push, 0);
790
791   nvc0_update_compute_invocations_counter(nvc0, info);
792
793out:
794   if (ret)
795      NOUVEAU_ERR("Failed to launch grid !\n");
796   nouveau_scratch_done(&nvc0->base);
797   nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_DESC);
798   nouveau_bufctx_reset(nvc0->bufctx_cp, NVC0_BIND_CP_BINDLESS);
799}
800
801
802#define NVE4_TIC_ENTRY_INVALID 0x000fffff
803
804static void
805nve4_compute_validate_textures(struct nvc0_context *nvc0)
806{
807   struct nouveau_bo *txc = nvc0->screen->txc;
808   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
809   const unsigned s = 5;
810   unsigned i;
811   uint32_t commands[2][32];
812   unsigned n[2] = { 0, 0 };
813
814   for (i = 0; i < nvc0->num_textures[s]; ++i) {
815      struct nv50_tic_entry *tic = nv50_tic_entry(nvc0->textures[s][i]);
816      struct nv04_resource *res;
817      const bool dirty = !!(nvc0->textures_dirty[s] & (1 << i));
818
819      if (!tic) {
820         nvc0->tex_handles[s][i] |= NVE4_TIC_ENTRY_INVALID;
821         continue;
822      }
823      res = nv04_resource(tic->pipe.texture);
824      nvc0_update_tic(nvc0, tic, res);
825
826      if (tic->id < 0) {
827         tic->id = nvc0_screen_tic_alloc(nvc0->screen, tic);
828
829         PUSH_SPACE(push, 16);
830         BEGIN_NVC0(push, NVE4_CP(UPLOAD_DST_ADDRESS_HIGH), 2);
831         PUSH_DATAh(push, txc->offset + (tic->id * 32));
832         PUSH_DATA (push, txc->offset + (tic->id * 32));
833         BEGIN_NVC0(push, NVE4_CP(UPLOAD_LINE_LENGTH_IN), 2);
834         PUSH_DATA (push, 32);
835         PUSH_DATA (push, 1);
836         BEGIN_1IC0(push, NVE4_CP(UPLOAD_EXEC), 9);
837         PUSH_DATA (push, NVE4_COMPUTE_UPLOAD_EXEC_LINEAR | (0x20 << 1));
838         PUSH_DATAp(push, &tic->tic[0], 8);
839
840         commands[0][n[0]++] = (tic->id << 4) | 1;
841      } else
842      if (res->status & NOUVEAU_BUFFER_STATUS_GPU_WRITING) {
843         commands[1][n[1]++] = (tic->id << 4) | 1;
844      }
845      nvc0->screen->tic.lock[tic->id / 32] |= 1 << (tic->id % 32);
846
847      res->status &= ~NOUVEAU_BUFFER_STATUS_GPU_WRITING;
848      res->status |=  NOUVEAU_BUFFER_STATUS_GPU_READING;
849
850      nvc0->tex_handles[s][i] &= ~NVE4_TIC_ENTRY_INVALID;
851      nvc0->tex_handles[s][i] |= tic->id;
852      if (dirty)
853         BCTX_REFN(nvc0->bufctx_cp, CP_TEX(i), res, RD);
854   }
855   for (; i < nvc0->state.num_textures[s]; ++i) {
856      nvc0->tex_handles[s][i] |= NVE4_TIC_ENTRY_INVALID;
857      nvc0->textures_dirty[s] |= 1 << i;
858   }
859
860   if (n[0]) {
861      BEGIN_NIC0(push, NVE4_CP(TIC_FLUSH), n[0]);
862      PUSH_DATAp(push, commands[0], n[0]);
863   }
864   if (n[1]) {
865      BEGIN_NIC0(push, NVE4_CP(TEX_CACHE_CTL), n[1]);
866      PUSH_DATAp(push, commands[1], n[1]);
867   }
868
869   nvc0->state.num_textures[s] = nvc0->num_textures[s];
870
871   /* Invalidate all 3D textures because they are aliased. */
872   for (int s = 0; s < 5; s++) {
873      for (int i = 0; i < nvc0->num_textures[s]; i++)
874         nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_3D_TEX(s, i));
875      nvc0->textures_dirty[s] = ~0;
876   }
877   nvc0->dirty_3d |= NVC0_NEW_3D_TEXTURES;
878}
879
880
881#ifdef DEBUG
882static const char *nve4_cache_split_name(unsigned value)
883{
884   switch (value) {
885   case NVC1_3D_CACHE_SPLIT_16K_SHARED_48K_L1: return "16K_SHARED_48K_L1";
886   case NVE4_3D_CACHE_SPLIT_32K_SHARED_32K_L1: return "32K_SHARED_32K_L1";
887   case NVC0_3D_CACHE_SPLIT_48K_SHARED_16K_L1: return "48K_SHARED_16K_L1";
888   default:
889      return "(invalid)";
890   }
891}
892
893static void
894nve4_compute_dump_launch_desc(const struct nve4_cp_launch_desc *desc)
895{
896   const uint32_t *data = (const uint32_t *)desc;
897   unsigned i;
898   bool zero = false;
899
900   debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
901
902   for (i = 0; i < sizeof(*desc); i += 4) {
903      if (data[i / 4]) {
904         debug_printf("[%x]: 0x%08x\n", i, data[i / 4]);
905         zero = false;
906      } else
907      if (!zero) {
908         debug_printf("...\n");
909         zero = true;
910      }
911   }
912
913   debug_printf("entry = 0x%x\n", desc->entry);
914   debug_printf("grid dimensions = %ux%ux%u\n",
915                desc->griddim_x, desc->griddim_y, desc->griddim_z);
916   debug_printf("block dimensions = %ux%ux%u\n",
917                desc->blockdim_x, desc->blockdim_y, desc->blockdim_z);
918   debug_printf("s[] size: 0x%x\n", desc->shared_size);
919   debug_printf("l[] size: -0x%x / +0x%x\n",
920                desc->local_size_n, desc->local_size_p);
921   debug_printf("stack size: 0x%x\n", desc->cstack_size);
922   debug_printf("barrier count: %u\n", desc->bar_alloc);
923   debug_printf("$r count: %u\n", desc->gpr_alloc);
924   debug_printf("cache split: %s\n", nve4_cache_split_name(desc->cache_split));
925   debug_printf("linked tsc: %d\n", desc->linked_tsc);
926
927   for (i = 0; i < 8; ++i) {
928      uint64_t address;
929      uint32_t size = desc->cb[i].size;
930      bool valid = !!(desc->cb_mask & (1 << i));
931
932      address = ((uint64_t)desc->cb[i].address_h << 32) | desc->cb[i].address_l;
933
934      if (!valid && !address && !size)
935         continue;
936      debug_printf("CB[%u]: address = 0x%"PRIx64", size 0x%x%s\n",
937                   i, address, size, valid ? "" : "  (invalid)");
938   }
939}
940
941static void
942gp100_compute_dump_launch_desc(const struct gp100_cp_launch_desc *desc)
943{
944   const uint32_t *data = (const uint32_t *)desc;
945   unsigned i;
946   bool zero = false;
947
948   debug_printf("COMPUTE LAUNCH DESCRIPTOR:\n");
949
950   for (i = 0; i < sizeof(*desc); i += 4) {
951      if (data[i / 4]) {
952         debug_printf("[%x]: 0x%08x\n", i, data[i / 4]);
953         zero = false;
954      } else
955      if (!zero) {
956         debug_printf("...\n");
957         zero = true;
958      }
959   }
960
961   debug_printf("entry = 0x%x\n", desc->entry);
962   debug_printf("grid dimensions = %ux%ux%u\n",
963                desc->griddim_x, desc->griddim_y, desc->griddim_z);
964   debug_printf("block dimensions = %ux%ux%u\n",
965                desc->blockdim_x, desc->blockdim_y, desc->blockdim_z);
966   debug_printf("s[] size: 0x%x\n", desc->shared_size);
967   debug_printf("l[] size: -0x%x / +0x%x\n",
968                desc->local_size_n, desc->local_size_p);
969   debug_printf("stack size: 0x%x\n", desc->cstack_size);
970   debug_printf("barrier count: %u\n", desc->bar_alloc);
971   debug_printf("$r count: %u\n", desc->gpr_alloc);
972   debug_printf("linked tsc: %d\n", desc->linked_tsc);
973
974   for (i = 0; i < 8; ++i) {
975      uint64_t address;
976      uint32_t size = desc->cb[i].size_sh4 << 4;
977      bool valid = !!(desc->cb_mask & (1 << i));
978
979      address = ((uint64_t)desc->cb[i].address_h << 32) | desc->cb[i].address_l;
980
981      if (!valid && !address && !size)
982         continue;
983      debug_printf("CB[%u]: address = 0x%"PRIx64", size 0x%x%s\n",
984                   i, address, size, valid ? "" : "  (invalid)");
985   }
986}
987#endif
988
989#ifdef NOUVEAU_NVE4_MP_TRAP_HANDLER
990static void
991nve4_compute_trap_info(struct nvc0_context *nvc0)
992{
993   struct nvc0_screen *screen = nvc0->screen;
994   struct nouveau_bo *bo = screen->parm;
995   int ret, i;
996   volatile struct nve4_mp_trap_info *info;
997   uint8_t *map;
998
999   ret = nouveau_bo_map(bo, NOUVEAU_BO_RDWR, nvc0->base.client);
1000   if (ret)
1001      return;
1002   map = (uint8_t *)bo->map;
1003   info = (volatile struct nve4_mp_trap_info *)(map + NVE4_CP_PARAM_TRAP_INFO);
1004
1005   if (info->lock) {
1006      debug_printf("trapstat = %08x\n", info->trapstat);
1007      debug_printf("warperr = %08x\n", info->warperr);
1008      debug_printf("PC = %x\n", info->pc);
1009      debug_printf("tid = %u %u %u\n",
1010                   info->tid[0], info->tid[1], info->tid[2]);
1011      debug_printf("ctaid = %u %u %u\n",
1012                   info->ctaid[0], info->ctaid[1], info->ctaid[2]);
1013      for (i = 0; i <= 63; ++i)
1014         debug_printf("$r%i = %08x\n", i, info->r[i]);
1015      for (i = 0; i <= 6; ++i)
1016         debug_printf("$p%i = %i\n", i, (info->flags >> i) & 1);
1017      debug_printf("$c = %x\n", info->flags >> 12);
1018   }
1019   info->lock = 0;
1020}
1021#endif
1022