1/*
2 * Copyright 2012 Francisco Jerez
3 * Copyright 2015 Samuel Pitoiset
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sublicense, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial
15 * portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
21 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
22 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
23 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 *
25 */
26
27#include "util/format/u_format.h"
28#include "nv50/nv50_context.h"
29#include "nv50/nv50_compute.xml.h"
30
31#include "nv50_ir_driver.h"
32
33int
34nv50_screen_compute_setup(struct nv50_screen *screen,
35                          struct nouveau_pushbuf *push)
36{
37   struct nouveau_device *dev = screen->base.device;
38   struct nouveau_object *chan = screen->base.channel;
39   struct nv04_fifo *fifo = (struct nv04_fifo *)chan->data;
40   unsigned obj_class;
41   int i, ret;
42
43   switch (dev->chipset & 0xf0) {
44   case 0x50:
45   case 0x80:
46   case 0x90:
47      obj_class = NV50_COMPUTE_CLASS;
48      break;
49   case 0xa0:
50      switch (dev->chipset) {
51      case 0xa3:
52      case 0xa5:
53      case 0xa8:
54         obj_class = NVA3_COMPUTE_CLASS;
55         break;
56      default:
57         obj_class = NV50_COMPUTE_CLASS;
58         break;
59      }
60      break;
61   default:
62      NOUVEAU_ERR("unsupported chipset: NV%02x\n", dev->chipset);
63      return -1;
64   }
65
66   ret = nouveau_object_new(chan, 0xbeef50c0, obj_class, NULL, 0,
67                            &screen->compute);
68   if (ret)
69      return ret;
70
71   BEGIN_NV04(push, SUBC_CP(NV01_SUBCHAN_OBJECT), 1);
72   PUSH_DATA (push, screen->compute->handle);
73
74   BEGIN_NV04(push, NV50_CP(UNK02A0), 1);
75   PUSH_DATA (push, 1);
76   BEGIN_NV04(push, NV50_CP(DMA_STACK), 1);
77   PUSH_DATA (push, fifo->vram);
78   BEGIN_NV04(push, NV50_CP(STACK_ADDRESS_HIGH), 2);
79   PUSH_DATAh(push, screen->stack_bo->offset);
80   PUSH_DATA (push, screen->stack_bo->offset);
81   BEGIN_NV04(push, NV50_CP(STACK_SIZE_LOG), 1);
82   PUSH_DATA (push, 4);
83
84   BEGIN_NV04(push, NV50_CP(UNK0290), 1);
85   PUSH_DATA (push, 1);
86   BEGIN_NV04(push, NV50_CP(LANES32_ENABLE), 1);
87   PUSH_DATA (push, 1);
88   BEGIN_NV04(push, NV50_CP(REG_MODE), 1);
89   PUSH_DATA (push, NV50_COMPUTE_REG_MODE_STRIPED);
90   BEGIN_NV04(push, NV50_CP(UNK0384), 1);
91   PUSH_DATA (push, 0x100);
92   BEGIN_NV04(push, NV50_CP(DMA_GLOBAL), 1);
93   PUSH_DATA (push, fifo->vram);
94
95   for (i = 0; i < 15; i++) {
96      BEGIN_NV04(push, NV50_CP(GLOBAL_ADDRESS_HIGH(i)), 2);
97      PUSH_DATA (push, 0);
98      PUSH_DATA (push, 0);
99      BEGIN_NV04(push, NV50_CP(GLOBAL_LIMIT(i)), 1);
100      PUSH_DATA (push, 0);
101      BEGIN_NV04(push, NV50_CP(GLOBAL_MODE(i)), 1);
102      PUSH_DATA (push, NV50_COMPUTE_GLOBAL_MODE_LINEAR);
103   }
104
105   BEGIN_NV04(push, NV50_CP(GLOBAL_ADDRESS_HIGH(15)), 2);
106   PUSH_DATA (push, 0);
107   PUSH_DATA (push, 0);
108   BEGIN_NV04(push, NV50_CP(GLOBAL_LIMIT(15)), 1);
109   PUSH_DATA (push, ~0);
110   BEGIN_NV04(push, NV50_CP(GLOBAL_MODE(15)), 1);
111   PUSH_DATA (push, NV50_COMPUTE_GLOBAL_MODE_LINEAR);
112
113   BEGIN_NV04(push, NV50_CP(LOCAL_WARPS_LOG_ALLOC), 1);
114   PUSH_DATA (push, 7);
115   BEGIN_NV04(push, NV50_CP(LOCAL_WARPS_NO_CLAMP), 1);
116   PUSH_DATA (push, 1);
117   BEGIN_NV04(push, NV50_CP(STACK_WARPS_LOG_ALLOC), 1);
118   PUSH_DATA (push, 7);
119   BEGIN_NV04(push, NV50_CP(STACK_WARPS_NO_CLAMP), 1);
120   PUSH_DATA (push, 1);
121   BEGIN_NV04(push, NV50_CP(USER_PARAM_COUNT), 1);
122   PUSH_DATA (push, 0);
123
124   BEGIN_NV04(push, NV50_CP(DMA_TEXTURE), 1);
125   PUSH_DATA (push, fifo->vram);
126   BEGIN_NV04(push, NV50_CP(TEX_LIMITS), 1);
127   PUSH_DATA (push, 0x54);
128   BEGIN_NV04(push, NV50_CP(LINKED_TSC), 1);
129   PUSH_DATA (push, 0);
130
131   BEGIN_NV04(push, NV50_CP(DMA_TIC), 1);
132   PUSH_DATA (push, fifo->vram);
133   BEGIN_NV04(push, NV50_CP(TIC_ADDRESS_HIGH), 3);
134   PUSH_DATAh(push, screen->txc->offset);
135   PUSH_DATA (push, screen->txc->offset);
136   PUSH_DATA (push, NV50_TIC_MAX_ENTRIES - 1);
137
138   BEGIN_NV04(push, NV50_CP(DMA_TSC), 1);
139   PUSH_DATA (push, fifo->vram);
140   BEGIN_NV04(push, NV50_CP(TSC_ADDRESS_HIGH), 3);
141   PUSH_DATAh(push, screen->txc->offset + 65536);
142   PUSH_DATA (push, screen->txc->offset + 65536);
143   PUSH_DATA (push, NV50_TSC_MAX_ENTRIES - 1);
144
145   BEGIN_NV04(push, NV50_CP(DMA_CODE_CB), 1);
146   PUSH_DATA (push, fifo->vram);
147
148   BEGIN_NV04(push, NV50_CP(DMA_LOCAL), 1);
149   PUSH_DATA (push, fifo->vram);
150   BEGIN_NV04(push, NV50_CP(LOCAL_ADDRESS_HIGH), 2);
151   PUSH_DATAh(push, screen->tls_bo->offset + 65536);
152   PUSH_DATA (push, screen->tls_bo->offset + 65536);
153   BEGIN_NV04(push, NV50_CP(LOCAL_SIZE_LOG), 1);
154   PUSH_DATA (push, util_logbase2((screen->max_tls_space / ONE_TEMP_SIZE) * 2));
155
156   BEGIN_NV04(push, NV50_CP(CB_DEF_ADDRESS_HIGH), 3);
157   PUSH_DATAh(push, screen->uniforms->offset + (3 << 16));
158   PUSH_DATA (push, screen->uniforms->offset + (3 << 16));
159   PUSH_DATA (push, (NV50_CB_PCP << 16) | 0x0000);
160
161   BEGIN_NV04(push, NV50_CP(QUERY_ADDRESS_HIGH), 2);
162   PUSH_DATAh(push, screen->fence.bo->offset + 16);
163   PUSH_DATA (push, screen->fence.bo->offset + 16);
164
165   return 0;
166}
167
168static void
169nv50_compute_validate_samplers(struct nv50_context *nv50)
170{
171   bool need_flush = nv50_validate_tsc(nv50, NV50_SHADER_STAGE_COMPUTE);
172   if (need_flush) {
173      BEGIN_NV04(nv50->base.pushbuf, NV50_CP(TSC_FLUSH), 1);
174      PUSH_DATA (nv50->base.pushbuf, 0);
175   }
176
177   /* Invalidate all 3D samplers because they are aliased. */
178   nv50->dirty_3d |= NV50_NEW_3D_SAMPLERS;
179}
180
181static void
182nv50_compute_validate_textures(struct nv50_context *nv50)
183{
184   bool need_flush = nv50_validate_tic(nv50, NV50_SHADER_STAGE_COMPUTE);
185   if (need_flush) {
186      BEGIN_NV04(nv50->base.pushbuf, NV50_CP(TIC_FLUSH), 1);
187      PUSH_DATA (nv50->base.pushbuf, 0);
188   }
189
190   /* Invalidate all 3D textures because they are aliased. */
191   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_3D_TEXTURES);
192   nv50->dirty_3d |= NV50_NEW_3D_TEXTURES;
193}
194
195static inline void
196nv50_compute_invalidate_constbufs(struct nv50_context *nv50)
197{
198   int s;
199
200   /* Invalidate all 3D constbufs because they are aliased with COMPUTE. */
201   for (s = 0; s < NV50_MAX_3D_SHADER_STAGES; s++) {
202      nv50->constbuf_dirty[s] |= nv50->constbuf_valid[s];
203      nv50->state.uniform_buffer_bound[s] = false;
204   }
205   nv50->dirty_3d |= NV50_NEW_3D_CONSTBUF;
206}
207
208static void
209nv50_compute_validate_constbufs(struct nv50_context *nv50)
210{
211   struct nouveau_pushbuf *push = nv50->base.pushbuf;
212   const int s = NV50_SHADER_STAGE_COMPUTE;
213
214   while (nv50->constbuf_dirty[s]) {
215      int i = ffs(nv50->constbuf_dirty[s]) - 1;
216      nv50->constbuf_dirty[s] &= ~(1 << i);
217
218      if (nv50->constbuf[s][i].user) {
219         const unsigned b = NV50_CB_PVP + s;
220         unsigned start = 0;
221         unsigned words = nv50->constbuf[s][0].size / 4;
222         if (i) {
223            NOUVEAU_ERR("user constbufs only supported in slot 0\n");
224            continue;
225         }
226         if (!nv50->state.uniform_buffer_bound[s]) {
227            nv50->state.uniform_buffer_bound[s] = true;
228            BEGIN_NV04(push, NV50_CP(SET_PROGRAM_CB), 1);
229            PUSH_DATA (push, (b << 12) | (i << 8) | 1);
230         }
231         while (words) {
232            unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN);
233
234            PUSH_SPACE(push, nr + 3);
235            BEGIN_NV04(push, NV50_CP(CB_ADDR), 1);
236            PUSH_DATA (push, (start << 8) | b);
237            BEGIN_NI04(push, NV50_CP(CB_DATA(0)), nr);
238            PUSH_DATAp(push, &nv50->constbuf[s][0].u.data[start * 4], nr);
239
240            start += nr;
241            words -= nr;
242         }
243      } else {
244         struct nv04_resource *res =
245            nv04_resource(nv50->constbuf[s][i].u.buf);
246         if (res) {
247            /* TODO: allocate persistent bindings */
248            const unsigned b = s * 16 + i;
249
250            assert(nouveau_resource_mapped_by_gpu(&res->base));
251
252            BEGIN_NV04(push, NV50_CP(CB_DEF_ADDRESS_HIGH), 3);
253            PUSH_DATAh(push, res->address + nv50->constbuf[s][i].offset);
254            PUSH_DATA (push, res->address + nv50->constbuf[s][i].offset);
255            PUSH_DATA (push, (b << 16) |
256                       (nv50->constbuf[s][i].size & 0xffff));
257            BEGIN_NV04(push, NV50_CP(SET_PROGRAM_CB), 1);
258            PUSH_DATA (push, (b << 12) | (i << 8) | 1);
259
260            BCTX_REFN(nv50->bufctx_cp, CP_CB(i), res, RD);
261
262            nv50->cb_dirty = 1; /* Force cache flush for UBO. */
263            res->cb_bindings[s] |= 1 << i;
264         } else {
265            BEGIN_NV04(push, NV50_CP(SET_PROGRAM_CB), 1);
266            PUSH_DATA (push, (i << 8) | 0);
267         }
268         if (i == 0)
269            nv50->state.uniform_buffer_bound[s] = false;
270      }
271   }
272
273   // TODO: Check if having orthogonal slots means the two don't trample over
274   // each other.
275   nv50_compute_invalidate_constbufs(nv50);
276}
277
278static void
279nv50_get_surface_dims(const struct pipe_image_view *view,
280                      int *width, int *height, int *depth)
281{
282   struct nv04_resource *res = nv04_resource(view->resource);
283   int level;
284
285   *width = *height = *depth = 1;
286   if (res->base.target == PIPE_BUFFER) {
287      *width = view->u.buf.size / util_format_get_blocksize(view->format);
288      return;
289   }
290
291   level = view->u.tex.level;
292   *width = u_minify(view->resource->width0, level);
293   *height = u_minify(view->resource->height0, level);
294   *depth = u_minify(view->resource->depth0, level);
295
296   switch (res->base.target) {
297   case PIPE_TEXTURE_1D_ARRAY:
298   case PIPE_TEXTURE_2D_ARRAY:
299   case PIPE_TEXTURE_CUBE:
300   case PIPE_TEXTURE_CUBE_ARRAY:
301      *depth = view->u.tex.last_layer - view->u.tex.first_layer + 1;
302      break;
303   case PIPE_TEXTURE_1D:
304   case PIPE_TEXTURE_2D:
305   case PIPE_TEXTURE_RECT:
306   case PIPE_TEXTURE_3D:
307      break;
308   default:
309      assert(!"unexpected texture target");
310      break;
311   }
312}
313
314static void
315nv50_mark_image_range_valid(const struct pipe_image_view *view)
316{
317   struct nv04_resource *res = (struct nv04_resource *)view->resource;
318
319   assert(view->resource->target == PIPE_BUFFER);
320
321   util_range_add(&res->base, &res->valid_buffer_range,
322                  view->u.buf.offset,
323                  view->u.buf.offset + view->u.buf.size);
324}
325
326static inline void
327nv50_set_surface_info(struct nouveau_pushbuf *push,
328                      const struct pipe_image_view *view,
329                      int width, int height, int depth)
330{
331   struct nv04_resource *res;
332   uint32_t *const info = push->cur;
333
334   push->cur += 12;
335
336   /* Make sure to always initialize the surface information area because it's
337    * used to check if the given image is bound or not. */
338   memset(info, 0, 12 * sizeof(*info));
339
340   if (!view || !view->resource)
341      return;
342   res = nv04_resource(view->resource);
343
344   /* Stick the image dimensions for the imageSize() builtin. */
345   info[0] = width;
346   info[1] = height;
347   info[2] = depth;
348
349   /* Stick the blockwidth (ie. number of bytes per pixel) to calculate pixel
350    * offset and to check if the format doesn't mismatch. */
351   info[3] = util_format_get_blocksize(view->format);
352
353   if (res->base.target != PIPE_BUFFER) {
354      struct nv50_miptree *mt = nv50_miptree(&res->base);
355      struct nv50_miptree_level *lvl = &mt->level[view->u.tex.level];
356      unsigned nby = align(util_format_get_nblocksy(view->format, height),
357                           NV50_TILE_SIZE_Y(lvl->tile_mode));
358
359      if (mt->layout_3d) {
360         info[4] = nby;
361         info[11] = view->u.tex.first_layer;
362      } else {
363         info[4] = mt->layer_stride / lvl->pitch;
364      }
365      info[6] = mt->ms_x;
366      info[7] = mt->ms_y;
367      info[8] = NV50_TILE_SHIFT_X(lvl->tile_mode);
368      info[9] = NV50_TILE_SHIFT_Y(lvl->tile_mode);
369      info[10] = NV50_TILE_SHIFT_Z(lvl->tile_mode);
370   }
371}
372
373static void
374nv50_compute_validate_surfaces(struct nv50_context *nv50)
375{
376   struct nouveau_pushbuf *push = nv50->base.pushbuf;
377   int i;
378
379   for (i = 0; i < NV50_MAX_GLOBALS - 1; i++) {
380      struct nv50_gmem_state *gmem = &nv50->compprog->cp.gmem[i];
381      int width, height, depth;
382      uint64_t address = 0;
383
384      BEGIN_NV04(push, NV50_CP(GLOBAL(i)), 5);
385
386      if (gmem->valid && !gmem->image && nv50->buffers[gmem->slot].buffer) {
387         struct pipe_shader_buffer *buffer = &nv50->buffers[gmem->slot];
388         struct nv04_resource *res = nv04_resource(buffer->buffer);
389         PUSH_DATAh(push, res->address + buffer->buffer_offset);
390         PUSH_DATA (push, res->address + buffer->buffer_offset);
391         PUSH_DATA (push, 0); /* pitch? */
392         PUSH_DATA (push, ALIGN(buffer->buffer_size, 256) - 1);
393         PUSH_DATA (push, NV50_COMPUTE_GLOBAL_MODE_LINEAR);
394         BCTX_REFN(nv50->bufctx_cp, CP_BUF, res, RDWR);
395         util_range_add(&res->base, &res->valid_buffer_range,
396                        buffer->buffer_offset,
397                        buffer->buffer_offset +
398                        buffer->buffer_size);
399
400         PUSH_SPACE(push, 1 + 3);
401         BEGIN_NV04(push, NV50_CP(CB_ADDR), 1);
402         PUSH_DATA (push, NV50_CB_AUX_BUF_INFO(i) << (8 - 2) | NV50_CB_AUX);
403         BEGIN_NI04(push, NV50_CP(CB_DATA(0)), 1);
404         PUSH_DATA (push, buffer->buffer_size);
405      } else if (gmem->valid && gmem->image && nv50->images[gmem->slot].resource) {
406         struct pipe_image_view *view = &nv50->images[gmem->slot];
407         struct nv04_resource *res = nv04_resource(view->resource);
408
409         /* get surface dimensions based on the target. */
410         nv50_get_surface_dims(view, &width, &height, &depth);
411
412         address = res->address;
413         if (res->base.target == PIPE_BUFFER) {
414            address += view->u.buf.offset;
415            assert(!(address & 0xff));
416
417            if (view->access & PIPE_IMAGE_ACCESS_WRITE)
418               nv50_mark_image_range_valid(view);
419
420            PUSH_DATAh(push, address);
421            PUSH_DATA (push, address);
422            PUSH_DATA (push, 0); /* pitch? */
423            PUSH_DATA (push, ALIGN(view->u.buf.size, 0x100) - 1);
424            PUSH_DATA (push, NV50_COMPUTE_GLOBAL_MODE_LINEAR);
425         } else {
426            struct nv50_miptree *mt = nv50_miptree(view->resource);
427            struct nv50_miptree_level *lvl = &mt->level[view->u.tex.level];
428            const unsigned z = view->u.tex.first_layer;
429            unsigned max_size;
430
431            if (mt->layout_3d) {
432               address += nv50_mt_zslice_offset(mt, view->u.tex.level, 0);
433               max_size = mt->total_size;
434            } else {
435               address += mt->layer_stride * z;
436               max_size = mt->layer_stride * (view->u.tex.last_layer - view->u.tex.first_layer + 1);
437            }
438            address += lvl->offset;
439
440            PUSH_DATAh(push, address);
441            PUSH_DATA (push, address);
442            if (mt->layout_3d) {
443               // We have to adjust the size of the 3d surface to be
444               // accessible within 2d limits. The size of each z tile goes
445               // into the x direction, while the number of z tiles goes into
446               // the y direction.
447               const unsigned nby = util_format_get_nblocksy(view->format, height);
448               const unsigned tsy = NV50_TILE_SIZE_Y(lvl->tile_mode);
449               const unsigned tsz = NV50_TILE_SIZE_Z(lvl->tile_mode);
450               const unsigned pitch = lvl->pitch * tsz;
451               const unsigned maxy = align(nby, tsy) * align(depth, tsz) >> NV50_TILE_SHIFT_Z(lvl->tile_mode);
452               PUSH_DATA (push, pitch * tsy);
453               PUSH_DATA (push, (maxy - 1) << 16 | (pitch - 1));
454               PUSH_DATA (push, (lvl->tile_mode & 0xff) << 4);
455            } else if (nouveau_bo_memtype(res->bo)) {
456               PUSH_DATA (push, lvl->pitch * NV50_TILE_SIZE_Y(lvl->tile_mode));
457               PUSH_DATA (push, (max_size / lvl->pitch - 1) << 16 | (lvl->pitch - 1));
458               PUSH_DATA (push, (lvl->tile_mode & 0xff) << 4);
459            } else {
460               PUSH_DATA (push, lvl->pitch);
461               PUSH_DATA (push, align(lvl->pitch * height, 0x100) - 1);
462               PUSH_DATA (push, NV50_COMPUTE_GLOBAL_MODE_LINEAR);
463            }
464         }
465
466         BCTX_REFN(nv50->bufctx_cp, CP_SUF, res, RDWR);
467
468         PUSH_SPACE(push, 12 + 3);
469         BEGIN_NV04(push, NV50_CP(CB_ADDR), 1);
470         PUSH_DATA (push, NV50_CB_AUX_BUF_INFO(i) << (8 - 2) | NV50_CB_AUX);
471         BEGIN_NI04(push, NV50_CP(CB_DATA(0)), 12);
472         nv50_set_surface_info(push, view, width, height, depth);
473      } else {
474         PUSH_DATA (push, 0);
475         PUSH_DATA (push, 0);
476         PUSH_DATA (push, 0);
477         PUSH_DATA (push, 0);
478         PUSH_DATA (push, 0);
479      }
480   }
481}
482
483static void
484nv50_compute_validate_globals(struct nv50_context *nv50)
485{
486   unsigned i;
487
488   for (i = 0; i < nv50->global_residents.size / sizeof(struct pipe_resource *);
489        ++i) {
490      struct pipe_resource *res = *util_dynarray_element(
491         &nv50->global_residents, struct pipe_resource *, i);
492      if (res)
493         nv50_add_bufctx_resident(nv50->bufctx_cp, NV50_BIND_CP_GLOBAL,
494                                  nv04_resource(res), NOUVEAU_BO_RDWR);
495   }
496}
497
498static struct nv50_state_validate
499validate_list_cp[] = {
500   { nv50_compprog_validate,              NV50_NEW_CP_PROGRAM     },
501   { nv50_compute_validate_constbufs,     NV50_NEW_CP_CONSTBUF    },
502   { nv50_compute_validate_surfaces,      NV50_NEW_CP_SURFACES |
503                                          NV50_NEW_CP_BUFFERS  |
504                                          NV50_NEW_CP_PROGRAM     },
505   { nv50_compute_validate_textures,      NV50_NEW_CP_TEXTURES    },
506   { nv50_compute_validate_samplers,      NV50_NEW_CP_SAMPLERS    },
507   { nv50_compute_validate_globals,       NV50_NEW_CP_GLOBALS     },
508};
509
510static bool
511nv50_state_validate_cp(struct nv50_context *nv50, uint32_t mask)
512{
513   bool ret;
514
515   /* TODO: validate textures, samplers, surfaces */
516   ret = nv50_state_validate(nv50, mask, validate_list_cp,
517                             ARRAY_SIZE(validate_list_cp), &nv50->dirty_cp,
518                             nv50->bufctx_cp);
519
520   if (unlikely(nv50->state.flushed))
521      nv50_bufctx_fence(nv50->bufctx_cp, true);
522   return ret;
523}
524
525static void
526nv50_compute_upload_input(struct nv50_context *nv50, const uint32_t *input)
527{
528   struct nv50_screen *screen = nv50->screen;
529   struct nouveau_pushbuf *push = nv50->base.pushbuf;
530   unsigned size = align(nv50->compprog->parm_size, 0x4);
531
532   BEGIN_NV04(push, NV50_CP(USER_PARAM_COUNT), 1);
533   PUSH_DATA (push, (1 + (size / 4)) << 8);
534
535   if (size) {
536      struct nouveau_mm_allocation *mm;
537      struct nouveau_bo *bo = NULL;
538      unsigned offset;
539
540      mm = nouveau_mm_allocate(screen->base.mm_GART, size, &bo, &offset);
541      assert(mm);
542
543      nouveau_bo_map(bo, 0, nv50->base.client);
544      memcpy(bo->map + offset, input, size);
545
546      nouveau_bufctx_refn(nv50->bufctx, 0, bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
547      nouveau_pushbuf_bufctx(push, nv50->bufctx);
548      nouveau_pushbuf_validate(push);
549
550      nouveau_pushbuf_space(push, 0, 0, 1);
551
552      BEGIN_NV04(push, NV50_CP(USER_PARAM(1)), size / 4);
553      nouveau_pushbuf_data(push, bo, offset, size);
554
555      nouveau_fence_work(screen->base.fence.current, nouveau_mm_free_work, mm);
556      nouveau_bo_ref(NULL, &bo);
557      nouveau_bufctx_reset(nv50->bufctx, 0);
558   }
559}
560
561void
562nv50_launch_grid(struct pipe_context *pipe, const struct pipe_grid_info *info)
563{
564   struct nv50_context *nv50 = nv50_context(pipe);
565   struct nouveau_pushbuf *push = nv50->base.pushbuf;
566   unsigned block_size = info->block[0] * info->block[1] * info->block[2];
567   struct nv50_program *cp = nv50->compprog;
568   bool ret;
569
570   ret = !nv50_state_validate_cp(nv50, ~0);
571   if (ret) {
572      NOUVEAU_ERR("Failed to launch grid !\n");
573      return;
574   }
575
576   nv50_compute_upload_input(nv50, info->input);
577
578   BEGIN_NV04(push, NV50_CP(CP_START_ID), 1);
579   PUSH_DATA (push, cp->code_base);
580
581   BEGIN_NV04(push, NV50_CP(SHARED_SIZE), 1);
582   PUSH_DATA (push, align(cp->cp.smem_size + cp->parm_size + 0x14, 0x40));
583   BEGIN_NV04(push, NV50_CP(CP_REG_ALLOC_TEMP), 1);
584   PUSH_DATA (push, cp->max_gpr);
585
586   /* no indirect support - just read the parameters out */
587   uint32_t grid[3];
588   if (unlikely(info->indirect)) {
589      pipe_buffer_read(pipe, info->indirect, info->indirect_offset,
590                       sizeof(grid), grid);
591   } else {
592      memcpy(grid, info->grid, sizeof(grid));
593   }
594
595   /* grid/block setup */
596   BEGIN_NV04(push, NV50_CP(BLOCKDIM_XY), 2);
597   PUSH_DATA (push, info->block[1] << 16 | info->block[0]);
598   PUSH_DATA (push, info->block[2]);
599   BEGIN_NV04(push, NV50_CP(BLOCK_ALLOC), 1);
600   PUSH_DATA (push, 1 << 16 | block_size);
601   BEGIN_NV04(push, NV50_CP(BLOCKDIM_LATCH), 1);
602   PUSH_DATA (push, 1);
603   BEGIN_NV04(push, NV50_CP(GRIDDIM), 1);
604   PUSH_DATA (push, grid[1] << 16 | grid[0]);
605   BEGIN_NV04(push, NV50_CP(GRIDID), 1);
606   PUSH_DATA (push, 1);
607
608   for (int i = 0; i < grid[2]; i++) {
609      BEGIN_NV04(push, NV50_CP(USER_PARAM(0)), 1);
610      PUSH_DATA (push, grid[2] | i << 16);
611
612      /* kernel launching */
613      BEGIN_NV04(push, NV50_CP(LAUNCH), 1);
614      PUSH_DATA (push, 0);
615   }
616
617   BEGIN_NV04(push, SUBC_CP(NV50_GRAPH_SERIALIZE), 1);
618   PUSH_DATA (push, 0);
619
620   /* bind a compute shader clobbers fragment shader state */
621   nv50->dirty_3d |= NV50_NEW_3D_FRAGPROG;
622
623   nv50->compute_invocations += info->block[0] * info->block[1] * info->block[2] *
624      grid[0] * grid[1] * grid[2];
625}
626