src/gallium/drivers/nv50/nv50_state_validate.c

   1 /*
   2  * Copyright 2008 Ben Skeggs
   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 BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 #include "nv50_context.h"
  24 #include "nouveau/nouveau_stateobj.h"
  25
  26 #define NV50_CBUF_FORMAT_CASE(n) \
  27         case PIPE_FORMAT_##n: so_data(so, NV50TCL_RT_FORMAT_##n); break
  28
  29 #define NV50_ZETA_FORMAT_CASE(n) \
  30         case PIPE_FORMAT_##n: so_data(so, NV50TCL_ZETA_FORMAT_##n); break
  31
  32 static void
  33 nv50_state_validate_fb(struct nv50_context *nv50)
  34 {
  35         struct nouveau_grobj *tesla = nv50->screen->tesla;
  36         struct nouveau_stateobj *so = so_new(128, 18);
  37         struct pipe_framebuffer_state *fb = &nv50->framebuffer;
  38         unsigned i, w, h, gw = 0;
  39
  40         /* Set nr of active RTs and select RT for each colour output.
  41          * FP result 0 always goes to RT[0], bits 4 - 6 are ignored.
  42          * Ambiguous assignment results in no rendering (no DATA_ERROR).
  43          */
  44         so_method(so, tesla, 0x121c, 1);
  45         so_data  (so, fb->nr_cbufs |
  46                   (0 <<  4) | (1 <<  7) | (2 << 10) | (3 << 13) |
  47                   (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25));
  48
  49         for (i = 0; i < fb->nr_cbufs; i++) {
  50                 struct pipe_texture *pt = fb->cbufs[i]->texture;
  51                 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
  52
  53                 if (!gw) {
  54                         w = fb->cbufs[i]->width;
  55                         h = fb->cbufs[i]->height;
  56                         gw = 1;
  57                 } else {
  58                         assert(w == fb->cbufs[i]->width);
  59                         assert(h == fb->cbufs[i]->height);
  60                 }
  61
  62                 so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
  63                 so_data  (so, fb->cbufs[i]->width);
  64                 so_data  (so, fb->cbufs[i]->height);
  65
  66                 so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
  67                 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
  68                               NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
  69                 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM |
  70                               NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
  71                 switch (fb->cbufs[i]->format) {
  72                 NV50_CBUF_FORMAT_CASE(A8R8G8B8_UNORM);
  73                 NV50_CBUF_FORMAT_CASE(X8R8G8B8_UNORM);
  74                 NV50_CBUF_FORMAT_CASE(R5G6B5_UNORM);
  75                 NV50_CBUF_FORMAT_CASE(R16G16B16A16_SNORM);
  76                 NV50_CBUF_FORMAT_CASE(R16G16B16A16_UNORM);
  77                 NV50_CBUF_FORMAT_CASE(R32G32B32A32_FLOAT);
  78                 NV50_CBUF_FORMAT_CASE(R16G16_SNORM);
  79                 NV50_CBUF_FORMAT_CASE(R16G16_UNORM);
  80                 default:
  81                         NOUVEAU_ERR("AIIII unknown format %s\n",
  82                                     pf_name(fb->cbufs[i]->format));
  83                         so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM);
  84                         break;
  85                 }
  86                 so_data(so, nv50_miptree(pt)->
  87                                 level[fb->cbufs[i]->level].tile_mode << 4);
  88                 so_data(so, 0x00000000);
  89
  90                 so_method(so, tesla, 0x1224, 1);
  91                 so_data  (so, 1);
  92         }
  93
  94         if (fb->zsbuf) {
  95                 struct pipe_texture *pt = fb->zsbuf->texture;
  96                 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo;
  97
  98                 if (!gw) {
  99                         w = fb->zsbuf->width;
 100                         h = fb->zsbuf->height;
 101                         gw = 1;
 102                 } else {
 103                         assert(w == fb->zsbuf->width);
 104                         assert(h == fb->zsbuf->height);
 105                 }
 106
 107                 so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
 108                 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
 109                               NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0);
 110                 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM |
 111                               NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0);
 112                 switch (fb->zsbuf->format) {
 113                 NV50_ZETA_FORMAT_CASE(S8Z24_UNORM);
 114                 NV50_ZETA_FORMAT_CASE(X8Z24_UNORM);
 115                 NV50_ZETA_FORMAT_CASE(Z24S8_UNORM);
 116                 NV50_ZETA_FORMAT_CASE(Z32_FLOAT);
 117                 default:
 118                         NOUVEAU_ERR("AIIII unknown format %s\n",
 119                                     pf_name(fb->zsbuf->format));
 120                         so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM);
 121                         break;
 122                 }
 123                 so_data(so, nv50_miptree(pt)->
 124                                 level[fb->zsbuf->level].tile_mode << 4);
 125                 so_data(so, 0x00000000);
 126
 127                 so_method(so, tesla, 0x1538, 1);
 128                 so_data  (so, 1);
 129                 so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3);
 130                 so_data  (so, fb->zsbuf->width);
 131                 so_data  (so, fb->zsbuf->height);
 132                 so_data  (so, 0x00010001);
 133         } else {
 134                 so_method(so, tesla, 0x1538, 1);
 135                 so_data  (so, 0);
 136         }
 137
 138         so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ, 2);
 139         so_data  (so, w << 16);
 140         so_data  (so, h << 16);
 141         /* set window lower left corner */
 142         so_method(so, tesla, NV50TCL_WINDOW_LEFT, 2);
 143         so_data  (so, 0);
 144         so_data  (so, 0);
 145         /* set screen scissor rectangle */
 146         so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2);
 147         so_data  (so, w << 16);
 148         so_data  (so, h << 16);
 149
 150         /* we set scissors to framebuffer size when they're 'turned off' */
 151         nv50->dirty |= NV50_NEW_SCISSOR;
 152         so_ref(NULL, &nv50->state.scissor);
 153
 154         so_ref(so, &nv50->state.fb);
 155         so_ref(NULL, &so);
 156 }
 157
 158 static void
 159 nv50_state_emit(struct nv50_context *nv50)
 160 {
 161         struct nv50_screen *screen = nv50->screen;
 162         struct nouveau_channel *chan = screen->base.channel;
 163
 164         if (nv50->pctx_id != screen->cur_pctx) {
 165                 if (nv50->state.fb)
 166                         nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
 167                 if (nv50->state.blend)
 168                         nv50->state.dirty |= NV50_NEW_BLEND;
 169                 if (nv50->state.zsa)
 170                         nv50->state.dirty |= NV50_NEW_ZSA;
 171                 if (nv50->state.vertprog)
 172                         nv50->state.dirty |= NV50_NEW_VERTPROG;
 173                 if (nv50->state.fragprog)
 174                         nv50->state.dirty |= NV50_NEW_FRAGPROG;
 175                 if (nv50->state.rast)
 176                         nv50->state.dirty |= NV50_NEW_RASTERIZER;
 177                 if (nv50->state.blend_colour)
 178                         nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
 179                 if (nv50->state.stipple)
 180                         nv50->state.dirty |= NV50_NEW_STIPPLE;
 181                 if (nv50->state.scissor)
 182                         nv50->state.dirty |= NV50_NEW_SCISSOR;
 183                 if (nv50->state.viewport)
 184                         nv50->state.dirty |= NV50_NEW_VIEWPORT;
 185                 if (nv50->state.tsc_upload)
 186                         nv50->state.dirty |= NV50_NEW_SAMPLER;
 187                 if (nv50->state.tic_upload)
 188                         nv50->state.dirty |= NV50_NEW_TEXTURE;
 189                 if (nv50->state.vtxfmt && nv50->state.vtxbuf)
 190                         nv50->state.dirty |= NV50_NEW_ARRAYS;
 191                 screen->cur_pctx = nv50->pctx_id;
 192         }
 193
 194         if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
 195                 so_emit(chan, nv50->state.fb);
 196         if (nv50->state.dirty & NV50_NEW_BLEND)
 197                 so_emit(chan, nv50->state.blend);
 198         if (nv50->state.dirty & NV50_NEW_ZSA)
 199                 so_emit(chan, nv50->state.zsa);
 200         if (nv50->state.dirty & NV50_NEW_VERTPROG)
 201                 so_emit(chan, nv50->state.vertprog);
 202         if (nv50->state.dirty & NV50_NEW_FRAGPROG)
 203                 so_emit(chan, nv50->state.fragprog);
 204         if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
 205                                  NV50_NEW_RASTERIZER))
 206                 so_emit(chan, nv50->state.programs);
 207         if (nv50->state.dirty & NV50_NEW_RASTERIZER)
 208                 so_emit(chan, nv50->state.rast);
 209         if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
 210                 so_emit(chan, nv50->state.blend_colour);
 211         if (nv50->state.dirty & NV50_NEW_STIPPLE)
 212                 so_emit(chan, nv50->state.stipple);
 213         if (nv50->state.dirty & NV50_NEW_SCISSOR)
 214                 so_emit(chan, nv50->state.scissor);
 215         if (nv50->state.dirty & NV50_NEW_VIEWPORT)
 216                 so_emit(chan, nv50->state.viewport);
 217         if (nv50->state.dirty & NV50_NEW_SAMPLER)
 218                 so_emit(chan, nv50->state.tsc_upload);
 219         if (nv50->state.dirty & NV50_NEW_TEXTURE)
 220                 so_emit(chan, nv50->state.tic_upload);
 221         if (nv50->state.dirty & NV50_NEW_ARRAYS) {
 222                 so_emit(chan, nv50->state.vtxfmt);
 223                 so_emit(chan, nv50->state.vtxbuf);
 224                 if (nv50->state.vtxattr)
 225                         so_emit(chan, nv50->state.vtxattr);
 226         }
 227         nv50->state.dirty = 0;
 228 }
 229
 230 void
 231 nv50_state_flush_notify(struct nouveau_channel *chan)
 232 {
 233         struct nv50_context *nv50 = chan->user_private;
 234
 235         if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
 236                 so_emit(chan, nv50->state.tic_upload);
 237
 238         so_emit_reloc_markers(chan, nv50->state.fb);
 239         so_emit_reloc_markers(chan, nv50->state.vertprog);
 240         so_emit_reloc_markers(chan, nv50->state.fragprog);
 241         so_emit_reloc_markers(chan, nv50->state.vtxbuf);
 242         so_emit_reloc_markers(chan, nv50->screen->static_init);
 243 }
 244
 245 boolean
 246 nv50_state_validate(struct nv50_context *nv50)
 247 {
 248         struct nouveau_grobj *tesla = nv50->screen->tesla;
 249         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 250         struct nouveau_stateobj *so;
 251         unsigned i;
 252
 253         if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
 254                 nv50_state_validate_fb(nv50);
 255
 256         if (nv50->dirty & NV50_NEW_BLEND)
 257                 so_ref(nv50->blend->so, &nv50->state.blend);
 258
 259         if (nv50->dirty & NV50_NEW_ZSA)
 260                 so_ref(nv50->zsa->so, &nv50->state.zsa);
 261
 262         if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
 263                 nv50_vertprog_validate(nv50);
 264
 265         if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
 266                 nv50_fragprog_validate(nv50);
 267
 268         if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
 269                            NV50_NEW_RASTERIZER))
 270                 nv50_linkage_validate(nv50);
 271
 272         if (nv50->dirty & NV50_NEW_RASTERIZER)
 273                 so_ref(nv50->rasterizer->so, &nv50->state.rast);
 274
 275         if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
 276                 so = so_new(5, 0);
 277                 so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
 278                 so_data  (so, fui(nv50->blend_colour.color[0]));
 279                 so_data  (so, fui(nv50->blend_colour.color[1]));
 280                 so_data  (so, fui(nv50->blend_colour.color[2]));
 281                 so_data  (so, fui(nv50->blend_colour.color[3]));
 282                 so_ref(so, &nv50->state.blend_colour);
 283                 so_ref(NULL, &so);
 284         }
 285
 286         if (nv50->dirty & NV50_NEW_STIPPLE) {
 287                 so = so_new(33, 0);
 288                 so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
 289                 for (i = 0; i < 32; i++)
 290                         so_data(so, util_bswap32(nv50->stipple.stipple[i]));
 291                 so_ref(so, &nv50->state.stipple);
 292                 so_ref(NULL, &so);
 293         }
 294
 295         if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
 296                 struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
 297                 struct pipe_scissor_state *s = &nv50->scissor;
 298
 299                 if (nv50->state.scissor &&
 300                     (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
 301                         goto scissor_uptodate;
 302                 nv50->state.scissor_enabled = rast->scissor;
 303
 304                 so = so_new(3, 0);
 305                 so_method(so, tesla, NV50TCL_SCISSOR_HORIZ, 2);
 306                 if (nv50->state.scissor_enabled) {
 307                         so_data(so, (s->maxx << 16) | s->minx);
 308                         so_data(so, (s->maxy << 16) | s->miny);
 309                 } else {
 310                         so_data(so, (nv50->framebuffer.width << 16));
 311                         so_data(so, (nv50->framebuffer.height << 16));
 312                 }
 313                 so_ref(so, &nv50->state.scissor);
 314                 so_ref(NULL, &so);
 315                 nv50->state.dirty |= NV50_NEW_SCISSOR;
 316         }
 317 scissor_uptodate:
 318
 319         if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
 320                 unsigned bypass;
 321
 322                 if (!nv50->rasterizer->pipe.bypass_vs_clip_and_viewport)
 323                         bypass = 0;
 324                 else
 325                         bypass = 1;
 326
 327                 if (nv50->state.viewport &&
 328                     (bypass || !(nv50->dirty & NV50_NEW_VIEWPORT)) &&
 329                     nv50->state.viewport_bypass == bypass)
 330                         goto viewport_uptodate;
 331                 nv50->state.viewport_bypass = bypass;
 332
 333                 so = so_new(14, 0);
 334                 if (!bypass) {
 335                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE(0), 3);
 336                         so_data  (so, fui(nv50->viewport.translate[0]));
 337                         so_data  (so, fui(nv50->viewport.translate[1]));
 338                         so_data  (so, fui(nv50->viewport.translate[2]));
 339                         so_method(so, tesla, NV50TCL_VIEWPORT_SCALE(0), 3);
 340                         so_data  (so, fui(nv50->viewport.scale[0]));
 341                         so_data  (so, fui(nv50->viewport.scale[1]));
 342                         so_data  (so, fui(nv50->viewport.scale[2]));
 343
 344                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
 345                         so_data  (so, 1);
 346                         /* 0x0000 = remove whole primitive only (xyz)
 347                          * 0x1018 = remove whole primitive only (xy), clamp z
 348                          * 0x1080 = clip primitive (xyz)
 349                          * 0x1098 = clip primitive (xy), clamp z
 350                          */
 351                         so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
 352                         so_data  (so, 0x1080);
 353                         /* no idea what 0f90 does */
 354                         so_method(so, tesla, 0x0f90, 1);
 355                         so_data  (so, 0);
 356                 } else {
 357                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
 358                         so_data  (so, 0);
 359                         so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
 360                         so_data  (so, 0x0000);
 361                         so_method(so, tesla, 0x0f90, 1);
 362                         so_data  (so, 1);
 363                 }
 364
 365                 so_ref(so, &nv50->state.viewport);
 366                 so_ref(NULL, &so);
 367                 nv50->state.dirty |= NV50_NEW_VIEWPORT;
 368         }
 369 viewport_uptodate:
 370
 371         if (nv50->dirty & NV50_NEW_SAMPLER) {
 372                 unsigned i;
 373
 374                 so = so_new(nv50->sampler_nr * 9 + 23 + 4, 2);
 375
 376                 nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
 377                                   nv50->sampler_nr * 8 * 4);
 378
 379                 for (i = 0; i < nv50->sampler_nr; i++) {
 380                         if (!nv50->sampler[i])
 381                                 continue;
 382                         so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), 8);
 383                         so_datap (so, nv50->sampler[i]->tsc, 8);
 384                 }
 385
 386                 so_method(so, tesla, 0x1440, 1); /* sync SIFC */
 387                 so_data  (so, 0);
 388                 so_method(so, tesla, 0x1334, 1); /* flush TSC */
 389                 so_data  (so, 0);
 390
 391                 so_ref(so, &nv50->state.tsc_upload);
 392                 so_ref(NULL, &so);
 393         }
 394
 395         if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
 396                 nv50_tex_validate(nv50);
 397
 398         if (nv50->dirty & NV50_NEW_ARRAYS)
 399                 nv50_vbo_validate(nv50);
 400
 401         nv50->state.dirty |= nv50->dirty;
 402         nv50->dirty = 0;
 403         nv50_state_emit(nv50);
 404
 405         return TRUE;
 406 }
 407
 408 void nv50_so_init_sifc(struct nv50_context *nv50,
 409                        struct nouveau_stateobj *so,
 410                        struct nouveau_bo *bo, unsigned reloc, unsigned size)
 411 {
 412         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 413
 414         so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
 415         so_data  (so, NV50_2D_DST_FORMAT_R8_UNORM);
 416         so_data  (so, 1);
 417         so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
 418         so_data  (so, 262144);
 419         so_data  (so, 65536);
 420         so_data  (so, 1);
 421         so_reloc (so, bo, 0, reloc | NOUVEAU_BO_WR | NOUVEAU_BO_HIGH, 0, 0);
 422         so_reloc (so, bo, 0, reloc | NOUVEAU_BO_WR | NOUVEAU_BO_LOW, 0, 0);
 423         so_method(so, eng2d, NV50_2D_SIFC_UNK0800, 2);
 424         so_data  (so, 0);
 425         so_data  (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
 426         so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
 427         so_data  (so, size);
 428         so_data  (so, 1);
 429         so_data  (so, 0);
 430         so_data  (so, 1);
 431         so_data  (so, 0);
 432         so_data  (so, 1);
 433         so_data  (so, 0);
 434         so_data  (so, 0);
 435         so_data  (so, 0);
 436         so_data  (so, 0);
 437 }