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, NV50TCL_RT_CONTROL, 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, NV50TCL_RT_ARRAY_MODE, 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, NV50TCL_ZETA_ENABLE, 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, NV50TCL_ZETA_ENABLE, 1);
 135                 so_data  (so, 0);
 136         }
 137
 138         so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 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_OFFSET_X, 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_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
 160                        unsigned p)
 161 {
 162         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 163         unsigned i, j, dw = nv50->sampler_nr[p] * 8;
 164
 165         if (!dw)
 166                 return;
 167         nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM,
 168                           p * (32 * 8 * 4), dw * 4);
 169
 170         so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw);
 171
 172         for (i = 0; i < nv50->sampler_nr[p]; ++i) {
 173                 if (nv50->sampler[p][i])
 174                         so_datap(so, nv50->sampler[p][i]->tsc, 8);
 175                 else {
 176                         for (j = 0; j < 8; ++j) /* you get punished */
 177                                 so_data(so, 0); /* ... for leaving holes */
 178                 }
 179         }
 180 }
 181
 182 static void
 183 nv50_state_emit(struct nv50_context *nv50)
 184 {
 185         struct nv50_screen *screen = nv50->screen;
 186         struct nouveau_channel *chan = screen->base.channel;
 187
 188         if (nv50->pctx_id != screen->cur_pctx) {
 189                 if (nv50->state.fb)
 190                         nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
 191                 if (nv50->state.blend)
 192                         nv50->state.dirty |= NV50_NEW_BLEND;
 193                 if (nv50->state.zsa)
 194                         nv50->state.dirty |= NV50_NEW_ZSA;
 195                 if (nv50->state.vertprog)
 196                         nv50->state.dirty |= NV50_NEW_VERTPROG;
 197                 if (nv50->state.fragprog)
 198                         nv50->state.dirty |= NV50_NEW_FRAGPROG;
 199                 if (nv50->state.rast)
 200                         nv50->state.dirty |= NV50_NEW_RASTERIZER;
 201                 if (nv50->state.blend_colour)
 202                         nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
 203                 if (nv50->state.stipple)
 204                         nv50->state.dirty |= NV50_NEW_STIPPLE;
 205                 if (nv50->state.scissor)
 206                         nv50->state.dirty |= NV50_NEW_SCISSOR;
 207                 if (nv50->state.viewport)
 208                         nv50->state.dirty |= NV50_NEW_VIEWPORT;
 209                 if (nv50->state.tsc_upload)
 210                         nv50->state.dirty |= NV50_NEW_SAMPLER;
 211                 if (nv50->state.tic_upload)
 212                         nv50->state.dirty |= NV50_NEW_TEXTURE;
 213                 if (nv50->state.vtxfmt && nv50->state.vtxbuf)
 214                         nv50->state.dirty |= NV50_NEW_ARRAYS;
 215                 screen->cur_pctx = nv50->pctx_id;
 216         }
 217
 218         if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
 219                 so_emit(chan, nv50->state.fb);
 220         if (nv50->state.dirty & NV50_NEW_BLEND)
 221                 so_emit(chan, nv50->state.blend);
 222         if (nv50->state.dirty & NV50_NEW_ZSA)
 223                 so_emit(chan, nv50->state.zsa);
 224         if (nv50->state.dirty & NV50_NEW_VERTPROG)
 225                 so_emit(chan, nv50->state.vertprog);
 226         if (nv50->state.dirty & NV50_NEW_FRAGPROG)
 227                 so_emit(chan, nv50->state.fragprog);
 228         if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
 229                                  NV50_NEW_RASTERIZER))
 230                 so_emit(chan, nv50->state.programs);
 231         if (nv50->state.dirty & NV50_NEW_RASTERIZER)
 232                 so_emit(chan, nv50->state.rast);
 233         if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
 234                 so_emit(chan, nv50->state.blend_colour);
 235         if (nv50->state.dirty & NV50_NEW_STIPPLE)
 236                 so_emit(chan, nv50->state.stipple);
 237         if (nv50->state.dirty & NV50_NEW_SCISSOR)
 238                 so_emit(chan, nv50->state.scissor);
 239         if (nv50->state.dirty & NV50_NEW_VIEWPORT)
 240                 so_emit(chan, nv50->state.viewport);
 241         if (nv50->state.dirty & NV50_NEW_SAMPLER)
 242                 so_emit(chan, nv50->state.tsc_upload);
 243         if (nv50->state.dirty & NV50_NEW_TEXTURE)
 244                 so_emit(chan, nv50->state.tic_upload);
 245         if (nv50->state.dirty & NV50_NEW_ARRAYS) {
 246                 so_emit(chan, nv50->state.vtxfmt);
 247                 so_emit(chan, nv50->state.vtxbuf);
 248                 if (nv50->state.vtxattr)
 249                         so_emit(chan, nv50->state.vtxattr);
 250         }
 251         nv50->state.dirty = 0;
 252 }
 253
 254 void
 255 nv50_state_flush_notify(struct nouveau_channel *chan)
 256 {
 257         struct nv50_context *nv50 = chan->user_private;
 258
 259         if (nv50->state.tic_upload && !(nv50->dirty & NV50_NEW_TEXTURE))
 260                 so_emit(chan, nv50->state.tic_upload);
 261
 262         so_emit_reloc_markers(chan, nv50->state.fb);
 263         so_emit_reloc_markers(chan, nv50->state.vertprog);
 264         so_emit_reloc_markers(chan, nv50->state.fragprog);
 265         so_emit_reloc_markers(chan, nv50->state.vtxbuf);
 266         so_emit_reloc_markers(chan, nv50->screen->static_init);
 267 }
 268
 269 boolean
 270 nv50_state_validate(struct nv50_context *nv50)
 271 {
 272         struct nouveau_grobj *tesla = nv50->screen->tesla;
 273         struct nouveau_stateobj *so;
 274         unsigned i;
 275
 276         if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
 277                 nv50_state_validate_fb(nv50);
 278
 279         if (nv50->dirty & NV50_NEW_BLEND)
 280                 so_ref(nv50->blend->so, &nv50->state.blend);
 281
 282         if (nv50->dirty & NV50_NEW_ZSA)
 283                 so_ref(nv50->zsa->so, &nv50->state.zsa);
 284
 285         if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
 286                 nv50_vertprog_validate(nv50);
 287
 288         if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
 289                 nv50_fragprog_validate(nv50);
 290
 291         if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
 292                            NV50_NEW_RASTERIZER))
 293                 nv50_linkage_validate(nv50);
 294
 295         if (nv50->dirty & NV50_NEW_RASTERIZER)
 296                 so_ref(nv50->rasterizer->so, &nv50->state.rast);
 297
 298         if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
 299                 so = so_new(5, 0);
 300                 so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
 301                 so_data  (so, fui(nv50->blend_colour.color[0]));
 302                 so_data  (so, fui(nv50->blend_colour.color[1]));
 303                 so_data  (so, fui(nv50->blend_colour.color[2]));
 304                 so_data  (so, fui(nv50->blend_colour.color[3]));
 305                 so_ref(so, &nv50->state.blend_colour);
 306                 so_ref(NULL, &so);
 307         }
 308
 309         if (nv50->dirty & NV50_NEW_STIPPLE) {
 310                 so = so_new(33, 0);
 311                 so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
 312                 for (i = 0; i < 32; i++)
 313                         so_data(so, util_bswap32(nv50->stipple.stipple[i]));
 314                 so_ref(so, &nv50->state.stipple);
 315                 so_ref(NULL, &so);
 316         }
 317
 318         if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
 319                 struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
 320                 struct pipe_scissor_state *s = &nv50->scissor;
 321
 322                 if (nv50->state.scissor &&
 323                     (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
 324                         goto scissor_uptodate;
 325                 nv50->state.scissor_enabled = rast->scissor;
 326
 327                 so = so_new(3, 0);
 328                 so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
 329                 if (nv50->state.scissor_enabled) {
 330                         so_data(so, (s->maxx << 16) | s->minx);
 331                         so_data(so, (s->maxy << 16) | s->miny);
 332                 } else {
 333                         so_data(so, (nv50->framebuffer.width << 16));
 334                         so_data(so, (nv50->framebuffer.height << 16));
 335                 }
 336                 so_ref(so, &nv50->state.scissor);
 337                 so_ref(NULL, &so);
 338                 nv50->state.dirty |= NV50_NEW_SCISSOR;
 339         }
 340 scissor_uptodate:
 341
 342         if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
 343                 unsigned bypass;
 344
 345                 if (!nv50->rasterizer->pipe.bypass_vs_clip_and_viewport)
 346                         bypass = 0;
 347                 else
 348                         bypass = 1;
 349
 350                 if (nv50->state.viewport &&
 351                     (bypass || !(nv50->dirty & NV50_NEW_VIEWPORT)) &&
 352                     nv50->state.viewport_bypass == bypass)
 353                         goto viewport_uptodate;
 354                 nv50->state.viewport_bypass = bypass;
 355
 356                 so = so_new(14, 0);
 357                 if (!bypass) {
 358                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
 359                         so_data  (so, fui(nv50->viewport.translate[0]));
 360                         so_data  (so, fui(nv50->viewport.translate[1]));
 361                         so_data  (so, fui(nv50->viewport.translate[2]));
 362                         so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
 363                         so_data  (so, fui(nv50->viewport.scale[0]));
 364                         so_data  (so, fui(nv50->viewport.scale[1]));
 365                         so_data  (so, fui(nv50->viewport.scale[2]));
 366
 367                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
 368                         so_data  (so, 1);
 369                         /* 0x0000 = remove whole primitive only (xyz)
 370                          * 0x1018 = remove whole primitive only (xy), clamp z
 371                          * 0x1080 = clip primitive (xyz)
 372                          * 0x1098 = clip primitive (xy), clamp z
 373                          */
 374                         so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
 375                         so_data  (so, 0x1080);
 376                         /* no idea what 0f90 does */
 377                         so_method(so, tesla, 0x0f90, 1);
 378                         so_data  (so, 0);
 379                 } else {
 380                         so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
 381                         so_data  (so, 0);
 382                         so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
 383                         so_data  (so, 0x0000);
 384                         so_method(so, tesla, 0x0f90, 1);
 385                         so_data  (so, 1);
 386                 }
 387
 388                 so_ref(so, &nv50->state.viewport);
 389                 so_ref(NULL, &so);
 390                 nv50->state.dirty |= NV50_NEW_VIEWPORT;
 391         }
 392 viewport_uptodate:
 393
 394         if (nv50->dirty & NV50_NEW_SAMPLER) {
 395                 unsigned nr = 0;
 396
 397                 for (i = 0; i < PIPE_SHADER_TYPES; ++i)
 398                         nr += nv50->sampler_nr[i];
 399
 400                 so = so_new(nr * 8 + 24 * PIPE_SHADER_TYPES + 2, 4);
 401
 402                 nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
 403                 nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
 404
 405                 so_method(so, tesla, 0x1334, 1); /* flush TSC */
 406                 so_data  (so, 0);
 407
 408                 so_ref(so, &nv50->state.tsc_upload);
 409                 so_ref(NULL, &so);
 410         }
 411
 412         if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
 413                 nv50_tex_validate(nv50);
 414
 415         if (nv50->dirty & NV50_NEW_ARRAYS)
 416                 nv50_vbo_validate(nv50);
 417
 418         nv50->state.dirty |= nv50->dirty;
 419         nv50->dirty = 0;
 420         nv50_state_emit(nv50);
 421
 422         return TRUE;
 423 }
 424
 425 void nv50_so_init_sifc(struct nv50_context *nv50,
 426                        struct nouveau_stateobj *so,
 427                        struct nouveau_bo *bo, unsigned reloc,
 428                        unsigned offset, unsigned size)
 429 {
 430         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 431
 432         reloc |= NOUVEAU_BO_WR;
 433
 434         so_method(so, eng2d, NV50_2D_DST_FORMAT, 2);
 435         so_data  (so, NV50_2D_DST_FORMAT_R8_UNORM);
 436         so_data  (so, 1);
 437         so_method(so, eng2d, NV50_2D_DST_PITCH, 5);
 438         so_data  (so, 262144);
 439         so_data  (so, 65536);
 440         so_data  (so, 1);
 441         so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0);
 442         so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0);
 443         so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
 444         so_data  (so, 0);
 445         so_data  (so, NV50_2D_SIFC_FORMAT_R8_UNORM);
 446         so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10);
 447         so_data  (so, size);
 448         so_data  (so, 1);
 449         so_data  (so, 0);
 450         so_data  (so, 1);
 451         so_data  (so, 0);
 452         so_data  (so, 1);
 453         so_data  (so, 0);
 454         so_data  (so, 0);
 455         so_data  (so, 0);
 456         so_data  (so, 0);
 457 }