nv50_state_validate.c revision c926ed0e76bdbf842170f9396856e9cedb6e718f
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 "util/u_format.h" 24 25#include "nv50_context.h" 26#include "nouveau/nouveau_stateobj.h" 27 28static struct nouveau_stateobj * 29validate_fb(struct nv50_context *nv50) 30{ 31 struct nouveau_grobj *tesla = nv50->screen->tesla; 32 struct nouveau_stateobj *so = so_new(32, 79, 18); 33 struct pipe_framebuffer_state *fb = &nv50->framebuffer; 34 unsigned i, w, h, gw = 0; 35 36 /* Set nr of active RTs and select RT for each colour output. 37 * FP result 0 always goes to RT[0], bits 4 - 6 are ignored. 38 * Ambiguous assignment results in no rendering (no DATA_ERROR). 39 */ 40 so_method(so, tesla, NV50TCL_RT_CONTROL, 1); 41 so_data (so, fb->nr_cbufs | 42 (0 << 4) | (1 << 7) | (2 << 10) | (3 << 13) | 43 (4 << 16) | (5 << 19) | (6 << 22) | (7 << 25)); 44 45 for (i = 0; i < fb->nr_cbufs; i++) { 46 struct pipe_texture *pt = fb->cbufs[i]->texture; 47 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo; 48 49 if (!gw) { 50 w = fb->cbufs[i]->width; 51 h = fb->cbufs[i]->height; 52 gw = 1; 53 } else { 54 assert(w == fb->cbufs[i]->width); 55 assert(h == fb->cbufs[i]->height); 56 } 57 58 so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2); 59 so_data (so, fb->cbufs[i]->width); 60 so_data (so, fb->cbufs[i]->height); 61 62 so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5); 63 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM | 64 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0); 65 so_reloc (so, bo, fb->cbufs[i]->offset, NOUVEAU_BO_VRAM | 66 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0); 67 switch (fb->cbufs[i]->format) { 68 case PIPE_FORMAT_B8G8R8A8_UNORM: 69 so_data(so, NV50TCL_RT_FORMAT_A8R8G8B8_UNORM); 70 break; 71 case PIPE_FORMAT_B8G8R8X8_UNORM: 72 so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM); 73 break; 74 case PIPE_FORMAT_B5G6R5_UNORM: 75 so_data(so, NV50TCL_RT_FORMAT_R5G6B5_UNORM); 76 break; 77 case PIPE_FORMAT_R16G16B16A16_SNORM: 78 so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_SNORM); 79 break; 80 case PIPE_FORMAT_R16G16B16A16_UNORM: 81 so_data(so, NV50TCL_RT_FORMAT_R16G16B16A16_UNORM); 82 break; 83 case PIPE_FORMAT_R32G32B32A32_FLOAT: 84 so_data(so, NV50TCL_RT_FORMAT_R32G32B32A32_FLOAT); 85 break; 86 case PIPE_FORMAT_R16G16_SNORM: 87 so_data(so, NV50TCL_RT_FORMAT_R16G16_SNORM); 88 break; 89 case PIPE_FORMAT_R16G16_UNORM: 90 so_data(so, NV50TCL_RT_FORMAT_R16G16_UNORM); 91 break; 92 default: 93 NOUVEAU_ERR("AIIII unknown format %s\n", 94 util_format_name(fb->cbufs[i]->format)); 95 so_data(so, NV50TCL_RT_FORMAT_X8R8G8B8_UNORM); 96 break; 97 } 98 so_data(so, nv50_miptree(pt)-> 99 level[fb->cbufs[i]->level].tile_mode << 4); 100 so_data(so, 0x00000000); 101 102 so_method(so, tesla, NV50TCL_RT_ARRAY_MODE, 1); 103 so_data (so, 1); 104 } 105 106 if (fb->zsbuf) { 107 struct pipe_texture *pt = fb->zsbuf->texture; 108 struct nouveau_bo *bo = nv50_miptree(pt)->base.bo; 109 110 if (!gw) { 111 w = fb->zsbuf->width; 112 h = fb->zsbuf->height; 113 gw = 1; 114 } else { 115 assert(w == fb->zsbuf->width); 116 assert(h == fb->zsbuf->height); 117 } 118 119 so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5); 120 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM | 121 NOUVEAU_BO_HIGH | NOUVEAU_BO_RDWR, 0, 0); 122 so_reloc (so, bo, fb->zsbuf->offset, NOUVEAU_BO_VRAM | 123 NOUVEAU_BO_LOW | NOUVEAU_BO_RDWR, 0, 0); 124 switch (fb->zsbuf->format) { 125 case PIPE_FORMAT_Z24S8_UNORM: 126 so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM); 127 break; 128 case PIPE_FORMAT_Z24X8_UNORM: 129 so_data(so, NV50TCL_ZETA_FORMAT_X8Z24_UNORM); 130 break; 131 case PIPE_FORMAT_S8Z24_UNORM: 132 so_data(so, NV50TCL_ZETA_FORMAT_Z24S8_UNORM); 133 break; 134 case PIPE_FORMAT_Z32_FLOAT: 135 so_data(so, NV50TCL_ZETA_FORMAT_Z32_FLOAT); 136 break; 137 default: 138 NOUVEAU_ERR("AIIII unknown format %s\n", 139 util_format_name(fb->zsbuf->format)); 140 so_data(so, NV50TCL_ZETA_FORMAT_S8Z24_UNORM); 141 break; 142 } 143 so_data(so, nv50_miptree(pt)-> 144 level[fb->zsbuf->level].tile_mode << 4); 145 so_data(so, 0x00000000); 146 147 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1); 148 so_data (so, 1); 149 so_method(so, tesla, NV50TCL_ZETA_HORIZ, 3); 150 so_data (so, fb->zsbuf->width); 151 so_data (so, fb->zsbuf->height); 152 so_data (so, 0x00010001); 153 } else { 154 so_method(so, tesla, NV50TCL_ZETA_ENABLE, 1); 155 so_data (so, 0); 156 } 157 158 so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ(0), 2); 159 so_data (so, w << 16); 160 so_data (so, h << 16); 161 /* set window lower left corner */ 162 so_method(so, tesla, NV50TCL_WINDOW_OFFSET_X, 2); 163 so_data (so, 0); 164 so_data (so, 0); 165 /* set screen scissor rectangle */ 166 so_method(so, tesla, NV50TCL_SCREEN_SCISSOR_HORIZ, 2); 167 so_data (so, w << 16); 168 so_data (so, h << 16); 169 170 /* we set scissors to framebuffer size when they're 'turned off' */ 171 nv50->dirty |= NV50_NEW_SCISSOR; 172 so_ref(NULL, &nv50->state.hw[12]); 173 return so; 174} 175 176static void 177nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so, 178 unsigned p) 179{ 180 struct nouveau_grobj *eng2d = nv50->screen->eng2d; 181 unsigned i, j, dw = nv50->sampler_nr[p] * 8; 182 183 if (!dw) 184 return; 185 nv50_so_init_sifc(nv50, so, nv50->screen->tsc, NOUVEAU_BO_VRAM, 186 p * (32 * 8 * 4), dw * 4); 187 188 so_method(so, eng2d, NV50_2D_SIFC_DATA | (2 << 29), dw); 189 190 for (i = 0; i < nv50->sampler_nr[p]; ++i) { 191 if (nv50->sampler[p][i]) 192 so_datap(so, nv50->sampler[p][i]->tsc, 8); 193 else { 194 for (j = 0; j < 8; ++j) /* you get punished */ 195 so_data(so, 0); /* ... for leaving holes */ 196 } 197 } 198} 199 200static struct nouveau_stateobj * 201validate_blend(struct nv50_context *nv50) 202{ 203 struct nouveau_stateobj *so = NULL; 204 so_ref(nv50->blend->so, &so); 205 return so; 206} 207 208static struct nouveau_stateobj * 209validate_zsa(struct nv50_context *nv50) 210{ 211 struct nouveau_stateobj *so = NULL; 212 so_ref(nv50->zsa->so, &so); 213 return so; 214} 215 216static struct nouveau_stateobj * 217validate_rast(struct nv50_context *nv50) 218{ 219 struct nouveau_stateobj *so = NULL; 220 so_ref(nv50->rasterizer->so, &so); 221 return so; 222} 223 224static struct nouveau_stateobj * 225validate_blend_colour(struct nv50_context *nv50) 226{ 227 struct nouveau_grobj *tesla = nv50->screen->tesla; 228 struct nouveau_stateobj *so = so_new(1, 4, 0); 229 230 so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4); 231 so_data (so, fui(nv50->blend_colour.color[0])); 232 so_data (so, fui(nv50->blend_colour.color[1])); 233 so_data (so, fui(nv50->blend_colour.color[2])); 234 so_data (so, fui(nv50->blend_colour.color[3])); 235 return so; 236} 237 238static struct nouveau_stateobj * 239validate_stencil_ref(struct nv50_context *nv50) 240{ 241 struct nouveau_grobj *tesla = nv50->screen->tesla; 242 struct nouveau_stateobj *so = so = so_new(2, 2, 0); 243 244 so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1); 245 so_data (so, nv50->stencil_ref.ref_value[0]); 246 so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1); 247 so_data (so, nv50->stencil_ref.ref_value[1]); 248 return so; 249} 250 251static struct nouveau_stateobj * 252validate_stipple(struct nv50_context *nv50) 253{ 254 struct nouveau_grobj *tesla = nv50->screen->tesla; 255 struct nouveau_stateobj *so = so_new(1, 32, 0); 256 int i; 257 258 so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32); 259 for (i = 0; i < 32; i++) 260 so_data(so, util_bswap32(nv50->stipple.stipple[i])); 261 return so; 262} 263 264static struct nouveau_stateobj * 265validate_scissor(struct nv50_context *nv50) 266{ 267 struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe; 268 struct pipe_scissor_state *s = &nv50->scissor; 269 struct nouveau_grobj *tesla = nv50->screen->tesla; 270 struct nouveau_stateobj *so; 271 272 if (nv50->state.hw[12] && 273 (rast->scissor == 0 && nv50->state.scissor_enabled == 0)) 274 return NULL; 275 nv50->state.scissor_enabled = rast->scissor; 276 277 so = so_new(1, 2, 0); 278 so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2); 279 if (nv50->state.scissor_enabled) { 280 so_data(so, (s->maxx << 16) | s->minx); 281 so_data(so, (s->maxy << 16) | s->miny); 282 } else { 283 so_data(so, (nv50->framebuffer.width << 16)); 284 so_data(so, (nv50->framebuffer.height << 16)); 285 } 286 287 return so; 288} 289 290static struct nouveau_stateobj * 291validate_viewport(struct nv50_context *nv50) 292{ 293 struct nouveau_grobj *tesla = nv50->screen->tesla; 294 struct nouveau_stateobj *so = so_new(5, 9, 0); 295 296 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3); 297 so_data (so, fui(nv50->viewport.translate[0])); 298 so_data (so, fui(nv50->viewport.translate[1])); 299 so_data (so, fui(nv50->viewport.translate[2])); 300 so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3); 301 so_data (so, fui(nv50->viewport.scale[0])); 302 so_data (so, fui(nv50->viewport.scale[1])); 303 so_data (so, fui(nv50->viewport.scale[2])); 304 305 so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1); 306 so_data (so, 1); 307 /* 0x0000 = remove whole primitive only (xyz) 308 * 0x1018 = remove whole primitive only (xy), clamp z 309 * 0x1080 = clip primitive (xyz) 310 * 0x1098 = clip primitive (xy), clamp z 311 */ 312 so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1); 313 so_data (so, 0x1080); 314 /* no idea what 0f90 does */ 315 so_method(so, tesla, 0x0f90, 1); 316 so_data (so, 0); 317 318 return so; 319} 320 321static struct nouveau_stateobj * 322validate_sampler(struct nv50_context *nv50) 323{ 324 struct nouveau_grobj *tesla = nv50->screen->tesla; 325 struct nouveau_stateobj *so; 326 unsigned nr = 0, i; 327 328 for (i = 0; i < PIPE_SHADER_TYPES; ++i) 329 nr += nv50->sampler_nr[i]; 330 331 so = so_new(1 + 5 * PIPE_SHADER_TYPES, 332 1 + 19 * PIPE_SHADER_TYPES + nr * 8, 333 PIPE_SHADER_TYPES * 2); 334 335 nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX); 336 nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT); 337 338 so_method(so, tesla, 0x1334, 1); /* flush TSC */ 339 so_data (so, 0); 340 341 return so; 342} 343 344static struct nouveau_stateobj * 345validate_vtxbuf(struct nv50_context *nv50) 346{ 347 struct nouveau_stateobj *so = NULL; 348 so_ref(nv50->state.vtxbuf, &so); 349 return so; 350} 351 352static struct nouveau_stateobj * 353validate_vtxattr(struct nv50_context *nv50) 354{ 355 struct nouveau_stateobj *so = NULL; 356 so_ref(nv50->state.vtxattr, &so); 357 return so; 358} 359 360struct state_validate { 361 struct nouveau_stateobj *(*func)(struct nv50_context *nv50); 362 unsigned states; 363} validate_list[] = { 364 { validate_fb , NV50_NEW_FRAMEBUFFER }, 365 { validate_blend , NV50_NEW_BLEND }, 366 { validate_zsa , NV50_NEW_ZSA }, 367 { nv50_vertprog_validate , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB }, 368 { nv50_fragprog_validate , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB }, 369 { nv50_geomprog_validate , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB }, 370 { nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG | 371 NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER }, 372 { nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG }, 373 { validate_rast , NV50_NEW_RASTERIZER }, 374 { validate_blend_colour , NV50_NEW_BLEND_COLOUR }, 375 { validate_stencil_ref , NV50_NEW_STENCIL_REF }, 376 { validate_stipple , NV50_NEW_STIPPLE }, 377 { validate_scissor , NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER }, 378 { validate_viewport , NV50_NEW_VIEWPORT }, 379 { validate_sampler , NV50_NEW_SAMPLER }, 380 { nv50_tex_validate , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER }, 381 { nv50_vbo_validate , NV50_NEW_ARRAYS }, 382 { validate_vtxbuf , NV50_NEW_ARRAYS }, 383 { validate_vtxattr , NV50_NEW_ARRAYS }, 384 {} 385}; 386#define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0])) 387 388void 389nv50_state_flush_notify(struct nouveau_channel *chan) 390{ 391 struct nv50_context *nv50 = chan->user_private; 392 393 nv50_tex_relocs(nv50); 394 395 so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */ 396 so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */ 397 so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */ 398 so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */ 399 so_emit_reloc_markers(chan, nv50->screen->static_init); 400 401#if 0 402 if (nv50->state.instbuf) 403 so_emit_reloc_markers(chan, nv50->state.instbuf); 404#endif 405} 406 407boolean 408nv50_state_validate(struct nv50_context *nv50, unsigned wait_dwords) 409{ 410 struct nouveau_channel *chan = nv50->screen->base.channel; 411 unsigned nr_relocs = 128, nr_dwords = wait_dwords + 128; 412 int ret, i; 413 414 for (i = 0; i < validate_list_len; i++) { 415 struct state_validate *validate = &validate_list[i]; 416 struct nouveau_stateobj *so; 417 418 if (!(nv50->dirty & validate->states)) 419 continue; 420 421 so = validate->func(nv50); 422 if (!so) 423 continue; 424 425 nr_dwords += (so->total + so->cur); 426 nr_relocs += so->cur_reloc; 427 428 so_ref(so, &nv50->state.hw[i]); 429 so_ref(NULL, &so); 430 nv50->state.hw_dirty |= (1 << i); 431 } 432 nv50->dirty = 0; 433 434 if (nv50->screen->cur_ctx != nv50) { 435 for (i = 0; i < validate_list_len; i++) { 436 if (!nv50->state.hw[i] || 437 (nv50->state.hw_dirty & (1 << i))) 438 continue; 439 440 nr_dwords += (nv50->state.hw[i]->total + 441 nv50->state.hw[i]->cur); 442 nr_relocs += nv50->state.hw[i]->cur_reloc; 443 nv50->state.hw_dirty |= (1 << i); 444 } 445 446 nv50->screen->cur_ctx = nv50; 447 } 448 449 ret = MARK_RING(chan, nr_dwords, nr_relocs); 450 if (ret) { 451 debug_printf("MARK_RING(%d, %d) failed: %d\n", 452 nr_dwords, nr_relocs, ret); 453 return FALSE; 454 } 455 456 while (nv50->state.hw_dirty) { 457 i = ffs(nv50->state.hw_dirty) - 1; 458 nv50->state.hw_dirty &= ~(1 << i); 459 460 so_emit(chan, nv50->state.hw[i]); 461 } 462 463 /* Yes, really, we need to do this. If a buffer that is referenced 464 * on the hardware isn't part of changed state above, without doing 465 * this the kernel is given no clue that the buffer is being used 466 * still. This can cause all sorts of fun issues. 467 */ 468 nv50_state_flush_notify(chan); 469 return TRUE; 470} 471 472void nv50_so_init_sifc(struct nv50_context *nv50, 473 struct nouveau_stateobj *so, 474 struct nouveau_bo *bo, unsigned reloc, 475 unsigned offset, unsigned size) 476{ 477 struct nouveau_grobj *eng2d = nv50->screen->eng2d; 478 479 reloc |= NOUVEAU_BO_WR; 480 481 so_method(so, eng2d, NV50_2D_DST_FORMAT, 2); 482 so_data (so, NV50_2D_DST_FORMAT_R8_UNORM); 483 so_data (so, 1); 484 so_method(so, eng2d, NV50_2D_DST_PITCH, 5); 485 so_data (so, 262144); 486 so_data (so, 65536); 487 so_data (so, 1); 488 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_HIGH, 0, 0); 489 so_reloc (so, bo, offset, reloc | NOUVEAU_BO_LOW, 0, 0); 490 so_method(so, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2); 491 so_data (so, 0); 492 so_data (so, NV50_2D_SIFC_FORMAT_R8_UNORM); 493 so_method(so, eng2d, NV50_2D_SIFC_WIDTH, 10); 494 so_data (so, size); 495 so_data (so, 1); 496 so_data (so, 0); 497 so_data (so, 1); 498 so_data (so, 0); 499 so_data (so, 1); 500 so_data (so, 0); 501 so_data (so, 0); 502 so_data (so, 0); 503 so_data (so, 0); 504} 505