r300_state.c revision c8b34616b345d7a64d39e0ea3d4f4006af0b9740
1/* 2 * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com> 3 * Copyright 2009 Marek Olšák <maraeo@gmail.com> 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * on the rights to use, copy, modify, merge, publish, distribute, sub 9 * license, and/or sell copies of the Software, and to permit persons to whom 10 * the Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice (including the next 13 * paragraph) shall be included in all copies or substantial portions of the 14 * Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, 20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 22 * USE OR OTHER DEALINGS IN THE SOFTWARE. */ 23 24#include "draw/draw_context.h" 25 26#include "util/u_math.h" 27#include "util/u_memory.h" 28#include "util/u_pack_color.h" 29 30#include "tgsi/tgsi_parse.h" 31 32#include "pipe/p_config.h" 33 34#include "r300_context.h" 35#include "r300_reg.h" 36#include "r300_screen.h" 37#include "r300_state_inlines.h" 38#include "r300_fs.h" 39#include "r300_vs.h" 40 41/* r300_state: Functions used to intialize state context by translating 42 * Gallium state objects into semi-native r300 state objects. */ 43 44static boolean blend_discard_if_src_alpha_0(unsigned srcRGB, unsigned srcA, 45 unsigned dstRGB, unsigned dstA) 46{ 47 /* If the blend equation is ADD or REVERSE_SUBTRACT, 48 * SRC_ALPHA == 0, and the following state is set, the colorbuffer 49 * will not be changed. 50 * Notice that the dst factors are the src factors inverted. */ 51 return (srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA || 52 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE || 53 srcRGB == PIPE_BLENDFACTOR_ZERO) && 54 (srcA == PIPE_BLENDFACTOR_SRC_COLOR || 55 srcA == PIPE_BLENDFACTOR_SRC_ALPHA || 56 srcA == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE || 57 srcA == PIPE_BLENDFACTOR_ZERO) && 58 (dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 59 dstRGB == PIPE_BLENDFACTOR_ONE) && 60 (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR || 61 dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 62 dstA == PIPE_BLENDFACTOR_ONE); 63} 64 65static boolean blend_discard_if_src_alpha_1(unsigned srcRGB, unsigned srcA, 66 unsigned dstRGB, unsigned dstA) 67{ 68 /* If the blend equation is ADD or REVERSE_SUBTRACT, 69 * SRC_ALPHA == 1, and the following state is set, the colorbuffer 70 * will not be changed. 71 * Notice that the dst factors are the src factors inverted. */ 72 return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 73 srcRGB == PIPE_BLENDFACTOR_ZERO) && 74 (srcA == PIPE_BLENDFACTOR_INV_SRC_COLOR || 75 srcA == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 76 srcA == PIPE_BLENDFACTOR_ZERO) && 77 (dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA || 78 dstRGB == PIPE_BLENDFACTOR_ONE) && 79 (dstA == PIPE_BLENDFACTOR_SRC_COLOR || 80 dstA == PIPE_BLENDFACTOR_SRC_ALPHA || 81 dstA == PIPE_BLENDFACTOR_ONE); 82} 83 84static boolean blend_discard_if_src_color_0(unsigned srcRGB, unsigned srcA, 85 unsigned dstRGB, unsigned dstA) 86{ 87 /* If the blend equation is ADD or REVERSE_SUBTRACT, 88 * SRC_COLOR == (0,0,0), and the following state is set, the colorbuffer 89 * will not be changed. 90 * Notice that the dst factors are the src factors inverted. */ 91 return (srcRGB == PIPE_BLENDFACTOR_SRC_COLOR || 92 srcRGB == PIPE_BLENDFACTOR_ZERO) && 93 (srcA == PIPE_BLENDFACTOR_ZERO) && 94 (dstRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR || 95 dstRGB == PIPE_BLENDFACTOR_ONE) && 96 (dstA == PIPE_BLENDFACTOR_ONE); 97} 98 99static boolean blend_discard_if_src_color_1(unsigned srcRGB, unsigned srcA, 100 unsigned dstRGB, unsigned dstA) 101{ 102 /* If the blend equation is ADD or REVERSE_SUBTRACT, 103 * SRC_COLOR == (1,1,1), and the following state is set, the colorbuffer 104 * will not be changed. 105 * Notice that the dst factors are the src factors inverted. */ 106 return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR || 107 srcRGB == PIPE_BLENDFACTOR_ZERO) && 108 (srcA == PIPE_BLENDFACTOR_ZERO) && 109 (dstRGB == PIPE_BLENDFACTOR_SRC_COLOR || 110 dstRGB == PIPE_BLENDFACTOR_ONE) && 111 (dstA == PIPE_BLENDFACTOR_ONE); 112} 113 114static boolean blend_discard_if_src_alpha_color_0(unsigned srcRGB, unsigned srcA, 115 unsigned dstRGB, unsigned dstA) 116{ 117 /* If the blend equation is ADD or REVERSE_SUBTRACT, 118 * SRC_ALPHA_COLOR == (0,0,0,0), and the following state is set, 119 * the colorbuffer will not be changed. 120 * Notice that the dst factors are the src factors inverted. */ 121 return (srcRGB == PIPE_BLENDFACTOR_SRC_COLOR || 122 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA || 123 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE || 124 srcRGB == PIPE_BLENDFACTOR_ZERO) && 125 (srcA == PIPE_BLENDFACTOR_SRC_COLOR || 126 srcA == PIPE_BLENDFACTOR_SRC_ALPHA || 127 srcA == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE || 128 srcA == PIPE_BLENDFACTOR_ZERO) && 129 (dstRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR || 130 dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 131 dstRGB == PIPE_BLENDFACTOR_ONE) && 132 (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR || 133 dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 134 dstA == PIPE_BLENDFACTOR_ONE); 135} 136 137static boolean blend_discard_if_src_alpha_color_1(unsigned srcRGB, unsigned srcA, 138 unsigned dstRGB, unsigned dstA) 139{ 140 /* If the blend equation is ADD or REVERSE_SUBTRACT, 141 * SRC_ALPHA_COLOR == (1,1,1,1), and the following state is set, 142 * the colorbuffer will not be changed. 143 * Notice that the dst factors are the src factors inverted. */ 144 return (srcRGB == PIPE_BLENDFACTOR_INV_SRC_COLOR || 145 srcRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 146 srcRGB == PIPE_BLENDFACTOR_ZERO) && 147 (srcA == PIPE_BLENDFACTOR_INV_SRC_COLOR || 148 srcA == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 149 srcA == PIPE_BLENDFACTOR_ZERO) && 150 (dstRGB == PIPE_BLENDFACTOR_SRC_COLOR || 151 dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA || 152 dstRGB == PIPE_BLENDFACTOR_ONE) && 153 (dstA == PIPE_BLENDFACTOR_SRC_COLOR || 154 dstA == PIPE_BLENDFACTOR_SRC_ALPHA || 155 dstA == PIPE_BLENDFACTOR_ONE); 156} 157 158static unsigned bgra_cmask(unsigned mask) 159{ 160 /* Gallium uses RGBA color ordering while R300 expects BGRA. */ 161 162 return ((mask & PIPE_MASK_R) << 2) | 163 ((mask & PIPE_MASK_B) >> 2) | 164 (mask & (PIPE_MASK_G | PIPE_MASK_A)); 165} 166 167/* Create a new blend state based on the CSO blend state. 168 * 169 * This encompasses alpha blending, logic/raster ops, and blend dithering. */ 170static void* r300_create_blend_state(struct pipe_context* pipe, 171 const struct pipe_blend_state* state) 172{ 173 struct r300_screen* r300screen = r300_screen(pipe->screen); 174 struct r300_blend_state* blend = CALLOC_STRUCT(r300_blend_state); 175 176 if (state->rt[0].blend_enable) 177 { 178 unsigned eqRGB = state->rt[0].rgb_func; 179 unsigned srcRGB = state->rt[0].rgb_src_factor; 180 unsigned dstRGB = state->rt[0].rgb_dst_factor; 181 182 unsigned eqA = state->rt[0].alpha_func; 183 unsigned srcA = state->rt[0].alpha_src_factor; 184 unsigned dstA = state->rt[0].alpha_dst_factor; 185 186 /* despite the name, ALPHA_BLEND_ENABLE has nothing to do with alpha, 187 * this is just the crappy D3D naming */ 188 blend->blend_control = R300_ALPHA_BLEND_ENABLE | 189 r300_translate_blend_function(eqRGB) | 190 ( r300_translate_blend_factor(srcRGB) << R300_SRC_BLEND_SHIFT) | 191 ( r300_translate_blend_factor(dstRGB) << R300_DST_BLEND_SHIFT); 192 193 /* Optimization: some operations do not require the destination color. 194 * 195 * When SRC_ALPHA_SATURATE is used, colorbuffer reads must be enabled, 196 * otherwise blending gives incorrect results. It seems to be 197 * a hardware bug. */ 198 if (eqRGB == PIPE_BLEND_MIN || eqA == PIPE_BLEND_MIN || 199 eqRGB == PIPE_BLEND_MAX || eqA == PIPE_BLEND_MAX || 200 dstRGB != PIPE_BLENDFACTOR_ZERO || 201 dstA != PIPE_BLENDFACTOR_ZERO || 202 srcRGB == PIPE_BLENDFACTOR_DST_COLOR || 203 srcRGB == PIPE_BLENDFACTOR_DST_ALPHA || 204 srcRGB == PIPE_BLENDFACTOR_INV_DST_COLOR || 205 srcRGB == PIPE_BLENDFACTOR_INV_DST_ALPHA || 206 srcA == PIPE_BLENDFACTOR_DST_COLOR || 207 srcA == PIPE_BLENDFACTOR_DST_ALPHA || 208 srcA == PIPE_BLENDFACTOR_INV_DST_COLOR || 209 srcA == PIPE_BLENDFACTOR_INV_DST_ALPHA || 210 srcRGB == PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE) { 211 /* Enable reading from the colorbuffer. */ 212 blend->blend_control |= R300_READ_ENABLE; 213 214 if (r300_screen(r300_context(pipe)->context.screen)->caps->is_r500) { 215 /* Optimization: Depending on incoming pixels, we can 216 * conditionally disable the reading in hardware... */ 217 if (eqRGB != PIPE_BLEND_MIN && eqA != PIPE_BLEND_MIN && 218 eqRGB != PIPE_BLEND_MAX && eqA != PIPE_BLEND_MAX) { 219 /* Disable reading if SRC_ALPHA == 0. */ 220 if ((dstRGB == PIPE_BLENDFACTOR_SRC_ALPHA || 221 dstRGB == PIPE_BLENDFACTOR_ZERO) && 222 (dstA == PIPE_BLENDFACTOR_SRC_COLOR || 223 dstA == PIPE_BLENDFACTOR_SRC_ALPHA || 224 dstA == PIPE_BLENDFACTOR_ZERO)) { 225 blend->blend_control |= R500_SRC_ALPHA_0_NO_READ; 226 } 227 228 /* Disable reading if SRC_ALPHA == 1. */ 229 if ((dstRGB == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 230 dstRGB == PIPE_BLENDFACTOR_ZERO) && 231 (dstA == PIPE_BLENDFACTOR_INV_SRC_COLOR || 232 dstA == PIPE_BLENDFACTOR_INV_SRC_ALPHA || 233 dstA == PIPE_BLENDFACTOR_ZERO)) { 234 blend->blend_control |= R500_SRC_ALPHA_1_NO_READ; 235 } 236 } 237 } 238 } 239 240 /* Optimization: discard pixels which don't change the colorbuffer. 241 * 242 * The code below is non-trivial and some math is involved. 243 * 244 * Discarding pixels must be disabled when FP16 AA is enabled. 245 * This is a hardware bug. Also, this implementation wouldn't work 246 * with FP blending enabled and equation clamping disabled. 247 * 248 * Equations other than ADD are rarely used and therefore won't be 249 * optimized. */ 250 if ((eqRGB == PIPE_BLEND_ADD || eqRGB == PIPE_BLEND_REVERSE_SUBTRACT) && 251 (eqA == PIPE_BLEND_ADD || eqA == PIPE_BLEND_REVERSE_SUBTRACT)) { 252 /* ADD: X+Y 253 * REVERSE_SUBTRACT: Y-X 254 * 255 * The idea is: 256 * If X = src*srcFactor = 0 and Y = dst*dstFactor = 1, 257 * then CB will not be changed. 258 * 259 * Given the srcFactor and dstFactor variables, we can derive 260 * what src and dst should be equal to and discard appropriate 261 * pixels. 262 */ 263 if (blend_discard_if_src_alpha_0(srcRGB, srcA, dstRGB, dstA)) { 264 blend->blend_control |= R300_DISCARD_SRC_PIXELS_SRC_ALPHA_0; 265 } else if (blend_discard_if_src_alpha_1(srcRGB, srcA, 266 dstRGB, dstA)) { 267 blend->blend_control |= R300_DISCARD_SRC_PIXELS_SRC_ALPHA_1; 268 } else if (blend_discard_if_src_color_0(srcRGB, srcA, 269 dstRGB, dstA)) { 270 blend->blend_control |= R300_DISCARD_SRC_PIXELS_SRC_COLOR_0; 271 } else if (blend_discard_if_src_color_1(srcRGB, srcA, 272 dstRGB, dstA)) { 273 blend->blend_control |= R300_DISCARD_SRC_PIXELS_SRC_COLOR_1; 274 } else if (blend_discard_if_src_alpha_color_0(srcRGB, srcA, 275 dstRGB, dstA)) { 276 blend->blend_control |= 277 R300_DISCARD_SRC_PIXELS_SRC_ALPHA_COLOR_0; 278 } else if (blend_discard_if_src_alpha_color_1(srcRGB, srcA, 279 dstRGB, dstA)) { 280 blend->blend_control |= 281 R300_DISCARD_SRC_PIXELS_SRC_ALPHA_COLOR_1; 282 } 283 } 284 285 /* separate alpha */ 286 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) { 287 blend->blend_control |= R300_SEPARATE_ALPHA_ENABLE; 288 blend->alpha_blend_control = 289 r300_translate_blend_function(eqA) | 290 (r300_translate_blend_factor(srcA) << R300_SRC_BLEND_SHIFT) | 291 (r300_translate_blend_factor(dstA) << R300_DST_BLEND_SHIFT); 292 } 293 } 294 295 /* PIPE_LOGICOP_* don't need to be translated, fortunately. */ 296 if (state->logicop_enable) { 297 blend->rop = R300_RB3D_ROPCNTL_ROP_ENABLE | 298 (state->logicop_func) << R300_RB3D_ROPCNTL_ROP_SHIFT; 299 } 300 301 /* Color channel masks for all MRTs. */ 302 blend->color_channel_mask = bgra_cmask(state->rt[0].colormask); 303 if (r300screen->caps->is_r500 && state->independent_blend_enable) { 304 if (state->rt[1].blend_enable) { 305 blend->color_channel_mask |= bgra_cmask(state->rt[1].colormask) << 4; 306 } 307 if (state->rt[2].blend_enable) { 308 blend->color_channel_mask |= bgra_cmask(state->rt[2].colormask) << 8; 309 } 310 if (state->rt[3].blend_enable) { 311 blend->color_channel_mask |= bgra_cmask(state->rt[3].colormask) << 12; 312 } 313 } 314 315 if (state->dither) { 316 blend->dither = R300_RB3D_DITHER_CTL_DITHER_MODE_LUT | 317 R300_RB3D_DITHER_CTL_ALPHA_DITHER_MODE_LUT; 318 } 319 320 return (void*)blend; 321} 322 323/* Bind blend state. */ 324static void r300_bind_blend_state(struct pipe_context* pipe, 325 void* state) 326{ 327 struct r300_context* r300 = r300_context(pipe); 328 329 r300->blend_state.state = state; 330 r300->blend_state.dirty = TRUE; 331} 332 333/* Free blend state. */ 334static void r300_delete_blend_state(struct pipe_context* pipe, 335 void* state) 336{ 337 FREE(state); 338} 339 340/* Convert float to 10bit integer */ 341static unsigned float_to_fixed10(float f) 342{ 343 return CLAMP((unsigned)(f * 1023.9f), 0, 1023); 344} 345 346/* Set blend color. 347 * Setup both R300 and R500 registers, figure out later which one to write. */ 348static void r300_set_blend_color(struct pipe_context* pipe, 349 const struct pipe_blend_color* color) 350{ 351 struct r300_context* r300 = r300_context(pipe); 352 struct r300_screen* r300screen = r300_screen(pipe->screen); 353 struct r300_blend_color_state* state = 354 (struct r300_blend_color_state*)r300->blend_color_state.state; 355 union util_color uc; 356 357 util_pack_color(color->color, PIPE_FORMAT_A8R8G8B8_UNORM, &uc); 358 state->blend_color = uc.ui; 359 360 /* XXX if FP16 blending is enabled, we should use the FP16 format */ 361 state->blend_color_red_alpha = 362 float_to_fixed10(color->color[0]) | 363 (float_to_fixed10(color->color[3]) << 16); 364 state->blend_color_green_blue = 365 float_to_fixed10(color->color[2]) | 366 (float_to_fixed10(color->color[1]) << 16); 367 368 r300->blend_color_state.size = r300screen->caps->is_r500 ? 3 : 2; 369 r300->blend_color_state.dirty = TRUE; 370} 371 372static void r300_set_clip_state(struct pipe_context* pipe, 373 const struct pipe_clip_state* state) 374{ 375 struct r300_context* r300 = r300_context(pipe); 376 377 if (r300_screen(pipe->screen)->caps->has_tcl) { 378 memcpy(r300->clip_state.state, state, sizeof(struct pipe_clip_state)); 379 r300->clip_state.size = 29; 380 } else { 381 draw_flush(r300->draw); 382 draw_set_clip_state(r300->draw, state); 383 r300->clip_state.size = 2; 384 } 385 386 r300->clip_state.dirty = TRUE; 387} 388 389/* Create a new depth, stencil, and alpha state based on the CSO dsa state. 390 * 391 * This contains the depth buffer, stencil buffer, alpha test, and such. 392 * On the Radeon, depth and stencil buffer setup are intertwined, which is 393 * the reason for some of the strange-looking assignments across registers. */ 394static void* 395 r300_create_dsa_state(struct pipe_context* pipe, 396 const struct pipe_depth_stencil_alpha_state* state) 397{ 398 struct r300_capabilities *caps = 399 r300_screen(r300_context(pipe)->context.screen)->caps; 400 struct r300_dsa_state* dsa = CALLOC_STRUCT(r300_dsa_state); 401 402 /* Depth test setup. */ 403 if (state->depth.enabled) { 404 dsa->z_buffer_control |= R300_Z_ENABLE; 405 406 if (state->depth.writemask) { 407 dsa->z_buffer_control |= R300_Z_WRITE_ENABLE; 408 } 409 410 dsa->z_stencil_control |= 411 (r300_translate_depth_stencil_function(state->depth.func) << 412 R300_Z_FUNC_SHIFT); 413 } 414 415 /* Stencil buffer setup. */ 416 if (state->stencil[0].enabled) { 417 dsa->z_buffer_control |= R300_STENCIL_ENABLE; 418 dsa->z_stencil_control |= 419 (r300_translate_depth_stencil_function(state->stencil[0].func) << 420 R300_S_FRONT_FUNC_SHIFT) | 421 (r300_translate_stencil_op(state->stencil[0].fail_op) << 422 R300_S_FRONT_SFAIL_OP_SHIFT) | 423 (r300_translate_stencil_op(state->stencil[0].zpass_op) << 424 R300_S_FRONT_ZPASS_OP_SHIFT) | 425 (r300_translate_stencil_op(state->stencil[0].zfail_op) << 426 R300_S_FRONT_ZFAIL_OP_SHIFT); 427 428 dsa->stencil_ref_mask = (state->stencil[0].ref_value) | 429 (state->stencil[0].valuemask << R300_STENCILMASK_SHIFT) | 430 (state->stencil[0].writemask << R300_STENCILWRITEMASK_SHIFT); 431 432 if (state->stencil[1].enabled) { 433 dsa->z_buffer_control |= R300_STENCIL_FRONT_BACK; 434 dsa->z_stencil_control |= 435 (r300_translate_depth_stencil_function(state->stencil[1].func) << 436 R300_S_BACK_FUNC_SHIFT) | 437 (r300_translate_stencil_op(state->stencil[1].fail_op) << 438 R300_S_BACK_SFAIL_OP_SHIFT) | 439 (r300_translate_stencil_op(state->stencil[1].zpass_op) << 440 R300_S_BACK_ZPASS_OP_SHIFT) | 441 (r300_translate_stencil_op(state->stencil[1].zfail_op) << 442 R300_S_BACK_ZFAIL_OP_SHIFT); 443 444 /* XXX it seems r3xx doesn't support STENCILREFMASK_BF */ 445 if (caps->is_r500) 446 { 447 dsa->z_buffer_control |= R500_STENCIL_REFMASK_FRONT_BACK; 448 dsa->stencil_ref_bf = (state->stencil[1].ref_value) | 449 (state->stencil[1].valuemask << 450 R300_STENCILMASK_SHIFT) | 451 (state->stencil[1].writemask << 452 R300_STENCILWRITEMASK_SHIFT); 453 } 454 } 455 } 456 457 /* Alpha test setup. */ 458 if (state->alpha.enabled) { 459 dsa->alpha_function = 460 r300_translate_alpha_function(state->alpha.func) | 461 R300_FG_ALPHA_FUNC_ENABLE; 462 463 /* XXX figure out why emitting 10bit alpha ref causes CS to dump */ 464 /* always use 8bit alpha ref */ 465 dsa->alpha_function |= float_to_ubyte(state->alpha.ref_value); 466 467 if (caps->is_r500) 468 dsa->alpha_function |= R500_FG_ALPHA_FUNC_8BIT; 469 } 470 471 return (void*)dsa; 472} 473 474/* Bind DSA state. */ 475static void r300_bind_dsa_state(struct pipe_context* pipe, 476 void* state) 477{ 478 struct r300_context* r300 = r300_context(pipe); 479 struct r300_screen* r300screen = r300_screen(pipe->screen); 480 481 r300->dsa_state.state = state; 482 r300->dsa_state.size = r300screen->caps->is_r500 ? 8 : 6; 483 r300->dsa_state.dirty = TRUE; 484} 485 486/* Free DSA state. */ 487static void r300_delete_dsa_state(struct pipe_context* pipe, 488 void* state) 489{ 490 FREE(state); 491} 492 493static void 494 r300_set_framebuffer_state(struct pipe_context* pipe, 495 const struct pipe_framebuffer_state* state) 496{ 497 struct r300_context* r300 = r300_context(pipe); 498 uint32_t zbuffer_bpp = 0; 499 500 r300->fb_state.size = (10 * state->nr_cbufs) + 501 (2 * (4 - state->nr_cbufs)) + 502 (state->zsbuf ? 10 : 0) + 6; 503 504 if (state->nr_cbufs > 4) { 505 debug_printf("r300: Implementation error: Too many MRTs in %s, " 506 "refusing to bind framebuffer state!\n", __FUNCTION__); 507 return; 508 } 509 510 if (r300->draw) { 511 draw_flush(r300->draw); 512 } 513 514 r300->fb_state.state = state; 515 516 /* Don't rely on the order of states being set for the first time. */ 517 /* XXX wait what */ 518 r300->blend_state.dirty = TRUE; 519 r300->dsa_state.dirty = TRUE; 520 r300->fb_state.dirty = TRUE; 521 r300->scissor_state.dirty = TRUE; 522 523 /* Polygon offset depends on the zbuffer bit depth. */ 524 if (state->zsbuf && r300->polygon_offset_enabled) { 525 switch (util_format_get_blocksize(state->zsbuf->texture->format)) { 526 case 2: 527 zbuffer_bpp = 16; 528 break; 529 case 4: 530 zbuffer_bpp = 24; 531 break; 532 } 533 534 if (r300->zbuffer_bpp != zbuffer_bpp) { 535 r300->zbuffer_bpp = zbuffer_bpp; 536 r300->rs_state.dirty = TRUE; 537 } 538 } 539} 540 541/* Create fragment shader state. */ 542static void* r300_create_fs_state(struct pipe_context* pipe, 543 const struct pipe_shader_state* shader) 544{ 545 struct r300_fragment_shader* fs = NULL; 546 547 fs = (struct r300_fragment_shader*)CALLOC_STRUCT(r300_fragment_shader); 548 549 /* Copy state directly into shader. */ 550 fs->state = *shader; 551 fs->state.tokens = tgsi_dup_tokens(shader->tokens); 552 553 tgsi_scan_shader(shader->tokens, &fs->info); 554 r300_shader_read_fs_inputs(&fs->info, &fs->inputs); 555 556 return (void*)fs; 557} 558 559/* Bind fragment shader state. */ 560static void r300_bind_fs_state(struct pipe_context* pipe, void* shader) 561{ 562 struct r300_context* r300 = r300_context(pipe); 563 struct r300_fragment_shader* fs = (struct r300_fragment_shader*)shader; 564 565 if (fs == NULL) { 566 r300->fs = NULL; 567 return; 568 } 569 570 r300->fs = fs; 571 r300_pick_fragment_shader(r300); 572 573 if (r300->vs && r300_vertex_shader_setup_wpos(r300)) { 574 r300->vertex_format_state.dirty = TRUE; 575 } 576 577 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER | R300_NEW_FRAGMENT_SHADER_CONSTANTS; 578} 579 580/* Delete fragment shader state. */ 581static void r300_delete_fs_state(struct pipe_context* pipe, void* shader) 582{ 583 struct r300_fragment_shader* fs = (struct r300_fragment_shader*)shader; 584 struct r300_fragment_shader_code *tmp, *ptr = fs->first; 585 586 while (ptr) { 587 tmp = ptr; 588 ptr = ptr->next; 589 rc_constants_destroy(&tmp->code.constants); 590 FREE(tmp); 591 } 592 FREE((void*)fs->state.tokens); 593 FREE(shader); 594} 595 596static void r300_set_polygon_stipple(struct pipe_context* pipe, 597 const struct pipe_poly_stipple* state) 598{ 599 /* XXX no idea how to set this up, but not terribly important */ 600} 601 602/* Create a new rasterizer state based on the CSO rasterizer state. 603 * 604 * This is a very large chunk of state, and covers most of the graphics 605 * backend (GB), geometry assembly (GA), and setup unit (SU) blocks. 606 * 607 * In a not entirely unironic sidenote, this state has nearly nothing to do 608 * with the actual block on the Radeon called the rasterizer (RS). */ 609static void* r300_create_rs_state(struct pipe_context* pipe, 610 const struct pipe_rasterizer_state* state) 611{ 612 struct r300_rs_state* rs = CALLOC_STRUCT(r300_rs_state); 613 614 /* Copy rasterizer state for Draw. */ 615 rs->rs = *state; 616 617#ifdef PIPE_ARCH_LITTLE_ENDIAN 618 rs->vap_control_status = R300_VC_NO_SWAP; 619#else 620 rs->vap_control_status = R300_VC_32BIT_SWAP; 621#endif 622 623 /* If bypassing TCL, or if no TCL engine is present, turn off the HW TCL. 624 * Else, enable HW TCL and force Draw's TCL off. */ 625 if (state->bypass_vs_clip_and_viewport || 626 !r300_screen(pipe->screen)->caps->has_tcl) { 627 rs->vap_control_status |= R300_VAP_TCL_BYPASS; 628 } 629 630 rs->point_size = pack_float_16_6x(state->point_size) | 631 (pack_float_16_6x(state->point_size) << R300_POINTSIZE_X_SHIFT); 632 633 rs->point_minmax = 634 ((int)(state->point_size_min * 6.0) << 635 R300_GA_POINT_MINMAX_MIN_SHIFT) | 636 ((int)(state->point_size_max * 6.0) << 637 R300_GA_POINT_MINMAX_MAX_SHIFT); 638 639 rs->line_control = pack_float_16_6x(state->line_width) | 640 R300_GA_LINE_CNTL_END_TYPE_COMP; 641 642 /* Enable polygon mode */ 643 if (state->fill_cw != PIPE_POLYGON_MODE_FILL || 644 state->fill_ccw != PIPE_POLYGON_MODE_FILL) { 645 rs->polygon_mode = R300_GA_POLY_MODE_DUAL; 646 } 647 648 /* Radeons don't think in "CW/CCW", they think in "front/back". */ 649 if (state->front_winding == PIPE_WINDING_CW) { 650 rs->cull_mode = R300_FRONT_FACE_CW; 651 652 /* Polygon offset */ 653 if (state->offset_cw) { 654 rs->polygon_offset_enable |= R300_FRONT_ENABLE; 655 } 656 if (state->offset_ccw) { 657 rs->polygon_offset_enable |= R300_BACK_ENABLE; 658 } 659 660 /* Polygon mode */ 661 if (rs->polygon_mode) { 662 rs->polygon_mode |= 663 r300_translate_polygon_mode_front(state->fill_cw); 664 rs->polygon_mode |= 665 r300_translate_polygon_mode_back(state->fill_ccw); 666 } 667 } else { 668 rs->cull_mode = R300_FRONT_FACE_CCW; 669 670 /* Polygon offset */ 671 if (state->offset_ccw) { 672 rs->polygon_offset_enable |= R300_FRONT_ENABLE; 673 } 674 if (state->offset_cw) { 675 rs->polygon_offset_enable |= R300_BACK_ENABLE; 676 } 677 678 /* Polygon mode */ 679 if (rs->polygon_mode) { 680 rs->polygon_mode |= 681 r300_translate_polygon_mode_front(state->fill_ccw); 682 rs->polygon_mode |= 683 r300_translate_polygon_mode_back(state->fill_cw); 684 } 685 } 686 if (state->front_winding & state->cull_mode) { 687 rs->cull_mode |= R300_CULL_FRONT; 688 } 689 if (~(state->front_winding) & state->cull_mode) { 690 rs->cull_mode |= R300_CULL_BACK; 691 } 692 693 if (rs->polygon_offset_enable) { 694 rs->depth_offset = state->offset_units; 695 rs->depth_scale = state->offset_scale; 696 } 697 698 if (state->line_stipple_enable) { 699 rs->line_stipple_config = 700 R300_GA_LINE_STIPPLE_CONFIG_LINE_RESET_LINE | 701 (fui((float)state->line_stipple_factor) & 702 R300_GA_LINE_STIPPLE_CONFIG_STIPPLE_SCALE_MASK); 703 /* XXX this might need to be scaled up */ 704 rs->line_stipple_value = state->line_stipple_pattern; 705 } 706 707 if (state->flatshade) { 708 rs->color_control = R300_SHADE_MODEL_FLAT; 709 } else { 710 rs->color_control = R300_SHADE_MODEL_SMOOTH; 711 } 712 713 return (void*)rs; 714} 715 716/* Bind rasterizer state. */ 717static void r300_bind_rs_state(struct pipe_context* pipe, void* state) 718{ 719 struct r300_context* r300 = r300_context(pipe); 720 struct r300_rs_state* rs = (struct r300_rs_state*)state; 721 722 if (r300->draw) { 723 draw_flush(r300->draw); 724 draw_set_rasterizer_state(r300->draw, &rs->rs); 725 } 726 727 if (rs) { 728 r300->tcl_bypass = rs->rs.bypass_vs_clip_and_viewport; 729 r300->polygon_offset_enabled = rs->rs.offset_cw || rs->rs.offset_ccw; 730 } else { 731 r300->tcl_bypass = FALSE; 732 r300->polygon_offset_enabled = FALSE; 733 } 734 735 r300->rs_state.state = rs; 736 r300->rs_state.dirty = TRUE; 737 /* XXX Why is this still needed, dammit!? */ 738 r300->scissor_state.dirty = TRUE; 739 r300->viewport_state.dirty = TRUE; 740 741 /* XXX Clean these up when we move to atom emits */ 742 if (r300->fs && r300->fs->inputs.wpos != ATTR_UNUSED) { 743 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS; 744 } 745} 746 747/* Free rasterizer state. */ 748static void r300_delete_rs_state(struct pipe_context* pipe, void* state) 749{ 750 FREE(state); 751} 752 753static void* 754 r300_create_sampler_state(struct pipe_context* pipe, 755 const struct pipe_sampler_state* state) 756{ 757 struct r300_context* r300 = r300_context(pipe); 758 struct r300_sampler_state* sampler = CALLOC_STRUCT(r300_sampler_state); 759 int lod_bias; 760 union util_color uc; 761 762 sampler->state = *state; 763 764 sampler->filter0 |= 765 (r300_translate_wrap(state->wrap_s) << R300_TX_WRAP_S_SHIFT) | 766 (r300_translate_wrap(state->wrap_t) << R300_TX_WRAP_T_SHIFT) | 767 (r300_translate_wrap(state->wrap_r) << R300_TX_WRAP_R_SHIFT); 768 769 sampler->filter0 |= r300_translate_tex_filters(state->min_img_filter, 770 state->mag_img_filter, 771 state->min_mip_filter, 772 state->max_anisotropy > 1.0); 773 774 /* Unfortunately, r300-r500 don't support floating-point mipmap lods. */ 775 /* We must pass these to the emit function to clamp them properly. */ 776 sampler->min_lod = MAX2((unsigned)state->min_lod, 0); 777 sampler->max_lod = MAX2((unsigned)ceilf(state->max_lod), 0); 778 779 lod_bias = CLAMP((int)(state->lod_bias * 32), -(1 << 9), (1 << 9) - 1); 780 781 sampler->filter1 |= lod_bias << R300_LOD_BIAS_SHIFT; 782 783 sampler->filter1 |= r300_anisotropy(state->max_anisotropy); 784 785 util_pack_color(state->border_color, PIPE_FORMAT_A8R8G8B8_UNORM, &uc); 786 sampler->border_color = uc.ui; 787 788 /* R500-specific fixups and optimizations */ 789 if (r300_screen(r300->context.screen)->caps->is_r500) { 790 sampler->filter1 |= R500_BORDER_FIX; 791 } 792 793 return (void*)sampler; 794} 795 796static void r300_bind_sampler_states(struct pipe_context* pipe, 797 unsigned count, 798 void** states) 799{ 800 struct r300_context* r300 = r300_context(pipe); 801 int i; 802 803 if (count > 8) { 804 return; 805 } 806 807 for (i = 0; i < count; i++) { 808 if (r300->sampler_states[i] != states[i]) { 809 r300->sampler_states[i] = (struct r300_sampler_state*)states[i]; 810 r300->dirty_state |= (R300_NEW_SAMPLER << i); 811 } 812 } 813 814 r300->sampler_count = count; 815 816 /* Pick a fragment shader based on the texture compare state. */ 817 if (r300->fs && (r300->dirty_state & R300_ANY_NEW_SAMPLERS)) { 818 if (r300_pick_fragment_shader(r300)) { 819 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER | 820 R300_NEW_FRAGMENT_SHADER_CONSTANTS; 821 } 822 } 823} 824 825static void r300_lacks_vertex_textures(struct pipe_context* pipe, 826 unsigned count, 827 void** states) 828{ 829} 830 831static void r300_delete_sampler_state(struct pipe_context* pipe, void* state) 832{ 833 FREE(state); 834} 835 836static void r300_set_sampler_textures(struct pipe_context* pipe, 837 unsigned count, 838 struct pipe_texture** texture) 839{ 840 struct r300_context* r300 = r300_context(pipe); 841 boolean is_r500 = r300_screen(r300->context.screen)->caps->is_r500; 842 int i; 843 844 /* XXX magic num */ 845 if (count > 8) { 846 return; 847 } 848 849 for (i = 0; i < count; i++) { 850 if (r300->textures[i] != (struct r300_texture*)texture[i]) { 851 pipe_texture_reference((struct pipe_texture**)&r300->textures[i], 852 texture[i]); 853 r300->dirty_state |= (R300_NEW_TEXTURE << i); 854 855 /* R300-specific - set the texrect factor in a fragment shader */ 856 if (!is_r500 && r300->textures[i]->is_npot) { 857 /* XXX It would be nice to re-emit just 1 constant, 858 * XXX not all of them */ 859 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS; 860 } 861 } 862 } 863 864 for (i = count; i < 8; i++) { 865 if (r300->textures[i]) { 866 pipe_texture_reference((struct pipe_texture**)&r300->textures[i], 867 NULL); 868 r300->dirty_state |= (R300_NEW_TEXTURE << i); 869 } 870 } 871 872 r300->texture_count = count; 873} 874 875static void r300_set_scissor_state(struct pipe_context* pipe, 876 const struct pipe_scissor_state* state) 877{ 878 struct r300_context* r300 = r300_context(pipe); 879 880 memcpy(r300->scissor_state.state, state, 881 sizeof(struct pipe_scissor_state)); 882 883 r300->scissor_state.dirty = TRUE; 884} 885 886static void r300_set_viewport_state(struct pipe_context* pipe, 887 const struct pipe_viewport_state* state) 888{ 889 struct r300_context* r300 = r300_context(pipe); 890 struct r300_viewport_state* viewport = 891 (struct r300_viewport_state*)r300->viewport_state.state; 892 893 /* Do the transform in HW. */ 894 viewport->vte_control = R300_VTX_W0_FMT; 895 896 if (state->scale[0] != 1.0f) { 897 viewport->xscale = state->scale[0]; 898 viewport->vte_control |= R300_VPORT_X_SCALE_ENA; 899 } 900 if (state->scale[1] != 1.0f) { 901 viewport->yscale = state->scale[1]; 902 viewport->vte_control |= R300_VPORT_Y_SCALE_ENA; 903 } 904 if (state->scale[2] != 1.0f) { 905 viewport->zscale = state->scale[2]; 906 viewport->vte_control |= R300_VPORT_Z_SCALE_ENA; 907 } 908 if (state->translate[0] != 0.0f) { 909 viewport->xoffset = state->translate[0]; 910 viewport->vte_control |= R300_VPORT_X_OFFSET_ENA; 911 } 912 if (state->translate[1] != 0.0f) { 913 viewport->yoffset = state->translate[1]; 914 viewport->vte_control |= R300_VPORT_Y_OFFSET_ENA; 915 } 916 if (state->translate[2] != 0.0f) { 917 viewport->zoffset = state->translate[2]; 918 viewport->vte_control |= R300_VPORT_Z_OFFSET_ENA; 919 } 920 921 r300->viewport_state.dirty = TRUE; 922 if (r300->fs && r300->fs->inputs.wpos != ATTR_UNUSED) { 923 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS; 924 } 925} 926 927static void r300_set_vertex_buffers(struct pipe_context* pipe, 928 unsigned count, 929 const struct pipe_vertex_buffer* buffers) 930{ 931 struct r300_context* r300 = r300_context(pipe); 932 933 memcpy(r300->vertex_buffer, buffers, 934 sizeof(struct pipe_vertex_buffer) * count); 935 r300->vertex_buffer_count = count; 936 937 if (r300->draw) { 938 draw_flush(r300->draw); 939 draw_set_vertex_buffers(r300->draw, count, buffers); 940 } 941 942 r300->vertex_format_state.dirty = TRUE; 943} 944 945static boolean r300_validate_aos(struct r300_context *r300) 946{ 947 struct pipe_vertex_buffer *vbuf = r300->vertex_buffer; 948 struct pipe_vertex_element *velem = r300->vertex_element; 949 int i; 950 951 /* Check if formats and strides are aligned to the size of DWORD. */ 952 for (i = 0; i < r300->vertex_element_count; i++) { 953 if (vbuf[velem[i].vertex_buffer_index].stride % 4 != 0 || 954 util_format_get_blocksize(velem[i].src_format) % 4 != 0) { 955 return FALSE; 956 } 957 } 958 return TRUE; 959} 960 961static void r300_set_vertex_elements(struct pipe_context* pipe, 962 unsigned count, 963 const struct pipe_vertex_element* elements) 964{ 965 struct r300_context* r300 = r300_context(pipe); 966 967 memcpy(r300->vertex_element, 968 elements, 969 sizeof(struct pipe_vertex_element) * count); 970 r300->vertex_element_count = count; 971 972 if (r300->draw) { 973 draw_flush(r300->draw); 974 draw_set_vertex_elements(r300->draw, count, elements); 975 } 976 977 if (!r300_validate_aos(r300)) { 978 /* XXX We should fallback using draw. */ 979 assert(0); 980 abort(); 981 } 982} 983 984static void* r300_create_vs_state(struct pipe_context* pipe, 985 const struct pipe_shader_state* shader) 986{ 987 struct r300_context* r300 = r300_context(pipe); 988 989 if (r300_screen(pipe->screen)->caps->has_tcl) { 990 struct r300_vertex_shader* vs = CALLOC_STRUCT(r300_vertex_shader); 991 /* Copy state directly into shader. */ 992 vs->state = *shader; 993 vs->state.tokens = tgsi_dup_tokens(shader->tokens); 994 995 tgsi_scan_shader(shader->tokens, &vs->info); 996 997 return (void*)vs; 998 } else { 999 return draw_create_vertex_shader(r300->draw, shader); 1000 } 1001} 1002 1003static void r300_bind_vs_state(struct pipe_context* pipe, void* shader) 1004{ 1005 struct r300_context* r300 = r300_context(pipe); 1006 1007 if (r300_screen(pipe->screen)->caps->has_tcl) { 1008 struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader; 1009 1010 if (vs == NULL) { 1011 r300->vs = NULL; 1012 return; 1013 } else if (!vs->translated) { 1014 r300_translate_vertex_shader(r300, vs); 1015 } 1016 1017 r300->vs = vs; 1018 if (r300->fs) { 1019 r300_vertex_shader_setup_wpos(r300); 1020 } 1021 1022 r300->vertex_format_state.dirty = TRUE; 1023 1024 r300->dirty_state |= 1025 R300_NEW_VERTEX_SHADER | R300_NEW_VERTEX_SHADER_CONSTANTS; 1026 } else { 1027 draw_flush(r300->draw); 1028 draw_bind_vertex_shader(r300->draw, 1029 (struct draw_vertex_shader*)shader); 1030 } 1031} 1032 1033static void r300_delete_vs_state(struct pipe_context* pipe, void* shader) 1034{ 1035 struct r300_context* r300 = r300_context(pipe); 1036 1037 if (r300_screen(pipe->screen)->caps->has_tcl) { 1038 struct r300_vertex_shader* vs = (struct r300_vertex_shader*)shader; 1039 1040 rc_constants_destroy(&vs->code.constants); 1041 FREE((void*)vs->state.tokens); 1042 FREE(shader); 1043 } else { 1044 draw_delete_vertex_shader(r300->draw, 1045 (struct draw_vertex_shader*)shader); 1046 } 1047} 1048 1049static void r300_set_constant_buffer(struct pipe_context *pipe, 1050 uint shader, uint index, 1051 struct pipe_buffer *buf) 1052{ 1053 struct r300_context* r300 = r300_context(pipe); 1054 void *mapped; 1055 1056 if (buf == NULL || buf->size == 0 || 1057 (mapped = pipe_buffer_map(pipe->screen, buf, PIPE_BUFFER_USAGE_CPU_READ)) == NULL) 1058 { 1059 r300->shader_constants[shader].count = 0; 1060 return; 1061 } 1062 1063 assert((buf->size % 4 * sizeof(float)) == 0); 1064 memcpy(r300->shader_constants[shader].constants, mapped, buf->size); 1065 r300->shader_constants[shader].count = buf->size / (4 * sizeof(float)); 1066 pipe_buffer_unmap(pipe->screen, buf); 1067 1068 if (shader == PIPE_SHADER_VERTEX) 1069 r300->dirty_state |= R300_NEW_VERTEX_SHADER_CONSTANTS; 1070 else if (shader == PIPE_SHADER_FRAGMENT) 1071 r300->dirty_state |= R300_NEW_FRAGMENT_SHADER_CONSTANTS; 1072} 1073 1074void r300_init_state_functions(struct r300_context* r300) 1075{ 1076 r300->context.create_blend_state = r300_create_blend_state; 1077 r300->context.bind_blend_state = r300_bind_blend_state; 1078 r300->context.delete_blend_state = r300_delete_blend_state; 1079 1080 r300->context.set_blend_color = r300_set_blend_color; 1081 1082 r300->context.set_clip_state = r300_set_clip_state; 1083 1084 r300->context.set_constant_buffer = r300_set_constant_buffer; 1085 1086 r300->context.create_depth_stencil_alpha_state = r300_create_dsa_state; 1087 r300->context.bind_depth_stencil_alpha_state = r300_bind_dsa_state; 1088 r300->context.delete_depth_stencil_alpha_state = r300_delete_dsa_state; 1089 1090 r300->context.set_framebuffer_state = r300_set_framebuffer_state; 1091 1092 r300->context.create_fs_state = r300_create_fs_state; 1093 r300->context.bind_fs_state = r300_bind_fs_state; 1094 r300->context.delete_fs_state = r300_delete_fs_state; 1095 1096 r300->context.set_polygon_stipple = r300_set_polygon_stipple; 1097 1098 r300->context.create_rasterizer_state = r300_create_rs_state; 1099 r300->context.bind_rasterizer_state = r300_bind_rs_state; 1100 r300->context.delete_rasterizer_state = r300_delete_rs_state; 1101 1102 r300->context.create_sampler_state = r300_create_sampler_state; 1103 r300->context.bind_fragment_sampler_states = r300_bind_sampler_states; 1104 r300->context.bind_vertex_sampler_states = r300_lacks_vertex_textures; 1105 r300->context.delete_sampler_state = r300_delete_sampler_state; 1106 1107 r300->context.set_fragment_sampler_textures = r300_set_sampler_textures; 1108 1109 r300->context.set_scissor_state = r300_set_scissor_state; 1110 1111 r300->context.set_viewport_state = r300_set_viewport_state; 1112 1113 r300->context.set_vertex_buffers = r300_set_vertex_buffers; 1114 r300->context.set_vertex_elements = r300_set_vertex_elements; 1115 1116 r300->context.create_vs_state = r300_create_vs_state; 1117 r300->context.bind_vs_state = r300_bind_vs_state; 1118 r300->context.delete_vs_state = r300_delete_vs_state; 1119} 1120