lp_state_fs.c revision 078eff659a7ef90691966d983f35ed9e4ce63901
1/************************************************************************** 2 * 3 * Copyright 2009 VMware, Inc. 4 * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas. 5 * All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the 9 * "Software"), to deal in the Software without restriction, including 10 * without limitation the rights to use, copy, modify, merge, publish, 11 * distribute, sub license, and/or sell copies of the Software, and to 12 * permit persons to whom the Software is furnished to do so, subject to 13 * the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the 16 * next paragraph) shall be included in all copies or substantial portions 17 * of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 22 * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR 23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 26 * 27 **************************************************************************/ 28 29/** 30 * @file 31 * Code generate the whole fragment pipeline. 32 * 33 * The fragment pipeline consists of the following stages: 34 * - early depth test 35 * - fragment shader 36 * - alpha test 37 * - depth/stencil test 38 * - blending 39 * 40 * This file has only the glue to assemble the fragment pipeline. The actual 41 * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the 42 * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we 43 * muster the LLVM JIT execution engine to create a function that follows an 44 * established binary interface and that can be called from C directly. 45 * 46 * A big source of complexity here is that we often want to run different 47 * stages with different precisions and data types and precisions. For example, 48 * the fragment shader needs typically to be done in floats, but the 49 * depth/stencil test and blending is better done in the type that most closely 50 * matches the depth/stencil and color buffer respectively. 51 * 52 * Since the width of a SIMD vector register stays the same regardless of the 53 * element type, different types imply different number of elements, so we must 54 * code generate more instances of the stages with larger types to be able to 55 * feed/consume the stages with smaller types. 56 * 57 * @author Jose Fonseca <jfonseca@vmware.com> 58 */ 59 60#include <limits.h> 61#include "pipe/p_defines.h" 62#include "util/u_inlines.h" 63#include "util/u_memory.h" 64#include "util/u_pointer.h" 65#include "util/u_format.h" 66#include "util/u_dump.h" 67#include "util/u_string.h" 68#include "util/u_simple_list.h" 69#include "os/os_time.h" 70#include "pipe/p_shader_tokens.h" 71#include "draw/draw_context.h" 72#include "tgsi/tgsi_dump.h" 73#include "tgsi/tgsi_scan.h" 74#include "tgsi/tgsi_parse.h" 75#include "gallivm/lp_bld_type.h" 76#include "gallivm/lp_bld_const.h" 77#include "gallivm/lp_bld_conv.h" 78#include "gallivm/lp_bld_init.h" 79#include "gallivm/lp_bld_intr.h" 80#include "gallivm/lp_bld_logic.h" 81#include "gallivm/lp_bld_tgsi.h" 82#include "gallivm/lp_bld_swizzle.h" 83#include "gallivm/lp_bld_flow.h" 84#include "gallivm/lp_bld_debug.h" 85 86#include "lp_bld_alpha.h" 87#include "lp_bld_blend.h" 88#include "lp_bld_depth.h" 89#include "lp_bld_interp.h" 90#include "lp_context.h" 91#include "lp_debug.h" 92#include "lp_perf.h" 93#include "lp_screen.h" 94#include "lp_setup.h" 95#include "lp_state.h" 96#include "lp_tex_sample.h" 97#include "lp_flush.h" 98#include "lp_state_fs.h" 99 100 101#include <llvm-c/Analysis.h> 102 103 104static unsigned fs_no = 0; 105 106 107/** 108 * Generate the depth /stencil test code. 109 */ 110static void 111generate_depth_stencil(LLVMBuilderRef builder, 112 const struct lp_fragment_shader_variant_key *key, 113 struct lp_type src_type, 114 struct lp_build_mask_context *mask, 115 LLVMValueRef stencil_refs[2], 116 LLVMValueRef src, 117 LLVMValueRef dst_ptr, 118 LLVMValueRef facing, 119 LLVMValueRef counter) 120{ 121 const struct util_format_description *format_desc; 122 struct lp_type dst_type; 123 124 if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled) 125 return; 126 127 format_desc = util_format_description(key->zsbuf_format); 128 assert(format_desc); 129 130 /* 131 * Depths are expected to be between 0 and 1, even if they are stored in 132 * floats. Setting these bits here will ensure that the lp_build_conv() call 133 * below won't try to unnecessarily clamp the incoming values. 134 */ 135 if(src_type.floating) { 136 src_type.sign = FALSE; 137 src_type.norm = TRUE; 138 } 139 else { 140 assert(!src_type.sign); 141 assert(src_type.norm); 142 } 143 144 /* Pick the depth type. */ 145 dst_type = lp_depth_type(format_desc, src_type.width*src_type.length); 146 147 /* FIXME: Cope with a depth test type with a different bit width. */ 148 assert(dst_type.width == src_type.width); 149 assert(dst_type.length == src_type.length); 150 151 /* Convert fragment Z from float to integer */ 152 lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1); 153 154 dst_ptr = LLVMBuildBitCast(builder, 155 dst_ptr, 156 LLVMPointerType(lp_build_vec_type(dst_type), 0), ""); 157 lp_build_depth_stencil_test(builder, 158 &key->depth, 159 key->stencil, 160 dst_type, 161 format_desc, 162 mask, 163 stencil_refs, 164 src, 165 dst_ptr, 166 facing, 167 counter); 168} 169 170 171/** 172 * Expand the relevent bits of mask_input to a 4-dword mask for the 173 * four pixels in a 2x2 quad. This will set the four elements of the 174 * quad mask vector to 0 or ~0. 175 * 176 * \param quad which quad of the quad group to test, in [0,3] 177 * \param mask_input bitwise mask for the whole 4x4 stamp 178 */ 179static LLVMValueRef 180generate_quad_mask(LLVMBuilderRef builder, 181 struct lp_type fs_type, 182 unsigned quad, 183 LLVMValueRef mask_input) /* int32 */ 184{ 185 struct lp_type mask_type; 186 LLVMTypeRef i32t = LLVMInt32Type(); 187 LLVMValueRef bits[4]; 188 LLVMValueRef mask; 189 190 /* 191 * XXX: We'll need a different path for 16 x u8 192 */ 193 assert(fs_type.width == 32); 194 assert(fs_type.length == 4); 195 mask_type = lp_int_type(fs_type); 196 197 /* 198 * mask_input >>= (quad * 4) 199 */ 200 201 mask_input = LLVMBuildLShr(builder, 202 mask_input, 203 LLVMConstInt(i32t, quad * 4, 0), 204 ""); 205 206 /* 207 * mask = { mask_input & (1 << i), for i in [0,3] } 208 */ 209 210 mask = lp_build_broadcast(builder, lp_build_vec_type(mask_type), mask_input); 211 212 bits[0] = LLVMConstInt(i32t, 1 << 0, 0); 213 bits[1] = LLVMConstInt(i32t, 1 << 1, 0); 214 bits[2] = LLVMConstInt(i32t, 1 << 2, 0); 215 bits[3] = LLVMConstInt(i32t, 1 << 3, 0); 216 217 mask = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), ""); 218 219 /* 220 * mask = mask != 0 ? ~0 : 0 221 */ 222 223 mask = lp_build_compare(builder, 224 mask_type, PIPE_FUNC_NOTEQUAL, 225 mask, 226 lp_build_const_int_vec(mask_type, 0)); 227 228 return mask; 229} 230 231 232 233/** 234 * Generate the fragment shader, depth/stencil test, and alpha tests. 235 * \param i which quad in the tile, in range [0,3] 236 * \param partial_mask if 1, do mask_input testing 237 */ 238static void 239generate_fs(struct llvmpipe_context *lp, 240 struct lp_fragment_shader *shader, 241 const struct lp_fragment_shader_variant_key *key, 242 LLVMBuilderRef builder, 243 struct lp_type type, 244 LLVMValueRef context_ptr, 245 unsigned i, 246 const struct lp_build_interp_soa_context *interp, 247 struct lp_build_sampler_soa *sampler, 248 LLVMValueRef *pmask, 249 LLVMValueRef (*color)[4], 250 LLVMValueRef depth_ptr, 251 LLVMValueRef facing, 252 unsigned partial_mask, 253 LLVMValueRef mask_input, 254 LLVMValueRef counter) 255{ 256 const struct tgsi_token *tokens = shader->base.tokens; 257 LLVMTypeRef vec_type; 258 LLVMValueRef consts_ptr; 259 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS]; 260 LLVMValueRef z = interp->pos[2]; 261 LLVMValueRef stencil_refs[2]; 262 struct lp_build_flow_context *flow; 263 struct lp_build_mask_context mask; 264 boolean early_depth_stencil_test; 265 unsigned attrib; 266 unsigned chan; 267 unsigned cbuf; 268 269 assert(i < 4); 270 271 stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr); 272 stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr); 273 274 vec_type = lp_build_vec_type(type); 275 276 consts_ptr = lp_jit_context_constants(builder, context_ptr); 277 278 flow = lp_build_flow_create(builder); 279 280 memset(outputs, 0, sizeof outputs); 281 282 lp_build_flow_scope_begin(flow); 283 284 /* Declare the color and z variables */ 285 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 286 for(chan = 0; chan < NUM_CHANNELS; ++chan) { 287 color[cbuf][chan] = LLVMGetUndef(vec_type); 288 lp_build_flow_scope_declare(flow, &color[cbuf][chan]); 289 } 290 } 291 lp_build_flow_scope_declare(flow, &z); 292 293 /* do triangle edge testing */ 294 if (partial_mask) { 295 *pmask = generate_quad_mask(builder, type, 296 i, mask_input); 297 } 298 else { 299 *pmask = lp_build_const_int_vec(type, ~0); 300 } 301 302 /* 'mask' will control execution based on quad's pixel alive/killed state */ 303 lp_build_mask_begin(&mask, flow, type, *pmask); 304 305 early_depth_stencil_test = 306 (key->depth.enabled || key->stencil[0].enabled) && 307 !key->alpha.enabled && 308 !shader->info.uses_kill && 309 !shader->info.writes_z; 310 311 if (early_depth_stencil_test) 312 generate_depth_stencil(builder, key, 313 type, &mask, 314 stencil_refs, z, depth_ptr, facing, counter); 315 316 lp_build_tgsi_soa(builder, tokens, type, &mask, 317 consts_ptr, interp->pos, interp->inputs, 318 outputs, sampler, &shader->info); 319 320 /* loop over fragment shader outputs/results */ 321 for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) { 322 for(chan = 0; chan < NUM_CHANNELS; ++chan) { 323 if(outputs[attrib][chan]) { 324 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], ""); 325 lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]); 326 327 switch (shader->info.output_semantic_name[attrib]) { 328 case TGSI_SEMANTIC_COLOR: 329 { 330 unsigned cbuf = shader->info.output_semantic_index[attrib]; 331 332 lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]); 333 334 /* Alpha test */ 335 /* XXX: should the alpha reference value be passed separately? */ 336 /* XXX: should only test the final assignment to alpha */ 337 if(cbuf == 0 && chan == 3) { 338 LLVMValueRef alpha = out; 339 LLVMValueRef alpha_ref_value; 340 alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr); 341 alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value); 342 lp_build_alpha_test(builder, &key->alpha, type, 343 &mask, alpha, alpha_ref_value); 344 } 345 346 color[cbuf][chan] = out; 347 break; 348 } 349 350 case TGSI_SEMANTIC_POSITION: 351 if(chan == 2) 352 z = out; 353 break; 354 } 355 } 356 } 357 } 358 359 if (!early_depth_stencil_test) 360 generate_depth_stencil(builder, key, 361 type, &mask, 362 stencil_refs, z, depth_ptr, facing, counter); 363 364 lp_build_mask_end(&mask); 365 366 lp_build_flow_scope_end(flow); 367 368 lp_build_flow_destroy(flow); 369 370 *pmask = mask.value; 371 372} 373 374 375/** 376 * Generate color blending and color output. 377 * \param rt the render target index (to index blend, colormask state) 378 * \param type the pixel color type 379 * \param context_ptr pointer to the runtime JIT context 380 * \param mask execution mask (active fragment/pixel mask) 381 * \param src colors from the fragment shader 382 * \param dst_ptr the destination color buffer pointer 383 */ 384static void 385generate_blend(const struct pipe_blend_state *blend, 386 unsigned rt, 387 LLVMBuilderRef builder, 388 struct lp_type type, 389 LLVMValueRef context_ptr, 390 LLVMValueRef mask, 391 LLVMValueRef *src, 392 LLVMValueRef dst_ptr) 393{ 394 struct lp_build_context bld; 395 struct lp_build_flow_context *flow; 396 struct lp_build_mask_context mask_ctx; 397 LLVMTypeRef vec_type; 398 LLVMValueRef const_ptr; 399 LLVMValueRef con[4]; 400 LLVMValueRef dst[4]; 401 LLVMValueRef res[4]; 402 unsigned chan; 403 404 lp_build_context_init(&bld, builder, type); 405 406 flow = lp_build_flow_create(builder); 407 408 /* we'll use this mask context to skip blending if all pixels are dead */ 409 lp_build_mask_begin(&mask_ctx, flow, type, mask); 410 411 vec_type = lp_build_vec_type(type); 412 413 const_ptr = lp_jit_context_blend_color(builder, context_ptr); 414 const_ptr = LLVMBuildBitCast(builder, const_ptr, 415 LLVMPointerType(vec_type, 0), ""); 416 417 /* load constant blend color and colors from the dest color buffer */ 418 for(chan = 0; chan < 4; ++chan) { 419 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); 420 con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), ""); 421 422 dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), ""); 423 424 lp_build_name(con[chan], "con.%c", "rgba"[chan]); 425 lp_build_name(dst[chan], "dst.%c", "rgba"[chan]); 426 } 427 428 /* do blend */ 429 lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res); 430 431 /* store results to color buffer */ 432 for(chan = 0; chan < 4; ++chan) { 433 if(blend->rt[rt].colormask & (1 << chan)) { 434 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0); 435 lp_build_name(res[chan], "res.%c", "rgba"[chan]); 436 res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]); 437 LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, "")); 438 } 439 } 440 441 lp_build_mask_end(&mask_ctx); 442 lp_build_flow_destroy(flow); 443} 444 445 446/** 447 * Generate the runtime callable function for the whole fragment pipeline. 448 * Note that the function which we generate operates on a block of 16 449 * pixels at at time. The block contains 2x2 quads. Each quad contains 450 * 2x2 pixels. 451 */ 452static void 453generate_fragment(struct llvmpipe_context *lp, 454 struct lp_fragment_shader *shader, 455 struct lp_fragment_shader_variant *variant, 456 unsigned partial_mask) 457{ 458 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen); 459 const struct lp_fragment_shader_variant_key *key = &variant->key; 460 char func_name[256]; 461 struct lp_type fs_type; 462 struct lp_type blend_type; 463 LLVMTypeRef fs_elem_type; 464 LLVMTypeRef fs_int_vec_type; 465 LLVMTypeRef blend_vec_type; 466 LLVMTypeRef arg_types[11]; 467 LLVMTypeRef func_type; 468 LLVMValueRef context_ptr; 469 LLVMValueRef x; 470 LLVMValueRef y; 471 LLVMValueRef a0_ptr; 472 LLVMValueRef dadx_ptr; 473 LLVMValueRef dady_ptr; 474 LLVMValueRef color_ptr_ptr; 475 LLVMValueRef depth_ptr; 476 LLVMValueRef mask_input; 477 LLVMValueRef counter = NULL; 478 LLVMBasicBlockRef block; 479 LLVMBuilderRef builder; 480 struct lp_build_sampler_soa *sampler; 481 struct lp_build_interp_soa_context interp; 482 LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH]; 483 LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH]; 484 LLVMValueRef blend_mask; 485 LLVMValueRef function; 486 LLVMValueRef facing; 487 unsigned num_fs; 488 unsigned i; 489 unsigned chan; 490 unsigned cbuf; 491 492 493 /* TODO: actually pick these based on the fs and color buffer 494 * characteristics. */ 495 496 memset(&fs_type, 0, sizeof fs_type); 497 fs_type.floating = TRUE; /* floating point values */ 498 fs_type.sign = TRUE; /* values are signed */ 499 fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ 500 fs_type.width = 32; /* 32-bit float */ 501 fs_type.length = 4; /* 4 elements per vector */ 502 num_fs = 4; /* number of quads per block */ 503 504 memset(&blend_type, 0, sizeof blend_type); 505 blend_type.floating = FALSE; /* values are integers */ 506 blend_type.sign = FALSE; /* values are unsigned */ 507 blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */ 508 blend_type.width = 8; /* 8-bit ubyte values */ 509 blend_type.length = 16; /* 16 elements per vector */ 510 511 /* 512 * Generate the function prototype. Any change here must be reflected in 513 * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa. 514 */ 515 516 fs_elem_type = lp_build_elem_type(fs_type); 517 fs_int_vec_type = lp_build_int_vec_type(fs_type); 518 519 blend_vec_type = lp_build_vec_type(blend_type); 520 521 util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s", 522 shader->no, variant->no, partial_mask ? "partial" : "whole"); 523 524 arg_types[0] = screen->context_ptr_type; /* context */ 525 arg_types[1] = LLVMInt32Type(); /* x */ 526 arg_types[2] = LLVMInt32Type(); /* y */ 527 arg_types[3] = LLVMFloatType(); /* facing */ 528 arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */ 529 arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */ 530 arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */ 531 arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */ 532 arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */ 533 arg_types[9] = LLVMInt32Type(); /* mask_input */ 534 arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */ 535 536 func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0); 537 538 function = LLVMAddFunction(screen->module, func_name, func_type); 539 LLVMSetFunctionCallConv(function, LLVMCCallConv); 540 541 variant->function[partial_mask] = function; 542 543 544 /* XXX: need to propagate noalias down into color param now we are 545 * passing a pointer-to-pointer? 546 */ 547 for(i = 0; i < Elements(arg_types); ++i) 548 if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) 549 LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute); 550 551 context_ptr = LLVMGetParam(function, 0); 552 x = LLVMGetParam(function, 1); 553 y = LLVMGetParam(function, 2); 554 facing = LLVMGetParam(function, 3); 555 a0_ptr = LLVMGetParam(function, 4); 556 dadx_ptr = LLVMGetParam(function, 5); 557 dady_ptr = LLVMGetParam(function, 6); 558 color_ptr_ptr = LLVMGetParam(function, 7); 559 depth_ptr = LLVMGetParam(function, 8); 560 mask_input = LLVMGetParam(function, 9); 561 562 lp_build_name(context_ptr, "context"); 563 lp_build_name(x, "x"); 564 lp_build_name(y, "y"); 565 lp_build_name(a0_ptr, "a0"); 566 lp_build_name(dadx_ptr, "dadx"); 567 lp_build_name(dady_ptr, "dady"); 568 lp_build_name(color_ptr_ptr, "color_ptr_ptr"); 569 lp_build_name(depth_ptr, "depth"); 570 lp_build_name(mask_input, "mask_input"); 571 572 if (key->occlusion_count) { 573 counter = LLVMGetParam(function, 10); 574 lp_build_name(counter, "counter"); 575 } 576 577 /* 578 * Function body 579 */ 580 581 block = LLVMAppendBasicBlock(function, "entry"); 582 builder = LLVMCreateBuilder(); 583 LLVMPositionBuilderAtEnd(builder, block); 584 585 /* 586 * The shader input interpolation info is not explicitely baked in the 587 * shader key, but everything it derives from (TGSI, and flatshade) is 588 * already included in the shader key. 589 */ 590 lp_build_interp_soa_init(&interp, 591 lp->num_inputs, 592 lp->inputs, 593 builder, fs_type, 594 a0_ptr, dadx_ptr, dady_ptr, 595 x, y); 596 597 /* code generated texture sampling */ 598 sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr); 599 600 /* loop over quads in the block */ 601 for(i = 0; i < num_fs; ++i) { 602 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0); 603 LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS]; 604 LLVMValueRef depth_ptr_i; 605 606 if(i != 0) 607 lp_build_interp_soa_update(&interp, i); 608 609 depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, ""); 610 611 generate_fs(lp, shader, key, 612 builder, 613 fs_type, 614 context_ptr, 615 i, 616 &interp, 617 sampler, 618 &fs_mask[i], /* output */ 619 out_color, 620 depth_ptr_i, 621 facing, 622 partial_mask, 623 mask_input, 624 counter); 625 626 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) 627 for(chan = 0; chan < NUM_CHANNELS; ++chan) 628 fs_out_color[cbuf][chan][i] = out_color[cbuf][chan]; 629 } 630 631 sampler->destroy(sampler); 632 633 /* Loop over color outputs / color buffers to do blending. 634 */ 635 for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) { 636 LLVMValueRef color_ptr; 637 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0); 638 LLVMValueRef blend_in_color[NUM_CHANNELS]; 639 unsigned rt; 640 641 /* 642 * Convert the fs's output color and mask to fit to the blending type. 643 */ 644 for(chan = 0; chan < NUM_CHANNELS; ++chan) { 645 lp_build_conv(builder, fs_type, blend_type, 646 fs_out_color[cbuf][chan], num_fs, 647 &blend_in_color[chan], 1); 648 lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]); 649 } 650 651 if (partial_mask || !variant->opaque) { 652 lp_build_conv_mask(builder, fs_type, blend_type, 653 fs_mask, num_fs, 654 &blend_mask, 1); 655 } else { 656 blend_mask = lp_build_const_int_vec(blend_type, ~0); 657 } 658 659 color_ptr = LLVMBuildLoad(builder, 660 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""), 661 ""); 662 lp_build_name(color_ptr, "color_ptr%d", cbuf); 663 664 /* which blend/colormask state to use */ 665 rt = key->blend.independent_blend_enable ? cbuf : 0; 666 667 /* 668 * Blending. 669 */ 670 generate_blend(&key->blend, 671 rt, 672 builder, 673 blend_type, 674 context_ptr, 675 blend_mask, 676 blend_in_color, 677 color_ptr); 678 } 679 680#ifdef PIPE_ARCH_X86 681 /* Avoid corrupting the FPU stack on 32bit OSes. */ 682 lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0); 683#endif 684 685 LLVMBuildRetVoid(builder); 686 687 LLVMDisposeBuilder(builder); 688 689 690 /* Verify the LLVM IR. If invalid, dump and abort */ 691#ifdef DEBUG 692 if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) { 693 if (1) 694 lp_debug_dump_value(function); 695 abort(); 696 } 697#endif 698 699 /* Apply optimizations to LLVM IR */ 700 LLVMRunFunctionPassManager(screen->pass, function); 701 702 if (gallivm_debug & GALLIVM_DEBUG_IR) { 703 /* Print the LLVM IR to stderr */ 704 lp_debug_dump_value(function); 705 debug_printf("\n"); 706 } 707 708 /* 709 * Translate the LLVM IR into machine code. 710 */ 711 { 712 void *f = LLVMGetPointerToGlobal(screen->engine, function); 713 714 variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f); 715 716 if (gallivm_debug & GALLIVM_DEBUG_ASM) { 717 lp_disassemble(f); 718 } 719 lp_func_delete_body(function); 720 } 721} 722 723 724static void 725dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key) 726{ 727 unsigned i; 728 729 debug_printf("fs variant %p:\n", (void *) key); 730 731 if (key->depth.enabled) { 732 debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format)); 733 debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE)); 734 debug_printf("depth.writemask = %u\n", key->depth.writemask); 735 } 736 737 for (i = 0; i < 2; ++i) { 738 if (key->stencil[i].enabled) { 739 debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE)); 740 debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE)); 741 debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE)); 742 debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE)); 743 debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask); 744 debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask); 745 } 746 } 747 748 if (key->alpha.enabled) { 749 debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE)); 750 debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value); 751 } 752 753 if (key->blend.logicop_enable) { 754 debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE)); 755 } 756 else if (key->blend.rt[0].blend_enable) { 757 debug_printf("blend.rgb_func = %s\n", util_dump_blend_func (key->blend.rt[0].rgb_func, TRUE)); 758 debug_printf("blend.rgb_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE)); 759 debug_printf("blend.rgb_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE)); 760 debug_printf("blend.alpha_func = %s\n", util_dump_blend_func (key->blend.rt[0].alpha_func, TRUE)); 761 debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE)); 762 debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE)); 763 } 764 debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask); 765 for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) { 766 if (key->sampler[i].format) { 767 debug_printf("sampler[%u] = \n", i); 768 debug_printf(" .format = %s\n", 769 util_format_name(key->sampler[i].format)); 770 debug_printf(" .target = %s\n", 771 util_dump_tex_target(key->sampler[i].target, TRUE)); 772 debug_printf(" .pot = %u %u %u\n", 773 key->sampler[i].pot_width, 774 key->sampler[i].pot_height, 775 key->sampler[i].pot_depth); 776 debug_printf(" .wrap = %s %s %s\n", 777 util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE), 778 util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE), 779 util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE)); 780 debug_printf(" .min_img_filter = %s\n", 781 util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE)); 782 debug_printf(" .min_mip_filter = %s\n", 783 util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE)); 784 debug_printf(" .mag_img_filter = %s\n", 785 util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE)); 786 if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE) 787 debug_printf(" .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE)); 788 debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords); 789 } 790 } 791} 792 793 794 795static struct lp_fragment_shader_variant * 796generate_variant(struct llvmpipe_context *lp, 797 struct lp_fragment_shader *shader, 798 const struct lp_fragment_shader_variant_key *key) 799{ 800 struct lp_fragment_shader_variant *variant; 801 802 variant = CALLOC_STRUCT(lp_fragment_shader_variant); 803 if(!variant) 804 return NULL; 805 806 variant->shader = shader; 807 variant->list_item_global.base = variant; 808 variant->list_item_local.base = variant; 809 variant->no = shader->variants_created++; 810 811 memcpy(&variant->key, key, sizeof *key); 812 813 if (gallivm_debug & GALLIVM_DEBUG_IR) { 814 debug_printf("llvmpipe: Creating fragment shader #%u variant #%u:\n", 815 shader->no, variant->no); 816 tgsi_dump(shader->base.tokens, 0); 817 dump_fs_variant_key(key); 818 } 819 820 generate_fragment(lp, shader, variant, RAST_WHOLE); 821 generate_fragment(lp, shader, variant, RAST_EDGE_TEST); 822 823 /* TODO: most of these can be relaxed, in particular the colormask */ 824 variant->opaque = 825 !key->blend.logicop_enable && 826 !key->blend.rt[0].blend_enable && 827 key->blend.rt[0].colormask == 0xf && 828 !key->stencil[0].enabled && 829 !key->alpha.enabled && 830 !key->depth.enabled && 831 !shader->info.uses_kill 832 ? TRUE : FALSE; 833 834 return variant; 835} 836 837 838static void * 839llvmpipe_create_fs_state(struct pipe_context *pipe, 840 const struct pipe_shader_state *templ) 841{ 842 struct lp_fragment_shader *shader; 843 844 shader = CALLOC_STRUCT(lp_fragment_shader); 845 if (!shader) 846 return NULL; 847 848 shader->no = fs_no++; 849 make_empty_list(&shader->variants); 850 851 /* get/save the summary info for this shader */ 852 tgsi_scan_shader(templ->tokens, &shader->info); 853 854 /* we need to keep a local copy of the tokens */ 855 shader->base.tokens = tgsi_dup_tokens(templ->tokens); 856 857 if (LP_DEBUG & DEBUG_TGSI) { 858 unsigned attrib; 859 debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader); 860 tgsi_dump(templ->tokens, 0); 861 debug_printf("usage masks:\n"); 862 for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) { 863 unsigned usage_mask = shader->info.input_usage_mask[attrib]; 864 debug_printf(" IN[%u].%s%s%s%s\n", 865 attrib, 866 usage_mask & TGSI_WRITEMASK_X ? "x" : "", 867 usage_mask & TGSI_WRITEMASK_Y ? "y" : "", 868 usage_mask & TGSI_WRITEMASK_Z ? "z" : "", 869 usage_mask & TGSI_WRITEMASK_W ? "w" : ""); 870 } 871 debug_printf("\n"); 872 } 873 874 return shader; 875} 876 877 878static void 879llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs) 880{ 881 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 882 883 if (llvmpipe->fs == fs) 884 return; 885 886 draw_flush(llvmpipe->draw); 887 888 llvmpipe->fs = fs; 889 890 llvmpipe->dirty |= LP_NEW_FS; 891} 892 893static void 894remove_shader_variant(struct llvmpipe_context *lp, 895 struct lp_fragment_shader_variant *variant) 896{ 897 struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen); 898 unsigned i; 899 900 if (gallivm_debug & GALLIVM_DEBUG_IR) { 901 debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n", 902 variant->shader->no, variant->no, variant->shader->variants_created, 903 variant->shader->variants_cached, lp->nr_fs_variants); 904 } 905 for (i = 0; i < Elements(variant->function); i++) { 906 if (variant->function[i]) { 907 if (variant->jit_function[i]) 908 LLVMFreeMachineCodeForFunction(screen->engine, 909 variant->function[i]); 910 LLVMDeleteFunction(variant->function[i]); 911 } 912 } 913 remove_from_list(&variant->list_item_local); 914 variant->shader->variants_cached--; 915 remove_from_list(&variant->list_item_global); 916 lp->nr_fs_variants--; 917 FREE(variant); 918} 919 920static void 921llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs) 922{ 923 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 924 struct pipe_fence_handle *fence = NULL; 925 struct lp_fragment_shader *shader = fs; 926 struct lp_fs_variant_list_item *li; 927 928 assert(fs != llvmpipe->fs); 929 (void) llvmpipe; 930 931 /* 932 * XXX: we need to flush the context until we have some sort of reference 933 * counting in fragment shaders as they may still be binned 934 * Flushing alone might not sufficient we need to wait on it too. 935 */ 936 937 llvmpipe_flush(pipe, 0, &fence); 938 939 if (fence) { 940 pipe->screen->fence_finish(pipe->screen, fence, 0); 941 pipe->screen->fence_reference(pipe->screen, &fence, NULL); 942 } 943 944 li = first_elem(&shader->variants); 945 while(!at_end(&shader->variants, li)) { 946 struct lp_fs_variant_list_item *next = next_elem(li); 947 remove_shader_variant(llvmpipe, li->base); 948 li = next; 949 } 950 951 assert(shader->variants_cached == 0); 952 FREE((void *) shader->base.tokens); 953 FREE(shader); 954} 955 956 957 958static void 959llvmpipe_set_constant_buffer(struct pipe_context *pipe, 960 uint shader, uint index, 961 struct pipe_resource *constants) 962{ 963 struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe); 964 unsigned size = constants ? constants->width0 : 0; 965 const void *data = constants ? llvmpipe_resource_data(constants) : NULL; 966 967 assert(shader < PIPE_SHADER_TYPES); 968 assert(index < PIPE_MAX_CONSTANT_BUFFERS); 969 970 if(llvmpipe->constants[shader][index] == constants) 971 return; 972 973 draw_flush(llvmpipe->draw); 974 975 /* note: reference counting */ 976 pipe_resource_reference(&llvmpipe->constants[shader][index], constants); 977 978 if(shader == PIPE_SHADER_VERTEX || 979 shader == PIPE_SHADER_GEOMETRY) { 980 draw_set_mapped_constant_buffer(llvmpipe->draw, shader, 981 index, data, size); 982 } 983 984 llvmpipe->dirty |= LP_NEW_CONSTANTS; 985} 986 987 988/** 989 * Return the blend factor equivalent to a destination alpha of one. 990 */ 991static INLINE unsigned 992force_dst_alpha_one(unsigned factor, boolean alpha) 993{ 994 switch(factor) { 995 case PIPE_BLENDFACTOR_DST_ALPHA: 996 return PIPE_BLENDFACTOR_ONE; 997 case PIPE_BLENDFACTOR_INV_DST_ALPHA: 998 return PIPE_BLENDFACTOR_ZERO; 999 case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: 1000 return PIPE_BLENDFACTOR_ZERO; 1001 } 1002 1003 if (alpha) { 1004 switch(factor) { 1005 case PIPE_BLENDFACTOR_DST_COLOR: 1006 return PIPE_BLENDFACTOR_ONE; 1007 case PIPE_BLENDFACTOR_INV_DST_COLOR: 1008 return PIPE_BLENDFACTOR_ZERO; 1009 } 1010 } 1011 1012 return factor; 1013} 1014 1015 1016/** 1017 * We need to generate several variants of the fragment pipeline to match 1018 * all the combinations of the contributing state atoms. 1019 * 1020 * TODO: there is actually no reason to tie this to context state -- the 1021 * generated code could be cached globally in the screen. 1022 */ 1023static void 1024make_variant_key(struct llvmpipe_context *lp, 1025 struct lp_fragment_shader *shader, 1026 struct lp_fragment_shader_variant_key *key) 1027{ 1028 unsigned i; 1029 1030 memset(key, 0, sizeof *key); 1031 1032 if (lp->framebuffer.zsbuf) { 1033 if (lp->depth_stencil->depth.enabled) { 1034 key->zsbuf_format = lp->framebuffer.zsbuf->format; 1035 memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth); 1036 } 1037 if (lp->depth_stencil->stencil[0].enabled) { 1038 key->zsbuf_format = lp->framebuffer.zsbuf->format; 1039 memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil); 1040 } 1041 } 1042 1043 key->alpha.enabled = lp->depth_stencil->alpha.enabled; 1044 if(key->alpha.enabled) 1045 key->alpha.func = lp->depth_stencil->alpha.func; 1046 /* alpha.ref_value is passed in jit_context */ 1047 1048 key->flatshade = lp->rasterizer->flatshade; 1049 if (lp->active_query_count) { 1050 key->occlusion_count = TRUE; 1051 } 1052 1053 if (lp->framebuffer.nr_cbufs) { 1054 memcpy(&key->blend, lp->blend, sizeof key->blend); 1055 } 1056 1057 key->nr_cbufs = lp->framebuffer.nr_cbufs; 1058 for (i = 0; i < lp->framebuffer.nr_cbufs; i++) { 1059 struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i]; 1060 const struct util_format_description *format_desc; 1061 unsigned chan; 1062 1063 format_desc = util_format_description(lp->framebuffer.cbufs[i]->format); 1064 assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB || 1065 format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB); 1066 1067 blend_rt->colormask = lp->blend->rt[i].colormask; 1068 1069 /* mask out color channels not present in the color buffer. 1070 * Should be simple to incorporate per-cbuf writemasks: 1071 */ 1072 for(chan = 0; chan < 4; ++chan) { 1073 enum util_format_swizzle swizzle = format_desc->swizzle[chan]; 1074 1075 if(swizzle > UTIL_FORMAT_SWIZZLE_W) 1076 blend_rt->colormask &= ~(1 << chan); 1077 } 1078 1079 /* 1080 * Our swizzled render tiles always have an alpha channel, but the linear 1081 * render target format often does not, so force here the dst alpha to be 1082 * one. 1083 * 1084 * This is not a mere optimization. Wrong results will be produced if the 1085 * dst alpha is used, the dst format does not have alpha, and the previous 1086 * rendering was not flushed from the swizzled to linear buffer. For 1087 * example, NonPowTwo DCT. 1088 * 1089 * TODO: This should be generalized to all channels for better 1090 * performance, but only alpha causes correctness issues. 1091 */ 1092 if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) { 1093 blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE); 1094 blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE); 1095 blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE); 1096 blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE); 1097 } 1098 } 1099 1100 for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) 1101 if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) 1102 lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i], lp->sampler[i]); 1103} 1104 1105/** 1106 * Update fragment state. This is called just prior to drawing 1107 * something when some fragment-related state has changed. 1108 */ 1109void 1110llvmpipe_update_fs(struct llvmpipe_context *lp) 1111{ 1112 struct lp_fragment_shader *shader = lp->fs; 1113 struct lp_fragment_shader_variant_key key; 1114 struct lp_fragment_shader_variant *variant = NULL; 1115 struct lp_fs_variant_list_item *li; 1116 1117 make_variant_key(lp, shader, &key); 1118 1119 li = first_elem(&shader->variants); 1120 while(!at_end(&shader->variants, li)) { 1121 if(memcmp(&li->base->key, &key, sizeof key) == 0) { 1122 variant = li->base; 1123 break; 1124 } 1125 li = next_elem(li); 1126 } 1127 1128 if (variant) { 1129 move_to_head(&lp->fs_variants_list, &variant->list_item_global); 1130 } 1131 else { 1132 int64_t t0, t1; 1133 int64_t dt; 1134 unsigned i; 1135 if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) { 1136 struct pipe_context *pipe = &lp->pipe; 1137 struct pipe_fence_handle *fence = NULL; 1138 1139 /* 1140 * XXX: we need to flush the context until we have some sort of reference 1141 * counting in fragment shaders as they may still be binned 1142 * Flushing alone might not be sufficient we need to wait on it too. 1143 */ 1144 llvmpipe_flush(pipe, 0, &fence); 1145 1146 if (fence) { 1147 pipe->screen->fence_finish(pipe->screen, fence, 0); 1148 pipe->screen->fence_reference(pipe->screen, &fence, NULL); 1149 } 1150 for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) { 1151 struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list); 1152 remove_shader_variant(lp, item->base); 1153 } 1154 } 1155 t0 = os_time_get(); 1156 1157 variant = generate_variant(lp, shader, &key); 1158 1159 t1 = os_time_get(); 1160 dt = t1 - t0; 1161 LP_COUNT_ADD(llvm_compile_time, dt); 1162 LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */ 1163 1164 if (variant) { 1165 insert_at_head(&shader->variants, &variant->list_item_local); 1166 insert_at_head(&lp->fs_variants_list, &variant->list_item_global); 1167 lp->nr_fs_variants++; 1168 shader->variants_cached++; 1169 } 1170 } 1171 1172 lp_setup_set_fs_variant(lp->setup, variant); 1173} 1174 1175 1176 1177void 1178llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe) 1179{ 1180 llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state; 1181 llvmpipe->pipe.bind_fs_state = llvmpipe_bind_fs_state; 1182 llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state; 1183 1184 llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer; 1185} 1186