draw_llvm.c revision 6b63e25b3d7a6ac0bd738c139ead0c7e7ad84368
1/************************************************************************** 2 * 3 * Copyright 2010 VMware, Inc. 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR 22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 28#include "draw_llvm.h" 29 30#include "draw_context.h" 31#include "draw_vs.h" 32 33#include "gallivm/lp_bld_arit.h" 34#include "gallivm/lp_bld_logic.h" 35#include "gallivm/lp_bld_const.h" 36#include "gallivm/lp_bld_swizzle.h" 37#include "gallivm/lp_bld_struct.h" 38#include "gallivm/lp_bld_type.h" 39#include "gallivm/lp_bld_flow.h" 40#include "gallivm/lp_bld_debug.h" 41#include "gallivm/lp_bld_tgsi.h" 42#include "gallivm/lp_bld_printf.h" 43#include "gallivm/lp_bld_intr.h" 44#include "gallivm/lp_bld_init.h" 45#include "gallivm/lp_bld_type.h" 46 47#include "tgsi/tgsi_exec.h" 48#include "tgsi/tgsi_dump.h" 49 50#include "util/u_math.h" 51#include "util/u_pointer.h" 52#include "util/u_string.h" 53#include "util/u_simple_list.h" 54 55 56#define DEBUG_STORE 0 57 58 59/** 60 * This function is called by the gallivm "garbage collector" when 61 * the LLVM global data structures are freed. We must free all LLVM-related 62 * data. Specifically, all JIT'd shader variants. 63 */ 64static void 65draw_llvm_garbage_collect_callback(void *cb_data) 66{ 67 struct draw_llvm *llvm = (struct draw_llvm *) cb_data; 68 struct draw_context *draw = llvm->draw; 69 struct draw_llvm_variant_list_item *li; 70 71 /* Ensure prepare will be run and shaders recompiled */ 72 assert(!draw->suspend_flushing); 73 draw_do_flush(draw, DRAW_FLUSH_STATE_CHANGE); 74 75 /* free all shader variants */ 76 li = first_elem(&llvm->vs_variants_list); 77 while (!at_end(&llvm->vs_variants_list, li)) { 78 struct draw_llvm_variant_list_item *next = next_elem(li); 79 draw_llvm_destroy_variant(li->base); 80 li = next; 81 } 82 83 /* Null-out these pointers so they get remade next time they're needed. 84 * See the accessor functions below. 85 */ 86 llvm->context_ptr_type = NULL; 87 llvm->buffer_ptr_type = NULL; 88 llvm->vb_ptr_type = NULL; 89 llvm->vertex_header_ptr_type = NULL; 90} 91 92 93static void 94draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var, 95 boolean elts); 96 97 98/** 99 * Create LLVM type for struct draw_jit_texture 100 */ 101static LLVMTypeRef 102create_jit_texture_type(struct gallivm_state *gallivm, const char *struct_name) 103{ 104 LLVMTargetDataRef target = gallivm->target; 105 LLVMTypeRef texture_type; 106 LLVMTypeRef elem_types[DRAW_JIT_TEXTURE_NUM_FIELDS]; 107 LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context); 108 109 elem_types[DRAW_JIT_TEXTURE_WIDTH] = 110 elem_types[DRAW_JIT_TEXTURE_HEIGHT] = 111 elem_types[DRAW_JIT_TEXTURE_DEPTH] = 112 elem_types[DRAW_JIT_TEXTURE_FIRST_LEVEL] = 113 elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = int32_type; 114 elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] = 115 elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] = 116 LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS); 117 elem_types[DRAW_JIT_TEXTURE_DATA] = 118 LLVMArrayType(LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0), 119 PIPE_MAX_TEXTURE_LEVELS); 120 elem_types[DRAW_JIT_TEXTURE_MIN_LOD] = 121 elem_types[DRAW_JIT_TEXTURE_MAX_LOD] = 122 elem_types[DRAW_JIT_TEXTURE_LOD_BIAS] = LLVMFloatTypeInContext(gallivm->context); 123 elem_types[DRAW_JIT_TEXTURE_BORDER_COLOR] = 124 LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4); 125 126#if HAVE_LLVM >= 0x0300 127 texture_type = LLVMStructCreateNamed(gallivm->context, struct_name); 128 LLVMStructSetBody(texture_type, elem_types, 129 Elements(elem_types), 0); 130#else 131 texture_type = LLVMStructTypeInContext(gallivm->context, elem_types, 132 Elements(elem_types), 0); 133 134 LLVMAddTypeName(gallivm->module, struct_name, texture_type); 135 136 /* Make sure the target's struct layout cache doesn't return 137 * stale/invalid data. 138 */ 139 LLVMInvalidateStructLayout(gallivm->target, texture_type); 140#endif 141 142 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width, 143 target, texture_type, 144 DRAW_JIT_TEXTURE_WIDTH); 145 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, height, 146 target, texture_type, 147 DRAW_JIT_TEXTURE_HEIGHT); 148 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, depth, 149 target, texture_type, 150 DRAW_JIT_TEXTURE_DEPTH); 151 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, first_level, 152 target, texture_type, 153 DRAW_JIT_TEXTURE_FIRST_LEVEL); 154 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, last_level, 155 target, texture_type, 156 DRAW_JIT_TEXTURE_LAST_LEVEL); 157 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, row_stride, 158 target, texture_type, 159 DRAW_JIT_TEXTURE_ROW_STRIDE); 160 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, img_stride, 161 target, texture_type, 162 DRAW_JIT_TEXTURE_IMG_STRIDE); 163 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data, 164 target, texture_type, 165 DRAW_JIT_TEXTURE_DATA); 166 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, min_lod, 167 target, texture_type, 168 DRAW_JIT_TEXTURE_MIN_LOD); 169 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, max_lod, 170 target, texture_type, 171 DRAW_JIT_TEXTURE_MAX_LOD); 172 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, lod_bias, 173 target, texture_type, 174 DRAW_JIT_TEXTURE_LOD_BIAS); 175 LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, border_color, 176 target, texture_type, 177 DRAW_JIT_TEXTURE_BORDER_COLOR); 178 179 LP_CHECK_STRUCT_SIZE(struct draw_jit_texture, target, texture_type); 180 181 return texture_type; 182} 183 184 185/** 186 * Create LLVM type for struct draw_jit_texture 187 */ 188static LLVMTypeRef 189create_jit_context_type(struct gallivm_state *gallivm, 190 LLVMTypeRef texture_type, const char *struct_name) 191{ 192 LLVMTargetDataRef target = gallivm->target; 193 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context); 194 LLVMTypeRef elem_types[5]; 195 LLVMTypeRef context_type; 196 197 elem_types[0] = LLVMPointerType(float_type, 0); /* vs_constants */ 198 elem_types[1] = LLVMPointerType(float_type, 0); /* gs_constants */ 199 elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4), 200 DRAW_TOTAL_CLIP_PLANES), 0); 201 elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */ 202 elem_types[4] = LLVMArrayType(texture_type, 203 PIPE_MAX_VERTEX_SAMPLERS); /* textures */ 204#if HAVE_LLVM >= 0x0300 205 context_type = LLVMStructCreateNamed(gallivm->context, struct_name); 206 LLVMStructSetBody(context_type, elem_types, 207 Elements(elem_types), 0); 208#else 209 context_type = LLVMStructTypeInContext(gallivm->context, elem_types, 210 Elements(elem_types), 0); 211 LLVMAddTypeName(gallivm->module, struct_name, context_type); 212 213 LLVMInvalidateStructLayout(gallivm->target, context_type); 214#endif 215 216 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants, 217 target, context_type, 0); 218 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, gs_constants, 219 target, context_type, 1); 220 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes, 221 target, context_type, 2); 222 LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures, 223 target, context_type, 224 DRAW_JIT_CTX_TEXTURES); 225 LP_CHECK_STRUCT_SIZE(struct draw_jit_context, 226 target, context_type); 227 228 return context_type; 229} 230 231 232/** 233 * Create LLVM type for struct pipe_vertex_buffer 234 */ 235static LLVMTypeRef 236create_jit_vertex_buffer_type(struct gallivm_state *gallivm, const char *struct_name) 237{ 238 LLVMTargetDataRef target = gallivm->target; 239 LLVMTypeRef elem_types[3]; 240 LLVMTypeRef vb_type; 241 242 elem_types[0] = 243 elem_types[1] = LLVMInt32TypeInContext(gallivm->context); 244 elem_types[2] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); /* vs_constants */ 245 246#if HAVE_LLVM >= 0x0300 247 vb_type = LLVMStructCreateNamed(gallivm->context, struct_name); 248 LLVMStructSetBody(vb_type, elem_types, 249 Elements(elem_types), 0); 250#else 251 vb_type = LLVMStructTypeInContext(gallivm->context, elem_types, 252 Elements(elem_types), 0); 253 LLVMAddTypeName(gallivm->module, struct_name, vb_type); 254 255 LLVMInvalidateStructLayout(gallivm->target, vb_type); 256#endif 257 258 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride, 259 target, vb_type, 0); 260 LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset, 261 target, vb_type, 1); 262 263 LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type); 264 265 return vb_type; 266} 267 268 269/** 270 * Create LLVM type for struct vertex_header; 271 */ 272static LLVMTypeRef 273create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems) 274{ 275 LLVMTargetDataRef target = gallivm->target; 276 LLVMTypeRef elem_types[4]; 277 LLVMTypeRef vertex_header; 278 char struct_name[24]; 279 280 util_snprintf(struct_name, 23, "vertex_header%d", data_elems); 281 282 elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32); 283 elem_types[DRAW_JIT_VERTEX_CLIP] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4); 284 elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4); 285 elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems); 286 287#if HAVE_LLVM >= 0x0300 288 vertex_header = LLVMStructCreateNamed(gallivm->context, struct_name); 289 LLVMStructSetBody(vertex_header, elem_types, 290 Elements(elem_types), 0); 291#else 292 vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types, 293 Elements(elem_types), 0); 294 LLVMAddTypeName(gallivm->module, struct_name, vertex_header); 295 296 LLVMInvalidateStructLayout(gallivm->target, vertex_header); 297#endif 298 299 /* these are bit-fields and we can't take address of them 300 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask, 301 target, vertex_header, 302 DRAW_JIT_VERTEX_CLIPMASK); 303 LP_CHECK_MEMBER_OFFSET(struct vertex_header, edgeflag, 304 target, vertex_header, 305 DRAW_JIT_VERTEX_EDGEFLAG); 306 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pad, 307 target, vertex_header, 308 DRAW_JIT_VERTEX_PAD); 309 LP_CHECK_MEMBER_OFFSET(struct vertex_header, vertex_id, 310 target, vertex_header, 311 DRAW_JIT_VERTEX_VERTEX_ID); 312 */ 313 LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip, 314 target, vertex_header, 315 DRAW_JIT_VERTEX_CLIP); 316 LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos, 317 target, vertex_header, 318 DRAW_JIT_VERTEX_PRE_CLIP_POS); 319 LP_CHECK_MEMBER_OFFSET(struct vertex_header, data, 320 target, vertex_header, 321 DRAW_JIT_VERTEX_DATA); 322 323 return vertex_header; 324} 325 326 327/** 328 * Create LLVM types for various structures. 329 */ 330static void 331create_jit_types(struct draw_llvm *llvm) 332{ 333 struct gallivm_state *gallivm = llvm->gallivm; 334 LLVMTypeRef texture_type, context_type, buffer_type, vb_type; 335 336 texture_type = create_jit_texture_type(gallivm, "texture"); 337 338 context_type = create_jit_context_type(gallivm, texture_type, "draw_jit_context"); 339 llvm->context_ptr_type = LLVMPointerType(context_type, 0); 340 341 buffer_type = LLVMPointerType(LLVMIntTypeInContext(gallivm->context, 8), 0); 342 llvm->buffer_ptr_type = LLVMPointerType(buffer_type, 0); 343 344 vb_type = create_jit_vertex_buffer_type(gallivm, "pipe_vertex_buffer"); 345 llvm->vb_ptr_type = LLVMPointerType(vb_type, 0); 346} 347 348 349static LLVMTypeRef 350get_context_ptr_type(struct draw_llvm *llvm) 351{ 352 if (!llvm->context_ptr_type) 353 create_jit_types(llvm); 354 return llvm->context_ptr_type; 355} 356 357 358static LLVMTypeRef 359get_buffer_ptr_type(struct draw_llvm *llvm) 360{ 361 if (!llvm->buffer_ptr_type) 362 create_jit_types(llvm); 363 return llvm->buffer_ptr_type; 364} 365 366 367static LLVMTypeRef 368get_vb_ptr_type(struct draw_llvm *llvm) 369{ 370 if (!llvm->vb_ptr_type) 371 create_jit_types(llvm); 372 return llvm->vb_ptr_type; 373} 374 375static LLVMTypeRef 376get_vertex_header_ptr_type(struct draw_llvm *llvm) 377{ 378 if (!llvm->vertex_header_ptr_type) 379 create_jit_types(llvm); 380 return llvm->vertex_header_ptr_type; 381} 382 383 384/** 385 * Create per-context LLVM info. 386 */ 387struct draw_llvm * 388draw_llvm_create(struct draw_context *draw, struct gallivm_state *gallivm) 389{ 390 struct draw_llvm *llvm; 391 392 llvm = CALLOC_STRUCT( draw_llvm ); 393 if (!llvm) 394 return NULL; 395 396 lp_build_init(); 397 398 llvm->draw = draw; 399 llvm->gallivm = gallivm; 400 401 if (gallivm_debug & GALLIVM_DEBUG_IR) { 402 LLVMDumpModule(llvm->gallivm->module); 403 } 404 405 llvm->nr_variants = 0; 406 make_empty_list(&llvm->vs_variants_list); 407 408 gallivm_register_garbage_collector_callback( 409 draw_llvm_garbage_collect_callback, llvm); 410 411 return llvm; 412} 413 414 415/** 416 * Free per-context LLVM info. 417 */ 418void 419draw_llvm_destroy(struct draw_llvm *llvm) 420{ 421 gallivm_remove_garbage_collector_callback( 422 draw_llvm_garbage_collect_callback, llvm); 423 424 /* XXX free other draw_llvm data? */ 425 FREE(llvm); 426} 427 428 429/** 430 * Create LLVM-generated code for a vertex shader. 431 */ 432struct draw_llvm_variant * 433draw_llvm_create_variant(struct draw_llvm *llvm, 434 unsigned num_inputs, 435 const struct draw_llvm_variant_key *key) 436{ 437 struct draw_llvm_variant *variant; 438 struct llvm_vertex_shader *shader = 439 llvm_vertex_shader(llvm->draw->vs.vertex_shader); 440 LLVMTypeRef vertex_header; 441 442 variant = MALLOC(sizeof *variant + 443 shader->variant_key_size - 444 sizeof variant->key); 445 if (variant == NULL) 446 return NULL; 447 448 variant->llvm = llvm; 449 450 memcpy(&variant->key, key, shader->variant_key_size); 451 452 vertex_header = create_jit_vertex_header(llvm->gallivm, num_inputs); 453 454 llvm->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0); 455 456 draw_llvm_generate(llvm, variant, FALSE); /* linear */ 457 draw_llvm_generate(llvm, variant, TRUE); /* elts */ 458 459 variant->shader = shader; 460 variant->list_item_global.base = variant; 461 variant->list_item_local.base = variant; 462 /*variant->no = */shader->variants_created++; 463 variant->list_item_global.base = variant; 464 465 return variant; 466} 467 468 469static void 470generate_vs(struct draw_llvm *llvm, 471 LLVMBuilderRef builder, 472 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS], 473 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS], 474 LLVMValueRef system_values_array, 475 LLVMValueRef context_ptr, 476 struct lp_build_sampler_soa *draw_sampler, 477 boolean clamp_vertex_color) 478{ 479 const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens; 480 struct lp_type vs_type; 481 LLVMValueRef consts_ptr = draw_jit_context_vs_constants(llvm->gallivm, context_ptr); 482 struct lp_build_sampler_soa *sampler = 0; 483 484 memset(&vs_type, 0, sizeof vs_type); 485 vs_type.floating = TRUE; /* floating point values */ 486 vs_type.sign = TRUE; /* values are signed */ 487 vs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */ 488 vs_type.width = 32; /* 32-bit float */ 489 vs_type.length = 4; /* 4 elements per vector */ 490#if 0 491 num_vs = 4; /* number of vertices per block */ 492#endif 493 494 if (gallivm_debug & GALLIVM_DEBUG_IR) { 495 tgsi_dump(tokens, 0); 496 } 497 498 if (llvm->draw->num_sampler_views && llvm->draw->num_samplers) 499 sampler = draw_sampler; 500 501 lp_build_tgsi_soa(llvm->gallivm, 502 tokens, 503 vs_type, 504 NULL /*struct lp_build_mask_context *mask*/, 505 consts_ptr, 506 system_values_array, 507 NULL /*pos*/, 508 inputs, 509 outputs, 510 sampler, 511 &llvm->draw->vs.vertex_shader->info); 512 513 if (clamp_vertex_color) { 514 LLVMValueRef out; 515 unsigned chan, attrib; 516 struct lp_build_context bld; 517 struct tgsi_shader_info* info = &llvm->draw->vs.vertex_shader->info; 518 lp_build_context_init(&bld, llvm->gallivm, vs_type); 519 520 for (attrib = 0; attrib < info->num_outputs; ++attrib) { 521 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) { 522 if (outputs[attrib][chan]) { 523 switch (info->output_semantic_name[attrib]) { 524 case TGSI_SEMANTIC_COLOR: 525 case TGSI_SEMANTIC_BCOLOR: 526 out = LLVMBuildLoad(builder, outputs[attrib][chan], ""); 527 out = lp_build_clamp(&bld, out, bld.zero, bld.one); 528 LLVMBuildStore(builder, out, outputs[attrib][chan]); 529 break; 530 } 531 } 532 } 533 } 534 } 535} 536 537 538#if DEBUG_STORE 539static void print_vectorf(LLVMBuilderRef builder, 540 LLVMValueRef vec) 541{ 542 LLVMValueRef val[4]; 543 val[0] = LLVMBuildExtractElement(builder, vec, 544 lp_build_const_int32(gallivm, 0), ""); 545 val[1] = LLVMBuildExtractElement(builder, vec, 546 lp_build_const_int32(gallivm, 1), ""); 547 val[2] = LLVMBuildExtractElement(builder, vec, 548 lp_build_const_int32(gallivm, 2), ""); 549 val[3] = LLVMBuildExtractElement(builder, vec, 550 lp_build_const_int32(gallivm, 3), ""); 551 lp_build_printf(builder, "vector = [%f, %f, %f, %f]\n", 552 val[0], val[1], val[2], val[3]); 553} 554#endif 555 556 557static void 558generate_fetch(struct gallivm_state *gallivm, 559 LLVMValueRef vbuffers_ptr, 560 LLVMValueRef *res, 561 struct pipe_vertex_element *velem, 562 LLVMValueRef vbuf, 563 LLVMValueRef index, 564 LLVMValueRef instance_id) 565{ 566 LLVMBuilderRef builder = gallivm->builder; 567 LLVMValueRef indices = 568 LLVMConstInt(LLVMInt64TypeInContext(gallivm->context), 569 velem->vertex_buffer_index, 0); 570 LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, 571 &indices, 1, ""); 572 LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf); 573 LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf); 574 LLVMValueRef stride; 575 576 if (velem->instance_divisor) { 577 /* array index = instance_id / instance_divisor */ 578 index = LLVMBuildUDiv(builder, instance_id, 579 lp_build_const_int32(gallivm, velem->instance_divisor), 580 "instance_divisor"); 581 } 582 583 stride = LLVMBuildMul(builder, vb_stride, index, ""); 584 585 vbuffer_ptr = LLVMBuildLoad(builder, vbuffer_ptr, "vbuffer"); 586 587 stride = LLVMBuildAdd(builder, stride, 588 vb_buffer_offset, 589 ""); 590 stride = LLVMBuildAdd(builder, stride, 591 lp_build_const_int32(gallivm, velem->src_offset), 592 ""); 593 594 /*lp_build_printf(builder, "vbuf index = %d, stride is %d\n", indices, stride);*/ 595 vbuffer_ptr = LLVMBuildGEP(builder, vbuffer_ptr, &stride, 1, ""); 596 597 *res = draw_llvm_translate_from(gallivm, vbuffer_ptr, velem->src_format); 598} 599 600 601static LLVMValueRef 602aos_to_soa(struct gallivm_state *gallivm, 603 LLVMValueRef val0, 604 LLVMValueRef val1, 605 LLVMValueRef val2, 606 LLVMValueRef val3, 607 LLVMValueRef channel) 608{ 609 LLVMBuilderRef builder = gallivm->builder; 610 LLVMValueRef ex, res; 611 612 ex = LLVMBuildExtractElement(builder, val0, 613 channel, ""); 614 res = LLVMBuildInsertElement(builder, 615 LLVMConstNull(LLVMTypeOf(val0)), 616 ex, 617 lp_build_const_int32(gallivm, 0), 618 ""); 619 620 ex = LLVMBuildExtractElement(builder, val1, 621 channel, ""); 622 res = LLVMBuildInsertElement(builder, 623 res, ex, 624 lp_build_const_int32(gallivm, 1), 625 ""); 626 627 ex = LLVMBuildExtractElement(builder, val2, 628 channel, ""); 629 res = LLVMBuildInsertElement(builder, 630 res, ex, 631 lp_build_const_int32(gallivm, 2), 632 ""); 633 634 ex = LLVMBuildExtractElement(builder, val3, 635 channel, ""); 636 res = LLVMBuildInsertElement(builder, 637 res, ex, 638 lp_build_const_int32(gallivm, 3), 639 ""); 640 641 return res; 642} 643 644 645static void 646soa_to_aos(struct gallivm_state *gallivm, 647 LLVMValueRef soa[TGSI_NUM_CHANNELS], 648 LLVMValueRef aos[TGSI_NUM_CHANNELS]) 649{ 650 LLVMBuilderRef builder = gallivm->builder; 651 LLVMValueRef comp; 652 int i = 0; 653 654 debug_assert(TGSI_NUM_CHANNELS == 4); 655 656 aos[0] = LLVMConstNull(LLVMTypeOf(soa[0])); 657 aos[1] = aos[2] = aos[3] = aos[0]; 658 659 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) { 660 LLVMValueRef channel = lp_build_const_int32(gallivm, i); 661 662 comp = LLVMBuildExtractElement(builder, soa[i], 663 lp_build_const_int32(gallivm, 0), ""); 664 aos[0] = LLVMBuildInsertElement(builder, aos[0], comp, channel, ""); 665 666 comp = LLVMBuildExtractElement(builder, soa[i], 667 lp_build_const_int32(gallivm, 1), ""); 668 aos[1] = LLVMBuildInsertElement(builder, aos[1], comp, channel, ""); 669 670 comp = LLVMBuildExtractElement(builder, soa[i], 671 lp_build_const_int32(gallivm, 2), ""); 672 aos[2] = LLVMBuildInsertElement(builder, aos[2], comp, channel, ""); 673 674 comp = LLVMBuildExtractElement(builder, soa[i], 675 lp_build_const_int32(gallivm, 3), ""); 676 aos[3] = LLVMBuildInsertElement(builder, aos[3], comp, channel, ""); 677 678 } 679} 680 681 682static void 683convert_to_soa(struct gallivm_state *gallivm, 684 LLVMValueRef (*aos)[TGSI_NUM_CHANNELS], 685 LLVMValueRef (*soa)[TGSI_NUM_CHANNELS], 686 int num_attribs) 687{ 688 int i; 689 690 debug_assert(TGSI_NUM_CHANNELS == 4); 691 692 for (i = 0; i < num_attribs; ++i) { 693 LLVMValueRef val0 = aos[i][0]; 694 LLVMValueRef val1 = aos[i][1]; 695 LLVMValueRef val2 = aos[i][2]; 696 LLVMValueRef val3 = aos[i][3]; 697 698 soa[i][0] = aos_to_soa(gallivm, val0, val1, val2, val3, 699 lp_build_const_int32(gallivm, 0)); 700 soa[i][1] = aos_to_soa(gallivm, val0, val1, val2, val3, 701 lp_build_const_int32(gallivm, 1)); 702 soa[i][2] = aos_to_soa(gallivm, val0, val1, val2, val3, 703 lp_build_const_int32(gallivm, 2)); 704 soa[i][3] = aos_to_soa(gallivm, val0, val1, val2, val3, 705 lp_build_const_int32(gallivm, 3)); 706 } 707} 708 709 710static void 711store_aos(struct gallivm_state *gallivm, 712 LLVMValueRef io_ptr, 713 LLVMValueRef index, 714 LLVMValueRef value, 715 LLVMValueRef clipmask) 716{ 717 LLVMBuilderRef builder = gallivm->builder; 718 LLVMValueRef id_ptr = draw_jit_header_id(gallivm, io_ptr); 719 LLVMValueRef data_ptr = draw_jit_header_data(gallivm, io_ptr); 720 LLVMValueRef indices[3]; 721 LLVMValueRef val; 722 int vertex_id_pad_edgeflag; 723 724 indices[0] = lp_build_const_int32(gallivm, 0); 725 indices[1] = index; 726 indices[2] = lp_build_const_int32(gallivm, 0); 727 728 /* If this assertion fails, it means we need to update the bit twidding 729 * code here. See struct vertex_header in draw_private.h. 730 */ 731 assert(DRAW_TOTAL_CLIP_PLANES==14); 732 /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */ 733 vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES); 734 val = lp_build_const_int32(gallivm, vertex_id_pad_edgeflag); 735 /* OR with the clipmask */ 736 val = LLVMBuildOr(builder, val, clipmask, ""); 737 738 /* store vertex header */ 739 LLVMBuildStore(builder, val, id_ptr); 740 741 742#if DEBUG_STORE 743 lp_build_printf(builder, " ---- %p storing attribute %d (io = %p)\n", data_ptr, index, io_ptr); 744#endif 745#if 0 746 /*lp_build_printf(builder, " ---- %p storing at %d (%p) ", io_ptr, index, data_ptr); 747 print_vectorf(builder, value);*/ 748 data_ptr = LLVMBuildBitCast(builder, data_ptr, 749 LLVMPointerType(LLVMArrayType(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), 0), 0), 750 "datavec"); 751 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 2, ""); 752 753 LLVMBuildStore(builder, value, data_ptr); 754#else 755 { 756 LLVMValueRef x, y, z, w; 757 LLVMValueRef idx0, idx1, idx2, idx3; 758 LLVMValueRef gep0, gep1, gep2, gep3; 759 data_ptr = LLVMBuildGEP(builder, data_ptr, indices, 3, ""); 760 761 idx0 = lp_build_const_int32(gallivm, 0); 762 idx1 = lp_build_const_int32(gallivm, 1); 763 idx2 = lp_build_const_int32(gallivm, 2); 764 idx3 = lp_build_const_int32(gallivm, 3); 765 766 x = LLVMBuildExtractElement(builder, value, 767 idx0, ""); 768 y = LLVMBuildExtractElement(builder, value, 769 idx1, ""); 770 z = LLVMBuildExtractElement(builder, value, 771 idx2, ""); 772 w = LLVMBuildExtractElement(builder, value, 773 idx3, ""); 774 775 gep0 = LLVMBuildGEP(builder, data_ptr, &idx0, 1, ""); 776 gep1 = LLVMBuildGEP(builder, data_ptr, &idx1, 1, ""); 777 gep2 = LLVMBuildGEP(builder, data_ptr, &idx2, 1, ""); 778 gep3 = LLVMBuildGEP(builder, data_ptr, &idx3, 1, ""); 779 780 /*lp_build_printf(builder, "##### x = %f (%p), y = %f (%p), z = %f (%p), w = %f (%p)\n", 781 x, gep0, y, gep1, z, gep2, w, gep3);*/ 782 LLVMBuildStore(builder, x, gep0); 783 LLVMBuildStore(builder, y, gep1); 784 LLVMBuildStore(builder, z, gep2); 785 LLVMBuildStore(builder, w, gep3); 786 } 787#endif 788} 789 790 791static void 792store_aos_array(struct gallivm_state *gallivm, 793 LLVMValueRef io_ptr, 794 LLVMValueRef aos[TGSI_NUM_CHANNELS], 795 int attrib, 796 int num_outputs, 797 LLVMValueRef clipmask) 798{ 799 LLVMBuilderRef builder = gallivm->builder; 800 LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib); 801 LLVMValueRef ind0 = lp_build_const_int32(gallivm, 0); 802 LLVMValueRef ind1 = lp_build_const_int32(gallivm, 1); 803 LLVMValueRef ind2 = lp_build_const_int32(gallivm, 2); 804 LLVMValueRef ind3 = lp_build_const_int32(gallivm, 3); 805 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr; 806 LLVMValueRef clipmask0, clipmask1, clipmask2, clipmask3; 807 808 debug_assert(TGSI_NUM_CHANNELS == 4); 809 810 io0_ptr = LLVMBuildGEP(builder, io_ptr, 811 &ind0, 1, ""); 812 io1_ptr = LLVMBuildGEP(builder, io_ptr, 813 &ind1, 1, ""); 814 io2_ptr = LLVMBuildGEP(builder, io_ptr, 815 &ind2, 1, ""); 816 io3_ptr = LLVMBuildGEP(builder, io_ptr, 817 &ind3, 1, ""); 818 819 clipmask0 = LLVMBuildExtractElement(builder, clipmask, 820 ind0, ""); 821 clipmask1 = LLVMBuildExtractElement(builder, clipmask, 822 ind1, ""); 823 clipmask2 = LLVMBuildExtractElement(builder, clipmask, 824 ind2, ""); 825 clipmask3 = LLVMBuildExtractElement(builder, clipmask, 826 ind3, ""); 827 828#if DEBUG_STORE 829 lp_build_printf(builder, "io = %p, indexes[%d, %d, %d, %d]\n, clipmask0 = %x, clipmask1 = %x, clipmask2 = %x, clipmask3 = %x\n", 830 io_ptr, ind0, ind1, ind2, ind3, clipmask0, clipmask1, clipmask2, clipmask3); 831#endif 832 /* store for each of the 4 vertices */ 833 store_aos(gallivm, io0_ptr, attr_index, aos[0], clipmask0); 834 store_aos(gallivm, io1_ptr, attr_index, aos[1], clipmask1); 835 store_aos(gallivm, io2_ptr, attr_index, aos[2], clipmask2); 836 store_aos(gallivm, io3_ptr, attr_index, aos[3], clipmask3); 837} 838 839 840static void 841convert_to_aos(struct gallivm_state *gallivm, 842 LLVMValueRef io, 843 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS], 844 LLVMValueRef clipmask, 845 int num_outputs, 846 int max_vertices) 847{ 848 LLVMBuilderRef builder = gallivm->builder; 849 unsigned chan, attrib; 850 851#if DEBUG_STORE 852 lp_build_printf(builder, " # storing begin\n"); 853#endif 854 for (attrib = 0; attrib < num_outputs; ++attrib) { 855 LLVMValueRef soa[4]; 856 LLVMValueRef aos[4]; 857 for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) { 858 if (outputs[attrib][chan]) { 859 LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], ""); 860 lp_build_name(out, "output%u.%c", attrib, "xyzw"[chan]); 861 /*lp_build_printf(builder, "output %d : %d ", 862 LLVMConstInt(LLVMInt32Type(), attrib, 0), 863 LLVMConstInt(LLVMInt32Type(), chan, 0)); 864 print_vectorf(builder, out);*/ 865 soa[chan] = out; 866 } 867 else { 868 soa[chan] = 0; 869 } 870 } 871 soa_to_aos(gallivm, soa, aos); 872 store_aos_array(gallivm, 873 io, 874 aos, 875 attrib, 876 num_outputs, 877 clipmask); 878 } 879#if DEBUG_STORE 880 lp_build_printf(builder, " # storing end\n"); 881#endif 882} 883 884 885/** 886 * Stores original vertex positions in clip coordinates 887 * There is probably a more efficient way to do this, 4 floats at once 888 * rather than extracting each element one by one. 889 */ 890static void 891store_clip(struct gallivm_state *gallivm, 892 LLVMValueRef io_ptr, 893 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS], 894 boolean pre_clip_pos) 895{ 896 LLVMBuilderRef builder = gallivm->builder; 897 LLVMValueRef out[4]; 898 LLVMValueRef indices[2]; 899 LLVMValueRef io0_ptr, io1_ptr, io2_ptr, io3_ptr; 900 LLVMValueRef clip_ptr0, clip_ptr1, clip_ptr2, clip_ptr3; 901 LLVMValueRef clip0_ptr, clip1_ptr, clip2_ptr, clip3_ptr; 902 LLVMValueRef out0elem, out1elem, out2elem, out3elem; 903 int i; 904 905 LLVMValueRef ind0 = lp_build_const_int32(gallivm, 0); 906 LLVMValueRef ind1 = lp_build_const_int32(gallivm, 1); 907 LLVMValueRef ind2 = lp_build_const_int32(gallivm, 2); 908 LLVMValueRef ind3 = lp_build_const_int32(gallivm, 3); 909 910 indices[0] = 911 indices[1] = lp_build_const_int32(gallivm, 0); 912 913 out[0] = LLVMBuildLoad(builder, outputs[0][0], ""); /*x0 x1 x2 x3*/ 914 out[1] = LLVMBuildLoad(builder, outputs[0][1], ""); /*y0 y1 y2 y3*/ 915 out[2] = LLVMBuildLoad(builder, outputs[0][2], ""); /*z0 z1 z2 z3*/ 916 out[3] = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/ 917 918 io0_ptr = LLVMBuildGEP(builder, io_ptr, &ind0, 1, ""); 919 io1_ptr = LLVMBuildGEP(builder, io_ptr, &ind1, 1, ""); 920 io2_ptr = LLVMBuildGEP(builder, io_ptr, &ind2, 1, ""); 921 io3_ptr = LLVMBuildGEP(builder, io_ptr, &ind3, 1, ""); 922 923 /* FIXME: this needs updating for clip vertex support */ 924 if (!pre_clip_pos) { 925 clip_ptr0 = draw_jit_header_clip(gallivm, io0_ptr); 926 clip_ptr1 = draw_jit_header_clip(gallivm, io1_ptr); 927 clip_ptr2 = draw_jit_header_clip(gallivm, io2_ptr); 928 clip_ptr3 = draw_jit_header_clip(gallivm, io3_ptr); 929 } else { 930 clip_ptr0 = draw_jit_header_pre_clip_pos(gallivm, io0_ptr); 931 clip_ptr1 = draw_jit_header_pre_clip_pos(gallivm, io1_ptr); 932 clip_ptr2 = draw_jit_header_pre_clip_pos(gallivm, io2_ptr); 933 clip_ptr3 = draw_jit_header_pre_clip_pos(gallivm, io3_ptr); 934 } 935 936 for (i = 0; i<4; i++) { 937 clip0_ptr = LLVMBuildGEP(builder, clip_ptr0, indices, 2, ""); /* x0 */ 938 clip1_ptr = LLVMBuildGEP(builder, clip_ptr1, indices, 2, ""); /* x1 */ 939 clip2_ptr = LLVMBuildGEP(builder, clip_ptr2, indices, 2, ""); /* x2 */ 940 clip3_ptr = LLVMBuildGEP(builder, clip_ptr3, indices, 2, ""); /* x3 */ 941 942 out0elem = LLVMBuildExtractElement(builder, out[i], ind0, ""); /* x0 */ 943 out1elem = LLVMBuildExtractElement(builder, out[i], ind1, ""); /* x1 */ 944 out2elem = LLVMBuildExtractElement(builder, out[i], ind2, ""); /* x2 */ 945 out3elem = LLVMBuildExtractElement(builder, out[i], ind3, ""); /* x3 */ 946 947 LLVMBuildStore(builder, out0elem, clip0_ptr); 948 LLVMBuildStore(builder, out1elem, clip1_ptr); 949 LLVMBuildStore(builder, out2elem, clip2_ptr); 950 LLVMBuildStore(builder, out3elem, clip3_ptr); 951 952 indices[1]= LLVMBuildAdd(builder, indices[1], ind1, ""); 953 } 954 955} 956 957 958/** 959 * Equivalent of _mm_set1_ps(a) 960 */ 961static LLVMValueRef 962vec4f_from_scalar(struct gallivm_state *gallivm, 963 LLVMValueRef a, 964 const char *name) 965{ 966 LLVMTypeRef float_type = LLVMFloatTypeInContext(gallivm->context); 967 LLVMValueRef res = LLVMGetUndef(LLVMVectorType(float_type, 4)); 968 int i; 969 970 for (i = 0; i < 4; ++i) { 971 LLVMValueRef index = lp_build_const_int32(gallivm, i); 972 res = LLVMBuildInsertElement(gallivm->builder, res, a, 973 index, i == 3 ? name : ""); 974 } 975 976 return res; 977} 978 979 980/** 981 * Transforms the outputs for viewport mapping 982 */ 983static void 984generate_viewport(struct draw_llvm *llvm, 985 LLVMBuilderRef builder, 986 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS], 987 LLVMValueRef context_ptr) 988{ 989 int i; 990 struct gallivm_state *gallivm = llvm->gallivm; 991 struct lp_type f32_type = lp_type_float_vec(32); 992 LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/ 993 LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/ 994 LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr); 995 996 /* for 1/w convention*/ 997 out3 = LLVMBuildFDiv(builder, const1, out3, ""); 998 LLVMBuildStore(builder, out3, outputs[0][3]); 999 1000 /* Viewport Mapping */ 1001 for (i=0; i<3; i++) { 1002 LLVMValueRef out = LLVMBuildLoad(builder, outputs[0][i], ""); /*x0 x1 x2 x3*/ 1003 LLVMValueRef scale; 1004 LLVMValueRef trans; 1005 LLVMValueRef scale_i; 1006 LLVMValueRef trans_i; 1007 LLVMValueRef index; 1008 1009 index = lp_build_const_int32(gallivm, i); 1010 scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, ""); 1011 1012 index = lp_build_const_int32(gallivm, i+4); 1013 trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, ""); 1014 1015 scale = vec4f_from_scalar(gallivm, LLVMBuildLoad(builder, scale_i, ""), "scale"); 1016 trans = vec4f_from_scalar(gallivm, LLVMBuildLoad(builder, trans_i, ""), "trans"); 1017 1018 /* divide by w */ 1019 out = LLVMBuildFMul(builder, out, out3, ""); 1020 /* mult by scale */ 1021 out = LLVMBuildFMul(builder, out, scale, ""); 1022 /* add translation */ 1023 out = LLVMBuildFAdd(builder, out, trans, ""); 1024 1025 /* store transformed outputs */ 1026 LLVMBuildStore(builder, out, outputs[0][i]); 1027 } 1028 1029} 1030 1031 1032/** 1033 * Returns clipmask as 4xi32 bitmask for the 4 vertices 1034 */ 1035static LLVMValueRef 1036generate_clipmask(struct gallivm_state *gallivm, 1037 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS], 1038 boolean clip_xy, 1039 boolean clip_z, 1040 boolean clip_user, 1041 boolean clip_halfz, 1042 unsigned ucp_enable, 1043 LLVMValueRef context_ptr) 1044{ 1045 LLVMBuilderRef builder = gallivm->builder; 1046 LLVMValueRef mask; /* stores the <4xi32> clipmasks */ 1047 LLVMValueRef test, temp; 1048 LLVMValueRef zero, shift; 1049 LLVMValueRef pos_x, pos_y, pos_z, pos_w; 1050 LLVMValueRef plane1, planes, plane_ptr, sum; 1051 struct lp_type f32_type = lp_type_float_vec(32); 1052 1053 mask = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0); 1054 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0); 1055 zero = lp_build_const_vec(gallivm, f32_type, 0); /* 0.0f 0.0f 0.0f 0.0f */ 1056 shift = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 1); /* 1 1 1 1 */ 1057 1058 /* Assuming position stored at output[0] */ 1059 pos_x = LLVMBuildLoad(builder, outputs[0][0], ""); /*x0 x1 x2 x3*/ 1060 pos_y = LLVMBuildLoad(builder, outputs[0][1], ""); /*y0 y1 y2 y3*/ 1061 pos_z = LLVMBuildLoad(builder, outputs[0][2], ""); /*z0 z1 z2 z3*/ 1062 pos_w = LLVMBuildLoad(builder, outputs[0][3], ""); /*w0 w1 w2 w3*/ 1063 1064 /* Cliptest, for hardwired planes */ 1065 if (clip_xy) { 1066 /* plane 1 */ 1067 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w); 1068 temp = shift; 1069 test = LLVMBuildAnd(builder, test, temp, ""); 1070 mask = test; 1071 1072 /* plane 2 */ 1073 test = LLVMBuildFAdd(builder, pos_x, pos_w, ""); 1074 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test); 1075 temp = LLVMBuildShl(builder, temp, shift, ""); 1076 test = LLVMBuildAnd(builder, test, temp, ""); 1077 mask = LLVMBuildOr(builder, mask, test, ""); 1078 1079 /* plane 3 */ 1080 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_y, pos_w); 1081 temp = LLVMBuildShl(builder, temp, shift, ""); 1082 test = LLVMBuildAnd(builder, test, temp, ""); 1083 mask = LLVMBuildOr(builder, mask, test, ""); 1084 1085 /* plane 4 */ 1086 test = LLVMBuildFAdd(builder, pos_y, pos_w, ""); 1087 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test); 1088 temp = LLVMBuildShl(builder, temp, shift, ""); 1089 test = LLVMBuildAnd(builder, test, temp, ""); 1090 mask = LLVMBuildOr(builder, mask, test, ""); 1091 } 1092 1093 if (clip_z) { 1094 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 16); 1095 if (clip_halfz) { 1096 /* plane 5 */ 1097 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z); 1098 test = LLVMBuildAnd(builder, test, temp, ""); 1099 mask = LLVMBuildOr(builder, mask, test, ""); 1100 } 1101 else { 1102 /* plane 5 */ 1103 test = LLVMBuildFAdd(builder, pos_z, pos_w, ""); 1104 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, test); 1105 test = LLVMBuildAnd(builder, test, temp, ""); 1106 mask = LLVMBuildOr(builder, mask, test, ""); 1107 } 1108 /* plane 6 */ 1109 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_z, pos_w); 1110 temp = LLVMBuildShl(builder, temp, shift, ""); 1111 test = LLVMBuildAnd(builder, test, temp, ""); 1112 mask = LLVMBuildOr(builder, mask, test, ""); 1113 } 1114 1115 if (clip_user) { 1116 LLVMValueRef planes_ptr = draw_jit_context_planes(gallivm, context_ptr); 1117 LLVMValueRef indices[3]; 1118 1119 /* userclip planes */ 1120 while (ucp_enable) { 1121 unsigned plane_idx = ffs(ucp_enable)-1; 1122 ucp_enable &= ~(1 << plane_idx); 1123 plane_idx += 6; 1124 1125 indices[0] = lp_build_const_int32(gallivm, 0); 1126 indices[1] = lp_build_const_int32(gallivm, plane_idx); 1127 1128 indices[2] = lp_build_const_int32(gallivm, 0); 1129 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, ""); 1130 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_x"); 1131 planes = vec4f_from_scalar(gallivm, plane1, "plane4_x"); 1132 sum = LLVMBuildFMul(builder, planes, pos_x, ""); 1133 1134 indices[2] = lp_build_const_int32(gallivm, 1); 1135 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, ""); 1136 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y"); 1137 planes = vec4f_from_scalar(gallivm, plane1, "plane4_y"); 1138 test = LLVMBuildFMul(builder, planes, pos_y, ""); 1139 sum = LLVMBuildFAdd(builder, sum, test, ""); 1140 1141 indices[2] = lp_build_const_int32(gallivm, 2); 1142 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, ""); 1143 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z"); 1144 planes = vec4f_from_scalar(gallivm, plane1, "plane4_z"); 1145 test = LLVMBuildFMul(builder, planes, pos_z, ""); 1146 sum = LLVMBuildFAdd(builder, sum, test, ""); 1147 1148 indices[2] = lp_build_const_int32(gallivm, 3); 1149 plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, ""); 1150 plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w"); 1151 planes = vec4f_from_scalar(gallivm, plane1, "plane4_w"); 1152 test = LLVMBuildFMul(builder, planes, pos_w, ""); 1153 sum = LLVMBuildFAdd(builder, sum, test, ""); 1154 1155 test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum); 1156 temp = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 1 << plane_idx); 1157 test = LLVMBuildAnd(builder, test, temp, ""); 1158 mask = LLVMBuildOr(builder, mask, test, ""); 1159 } 1160 } 1161 return mask; 1162} 1163 1164 1165/** 1166 * Returns boolean if any clipping has occurred 1167 * Used zero/non-zero i32 value to represent boolean 1168 */ 1169static void 1170clipmask_bool(struct gallivm_state *gallivm, 1171 LLVMValueRef clipmask, 1172 LLVMValueRef ret_ptr) 1173{ 1174 LLVMBuilderRef builder = gallivm->builder; 1175 LLVMValueRef ret = LLVMBuildLoad(builder, ret_ptr, ""); 1176 LLVMValueRef temp; 1177 int i; 1178 1179 for (i=0; i<4; i++) { 1180 temp = LLVMBuildExtractElement(builder, clipmask, 1181 lp_build_const_int32(gallivm, i) , ""); 1182 ret = LLVMBuildOr(builder, ret, temp, ""); 1183 } 1184 1185 LLVMBuildStore(builder, ret, ret_ptr); 1186} 1187 1188 1189static void 1190draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant, 1191 boolean elts) 1192{ 1193 struct gallivm_state *gallivm = llvm->gallivm; 1194 LLVMContextRef context = gallivm->context; 1195 LLVMTypeRef int32_type = LLVMInt32TypeInContext(context); 1196 LLVMTypeRef arg_types[8]; 1197 LLVMTypeRef func_type; 1198 LLVMValueRef context_ptr; 1199 LLVMBasicBlockRef block; 1200 LLVMBuilderRef builder; 1201 LLVMValueRef end, start; 1202 LLVMValueRef count, fetch_elts, fetch_count; 1203 LLVMValueRef stride, step, io_itr; 1204 LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr; 1205 LLVMValueRef instance_id; 1206 LLVMValueRef system_values_array; 1207 LLVMValueRef zero = lp_build_const_int32(gallivm, 0); 1208 LLVMValueRef one = lp_build_const_int32(gallivm, 1); 1209 struct draw_context *draw = llvm->draw; 1210 const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info; 1211 unsigned i, j; 1212 struct lp_build_context bld; 1213 struct lp_build_loop_state lp_loop; 1214 const int max_vertices = 4; 1215 LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS]; 1216 LLVMValueRef fetch_max; 1217 void *code; 1218 struct lp_build_sampler_soa *sampler = 0; 1219 LLVMValueRef ret, ret_ptr; 1220 const boolean bypass_viewport = variant->key.bypass_viewport; 1221 const boolean enable_cliptest = variant->key.clip_xy || 1222 variant->key.clip_z || 1223 variant->key.clip_user; 1224 LLVMValueRef variant_func; 1225 1226 arg_types[0] = get_context_ptr_type(llvm); /* context */ 1227 arg_types[1] = get_vertex_header_ptr_type(llvm); /* vertex_header */ 1228 arg_types[2] = get_buffer_ptr_type(llvm); /* vbuffers */ 1229 if (elts) 1230 arg_types[3] = LLVMPointerType(int32_type, 0);/* fetch_elts * */ 1231 else 1232 arg_types[3] = int32_type; /* start */ 1233 arg_types[4] = int32_type; /* fetch_count / count */ 1234 arg_types[5] = int32_type; /* stride */ 1235 arg_types[6] = get_vb_ptr_type(llvm); /* pipe_vertex_buffer's */ 1236 arg_types[7] = int32_type; /* instance_id */ 1237 1238 func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0); 1239 1240 variant_func = LLVMAddFunction(gallivm->module, 1241 elts ? "draw_llvm_shader_elts" : "draw_llvm_shader", 1242 func_type); 1243 1244 if (elts) 1245 variant->function_elts = variant_func; 1246 else 1247 variant->function = variant_func; 1248 1249 LLVMSetFunctionCallConv(variant_func, LLVMCCallConv); 1250 for (i = 0; i < Elements(arg_types); ++i) 1251 if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) 1252 LLVMAddAttribute(LLVMGetParam(variant_func, i), 1253 LLVMNoAliasAttribute); 1254 1255 context_ptr = LLVMGetParam(variant_func, 0); 1256 io_ptr = LLVMGetParam(variant_func, 1); 1257 vbuffers_ptr = LLVMGetParam(variant_func, 2); 1258 stride = LLVMGetParam(variant_func, 5); 1259 vb_ptr = LLVMGetParam(variant_func, 6); 1260 instance_id = LLVMGetParam(variant_func, 7); 1261 1262 lp_build_name(context_ptr, "context"); 1263 lp_build_name(io_ptr, "io"); 1264 lp_build_name(vbuffers_ptr, "vbuffers"); 1265 lp_build_name(stride, "stride"); 1266 lp_build_name(vb_ptr, "vb"); 1267 lp_build_name(instance_id, "instance_id"); 1268 1269 if (elts) { 1270 fetch_elts = LLVMGetParam(variant_func, 3); 1271 fetch_count = LLVMGetParam(variant_func, 4); 1272 lp_build_name(fetch_elts, "fetch_elts"); 1273 lp_build_name(fetch_count, "fetch_count"); 1274 start = count = NULL; 1275 } 1276 else { 1277 start = LLVMGetParam(variant_func, 3); 1278 count = LLVMGetParam(variant_func, 4); 1279 lp_build_name(start, "start"); 1280 lp_build_name(count, "count"); 1281 fetch_elts = fetch_count = NULL; 1282 } 1283 1284 /* 1285 * Function body 1286 */ 1287 1288 block = LLVMAppendBasicBlockInContext(gallivm->context, variant_func, "entry"); 1289 builder = gallivm->builder; 1290 LLVMPositionBuilderAtEnd(builder, block); 1291 1292 lp_build_context_init(&bld, gallivm, lp_type_int(32)); 1293 1294 system_values_array = lp_build_system_values_array(gallivm, vs_info, 1295 instance_id, NULL); 1296 1297 /* function will return non-zero i32 value if any clipped vertices */ 1298 ret_ptr = lp_build_alloca(gallivm, int32_type, ""); 1299 LLVMBuildStore(builder, zero, ret_ptr); 1300 1301 /* code generated texture sampling */ 1302 sampler = draw_llvm_sampler_soa_create( 1303 draw_llvm_variant_key_samplers(&variant->key), 1304 context_ptr); 1305 1306 if (elts) { 1307 start = zero; 1308 end = fetch_count; 1309 } 1310 else { 1311 end = lp_build_add(&bld, start, count); 1312 } 1313 1314 step = lp_build_const_int32(gallivm, max_vertices); 1315 1316 fetch_max = LLVMBuildSub(builder, end, one, "fetch_max"); 1317 1318 lp_build_loop_begin(&lp_loop, gallivm, start); 1319 { 1320 LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS]; 1321 LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS] = { { 0 } }; 1322 LLVMValueRef io; 1323 LLVMValueRef clipmask; /* holds the clipmask value */ 1324 const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS]; 1325 1326 if (elts) 1327 io_itr = lp_loop.counter; 1328 else 1329 io_itr = LLVMBuildSub(builder, lp_loop.counter, start, ""); 1330 1331 io = LLVMBuildGEP(builder, io_ptr, &io_itr, 1, ""); 1332#if DEBUG_STORE 1333 lp_build_printf(builder, " --- io %d = %p, loop counter %d\n", 1334 io_itr, io, lp_loop.counter); 1335#endif 1336 for (i = 0; i < TGSI_NUM_CHANNELS; ++i) { 1337 LLVMValueRef true_index = 1338 LLVMBuildAdd(builder, 1339 lp_loop.counter, 1340 lp_build_const_int32(gallivm, i), ""); 1341 1342 /* make sure we're not out of bounds which can happen 1343 * if fetch_count % 4 != 0, because on the last iteration 1344 * a few of the 4 vertex fetches will be out of bounds */ 1345 true_index = lp_build_min(&bld, true_index, fetch_max); 1346 1347 if (elts) { 1348 LLVMValueRef fetch_ptr; 1349 fetch_ptr = LLVMBuildGEP(builder, fetch_elts, 1350 &true_index, 1, ""); 1351 true_index = LLVMBuildLoad(builder, fetch_ptr, "fetch_elt"); 1352 } 1353 1354 for (j = 0; j < draw->pt.nr_vertex_elements; ++j) { 1355 struct pipe_vertex_element *velem = &draw->pt.vertex_element[j]; 1356 LLVMValueRef vb_index = 1357 lp_build_const_int32(gallivm, velem->vertex_buffer_index); 1358 LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, ""); 1359 generate_fetch(gallivm, vbuffers_ptr, 1360 &aos_attribs[j][i], velem, vb, true_index, 1361 instance_id); 1362 } 1363 } 1364 convert_to_soa(gallivm, aos_attribs, inputs, 1365 draw->pt.nr_vertex_elements); 1366 1367 ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs; 1368 generate_vs(llvm, 1369 builder, 1370 outputs, 1371 ptr_aos, 1372 system_values_array, 1373 context_ptr, 1374 sampler, 1375 variant->key.clamp_vertex_color); 1376 1377 /* store original positions in clip before further manipulation */ 1378 store_clip(gallivm, io, outputs, 0); 1379 store_clip(gallivm, io, outputs, 1); 1380 1381 /* do cliptest */ 1382 if (enable_cliptest) { 1383 /* allocate clipmask, assign it integer type */ 1384 clipmask = generate_clipmask(gallivm, outputs, 1385 variant->key.clip_xy, 1386 variant->key.clip_z, 1387 variant->key.clip_user, 1388 variant->key.clip_halfz, 1389 variant->key.ucp_enable, 1390 context_ptr); 1391 /* return clipping boolean value for function */ 1392 clipmask_bool(gallivm, clipmask, ret_ptr); 1393 } 1394 else { 1395 clipmask = lp_build_const_int_vec(gallivm, lp_type_int_vec(32), 0); 1396 } 1397 1398 /* do viewport mapping */ 1399 if (!bypass_viewport) { 1400 generate_viewport(llvm, builder, outputs, context_ptr); 1401 } 1402 1403 /* store clipmask in vertex header, 1404 * original positions in clip 1405 * and transformed positions in data 1406 */ 1407 convert_to_aos(gallivm, io, outputs, clipmask, 1408 vs_info->num_outputs, max_vertices); 1409 } 1410 1411 lp_build_loop_end_cond(&lp_loop, end, step, LLVMIntUGE); 1412 1413 sampler->destroy(sampler); 1414 1415 ret = LLVMBuildLoad(builder, ret_ptr, ""); 1416 LLVMBuildRet(builder, ret); 1417 1418 /* 1419 * Translate the LLVM IR into machine code. 1420 */ 1421#ifdef DEBUG 1422 if (LLVMVerifyFunction(variant_func, LLVMPrintMessageAction)) { 1423 lp_debug_dump_value(variant_func); 1424 assert(0); 1425 } 1426#endif 1427 1428 LLVMRunFunctionPassManager(gallivm->passmgr, variant_func); 1429 1430 if (gallivm_debug & GALLIVM_DEBUG_IR) { 1431 lp_debug_dump_value(variant_func); 1432 debug_printf("\n"); 1433 } 1434 1435 code = LLVMGetPointerToGlobal(gallivm->engine, variant_func); 1436 if (elts) 1437 variant->jit_func_elts = (draw_jit_vert_func_elts) pointer_to_func(code); 1438 else 1439 variant->jit_func = (draw_jit_vert_func) pointer_to_func(code); 1440 1441 if (gallivm_debug & GALLIVM_DEBUG_ASM) { 1442 lp_disassemble(code); 1443 } 1444 lp_func_delete_body(variant_func); 1445} 1446 1447 1448struct draw_llvm_variant_key * 1449draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store) 1450{ 1451 unsigned i; 1452 struct draw_llvm_variant_key *key; 1453 struct lp_sampler_static_state *sampler; 1454 1455 key = (struct draw_llvm_variant_key *)store; 1456 1457 key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/ 1458 1459 /* Presumably all variants of the shader should have the same 1460 * number of vertex elements - ie the number of shader inputs. 1461 */ 1462 key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements; 1463 1464 /* will have to rig this up properly later */ 1465 key->clip_xy = llvm->draw->clip_xy; 1466 key->clip_z = llvm->draw->clip_z; 1467 key->clip_user = llvm->draw->clip_user; 1468 key->bypass_viewport = llvm->draw->identity_viewport; 1469 key->clip_halfz = !llvm->draw->rasterizer->gl_rasterization_rules; 1470 key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE); 1471 key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable; 1472 key->pad = 0; 1473 1474 /* All variants of this shader will have the same value for 1475 * nr_samplers. Not yet trying to compact away holes in the 1476 * sampler array. 1477 */ 1478 key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1; 1479 1480 sampler = draw_llvm_variant_key_samplers(key); 1481 1482 memcpy(key->vertex_element, 1483 llvm->draw->pt.vertex_element, 1484 sizeof(struct pipe_vertex_element) * key->nr_vertex_elements); 1485 1486 memset(sampler, 0, key->nr_samplers * sizeof *sampler); 1487 1488 for (i = 0 ; i < key->nr_samplers; i++) { 1489 lp_sampler_static_state(&sampler[i], 1490 llvm->draw->sampler_views[i], 1491 llvm->draw->samplers[i]); 1492 } 1493 1494 return key; 1495} 1496 1497 1498void 1499draw_llvm_set_mapped_texture(struct draw_context *draw, 1500 unsigned sampler_idx, 1501 uint32_t width, uint32_t height, uint32_t depth, 1502 uint32_t first_level, uint32_t last_level, 1503 uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS], 1504 uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS], 1505 const void *data[PIPE_MAX_TEXTURE_LEVELS]) 1506{ 1507 unsigned j; 1508 struct draw_jit_texture *jit_tex; 1509 1510 assert(sampler_idx < PIPE_MAX_VERTEX_SAMPLERS); 1511 1512 jit_tex = &draw->llvm->jit_context.textures[sampler_idx]; 1513 1514 jit_tex->width = width; 1515 jit_tex->height = height; 1516 jit_tex->depth = depth; 1517 jit_tex->first_level = first_level; 1518 jit_tex->last_level = last_level; 1519 1520 for (j = first_level; j <= last_level; j++) { 1521 jit_tex->data[j] = data[j]; 1522 jit_tex->row_stride[j] = row_stride[j]; 1523 jit_tex->img_stride[j] = img_stride[j]; 1524 } 1525} 1526 1527 1528void 1529draw_llvm_set_sampler_state(struct draw_context *draw) 1530{ 1531 unsigned i; 1532 1533 for (i = 0; i < draw->num_samplers; i++) { 1534 struct draw_jit_texture *jit_tex = &draw->llvm->jit_context.textures[i]; 1535 1536 if (draw->samplers[i]) { 1537 jit_tex->min_lod = draw->samplers[i]->min_lod; 1538 jit_tex->max_lod = draw->samplers[i]->max_lod; 1539 jit_tex->lod_bias = draw->samplers[i]->lod_bias; 1540 COPY_4V(jit_tex->border_color, draw->samplers[i]->border_color.f); 1541 } 1542 } 1543} 1544 1545 1546void 1547draw_llvm_destroy_variant(struct draw_llvm_variant *variant) 1548{ 1549 struct draw_llvm *llvm = variant->llvm; 1550 1551 if (variant->function_elts) { 1552 LLVMFreeMachineCodeForFunction(llvm->gallivm->engine, 1553 variant->function_elts); 1554 LLVMDeleteFunction(variant->function_elts); 1555 } 1556 1557 if (variant->function) { 1558 LLVMFreeMachineCodeForFunction(llvm->gallivm->engine, 1559 variant->function); 1560 LLVMDeleteFunction(variant->function); 1561 } 1562 1563 remove_from_list(&variant->list_item_local); 1564 variant->shader->variants_cached--; 1565 remove_from_list(&variant->list_item_global); 1566 llvm->nr_variants--; 1567 FREE(variant); 1568} 1569