1 /* 2 Copyright (C) Intel Corp. 2006. All Rights Reserved. 3 Intel funded Tungsten Graphics to 4 develop this 3D driver. 5 6 Permission is hereby granted, free of charge, to any person obtaining 7 a 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, sublicense, 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 16 portions of the Software. 17 18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. 21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE 22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION 23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION 24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 26 **********************************************************************/ 27 /* 28 * Authors: 29 * Keith Whitwell <keithw (at) vmware.com> 30 */ 31 32 33 #ifndef BRW_EU_H 34 #define BRW_EU_H 35 36 #include <stdbool.h> 37 #include "brw_inst.h" 38 #include "brw_defines.h" 39 #include "brw_reg.h" 40 #include "intel_asm_annotation.h" 41 42 #ifdef __cplusplus 43 extern "C" { 44 #endif 45 46 #define BRW_EU_MAX_INSN_STACK 5 47 48 /* A helper for accessing the last instruction emitted. This makes it easy 49 * to set various bits on an instruction without having to create temporary 50 * variable and assign the emitted instruction to those. 51 */ 52 #define brw_last_inst (&p->store[p->nr_insn - 1]) 53 54 struct brw_codegen { 55 brw_inst *store; 56 int store_size; 57 unsigned nr_insn; 58 unsigned int next_insn_offset; 59 60 void *mem_ctx; 61 62 /* Allow clients to push/pop instruction state: 63 */ 64 brw_inst stack[BRW_EU_MAX_INSN_STACK]; 65 bool compressed_stack[BRW_EU_MAX_INSN_STACK]; 66 brw_inst *current; 67 68 bool single_program_flow; 69 const struct gen_device_info *devinfo; 70 71 /* Control flow stacks: 72 * - if_stack contains IF and ELSE instructions which must be patched 73 * (and popped) once the matching ENDIF instruction is encountered. 74 * 75 * Just store the instruction pointer(an index). 76 */ 77 int *if_stack; 78 int if_stack_depth; 79 int if_stack_array_size; 80 81 /** 82 * loop_stack contains the instruction pointers of the starts of loops which 83 * must be patched (and popped) once the matching WHILE instruction is 84 * encountered. 85 */ 86 int *loop_stack; 87 /** 88 * pre-gen6, the BREAK and CONT instructions had to tell how many IF/ENDIF 89 * blocks they were popping out of, to fix up the mask stack. This tracks 90 * the IF/ENDIF nesting in each current nested loop level. 91 */ 92 int *if_depth_in_loop; 93 int loop_stack_depth; 94 int loop_stack_array_size; 95 }; 96 97 void brw_pop_insn_state( struct brw_codegen *p ); 98 void brw_push_insn_state( struct brw_codegen *p ); 99 void brw_set_default_exec_size(struct brw_codegen *p, unsigned value); 100 void brw_set_default_mask_control( struct brw_codegen *p, unsigned value ); 101 void brw_set_default_saturate( struct brw_codegen *p, bool enable ); 102 void brw_set_default_access_mode( struct brw_codegen *p, unsigned access_mode ); 103 void brw_inst_set_compression(const struct gen_device_info *devinfo, 104 brw_inst *inst, bool on); 105 void brw_set_default_compression(struct brw_codegen *p, bool on); 106 void brw_inst_set_group(const struct gen_device_info *devinfo, 107 brw_inst *inst, unsigned group); 108 void brw_set_default_group(struct brw_codegen *p, unsigned group); 109 void brw_set_default_compression_control(struct brw_codegen *p, enum brw_compression c); 110 void brw_set_default_predicate_control( struct brw_codegen *p, unsigned pc ); 111 void brw_set_default_predicate_inverse(struct brw_codegen *p, bool predicate_inverse); 112 void brw_set_default_flag_reg(struct brw_codegen *p, int reg, int subreg); 113 void brw_set_default_acc_write_control(struct brw_codegen *p, unsigned value); 114 115 void brw_init_codegen(const struct gen_device_info *, struct brw_codegen *p, 116 void *mem_ctx); 117 void brw_disassemble(const struct gen_device_info *devinfo, void *assembly, 118 int start, int end, FILE *out); 119 const unsigned *brw_get_program( struct brw_codegen *p, unsigned *sz ); 120 121 brw_inst *brw_next_insn(struct brw_codegen *p, unsigned opcode); 122 void brw_set_dest(struct brw_codegen *p, brw_inst *insn, struct brw_reg dest); 123 void brw_set_src0(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg); 124 125 void gen6_resolve_implied_move(struct brw_codegen *p, 126 struct brw_reg *src, 127 unsigned msg_reg_nr); 128 129 /* Helpers for regular instructions: 130 */ 131 #define ALU1(OP) \ 132 brw_inst *brw_##OP(struct brw_codegen *p, \ 133 struct brw_reg dest, \ 134 struct brw_reg src0); 135 136 #define ALU2(OP) \ 137 brw_inst *brw_##OP(struct brw_codegen *p, \ 138 struct brw_reg dest, \ 139 struct brw_reg src0, \ 140 struct brw_reg src1); 141 142 #define ALU3(OP) \ 143 brw_inst *brw_##OP(struct brw_codegen *p, \ 144 struct brw_reg dest, \ 145 struct brw_reg src0, \ 146 struct brw_reg src1, \ 147 struct brw_reg src2); 148 149 #define ROUND(OP) \ 150 void brw_##OP(struct brw_codegen *p, struct brw_reg dest, struct brw_reg src0); 151 152 ALU1(MOV) 153 ALU2(SEL) 154 ALU1(NOT) 155 ALU2(AND) 156 ALU2(OR) 157 ALU2(XOR) 158 ALU2(SHR) 159 ALU2(SHL) 160 ALU1(DIM) 161 ALU2(ASR) 162 ALU1(F32TO16) 163 ALU1(F16TO32) 164 ALU2(ADD) 165 ALU2(AVG) 166 ALU2(MUL) 167 ALU1(FRC) 168 ALU1(RNDD) 169 ALU2(MAC) 170 ALU2(MACH) 171 ALU1(LZD) 172 ALU2(DP4) 173 ALU2(DPH) 174 ALU2(DP3) 175 ALU2(DP2) 176 ALU2(LINE) 177 ALU2(PLN) 178 ALU3(MAD) 179 ALU3(LRP) 180 ALU1(BFREV) 181 ALU3(BFE) 182 ALU2(BFI1) 183 ALU3(BFI2) 184 ALU1(FBH) 185 ALU1(FBL) 186 ALU1(CBIT) 187 ALU2(ADDC) 188 ALU2(SUBB) 189 ALU2(MAC) 190 191 ROUND(RNDZ) 192 ROUND(RNDE) 193 194 #undef ALU1 195 #undef ALU2 196 #undef ALU3 197 #undef ROUND 198 199 200 /* Helpers for SEND instruction: 201 */ 202 void brw_set_sampler_message(struct brw_codegen *p, 203 brw_inst *insn, 204 unsigned binding_table_index, 205 unsigned sampler, 206 unsigned msg_type, 207 unsigned response_length, 208 unsigned msg_length, 209 unsigned header_present, 210 unsigned simd_mode, 211 unsigned return_format); 212 213 void brw_set_message_descriptor(struct brw_codegen *p, 214 brw_inst *inst, 215 enum brw_message_target sfid, 216 unsigned msg_length, 217 unsigned response_length, 218 bool header_present, 219 bool end_of_thread); 220 221 void brw_set_dp_read_message(struct brw_codegen *p, 222 brw_inst *insn, 223 unsigned binding_table_index, 224 unsigned msg_control, 225 unsigned msg_type, 226 unsigned target_cache, 227 unsigned msg_length, 228 bool header_present, 229 unsigned response_length); 230 231 void brw_set_dp_write_message(struct brw_codegen *p, 232 brw_inst *insn, 233 unsigned binding_table_index, 234 unsigned msg_control, 235 unsigned msg_type, 236 unsigned target_cache, 237 unsigned msg_length, 238 bool header_present, 239 unsigned last_render_target, 240 unsigned response_length, 241 unsigned end_of_thread, 242 unsigned send_commit_msg); 243 244 void brw_urb_WRITE(struct brw_codegen *p, 245 struct brw_reg dest, 246 unsigned msg_reg_nr, 247 struct brw_reg src0, 248 enum brw_urb_write_flags flags, 249 unsigned msg_length, 250 unsigned response_length, 251 unsigned offset, 252 unsigned swizzle); 253 254 /** 255 * Send message to shared unit \p sfid with a possibly indirect descriptor \p 256 * desc. If \p desc is not an immediate it will be transparently loaded to an 257 * address register using an OR instruction. The returned instruction can be 258 * passed as argument to the usual brw_set_*_message() functions in order to 259 * specify any additional descriptor bits -- If \p desc is an immediate this 260 * will be the SEND instruction itself, otherwise it will be the OR 261 * instruction. 262 */ 263 struct brw_inst * 264 brw_send_indirect_message(struct brw_codegen *p, 265 unsigned sfid, 266 struct brw_reg dst, 267 struct brw_reg payload, 268 struct brw_reg desc); 269 270 void brw_ff_sync(struct brw_codegen *p, 271 struct brw_reg dest, 272 unsigned msg_reg_nr, 273 struct brw_reg src0, 274 bool allocate, 275 unsigned response_length, 276 bool eot); 277 278 void brw_svb_write(struct brw_codegen *p, 279 struct brw_reg dest, 280 unsigned msg_reg_nr, 281 struct brw_reg src0, 282 unsigned binding_table_index, 283 bool send_commit_msg); 284 285 void brw_fb_WRITE(struct brw_codegen *p, 286 struct brw_reg payload, 287 struct brw_reg implied_header, 288 unsigned msg_control, 289 unsigned binding_table_index, 290 unsigned msg_length, 291 unsigned response_length, 292 bool eot, 293 bool last_render_target, 294 bool header_present); 295 296 brw_inst *gen9_fb_READ(struct brw_codegen *p, 297 struct brw_reg dst, 298 struct brw_reg payload, 299 unsigned binding_table_index, 300 unsigned msg_length, 301 unsigned response_length, 302 bool per_sample); 303 304 void brw_SAMPLE(struct brw_codegen *p, 305 struct brw_reg dest, 306 unsigned msg_reg_nr, 307 struct brw_reg src0, 308 unsigned binding_table_index, 309 unsigned sampler, 310 unsigned msg_type, 311 unsigned response_length, 312 unsigned msg_length, 313 unsigned header_present, 314 unsigned simd_mode, 315 unsigned return_format); 316 317 void brw_adjust_sampler_state_pointer(struct brw_codegen *p, 318 struct brw_reg header, 319 struct brw_reg sampler_index); 320 321 void gen4_math(struct brw_codegen *p, 322 struct brw_reg dest, 323 unsigned function, 324 unsigned msg_reg_nr, 325 struct brw_reg src, 326 unsigned precision ); 327 328 void gen6_math(struct brw_codegen *p, 329 struct brw_reg dest, 330 unsigned function, 331 struct brw_reg src0, 332 struct brw_reg src1); 333 334 void brw_oword_block_read(struct brw_codegen *p, 335 struct brw_reg dest, 336 struct brw_reg mrf, 337 uint32_t offset, 338 uint32_t bind_table_index); 339 340 unsigned brw_scratch_surface_idx(const struct brw_codegen *p); 341 342 void brw_oword_block_read_scratch(struct brw_codegen *p, 343 struct brw_reg dest, 344 struct brw_reg mrf, 345 int num_regs, 346 unsigned offset); 347 348 void brw_oword_block_write_scratch(struct brw_codegen *p, 349 struct brw_reg mrf, 350 int num_regs, 351 unsigned offset); 352 353 void gen7_block_read_scratch(struct brw_codegen *p, 354 struct brw_reg dest, 355 int num_regs, 356 unsigned offset); 357 358 void brw_shader_time_add(struct brw_codegen *p, 359 struct brw_reg payload, 360 uint32_t surf_index); 361 362 /** 363 * Return the generation-specific jump distance scaling factor. 364 * 365 * Given the number of instructions to jump, we need to scale by 366 * some number to obtain the actual jump distance to program in an 367 * instruction. 368 */ 369 static inline unsigned 370 brw_jump_scale(const struct gen_device_info *devinfo) 371 { 372 /* Broadwell measures jump targets in bytes. */ 373 if (devinfo->gen >= 8) 374 return 16; 375 376 /* Ironlake and later measure jump targets in 64-bit data chunks (in order 377 * (to support compaction), so each 128-bit instruction requires 2 chunks. 378 */ 379 if (devinfo->gen >= 5) 380 return 2; 381 382 /* Gen4 simply uses the number of 128-bit instructions. */ 383 return 1; 384 } 385 386 void brw_barrier(struct brw_codegen *p, struct brw_reg src); 387 388 /* If/else/endif. Works by manipulating the execution flags on each 389 * channel. 390 */ 391 brw_inst *brw_IF(struct brw_codegen *p, unsigned execute_size); 392 brw_inst *gen6_IF(struct brw_codegen *p, enum brw_conditional_mod conditional, 393 struct brw_reg src0, struct brw_reg src1); 394 395 void brw_ELSE(struct brw_codegen *p); 396 void brw_ENDIF(struct brw_codegen *p); 397 398 /* DO/WHILE loops: 399 */ 400 brw_inst *brw_DO(struct brw_codegen *p, unsigned execute_size); 401 402 brw_inst *brw_WHILE(struct brw_codegen *p); 403 404 brw_inst *brw_BREAK(struct brw_codegen *p); 405 brw_inst *brw_CONT(struct brw_codegen *p); 406 brw_inst *gen6_HALT(struct brw_codegen *p); 407 408 /* Forward jumps: 409 */ 410 void brw_land_fwd_jump(struct brw_codegen *p, int jmp_insn_idx); 411 412 brw_inst *brw_JMPI(struct brw_codegen *p, struct brw_reg index, 413 unsigned predicate_control); 414 415 void brw_NOP(struct brw_codegen *p); 416 417 void brw_WAIT(struct brw_codegen *p); 418 419 /* Special case: there is never a destination, execution size will be 420 * taken from src0: 421 */ 422 void brw_CMP(struct brw_codegen *p, 423 struct brw_reg dest, 424 unsigned conditional, 425 struct brw_reg src0, 426 struct brw_reg src1); 427 428 void 429 brw_untyped_atomic(struct brw_codegen *p, 430 struct brw_reg dst, 431 struct brw_reg payload, 432 struct brw_reg surface, 433 unsigned atomic_op, 434 unsigned msg_length, 435 bool response_expected); 436 437 void 438 brw_untyped_surface_read(struct brw_codegen *p, 439 struct brw_reg dst, 440 struct brw_reg payload, 441 struct brw_reg surface, 442 unsigned msg_length, 443 unsigned num_channels); 444 445 void 446 brw_untyped_surface_write(struct brw_codegen *p, 447 struct brw_reg payload, 448 struct brw_reg surface, 449 unsigned msg_length, 450 unsigned num_channels); 451 452 void 453 brw_typed_atomic(struct brw_codegen *p, 454 struct brw_reg dst, 455 struct brw_reg payload, 456 struct brw_reg surface, 457 unsigned atomic_op, 458 unsigned msg_length, 459 bool response_expected); 460 461 void 462 brw_typed_surface_read(struct brw_codegen *p, 463 struct brw_reg dst, 464 struct brw_reg payload, 465 struct brw_reg surface, 466 unsigned msg_length, 467 unsigned num_channels); 468 469 void 470 brw_typed_surface_write(struct brw_codegen *p, 471 struct brw_reg payload, 472 struct brw_reg surface, 473 unsigned msg_length, 474 unsigned num_channels); 475 476 void 477 brw_memory_fence(struct brw_codegen *p, 478 struct brw_reg dst); 479 480 void 481 brw_pixel_interpolator_query(struct brw_codegen *p, 482 struct brw_reg dest, 483 struct brw_reg mrf, 484 bool noperspective, 485 unsigned mode, 486 struct brw_reg data, 487 unsigned msg_length, 488 unsigned response_length); 489 490 void 491 brw_find_live_channel(struct brw_codegen *p, 492 struct brw_reg dst, 493 struct brw_reg mask); 494 495 void 496 brw_broadcast(struct brw_codegen *p, 497 struct brw_reg dst, 498 struct brw_reg src, 499 struct brw_reg idx); 500 501 /*********************************************************************** 502 * brw_eu_util.c: 503 */ 504 505 void brw_copy_indirect_to_indirect(struct brw_codegen *p, 506 struct brw_indirect dst_ptr, 507 struct brw_indirect src_ptr, 508 unsigned count); 509 510 void brw_copy_from_indirect(struct brw_codegen *p, 511 struct brw_reg dst, 512 struct brw_indirect ptr, 513 unsigned count); 514 515 void brw_copy4(struct brw_codegen *p, 516 struct brw_reg dst, 517 struct brw_reg src, 518 unsigned count); 519 520 void brw_copy8(struct brw_codegen *p, 521 struct brw_reg dst, 522 struct brw_reg src, 523 unsigned count); 524 525 void brw_math_invert( struct brw_codegen *p, 526 struct brw_reg dst, 527 struct brw_reg src); 528 529 void brw_set_src1(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg); 530 531 void brw_set_uip_jip(struct brw_codegen *p, int start_offset); 532 533 enum brw_conditional_mod brw_negate_cmod(uint32_t cmod); 534 enum brw_conditional_mod brw_swap_cmod(uint32_t cmod); 535 536 /* brw_eu_compact.c */ 537 void brw_init_compaction_tables(const struct gen_device_info *devinfo); 538 void brw_compact_instructions(struct brw_codegen *p, int start_offset, 539 int num_annotations, struct annotation *annotation); 540 void brw_uncompact_instruction(const struct gen_device_info *devinfo, 541 brw_inst *dst, brw_compact_inst *src); 542 bool brw_try_compact_instruction(const struct gen_device_info *devinfo, 543 brw_compact_inst *dst, brw_inst *src); 544 545 void brw_debug_compact_uncompact(const struct gen_device_info *devinfo, 546 brw_inst *orig, brw_inst *uncompacted); 547 548 /* brw_eu_validate.c */ 549 bool brw_validate_instructions(const struct brw_codegen *p, int start_offset, 550 struct annotation_info *annotation); 551 552 static inline int 553 next_offset(const struct gen_device_info *devinfo, void *store, int offset) 554 { 555 brw_inst *insn = (brw_inst *)((char *)store + offset); 556 557 if (brw_inst_cmpt_control(devinfo, insn)) 558 return offset + 8; 559 else 560 return offset + 16; 561 } 562 563 struct opcode_desc { 564 /* The union is an implementation detail used by brw_opcode_desc() to handle 565 * opcodes that have been reused for different instructions across hardware 566 * generations. 567 * 568 * The gens field acts as a tag. If it is non-zero, name points to a string 569 * containing the instruction mnemonic. If it is zero, the table field is 570 * valid and either points to a secondary opcode_desc table with 'size' 571 * elements or is NULL and no such instruction exists for the opcode. 572 */ 573 union { 574 struct { 575 char *name; 576 int nsrc; 577 }; 578 struct { 579 const struct opcode_desc *table; 580 unsigned size; 581 }; 582 }; 583 int ndst; 584 int gens; 585 }; 586 587 const struct opcode_desc * 588 brw_opcode_desc(const struct gen_device_info *devinfo, enum opcode opcode); 589 590 static inline bool 591 is_3src(const struct gen_device_info *devinfo, enum opcode opcode) 592 { 593 const struct opcode_desc *desc = brw_opcode_desc(devinfo, opcode); 594 return desc && desc->nsrc == 3; 595 } 596 597 /** Maximum SEND message length */ 598 #define BRW_MAX_MSG_LENGTH 15 599 600 /** First MRF register used by pull loads */ 601 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13) 602 603 /** First MRF register used by spills */ 604 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13) 605 606 #ifdef __cplusplus 607 } 608 #endif 609 610 #endif 611