1 /* 2 * Copyright 2014 Broadcom 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21 * IN THE SOFTWARE. 22 */ 23 24 #ifndef VC4_QIR_H 25 #define VC4_QIR_H 26 27 #include <assert.h> 28 #include <stdio.h> 29 #include <stdlib.h> 30 #include <stdbool.h> 31 #include <stdint.h> 32 #include <string.h> 33 34 #include "util/macros.h" 35 #include "compiler/nir/nir.h" 36 #include "util/list.h" 37 #include "util/u_math.h" 38 39 #include "vc4_screen.h" 40 #include "vc4_qpu_defines.h" 41 #include "vc4_qpu.h" 42 #include "kernel/vc4_packet.h" 43 #include "pipe/p_state.h" 44 45 struct nir_builder; 46 47 enum qfile { 48 QFILE_NULL, 49 QFILE_TEMP, 50 QFILE_VARY, 51 QFILE_UNIF, 52 QFILE_VPM, 53 QFILE_TLB_COLOR_WRITE, 54 QFILE_TLB_COLOR_WRITE_MS, 55 QFILE_TLB_Z_WRITE, 56 QFILE_TLB_STENCIL_SETUP, 57 58 /* If tex_s is written on its own without preceding t/r/b setup, it's 59 * a direct memory access using the input value, without the sideband 60 * uniform load. We represent these in QIR as a separate write 61 * destination so we can tell if the sideband uniform is present. 62 */ 63 QFILE_TEX_S_DIRECT, 64 65 QFILE_TEX_S, 66 QFILE_TEX_T, 67 QFILE_TEX_R, 68 QFILE_TEX_B, 69 70 /* Payload registers that aren't in the physical register file, so we 71 * can just use the corresponding qpu_reg at qpu_emit time. 72 */ 73 QFILE_FRAG_X, 74 QFILE_FRAG_Y, 75 QFILE_FRAG_REV_FLAG, 76 QFILE_QPU_ELEMENT, 77 78 /** 79 * Stores an immediate value in the index field that will be used 80 * directly by qpu_load_imm(). 81 */ 82 QFILE_LOAD_IMM, 83 84 /** 85 * Stores an immediate value in the index field that can be turned 86 * into a small immediate field by qpu_encode_small_immediate(). 87 */ 88 QFILE_SMALL_IMM, 89 }; 90 91 struct qreg { 92 enum qfile file; 93 uint32_t index; 94 int pack; 95 }; 96 97 static inline struct qreg qir_reg(enum qfile file, uint32_t index) 98 { 99 return (struct qreg){file, index}; 100 } 101 102 enum qop { 103 QOP_UNDEF, 104 QOP_MOV, 105 QOP_FMOV, 106 QOP_MMOV, 107 QOP_FADD, 108 QOP_FSUB, 109 QOP_FMUL, 110 QOP_V8MULD, 111 QOP_V8MIN, 112 QOP_V8MAX, 113 QOP_V8ADDS, 114 QOP_V8SUBS, 115 QOP_MUL24, 116 QOP_FMIN, 117 QOP_FMAX, 118 QOP_FMINABS, 119 QOP_FMAXABS, 120 QOP_ADD, 121 QOP_SUB, 122 QOP_SHL, 123 QOP_SHR, 124 QOP_ASR, 125 QOP_MIN, 126 QOP_MIN_NOIMM, 127 QOP_MAX, 128 QOP_AND, 129 QOP_OR, 130 QOP_XOR, 131 QOP_NOT, 132 133 QOP_FTOI, 134 QOP_ITOF, 135 QOP_RCP, 136 QOP_RSQ, 137 QOP_EXP2, 138 QOP_LOG2, 139 QOP_VW_SETUP, 140 QOP_VR_SETUP, 141 QOP_TLB_COLOR_READ, 142 QOP_MS_MASK, 143 QOP_VARY_ADD_C, 144 145 QOP_FRAG_Z, 146 QOP_FRAG_W, 147 148 /** 149 * Signal of texture read being necessary and then reading r4 into 150 * the destination 151 */ 152 QOP_TEX_RESULT, 153 154 /** 155 * Insert the signal for switching threads in a threaded fragment 156 * shader. No value can be live in an accumulator across a thrsw. 157 * 158 * At the QPU level, this will have several delay slots before the 159 * switch happens. Those slots are the responsibility of the 160 * scheduler. 161 */ 162 QOP_THRSW, 163 164 /* 32-bit immediate loaded to each SIMD channel */ 165 QOP_LOAD_IMM, 166 167 /* 32-bit immediate divided into 16 2-bit unsigned int values and 168 * loaded to each corresponding SIMD channel. 169 */ 170 QOP_LOAD_IMM_U2, 171 /* 32-bit immediate divided into 16 2-bit signed int values and 172 * loaded to each corresponding SIMD channel. 173 */ 174 QOP_LOAD_IMM_I2, 175 176 QOP_ROT_MUL, 177 178 /* Jumps to block->successor[0] if the qinst->cond (as a 179 * QPU_COND_BRANCH_*) passes, or block->successor[1] if not. Note 180 * that block->successor[1] may be unset if the condition is ALWAYS. 181 */ 182 QOP_BRANCH, 183 184 /* Emits an ADD from src[0] to src[1], where src[0] must be a 185 * QOP_LOAD_IMM result and src[1] is a QUNIFORM_UNIFORMS_ADDRESS, 186 * required by the kernel as part of its branch validation. 187 */ 188 QOP_UNIFORMS_RESET, 189 }; 190 191 struct queued_qpu_inst { 192 struct list_head link; 193 uint64_t inst; 194 }; 195 196 struct qinst { 197 struct list_head link; 198 199 enum qop op; 200 struct qreg dst; 201 struct qreg src[3]; 202 bool sf; 203 bool cond_is_exec_mask; 204 uint8_t cond; 205 }; 206 207 enum qstage { 208 /** 209 * Coordinate shader, runs during binning, before the VS, and just 210 * outputs position. 211 */ 212 QSTAGE_COORD, 213 QSTAGE_VERT, 214 QSTAGE_FRAG, 215 }; 216 217 enum quniform_contents { 218 /** 219 * Indicates that a constant 32-bit value is copied from the program's 220 * uniform contents. 221 */ 222 QUNIFORM_CONSTANT, 223 /** 224 * Indicates that the program's uniform contents are used as an index 225 * into the GL uniform storage. 226 */ 227 QUNIFORM_UNIFORM, 228 229 /** @{ 230 * Scaling factors from clip coordinates to relative to the viewport 231 * center. 232 * 233 * This is used by the coordinate and vertex shaders to produce the 234 * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed 235 * point offsets from the viewport ccenter. 236 */ 237 QUNIFORM_VIEWPORT_X_SCALE, 238 QUNIFORM_VIEWPORT_Y_SCALE, 239 /** @} */ 240 241 QUNIFORM_VIEWPORT_Z_OFFSET, 242 QUNIFORM_VIEWPORT_Z_SCALE, 243 244 QUNIFORM_USER_CLIP_PLANE, 245 246 /** 247 * A reference to a texture config parameter 0 uniform. 248 * 249 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which 250 * defines texture type, miplevels, and such. It will be found as a 251 * parameter to the first QOP_TEX_[STRB] instruction in a sequence. 252 */ 253 QUNIFORM_TEXTURE_CONFIG_P0, 254 255 /** 256 * A reference to a texture config parameter 1 uniform. 257 * 258 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which 259 * defines texture width, height, filters, and wrap modes. It will be 260 * found as a parameter to the second QOP_TEX_[STRB] instruction in a 261 * sequence. 262 */ 263 QUNIFORM_TEXTURE_CONFIG_P1, 264 265 /** A reference to a texture config parameter 2 cubemap stride uniform */ 266 QUNIFORM_TEXTURE_CONFIG_P2, 267 268 QUNIFORM_TEXTURE_FIRST_LEVEL, 269 270 QUNIFORM_TEXTURE_MSAA_ADDR, 271 272 QUNIFORM_UBO_ADDR, 273 274 QUNIFORM_TEXRECT_SCALE_X, 275 QUNIFORM_TEXRECT_SCALE_Y, 276 277 QUNIFORM_TEXTURE_BORDER_COLOR, 278 279 QUNIFORM_BLEND_CONST_COLOR_X, 280 QUNIFORM_BLEND_CONST_COLOR_Y, 281 QUNIFORM_BLEND_CONST_COLOR_Z, 282 QUNIFORM_BLEND_CONST_COLOR_W, 283 QUNIFORM_BLEND_CONST_COLOR_RGBA, 284 QUNIFORM_BLEND_CONST_COLOR_AAAA, 285 286 QUNIFORM_STENCIL, 287 288 QUNIFORM_ALPHA_REF, 289 QUNIFORM_SAMPLE_MASK, 290 291 /* Placeholder uniform that will be updated by the kernel when used by 292 * an instruction writing to QPU_W_UNIFORMS_ADDRESS. 293 */ 294 QUNIFORM_UNIFORMS_ADDRESS, 295 }; 296 297 struct vc4_varying_slot { 298 uint8_t slot; 299 uint8_t swizzle; 300 }; 301 302 struct vc4_compiler_ubo_range { 303 /** 304 * offset in bytes from the start of the ubo where this range is 305 * uploaded. 306 * 307 * Only set once used is set. 308 */ 309 uint32_t dst_offset; 310 311 /** 312 * offset in bytes from the start of the gallium uniforms where the 313 * data comes from. 314 */ 315 uint32_t src_offset; 316 317 /** size in bytes of this ubo range */ 318 uint32_t size; 319 320 /** 321 * Set if this range is used by the shader for indirect uniforms 322 * access. 323 */ 324 bool used; 325 }; 326 327 struct vc4_key { 328 struct vc4_uncompiled_shader *shader_state; 329 struct { 330 enum pipe_format format; 331 uint8_t swizzle[4]; 332 union { 333 struct { 334 unsigned compare_mode:1; 335 unsigned compare_func:3; 336 unsigned wrap_s:3; 337 unsigned wrap_t:3; 338 bool force_first_level:1; 339 }; 340 struct { 341 uint16_t msaa_width, msaa_height; 342 }; 343 }; 344 } tex[VC4_MAX_TEXTURE_SAMPLERS]; 345 uint8_t ucp_enables; 346 }; 347 348 struct vc4_fs_key { 349 struct vc4_key base; 350 enum pipe_format color_format; 351 bool depth_enabled; 352 bool stencil_enabled; 353 bool stencil_twoside; 354 bool stencil_full_writemasks; 355 bool is_points; 356 bool is_lines; 357 bool point_coord_upper_left; 358 bool light_twoside; 359 bool msaa; 360 bool sample_coverage; 361 bool sample_alpha_to_coverage; 362 bool sample_alpha_to_one; 363 uint8_t alpha_test_func; 364 uint8_t logicop_func; 365 uint32_t point_sprite_mask; 366 367 struct pipe_rt_blend_state blend; 368 }; 369 370 struct vc4_vs_key { 371 struct vc4_key base; 372 373 const struct vc4_fs_inputs *fs_inputs; 374 enum pipe_format attr_formats[8]; 375 bool is_coord; 376 bool per_vertex_point_size; 377 bool clamp_color; 378 }; 379 380 /** A basic block of QIR intructions. */ 381 struct qblock { 382 struct list_head link; 383 384 struct list_head instructions; 385 struct list_head qpu_inst_list; 386 387 struct set *predecessors; 388 struct qblock *successors[2]; 389 390 int index; 391 392 /* Instruction IPs for the first and last instruction of the block. 393 * Set by vc4_qpu_schedule.c. 394 */ 395 uint32_t start_qpu_ip; 396 uint32_t end_qpu_ip; 397 398 /* Instruction IP for the branch instruction of the block. Set by 399 * vc4_qpu_schedule.c. 400 */ 401 uint32_t branch_qpu_ip; 402 403 /** @{ used by vc4_qir_live_variables.c */ 404 BITSET_WORD *def; 405 BITSET_WORD *use; 406 BITSET_WORD *live_in; 407 BITSET_WORD *live_out; 408 int start_ip, end_ip; 409 /** @} */ 410 }; 411 412 struct vc4_compile { 413 struct vc4_context *vc4; 414 nir_shader *s; 415 nir_function_impl *impl; 416 struct exec_list *cf_node_list; 417 418 /** 419 * Mapping from nir_register * or nir_ssa_def * to array of struct 420 * qreg for the values. 421 */ 422 struct hash_table *def_ht; 423 424 /* For each temp, the instruction generating its value. */ 425 struct qinst **defs; 426 uint32_t defs_array_size; 427 428 /** 429 * Inputs to the shader, arranged by TGSI declaration order. 430 * 431 * Not all fragment shader QFILE_VARY reads are present in this array. 432 */ 433 struct qreg *inputs; 434 struct qreg *outputs; 435 bool msaa_per_sample_output; 436 struct qreg color_reads[VC4_MAX_SAMPLES]; 437 struct qreg sample_colors[VC4_MAX_SAMPLES]; 438 uint32_t inputs_array_size; 439 uint32_t outputs_array_size; 440 uint32_t uniforms_array_size; 441 442 struct vc4_compiler_ubo_range *ubo_ranges; 443 uint32_t ubo_ranges_array_size; 444 /** Number of uniform areas declared in ubo_ranges. */ 445 uint32_t num_uniform_ranges; 446 /** Number of uniform areas used for indirect addressed loads. */ 447 uint32_t num_ubo_ranges; 448 uint32_t next_ubo_dst_offset; 449 450 /* State for whether we're executing on each channel currently. 0 if 451 * yes, otherwise a block number + 1 that the channel jumped to. 452 */ 453 struct qreg execute; 454 455 struct qreg line_x, point_x, point_y; 456 /** boolean (~0 -> true) if the fragment has been discarded. */ 457 struct qreg discard; 458 struct qreg payload_FRAG_Z; 459 struct qreg payload_FRAG_W; 460 461 uint8_t vattr_sizes[8]; 462 463 /** 464 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads. 465 * 466 * This includes those that aren't part of the VPM varyings, like 467 * point/line coordinates. 468 */ 469 struct vc4_varying_slot *input_slots; 470 uint32_t num_input_slots; 471 uint32_t input_slots_array_size; 472 473 /** 474 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_* 475 * of the output is. Used to emit from the VS in the order that the 476 * FS needs. 477 */ 478 struct vc4_varying_slot *output_slots; 479 480 struct pipe_shader_state *shader_state; 481 struct vc4_key *key; 482 struct vc4_fs_key *fs_key; 483 struct vc4_vs_key *vs_key; 484 485 /* Live ranges of temps. */ 486 int *temp_start, *temp_end; 487 488 uint32_t *uniform_data; 489 enum quniform_contents *uniform_contents; 490 uint32_t uniform_array_size; 491 uint32_t num_uniforms; 492 uint32_t num_outputs; 493 uint32_t num_texture_samples; 494 uint32_t output_position_index; 495 uint32_t output_color_index; 496 uint32_t output_point_size_index; 497 uint32_t output_sample_mask_index; 498 499 struct qreg undef; 500 enum qstage stage; 501 uint32_t num_temps; 502 503 struct list_head blocks; 504 int next_block_index; 505 struct qblock *cur_block; 506 struct qblock *loop_cont_block; 507 struct qblock *loop_break_block; 508 struct qblock *last_top_block; 509 510 struct list_head qpu_inst_list; 511 512 /* Pre-QPU-scheduled instruction containing the last THRSW */ 513 uint64_t *last_thrsw; 514 515 uint64_t *qpu_insts; 516 uint32_t qpu_inst_count; 517 uint32_t qpu_inst_size; 518 uint32_t num_inputs; 519 520 /** 521 * Number of inputs from num_inputs remaining to be queued to the read 522 * FIFO in the VS/CS. 523 */ 524 uint32_t num_inputs_remaining; 525 526 /* Number of inputs currently in the read FIFO for the VS/CS */ 527 uint32_t num_inputs_in_fifo; 528 529 /** Next offset in the VPM to read from in the VS/CS */ 530 uint32_t vpm_read_offset; 531 532 uint32_t program_id; 533 uint32_t variant_id; 534 535 /* Set to compile program in threaded FS mode, where SIG_THREAD_SWITCH 536 * is used to hide texturing latency at the cost of limiting ourselves 537 * to the bottom half of physical reg space. 538 */ 539 bool fs_threaded; 540 541 bool last_thrsw_at_top_level; 542 543 bool failed; 544 }; 545 546 /* Special nir_load_input intrinsic index for loading the current TLB 547 * destination color. 548 */ 549 #define VC4_NIR_TLB_COLOR_READ_INPUT 2000000000 550 551 #define VC4_NIR_MS_MASK_OUTPUT 2000000000 552 553 struct vc4_compile *qir_compile_init(void); 554 void qir_compile_destroy(struct vc4_compile *c); 555 struct qblock *qir_new_block(struct vc4_compile *c); 556 void qir_set_emit_block(struct vc4_compile *c, struct qblock *block); 557 void qir_link_blocks(struct qblock *predecessor, struct qblock *successor); 558 struct qblock *qir_entry_block(struct vc4_compile *c); 559 struct qblock *qir_exit_block(struct vc4_compile *c); 560 struct qinst *qir_inst(enum qop op, struct qreg dst, 561 struct qreg src0, struct qreg src1); 562 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst); 563 struct qreg qir_uniform(struct vc4_compile *c, 564 enum quniform_contents contents, 565 uint32_t data); 566 void qir_schedule_instructions(struct vc4_compile *c); 567 void qir_reorder_uniforms(struct vc4_compile *c); 568 void qir_emit_uniform_stream_resets(struct vc4_compile *c); 569 570 struct qreg qir_emit_def(struct vc4_compile *c, struct qinst *inst); 571 struct qinst *qir_emit_nondef(struct vc4_compile *c, struct qinst *inst); 572 573 struct qreg qir_get_temp(struct vc4_compile *c); 574 void qir_calculate_live_intervals(struct vc4_compile *c); 575 int qir_get_nsrc(struct qinst *inst); 576 int qir_get_non_sideband_nsrc(struct qinst *inst); 577 int qir_get_tex_uniform_src(struct qinst *inst); 578 bool qir_reg_equals(struct qreg a, struct qreg b); 579 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst); 580 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst); 581 bool qir_has_uniform_read(struct qinst *inst); 582 bool qir_is_mul(struct qinst *inst); 583 bool qir_is_raw_mov(struct qinst *inst); 584 bool qir_is_tex(struct qinst *inst); 585 bool qir_has_implicit_tex_uniform(struct qinst *inst); 586 bool qir_is_float_input(struct qinst *inst); 587 bool qir_depends_on_flags(struct qinst *inst); 588 bool qir_writes_r4(struct qinst *inst); 589 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg); 590 uint8_t qir_channels_written(struct qinst *inst); 591 592 void qir_dump(struct vc4_compile *c); 593 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst); 594 const char *qir_get_stage_name(enum qstage stage); 595 596 void qir_validate(struct vc4_compile *c); 597 598 void qir_optimize(struct vc4_compile *c); 599 bool qir_opt_algebraic(struct vc4_compile *c); 600 bool qir_opt_coalesce_ff_writes(struct vc4_compile *c); 601 bool qir_opt_constant_folding(struct vc4_compile *c); 602 bool qir_opt_copy_propagation(struct vc4_compile *c); 603 bool qir_opt_dead_code(struct vc4_compile *c); 604 bool qir_opt_peephole_sf(struct vc4_compile *c); 605 bool qir_opt_small_immediates(struct vc4_compile *c); 606 bool qir_opt_vpm(struct vc4_compile *c); 607 void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c); 608 void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c); 609 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b, 610 nir_ssa_def **srcs, int swiz); 611 void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c); 612 void qir_lower_uniforms(struct vc4_compile *c); 613 614 uint32_t qpu_schedule_instructions(struct vc4_compile *c); 615 616 void qir_SF(struct vc4_compile *c, struct qreg src); 617 618 static inline struct qreg 619 qir_uniform_ui(struct vc4_compile *c, uint32_t ui) 620 { 621 return qir_uniform(c, QUNIFORM_CONSTANT, ui); 622 } 623 624 static inline struct qreg 625 qir_uniform_f(struct vc4_compile *c, float f) 626 { 627 return qir_uniform(c, QUNIFORM_CONSTANT, fui(f)); 628 } 629 630 #define QIR_ALU0(name) \ 631 static inline struct qreg \ 632 qir_##name(struct vc4_compile *c) \ 633 { \ 634 return qir_emit_def(c, qir_inst(QOP_##name, c->undef, \ 635 c->undef, c->undef)); \ 636 } \ 637 static inline struct qinst * \ 638 qir_##name##_dest(struct vc4_compile *c, struct qreg dest) \ 639 { \ 640 return qir_emit_nondef(c, qir_inst(QOP_##name, dest, \ 641 c->undef, c->undef)); \ 642 } 643 644 #define QIR_ALU1(name) \ 645 static inline struct qreg \ 646 qir_##name(struct vc4_compile *c, struct qreg a) \ 647 { \ 648 return qir_emit_def(c, qir_inst(QOP_##name, c->undef, \ 649 a, c->undef)); \ 650 } \ 651 static inline struct qinst * \ 652 qir_##name##_dest(struct vc4_compile *c, struct qreg dest, \ 653 struct qreg a) \ 654 { \ 655 return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, \ 656 c->undef)); \ 657 } 658 659 #define QIR_ALU2(name) \ 660 static inline struct qreg \ 661 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b) \ 662 { \ 663 return qir_emit_def(c, qir_inst(QOP_##name, c->undef, a, b)); \ 664 } \ 665 static inline struct qinst * \ 666 qir_##name##_dest(struct vc4_compile *c, struct qreg dest, \ 667 struct qreg a, struct qreg b) \ 668 { \ 669 return qir_emit_nondef(c, qir_inst(QOP_##name, dest, a, b)); \ 670 } 671 672 #define QIR_NODST_1(name) \ 673 static inline struct qinst * \ 674 qir_##name(struct vc4_compile *c, struct qreg a) \ 675 { \ 676 return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef, \ 677 a, c->undef)); \ 678 } 679 680 #define QIR_NODST_2(name) \ 681 static inline struct qinst * \ 682 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b) \ 683 { \ 684 return qir_emit_nondef(c, qir_inst(QOP_##name, c->undef, \ 685 a, b)); \ 686 } 687 688 #define QIR_PAYLOAD(name) \ 689 static inline struct qreg \ 690 qir_##name(struct vc4_compile *c) \ 691 { \ 692 struct qreg *payload = &c->payload_##name; \ 693 if (payload->file != QFILE_NULL) \ 694 return *payload; \ 695 *payload = qir_get_temp(c); \ 696 struct qinst *inst = qir_inst(QOP_##name, *payload, \ 697 c->undef, c->undef); \ 698 struct qblock *entry = qir_entry_block(c); \ 699 list_add(&inst->link, &entry->instructions); \ 700 c->defs[payload->index] = inst; \ 701 return *payload; \ 702 } 703 704 QIR_ALU1(MOV) 705 QIR_ALU1(FMOV) 706 QIR_ALU1(MMOV) 707 QIR_ALU2(FADD) 708 QIR_ALU2(FSUB) 709 QIR_ALU2(FMUL) 710 QIR_ALU2(V8MULD) 711 QIR_ALU2(V8MIN) 712 QIR_ALU2(V8MAX) 713 QIR_ALU2(V8ADDS) 714 QIR_ALU2(V8SUBS) 715 QIR_ALU2(MUL24) 716 QIR_ALU2(FMIN) 717 QIR_ALU2(FMAX) 718 QIR_ALU2(FMINABS) 719 QIR_ALU2(FMAXABS) 720 QIR_ALU1(FTOI) 721 QIR_ALU1(ITOF) 722 723 QIR_ALU2(ADD) 724 QIR_ALU2(SUB) 725 QIR_ALU2(SHL) 726 QIR_ALU2(SHR) 727 QIR_ALU2(ASR) 728 QIR_ALU2(MIN) 729 QIR_ALU2(MIN_NOIMM) 730 QIR_ALU2(MAX) 731 QIR_ALU2(AND) 732 QIR_ALU2(OR) 733 QIR_ALU2(XOR) 734 QIR_ALU1(NOT) 735 736 QIR_ALU1(RCP) 737 QIR_ALU1(RSQ) 738 QIR_ALU1(EXP2) 739 QIR_ALU1(LOG2) 740 QIR_ALU1(VARY_ADD_C) 741 QIR_PAYLOAD(FRAG_Z) 742 QIR_PAYLOAD(FRAG_W) 743 QIR_ALU0(TEX_RESULT) 744 QIR_ALU0(TLB_COLOR_READ) 745 QIR_NODST_1(MS_MASK) 746 747 static inline struct qreg 748 qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1) 749 { 750 struct qreg t = qir_get_temp(c); 751 qir_MOV_dest(c, t, src1); 752 qir_MOV_dest(c, t, src0)->cond = cond; 753 return t; 754 } 755 756 static inline struct qreg 757 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i) 758 { 759 struct qreg t = qir_FMOV(c, src); 760 c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i; 761 return t; 762 } 763 764 static inline struct qreg 765 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i) 766 { 767 struct qreg t = qir_MOV(c, src); 768 c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i; 769 return t; 770 } 771 772 static inline struct qreg 773 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i) 774 { 775 struct qreg t = qir_FMOV(c, src); 776 c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i; 777 return t; 778 } 779 780 static inline struct qreg 781 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i) 782 { 783 struct qreg t = qir_MOV(c, src); 784 c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i; 785 return t; 786 } 787 788 static inline void 789 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan) 790 { 791 assert(!dest.pack); 792 dest.pack = QPU_PACK_MUL_8A + chan; 793 qir_emit_nondef(c, qir_inst(QOP_MMOV, dest, val, c->undef)); 794 } 795 796 static inline struct qreg 797 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val) 798 { 799 struct qreg dest = qir_MMOV(c, val); 800 c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888; 801 return dest; 802 } 803 804 static inline struct qreg 805 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y) 806 { 807 return qir_EXP2(c, qir_FMUL(c, 808 y, 809 qir_LOG2(c, x))); 810 } 811 812 static inline void 813 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val) 814 { 815 qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val); 816 } 817 818 static inline struct qreg 819 qir_LOAD_IMM(struct vc4_compile *c, uint32_t val) 820 { 821 return qir_emit_def(c, qir_inst(QOP_LOAD_IMM, c->undef, 822 qir_reg(QFILE_LOAD_IMM, val), c->undef)); 823 } 824 825 static inline struct qreg 826 qir_LOAD_IMM_U2(struct vc4_compile *c, uint32_t val) 827 { 828 return qir_emit_def(c, qir_inst(QOP_LOAD_IMM_U2, c->undef, 829 qir_reg(QFILE_LOAD_IMM, val), 830 c->undef)); 831 } 832 833 static inline struct qreg 834 qir_LOAD_IMM_I2(struct vc4_compile *c, uint32_t val) 835 { 836 return qir_emit_def(c, qir_inst(QOP_LOAD_IMM_I2, c->undef, 837 qir_reg(QFILE_LOAD_IMM, val), 838 c->undef)); 839 } 840 841 /** Shifts the multiply output to the right by rot channels */ 842 static inline struct qreg 843 qir_ROT_MUL(struct vc4_compile *c, struct qreg val, uint32_t rot) 844 { 845 return qir_emit_def(c, qir_inst(QOP_ROT_MUL, c->undef, 846 val, 847 qir_reg(QFILE_LOAD_IMM, 848 QPU_SMALL_IMM_MUL_ROT + rot))); 849 } 850 851 static inline struct qinst * 852 qir_MOV_cond(struct vc4_compile *c, uint8_t cond, 853 struct qreg dest, struct qreg src) 854 { 855 struct qinst *mov = qir_MOV_dest(c, dest, src); 856 mov->cond = cond; 857 return mov; 858 } 859 860 static inline struct qinst * 861 qir_BRANCH(struct vc4_compile *c, uint8_t cond) 862 { 863 struct qinst *inst = qir_inst(QOP_BRANCH, c->undef, c->undef, c->undef); 864 inst->cond = cond; 865 qir_emit_nondef(c, inst); 866 return inst; 867 } 868 869 #define qir_for_each_block(block, c) \ 870 list_for_each_entry(struct qblock, block, &c->blocks, link) 871 872 #define qir_for_each_block_rev(block, c) \ 873 list_for_each_entry_rev(struct qblock, block, &c->blocks, link) 874 875 /* Loop over the non-NULL members of the successors array. */ 876 #define qir_for_each_successor(succ, block) \ 877 for (struct qblock *succ = block->successors[0]; \ 878 succ != NULL; \ 879 succ = (succ == block->successors[1] ? NULL : \ 880 block->successors[1])) 881 882 #define qir_for_each_inst(inst, block) \ 883 list_for_each_entry(struct qinst, inst, &block->instructions, link) 884 885 #define qir_for_each_inst_rev(inst, block) \ 886 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link) 887 888 #define qir_for_each_inst_safe(inst, block) \ 889 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link) 890 891 #define qir_for_each_inst_inorder(inst, c) \ 892 qir_for_each_block(_block, c) \ 893 qir_for_each_inst_safe(inst, _block) 894 895 #endif /* VC4_QIR_H */ 896
