src/gallium/drivers/vc4/vc4_qir.h

   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 "glsl/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 "kernel/vc4_packet.h"
  42 #include "pipe/p_state.h"
  43
  44 struct nir_builder;
  45
  46 enum qfile {
  47         QFILE_NULL,
  48         QFILE_TEMP,
  49         QFILE_VARY,
  50         QFILE_UNIF,
  51         QFILE_VPM,
  52
  53         /**
  54          * Stores an immediate value in the index field that can be turned
  55          * into a small immediate field by qpu_encode_small_immediate().
  56          */
  57         QFILE_SMALL_IMM,
  58 };
  59
  60 struct qreg {
  61         enum qfile file;
  62         uint32_t index;
  63         int pack;
  64 };
  65
  66 enum qop {
  67         QOP_UNDEF,
  68         QOP_MOV,
  69         QOP_FMOV,
  70         QOP_MMOV,
  71         QOP_FADD,
  72         QOP_FSUB,
  73         QOP_FMUL,
  74         QOP_V8MULD,
  75         QOP_V8MIN,
  76         QOP_V8MAX,
  77         QOP_V8ADDS,
  78         QOP_V8SUBS,
  79         QOP_MUL24,
  80         QOP_FMIN,
  81         QOP_FMAX,
  82         QOP_FMINABS,
  83         QOP_FMAXABS,
  84         QOP_ADD,
  85         QOP_SUB,
  86         QOP_SHL,
  87         QOP_SHR,
  88         QOP_ASR,
  89         QOP_MIN,
  90         QOP_MAX,
  91         QOP_AND,
  92         QOP_OR,
  93         QOP_XOR,
  94         QOP_NOT,
  95
  96         /* Note: Orderings of these compares must be the same as in
  97          * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
  98          * otherwise 0. */
  99         QOP_SEL_X_0_ZS,
 100         QOP_SEL_X_0_ZC,
 101         QOP_SEL_X_0_NS,
 102         QOP_SEL_X_0_NC,
 103         QOP_SEL_X_0_CS,
 104         QOP_SEL_X_0_CC,
 105         /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
 106         QOP_SEL_X_Y_ZS,
 107         QOP_SEL_X_Y_ZC,
 108         QOP_SEL_X_Y_NS,
 109         QOP_SEL_X_Y_NC,
 110         QOP_SEL_X_Y_CS,
 111         QOP_SEL_X_Y_CC,
 112
 113         QOP_FTOI,
 114         QOP_ITOF,
 115         QOP_RCP,
 116         QOP_RSQ,
 117         QOP_EXP2,
 118         QOP_LOG2,
 119         QOP_VW_SETUP,
 120         QOP_VR_SETUP,
 121         QOP_TLB_DISCARD_SETUP,
 122         QOP_TLB_STENCIL_SETUP,
 123         QOP_TLB_Z_WRITE,
 124         QOP_TLB_COLOR_WRITE,
 125         QOP_TLB_COLOR_WRITE_MS,
 126         QOP_TLB_COLOR_READ,
 127         QOP_MS_MASK,
 128         QOP_VARY_ADD_C,
 129
 130         QOP_FRAG_X,
 131         QOP_FRAG_Y,
 132         QOP_FRAG_Z,
 133         QOP_FRAG_W,
 134         QOP_FRAG_REV_FLAG,
 135
 136         /** Texture x coordinate parameter write */
 137         QOP_TEX_S,
 138         /** Texture y coordinate parameter write */
 139         QOP_TEX_T,
 140         /** Texture border color parameter or cube map z coordinate write */
 141         QOP_TEX_R,
 142         /** Texture LOD bias parameter write */
 143         QOP_TEX_B,
 144
 145         /**
 146          * Texture-unit 4-byte read with address provided direct in S
 147          * cooordinate.
 148          *
 149          * The first operand is the offset from the start of the UBO, and the
 150          * second is the uniform that has the UBO's base pointer.
 151          */
 152         QOP_TEX_DIRECT,
 153
 154         /**
 155          * Signal of texture read being necessary and then reading r4 into
 156          * the destination
 157          */
 158         QOP_TEX_RESULT,
 159 };
 160
 161 struct queued_qpu_inst {
 162         struct list_head link;
 163         uint64_t inst;
 164 };
 165
 166 struct qinst {
 167         struct list_head link;
 168
 169         enum qop op;
 170         struct qreg dst;
 171         struct qreg *src;
 172         bool sf;
 173 };
 174
 175 enum qstage {
 176         /**
 177          * Coordinate shader, runs during binning, before the VS, and just
 178          * outputs position.
 179          */
 180         QSTAGE_COORD,
 181         QSTAGE_VERT,
 182         QSTAGE_FRAG,
 183 };
 184
 185 enum quniform_contents {
 186         /**
 187          * Indicates that a constant 32-bit value is copied from the program's
 188          * uniform contents.
 189          */
 190         QUNIFORM_CONSTANT,
 191         /**
 192          * Indicates that the program's uniform contents are used as an index
 193          * into the GL uniform storage.
 194          */
 195         QUNIFORM_UNIFORM,
 196
 197         /** @{
 198          * Scaling factors from clip coordinates to relative to the viewport
 199          * center.
 200          *
 201          * This is used by the coordinate and vertex shaders to produce the
 202          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 203          * point offsets from the viewport ccenter.
 204          */
 205         QUNIFORM_VIEWPORT_X_SCALE,
 206         QUNIFORM_VIEWPORT_Y_SCALE,
 207         /** @} */
 208
 209         QUNIFORM_VIEWPORT_Z_OFFSET,
 210         QUNIFORM_VIEWPORT_Z_SCALE,
 211
 212         QUNIFORM_USER_CLIP_PLANE,
 213
 214         /**
 215          * A reference to a texture config parameter 0 uniform.
 216          *
 217          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 218          * defines texture type, miplevels, and such.  It will be found as a
 219          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 220          */
 221         QUNIFORM_TEXTURE_CONFIG_P0,
 222
 223         /**
 224          * A reference to a texture config parameter 1 uniform.
 225          *
 226          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 227          * defines texture width, height, filters, and wrap modes.  It will be
 228          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 229          * sequence.
 230          */
 231         QUNIFORM_TEXTURE_CONFIG_P1,
 232
 233         /** A reference to a texture config parameter 2 cubemap stride uniform */
 234         QUNIFORM_TEXTURE_CONFIG_P2,
 235
 236         QUNIFORM_UBO_ADDR,
 237
 238         QUNIFORM_TEXRECT_SCALE_X,
 239         QUNIFORM_TEXRECT_SCALE_Y,
 240
 241         QUNIFORM_TEXTURE_BORDER_COLOR,
 242
 243         QUNIFORM_BLEND_CONST_COLOR_X,
 244         QUNIFORM_BLEND_CONST_COLOR_Y,
 245         QUNIFORM_BLEND_CONST_COLOR_Z,
 246         QUNIFORM_BLEND_CONST_COLOR_W,
 247         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 248         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 249
 250         QUNIFORM_STENCIL,
 251
 252         QUNIFORM_ALPHA_REF,
 253         QUNIFORM_SAMPLE_MASK,
 254 };
 255
 256 struct vc4_varying_slot {
 257         uint8_t slot;
 258         uint8_t swizzle;
 259 };
 260
 261 struct vc4_compiler_ubo_range {
 262         /**
 263          * offset in bytes from the start of the ubo where this range is
 264          * uploaded.
 265          *
 266          * Only set once used is set.
 267          */
 268         uint32_t dst_offset;
 269
 270         /**
 271          * offset in bytes from the start of the gallium uniforms where the
 272          * data comes from.
 273          */
 274         uint32_t src_offset;
 275
 276         /** size in bytes of this ubo range */
 277         uint32_t size;
 278
 279         /**
 280          * Set if this range is used by the shader for indirect uniforms
 281          * access.
 282          */
 283         bool used;
 284 };
 285
 286 struct vc4_key {
 287         struct vc4_uncompiled_shader *shader_state;
 288         struct {
 289                 enum pipe_format format;
 290                 unsigned compare_mode:1;
 291                 unsigned compare_func:3;
 292                 unsigned wrap_s:3;
 293                 unsigned wrap_t:3;
 294                 uint8_t swizzle[4];
 295         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 296         uint8_t ucp_enables;
 297 };
 298
 299 struct vc4_fs_key {
 300         struct vc4_key base;
 301         enum pipe_format color_format;
 302         bool depth_enabled;
 303         bool stencil_enabled;
 304         bool stencil_twoside;
 305         bool stencil_full_writemasks;
 306         bool is_points;
 307         bool is_lines;
 308         bool alpha_test;
 309         bool point_coord_upper_left;
 310         bool light_twoside;
 311         bool msaa;
 312         bool sample_coverage;
 313         bool sample_alpha_to_coverage;
 314         bool sample_alpha_to_one;
 315         uint8_t alpha_test_func;
 316         uint8_t logicop_func;
 317         uint32_t point_sprite_mask;
 318
 319         struct pipe_rt_blend_state blend;
 320 };
 321
 322 struct vc4_vs_key {
 323         struct vc4_key base;
 324
 325         /**
 326          * This is a proxy for the array of FS input semantics, which is
 327          * larger than we would want to put in the key.
 328          */
 329         uint64_t compiled_fs_id;
 330
 331         enum pipe_format attr_formats[8];
 332         bool is_coord;
 333         bool per_vertex_point_size;
 334 };
 335
 336 struct vc4_compile {
 337         struct vc4_context *vc4;
 338         nir_shader *s;
 339         nir_function_impl *impl;
 340         struct exec_list *cf_node_list;
 341
 342         /**
 343          * Mapping from nir_register * or nir_ssa_def * to array of struct
 344          * qreg for the values.
 345          */
 346         struct hash_table *def_ht;
 347
 348         /* For each temp, the instruction generating its value. */
 349         struct qinst **defs;
 350         uint32_t defs_array_size;
 351
 352         /**
 353          * Inputs to the shader, arranged by TGSI declaration order.
 354          *
 355          * Not all fragment shader QFILE_VARY reads are present in this array.
 356          */
 357         struct qreg *inputs;
 358         struct qreg *outputs;
 359         bool msaa_per_sample_output;
 360         struct qreg color_reads[VC4_MAX_SAMPLES];
 361         struct qreg sample_colors[VC4_MAX_SAMPLES];
 362         uint32_t inputs_array_size;
 363         uint32_t outputs_array_size;
 364         uint32_t uniforms_array_size;
 365
 366         struct vc4_compiler_ubo_range *ubo_ranges;
 367         uint32_t ubo_ranges_array_size;
 368         /** Number of uniform areas declared in ubo_ranges. */
 369         uint32_t num_uniform_ranges;
 370         /** Number of uniform areas used for indirect addressed loads. */
 371         uint32_t num_ubo_ranges;
 372         uint32_t next_ubo_dst_offset;
 373
 374         struct qreg line_x, point_x, point_y;
 375         struct qreg discard;
 376
 377         uint8_t vattr_sizes[8];
 378
 379         /**
 380          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 381          *
 382          * This includes those that aren't part of the VPM varyings, like
 383          * point/line coordinates.
 384          */
 385         struct vc4_varying_slot *input_slots;
 386         uint32_t num_input_slots;
 387         uint32_t input_slots_array_size;
 388
 389         /**
 390          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 391          * of the output is.  Used to emit from the VS in the order that the
 392          * FS needs.
 393          */
 394         struct vc4_varying_slot *output_slots;
 395
 396         struct pipe_shader_state *shader_state;
 397         struct vc4_key *key;
 398         struct vc4_fs_key *fs_key;
 399         struct vc4_vs_key *vs_key;
 400
 401         uint32_t *uniform_data;
 402         enum quniform_contents *uniform_contents;
 403         uint32_t uniform_array_size;
 404         uint32_t num_uniforms;
 405         uint32_t num_outputs;
 406         uint32_t num_texture_samples;
 407         uint32_t output_position_index;
 408         uint32_t output_color_index;
 409         uint32_t output_point_size_index;
 410         uint32_t output_sample_mask_index;
 411
 412         struct qreg undef;
 413         enum qstage stage;
 414         uint32_t num_temps;
 415         struct list_head instructions;
 416         uint32_t immediates[1024];
 417
 418         struct list_head qpu_inst_list;
 419         uint64_t *qpu_insts;
 420         uint32_t qpu_inst_count;
 421         uint32_t qpu_inst_size;
 422         uint32_t num_inputs;
 423
 424         uint32_t program_id;
 425         uint32_t variant_id;
 426 };
 427
 428 /* Special nir_load_input intrinsic index for loading the current TLB
 429  * destination color.
 430  */
 431 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 432
 433 #define VC4_NIR_MS_MASK_OUTPUT                  2000000000
 434
 435 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 436  * state-dependent value.
 437  */
 438 #define VC4_NIR_STATE_UNIFORM_OFFSET            2000000000
 439
 440 struct vc4_compile *qir_compile_init(void);
 441 void qir_compile_destroy(struct vc4_compile *c);
 442 struct qinst *qir_inst(enum qop op, struct qreg dst,
 443                        struct qreg src0, struct qreg src1);
 444 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 445                         struct qreg a,
 446                         struct qreg b,
 447                         struct qreg c,
 448                         struct qreg d);
 449 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 450 struct qreg qir_uniform(struct vc4_compile *c,
 451                         enum quniform_contents contents,
 452                         uint32_t data);
 453 void qir_reorder_uniforms(struct vc4_compile *c);
 454
 455 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 456 static inline void qir_emit_nodef(struct vc4_compile *c, struct qinst *inst)
 457 {
 458         list_addtail(&inst->link, &c->instructions);
 459 }
 460
 461 struct qreg qir_get_temp(struct vc4_compile *c);
 462 int qir_get_op_nsrc(enum qop qop);
 463 bool qir_reg_equals(struct qreg a, struct qreg b);
 464 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 465 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 466 bool qir_is_multi_instruction(struct qinst *inst);
 467 bool qir_is_mul(struct qinst *inst);
 468 bool qir_is_raw_mov(struct qinst *inst);
 469 bool qir_is_tex(struct qinst *inst);
 470 bool qir_is_float_input(struct qinst *inst);
 471 bool qir_depends_on_flags(struct qinst *inst);
 472 bool qir_writes_r4(struct qinst *inst);
 473 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 474
 475 void qir_dump(struct vc4_compile *c);
 476 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 477 const char *qir_get_stage_name(enum qstage stage);
 478
 479 void qir_optimize(struct vc4_compile *c);
 480 bool qir_opt_algebraic(struct vc4_compile *c);
 481 bool qir_opt_constant_folding(struct vc4_compile *c);
 482 bool qir_opt_copy_propagation(struct vc4_compile *c);
 483 bool qir_opt_cse(struct vc4_compile *c);
 484 bool qir_opt_dead_code(struct vc4_compile *c);
 485 bool qir_opt_small_immediates(struct vc4_compile *c);
 486 bool qir_opt_vpm_writes(struct vc4_compile *c);
 487 void vc4_nir_lower_blend(struct vc4_compile *c);
 488 void vc4_nir_lower_io(struct vc4_compile *c);
 489 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 490                                        enum quniform_contents contents);
 491 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 492                                           nir_ssa_def **srcs, int swiz);
 493 void qir_lower_uniforms(struct vc4_compile *c);
 494
 495 void qpu_schedule_instructions(struct vc4_compile *c);
 496
 497 void qir_SF(struct vc4_compile *c, struct qreg src);
 498
 499 static inline struct qreg
 500 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 501 {
 502         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 503 }
 504
 505 static inline struct qreg
 506 qir_uniform_f(struct vc4_compile *c, float f)
 507 {
 508         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 509 }
 510
 511 #define QIR_ALU0(name)                                                   \
 512 static inline struct qreg                                                \
 513 qir_##name(struct vc4_compile *c)                                        \
 514 {                                                                        \
 515         struct qreg t = qir_get_temp(c);                                 \
 516         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 517         return t;                                                        \
 518 }
 519
 520 #define QIR_ALU1(name)                                                   \
 521 static inline struct qreg                                                \
 522 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 523 {                                                                        \
 524         struct qreg t = qir_get_temp(c);                                 \
 525         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 526         return t;                                                        \
 527 }                                                                        \
 528 static inline void                                                       \
 529 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 530                   struct qreg a)                                         \
 531 {                                                                        \
 532         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 533 }
 534
 535 #define QIR_ALU2(name)                                                   \
 536 static inline struct qreg                                                \
 537 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 538 {                                                                        \
 539         struct qreg t = qir_get_temp(c);                                 \
 540         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 541         return t;                                                        \
 542 }                                                                        \
 543 static inline void                                                       \
 544 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 545                   struct qreg a, struct qreg b)                          \
 546 {                                                                        \
 547         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, b));             \
 548 }
 549
 550 #define QIR_NODST_1(name)                                               \
 551 static inline void                                                      \
 552 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 553 {                                                                       \
 554         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 555 }
 556
 557 #define QIR_NODST_2(name)                                               \
 558 static inline void                                                      \
 559 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 560 {                                                                       \
 561         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 562 }
 563
 564 #define QIR_PACK(name)                                                   \
 565 static inline struct qreg                                                \
 566 qir_##name(struct vc4_compile *c, struct qreg dest, struct qreg a)       \
 567 {                                                                        \
 568         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 569         return dest;                                                     \
 570 }
 571
 572 QIR_ALU1(MOV)
 573 QIR_ALU1(FMOV)
 574 QIR_ALU1(MMOV)
 575 QIR_ALU2(FADD)
 576 QIR_ALU2(FSUB)
 577 QIR_ALU2(FMUL)
 578 QIR_ALU2(V8MULD)
 579 QIR_ALU2(V8MIN)
 580 QIR_ALU2(V8MAX)
 581 QIR_ALU2(V8ADDS)
 582 QIR_ALU2(V8SUBS)
 583 QIR_ALU2(MUL24)
 584 QIR_ALU1(SEL_X_0_ZS)
 585 QIR_ALU1(SEL_X_0_ZC)
 586 QIR_ALU1(SEL_X_0_NS)
 587 QIR_ALU1(SEL_X_0_NC)
 588 QIR_ALU1(SEL_X_0_CS)
 589 QIR_ALU1(SEL_X_0_CC)
 590 QIR_ALU2(SEL_X_Y_ZS)
 591 QIR_ALU2(SEL_X_Y_ZC)
 592 QIR_ALU2(SEL_X_Y_NS)
 593 QIR_ALU2(SEL_X_Y_NC)
 594 QIR_ALU2(SEL_X_Y_CS)
 595 QIR_ALU2(SEL_X_Y_CC)
 596 QIR_ALU2(FMIN)
 597 QIR_ALU2(FMAX)
 598 QIR_ALU2(FMINABS)
 599 QIR_ALU2(FMAXABS)
 600 QIR_ALU1(FTOI)
 601 QIR_ALU1(ITOF)
 602
 603 QIR_ALU2(ADD)
 604 QIR_ALU2(SUB)
 605 QIR_ALU2(SHL)
 606 QIR_ALU2(SHR)
 607 QIR_ALU2(ASR)
 608 QIR_ALU2(MIN)
 609 QIR_ALU2(MAX)
 610 QIR_ALU2(AND)
 611 QIR_ALU2(OR)
 612 QIR_ALU2(XOR)
 613 QIR_ALU1(NOT)
 614
 615 QIR_ALU1(RCP)
 616 QIR_ALU1(RSQ)
 617 QIR_ALU1(EXP2)
 618 QIR_ALU1(LOG2)
 619 QIR_ALU1(VARY_ADD_C)
 620 QIR_NODST_2(TEX_S)
 621 QIR_NODST_2(TEX_T)
 622 QIR_NODST_2(TEX_R)
 623 QIR_NODST_2(TEX_B)
 624 QIR_NODST_2(TEX_DIRECT)
 625 QIR_ALU0(FRAG_X)
 626 QIR_ALU0(FRAG_Y)
 627 QIR_ALU0(FRAG_Z)
 628 QIR_ALU0(FRAG_W)
 629 QIR_ALU0(FRAG_REV_FLAG)
 630 QIR_ALU0(TEX_RESULT)
 631 QIR_ALU0(TLB_COLOR_READ)
 632 QIR_NODST_1(TLB_COLOR_WRITE)
 633 QIR_NODST_1(TLB_COLOR_WRITE_MS)
 634 QIR_NODST_1(TLB_Z_WRITE)
 635 QIR_NODST_1(TLB_DISCARD_SETUP)
 636 QIR_NODST_1(TLB_STENCIL_SETUP)
 637 QIR_NODST_1(MS_MASK)
 638
 639 static inline struct qreg
 640 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 641 {
 642         struct qreg t = qir_FMOV(c, src);
 643         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 644         return t;
 645 }
 646
 647 static inline struct qreg
 648 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 649 {
 650         struct qreg t = qir_MOV(c, src);
 651         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 652         return t;
 653 }
 654
 655 static inline struct qreg
 656 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 657 {
 658         struct qreg t = qir_FMOV(c, src);
 659         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 660         return t;
 661 }
 662
 663 static inline struct qreg
 664 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 665 {
 666         struct qreg t = qir_MOV(c, src);
 667         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 668         return t;
 669 }
 670
 671 static inline void
 672 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 673 {
 674         assert(!dest.pack);
 675         dest.pack = QPU_PACK_MUL_8A + chan;
 676         qir_emit(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 677         if (dest.file == QFILE_TEMP)
 678                 c->defs[dest.index] = NULL;
 679 }
 680
 681 static inline struct qreg
 682 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 683 {
 684         struct qreg dest = qir_MMOV(c, val);
 685         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 686         return dest;
 687 }
 688
 689 static inline struct qreg
 690 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 691 {
 692         return qir_EXP2(c, qir_FMUL(c,
 693                                     y,
 694                                     qir_LOG2(c, x)));
 695 }
 696
 697 static inline void
 698 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 699 {
 700         static const struct qreg vpm = { QFILE_VPM, 0 };
 701         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 702 }
 703
 704 #endif /* VC4_QIR_H */