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 "pipe/p_state.h"
  41
  42 struct nir_builder;
  43
  44 enum qfile {
  45         QFILE_NULL,
  46         QFILE_TEMP,
  47         QFILE_VARY,
  48         QFILE_UNIF,
  49         QFILE_VPM,
  50
  51         /**
  52          * Stores an immediate value in the index field that can be turned
  53          * into a small immediate field by qpu_encode_small_immediate().
  54          */
  55         QFILE_SMALL_IMM,
  56 };
  57
  58 struct qreg {
  59         enum qfile file;
  60         uint32_t index;
  61         int pack;
  62 };
  63
  64 enum qop {
  65         QOP_UNDEF,
  66         QOP_MOV,
  67         QOP_FADD,
  68         QOP_FSUB,
  69         QOP_FMUL,
  70         QOP_MUL24,
  71         QOP_FMIN,
  72         QOP_FMAX,
  73         QOP_FMINABS,
  74         QOP_FMAXABS,
  75         QOP_ADD,
  76         QOP_SUB,
  77         QOP_SHL,
  78         QOP_SHR,
  79         QOP_ASR,
  80         QOP_MIN,
  81         QOP_MAX,
  82         QOP_AND,
  83         QOP_OR,
  84         QOP_XOR,
  85         QOP_NOT,
  86
  87         /* Note: Orderings of these compares must be the same as in
  88          * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
  89          * otherwise 0. */
  90         QOP_SEL_X_0_ZS,
  91         QOP_SEL_X_0_ZC,
  92         QOP_SEL_X_0_NS,
  93         QOP_SEL_X_0_NC,
  94         /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
  95         QOP_SEL_X_Y_ZS,
  96         QOP_SEL_X_Y_ZC,
  97         QOP_SEL_X_Y_NS,
  98         QOP_SEL_X_Y_NC,
  99
 100         QOP_FTOI,
 101         QOP_ITOF,
 102         QOP_RCP,
 103         QOP_RSQ,
 104         QOP_EXP2,
 105         QOP_LOG2,
 106         QOP_VW_SETUP,
 107         QOP_VR_SETUP,
 108         QOP_PACK_8888_F,
 109         QOP_PACK_8A_F,
 110         QOP_PACK_8B_F,
 111         QOP_PACK_8C_F,
 112         QOP_PACK_8D_F,
 113         QOP_TLB_DISCARD_SETUP,
 114         QOP_TLB_STENCIL_SETUP,
 115         QOP_TLB_Z_WRITE,
 116         QOP_TLB_COLOR_WRITE,
 117         QOP_TLB_COLOR_READ,
 118         QOP_VARY_ADD_C,
 119
 120         QOP_FRAG_X,
 121         QOP_FRAG_Y,
 122         QOP_FRAG_Z,
 123         QOP_FRAG_W,
 124         QOP_FRAG_REV_FLAG,
 125
 126         QOP_UNPACK_8A_F,
 127         QOP_UNPACK_8B_F,
 128         QOP_UNPACK_8C_F,
 129         QOP_UNPACK_8D_F,
 130         QOP_UNPACK_16A_F,
 131         QOP_UNPACK_16B_F,
 132
 133         QOP_UNPACK_8A_I,
 134         QOP_UNPACK_8B_I,
 135         QOP_UNPACK_8C_I,
 136         QOP_UNPACK_8D_I,
 137         QOP_UNPACK_16A_I,
 138         QOP_UNPACK_16B_I,
 139
 140         /** Texture x coordinate parameter write */
 141         QOP_TEX_S,
 142         /** Texture y coordinate parameter write */
 143         QOP_TEX_T,
 144         /** Texture border color parameter or cube map z coordinate write */
 145         QOP_TEX_R,
 146         /** Texture LOD bias parameter write */
 147         QOP_TEX_B,
 148
 149         /**
 150          * Texture-unit 4-byte read with address provided direct in S
 151          * cooordinate.
 152          *
 153          * The first operand is the offset from the start of the UBO, and the
 154          * second is the uniform that has the UBO's base pointer.
 155          */
 156         QOP_TEX_DIRECT,
 157
 158         /**
 159          * Signal of texture read being necessary and then reading r4 into
 160          * the destination
 161          */
 162         QOP_TEX_RESULT,
 163 };
 164
 165 struct queued_qpu_inst {
 166         struct list_head link;
 167         uint64_t inst;
 168 };
 169
 170 struct qinst {
 171         struct list_head link;
 172
 173         enum qop op;
 174         struct qreg dst;
 175         struct qreg *src;
 176         bool sf;
 177 };
 178
 179 enum qstage {
 180         /**
 181          * Coordinate shader, runs during binning, before the VS, and just
 182          * outputs position.
 183          */
 184         QSTAGE_COORD,
 185         QSTAGE_VERT,
 186         QSTAGE_FRAG,
 187 };
 188
 189 enum quniform_contents {
 190         /**
 191          * Indicates that a constant 32-bit value is copied from the program's
 192          * uniform contents.
 193          */
 194         QUNIFORM_CONSTANT,
 195         /**
 196          * Indicates that the program's uniform contents are used as an index
 197          * into the GL uniform storage.
 198          */
 199         QUNIFORM_UNIFORM,
 200
 201         /** @{
 202          * Scaling factors from clip coordinates to relative to the viewport
 203          * center.
 204          *
 205          * This is used by the coordinate and vertex shaders to produce the
 206          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 207          * point offsets from the viewport ccenter.
 208          */
 209         QUNIFORM_VIEWPORT_X_SCALE,
 210         QUNIFORM_VIEWPORT_Y_SCALE,
 211         /** @} */
 212
 213         QUNIFORM_VIEWPORT_Z_OFFSET,
 214         QUNIFORM_VIEWPORT_Z_SCALE,
 215
 216         QUNIFORM_USER_CLIP_PLANE,
 217
 218         /**
 219          * A reference to a texture config parameter 0 uniform.
 220          *
 221          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 222          * defines texture type, miplevels, and such.  It will be found as a
 223          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 224          */
 225         QUNIFORM_TEXTURE_CONFIG_P0,
 226
 227         /**
 228          * A reference to a texture config parameter 1 uniform.
 229          *
 230          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 231          * defines texture width, height, filters, and wrap modes.  It will be
 232          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 233          * sequence.
 234          */
 235         QUNIFORM_TEXTURE_CONFIG_P1,
 236
 237         /** A reference to a texture config parameter 2 cubemap stride uniform */
 238         QUNIFORM_TEXTURE_CONFIG_P2,
 239
 240         QUNIFORM_UBO_ADDR,
 241
 242         QUNIFORM_TEXRECT_SCALE_X,
 243         QUNIFORM_TEXRECT_SCALE_Y,
 244
 245         QUNIFORM_TEXTURE_BORDER_COLOR,
 246
 247         QUNIFORM_BLEND_CONST_COLOR_X,
 248         QUNIFORM_BLEND_CONST_COLOR_Y,
 249         QUNIFORM_BLEND_CONST_COLOR_Z,
 250         QUNIFORM_BLEND_CONST_COLOR_W,
 251
 252         QUNIFORM_STENCIL,
 253
 254         QUNIFORM_ALPHA_REF,
 255 };
 256
 257 struct vc4_varying_slot {
 258         uint8_t slot;
 259         uint8_t swizzle;
 260 };
 261
 262 struct vc4_compiler_ubo_range {
 263         /**
 264          * offset in bytes from the start of the ubo where this range is
 265          * uploaded.
 266          *
 267          * Only set once used is set.
 268          */
 269         uint32_t dst_offset;
 270
 271         /**
 272          * offset in bytes from the start of the gallium uniforms where the
 273          * data comes from.
 274          */
 275         uint32_t src_offset;
 276
 277         /** size in bytes of this ubo range */
 278         uint32_t size;
 279
 280         /**
 281          * Set if this range is used by the shader for indirect uniforms
 282          * access.
 283          */
 284         bool used;
 285 };
 286
 287 struct vc4_key {
 288         struct vc4_uncompiled_shader *shader_state;
 289         struct {
 290                 enum pipe_format format;
 291                 unsigned compare_mode:1;
 292                 unsigned compare_func:3;
 293                 unsigned wrap_s:3;
 294                 unsigned wrap_t:3;
 295                 uint8_t swizzle[4];
 296         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 297         uint8_t ucp_enables;
 298 };
 299
 300 struct vc4_fs_key {
 301         struct vc4_key base;
 302         enum pipe_format color_format;
 303         bool depth_enabled;
 304         bool stencil_enabled;
 305         bool stencil_twoside;
 306         bool stencil_full_writemasks;
 307         bool is_points;
 308         bool is_lines;
 309         bool alpha_test;
 310         bool point_coord_upper_left;
 311         bool light_twoside;
 312         uint8_t alpha_test_func;
 313         uint8_t logicop_func;
 314         uint32_t point_sprite_mask;
 315
 316         struct pipe_rt_blend_state blend;
 317 };
 318
 319 struct vc4_vs_key {
 320         struct vc4_key base;
 321
 322         /**
 323          * This is a proxy for the array of FS input semantics, which is
 324          * larger than we would want to put in the key.
 325          */
 326         uint64_t compiled_fs_id;
 327
 328         enum pipe_format attr_formats[8];
 329         bool is_coord;
 330         bool per_vertex_point_size;
 331 };
 332
 333 struct vc4_compile {
 334         struct vc4_context *vc4;
 335         nir_shader *s;
 336         nir_function_impl *impl;
 337         struct exec_list *cf_node_list;
 338
 339         /**
 340          * Mapping from nir_register * or nir_ssa_def * to array of struct
 341          * qreg for the values.
 342          */
 343         struct hash_table *def_ht;
 344
 345         /* For each temp, the instruction generating its value. */
 346         struct qinst **defs;
 347         uint32_t defs_array_size;
 348
 349         /**
 350          * Inputs to the shader, arranged by TGSI declaration order.
 351          *
 352          * Not all fragment shader QFILE_VARY reads are present in this array.
 353          */
 354         struct qreg *inputs;
 355         struct qreg *outputs;
 356         uint32_t inputs_array_size;
 357         uint32_t outputs_array_size;
 358         uint32_t uniforms_array_size;
 359
 360         struct vc4_compiler_ubo_range *ubo_ranges;
 361         uint32_t ubo_ranges_array_size;
 362         /** Number of uniform areas declared in ubo_ranges. */
 363         uint32_t num_uniform_ranges;
 364         /** Number of uniform areas used for indirect addressed loads. */
 365         uint32_t num_ubo_ranges;
 366         uint32_t next_ubo_dst_offset;
 367
 368         struct qreg line_x, point_x, point_y;
 369         struct qreg discard;
 370
 371         uint8_t vattr_sizes[8];
 372
 373         /**
 374          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 375          *
 376          * This includes those that aren't part of the VPM varyings, like
 377          * point/line coordinates.
 378          */
 379         struct vc4_varying_slot *input_slots;
 380         uint32_t num_input_slots;
 381         uint32_t input_slots_array_size;
 382
 383         /**
 384          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 385          * of the output is.  Used to emit from the VS in the order that the
 386          * FS needs.
 387          */
 388         struct vc4_varying_slot *output_slots;
 389
 390         struct pipe_shader_state *shader_state;
 391         struct vc4_key *key;
 392         struct vc4_fs_key *fs_key;
 393         struct vc4_vs_key *vs_key;
 394
 395         uint32_t *uniform_data;
 396         enum quniform_contents *uniform_contents;
 397         uint32_t uniform_array_size;
 398         uint32_t num_uniforms;
 399         uint32_t num_outputs;
 400         uint32_t num_texture_samples;
 401         uint32_t output_position_index;
 402         uint32_t output_clipvertex_index;
 403         uint32_t output_color_index;
 404         uint32_t output_point_size_index;
 405
 406         struct qreg undef;
 407         enum qstage stage;
 408         uint32_t num_temps;
 409         struct list_head instructions;
 410         uint32_t immediates[1024];
 411
 412         struct list_head qpu_inst_list;
 413         uint64_t *qpu_insts;
 414         uint32_t qpu_inst_count;
 415         uint32_t qpu_inst_size;
 416         uint32_t num_inputs;
 417
 418         uint32_t program_id;
 419         uint32_t variant_id;
 420 };
 421
 422 /* Special nir_load_input intrinsic index for loading the current TLB
 423  * destination color.
 424  */
 425 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 426
 427 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 428  * state-dependent value.
 429  */
 430 #define VC4_NIR_STATE_UNIFORM_OFFSET            2000000000
 431
 432 struct vc4_compile *qir_compile_init(void);
 433 void qir_compile_destroy(struct vc4_compile *c);
 434 struct qinst *qir_inst(enum qop op, struct qreg dst,
 435                        struct qreg src0, struct qreg src1);
 436 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 437                         struct qreg a,
 438                         struct qreg b,
 439                         struct qreg c,
 440                         struct qreg d);
 441 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 442 struct qreg qir_uniform(struct vc4_compile *c,
 443                         enum quniform_contents contents,
 444                         uint32_t data);
 445 void qir_reorder_uniforms(struct vc4_compile *c);
 446
 447 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 448 static inline void qir_emit_nodef(struct vc4_compile *c, struct qinst *inst)
 449 {
 450         list_addtail(&inst->link, &c->instructions);
 451 }
 452
 453 struct qreg qir_get_temp(struct vc4_compile *c);
 454 int qir_get_op_nsrc(enum qop qop);
 455 bool qir_reg_equals(struct qreg a, struct qreg b);
 456 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 457 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 458 bool qir_is_multi_instruction(struct qinst *inst);
 459 bool qir_is_mul(struct qinst *inst);
 460 bool qir_is_tex(struct qinst *inst);
 461 bool qir_depends_on_flags(struct qinst *inst);
 462 bool qir_writes_r4(struct qinst *inst);
 463 bool qir_src_needs_a_file(struct qinst *inst);
 464 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 465
 466 void qir_dump(struct vc4_compile *c);
 467 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 468 const char *qir_get_stage_name(enum qstage stage);
 469
 470 void qir_optimize(struct vc4_compile *c);
 471 bool qir_opt_algebraic(struct vc4_compile *c);
 472 bool qir_opt_constant_folding(struct vc4_compile *c);
 473 bool qir_opt_copy_propagation(struct vc4_compile *c);
 474 bool qir_opt_cse(struct vc4_compile *c);
 475 bool qir_opt_dead_code(struct vc4_compile *c);
 476 bool qir_opt_small_immediates(struct vc4_compile *c);
 477 bool qir_opt_vpm_writes(struct vc4_compile *c);
 478 void vc4_nir_lower_blend(struct vc4_compile *c);
 479 void vc4_nir_lower_io(struct vc4_compile *c);
 480 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 481                                        enum quniform_contents contents);
 482 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 483                                           nir_ssa_def **srcs, int swiz);
 484 void qir_lower_uniforms(struct vc4_compile *c);
 485
 486 void qpu_schedule_instructions(struct vc4_compile *c);
 487
 488 void qir_SF(struct vc4_compile *c, struct qreg src);
 489
 490 static inline struct qreg
 491 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 492 {
 493         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 494 }
 495
 496 static inline struct qreg
 497 qir_uniform_f(struct vc4_compile *c, float f)
 498 {
 499         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 500 }
 501
 502 #define QIR_ALU0(name)                                                   \
 503 static inline struct qreg                                                \
 504 qir_##name(struct vc4_compile *c)                                        \
 505 {                                                                        \
 506         struct qreg t = qir_get_temp(c);                                 \
 507         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 508         return t;                                                        \
 509 }
 510
 511 #define QIR_ALU1(name)                                                   \
 512 static inline struct qreg                                                \
 513 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 514 {                                                                        \
 515         struct qreg t = qir_get_temp(c);                                 \
 516         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 517         return t;                                                        \
 518 }                                                                        \
 519 static inline void                                                       \
 520 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 521                   struct qreg a)                                         \
 522 {                                                                        \
 523         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 524 }
 525
 526 #define QIR_ALU2(name)                                                   \
 527 static inline struct qreg                                                \
 528 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 529 {                                                                        \
 530         struct qreg t = qir_get_temp(c);                                 \
 531         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 532         return t;                                                        \
 533 }                                                                        \
 534 static inline void                                                       \
 535 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 536                   struct qreg a, struct qreg b)                          \
 537 {                                                                        \
 538         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, b));             \
 539 }
 540
 541 #define QIR_NODST_1(name)                                               \
 542 static inline void                                                      \
 543 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 544 {                                                                       \
 545         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 546 }
 547
 548 #define QIR_NODST_2(name)                                               \
 549 static inline void                                                      \
 550 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 551 {                                                                       \
 552         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 553 }
 554
 555 #define QIR_PACK(name)                                                   \
 556 static inline struct qreg                                                \
 557 qir_##name(struct vc4_compile *c, struct qreg dest, struct qreg a)       \
 558 {                                                                        \
 559         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 560         return dest;                                                     \
 561 }
 562
 563 QIR_ALU1(MOV)
 564 QIR_ALU2(FADD)
 565 QIR_ALU2(FSUB)
 566 QIR_ALU2(FMUL)
 567 QIR_ALU2(MUL24)
 568 QIR_ALU1(SEL_X_0_ZS)
 569 QIR_ALU1(SEL_X_0_ZC)
 570 QIR_ALU1(SEL_X_0_NS)
 571 QIR_ALU1(SEL_X_0_NC)
 572 QIR_ALU2(SEL_X_Y_ZS)
 573 QIR_ALU2(SEL_X_Y_ZC)
 574 QIR_ALU2(SEL_X_Y_NS)
 575 QIR_ALU2(SEL_X_Y_NC)
 576 QIR_ALU2(FMIN)
 577 QIR_ALU2(FMAX)
 578 QIR_ALU2(FMINABS)
 579 QIR_ALU2(FMAXABS)
 580 QIR_ALU1(FTOI)
 581 QIR_ALU1(ITOF)
 582
 583 QIR_ALU2(ADD)
 584 QIR_ALU2(SUB)
 585 QIR_ALU2(SHL)
 586 QIR_ALU2(SHR)
 587 QIR_ALU2(ASR)
 588 QIR_ALU2(MIN)
 589 QIR_ALU2(MAX)
 590 QIR_ALU2(AND)
 591 QIR_ALU2(OR)
 592 QIR_ALU2(XOR)
 593 QIR_ALU1(NOT)
 594
 595 QIR_ALU1(RCP)
 596 QIR_ALU1(RSQ)
 597 QIR_ALU1(EXP2)
 598 QIR_ALU1(LOG2)
 599 QIR_ALU1(PACK_8888_F)
 600 QIR_PACK(PACK_8A_F)
 601 QIR_PACK(PACK_8B_F)
 602 QIR_PACK(PACK_8C_F)
 603 QIR_PACK(PACK_8D_F)
 604 QIR_ALU1(VARY_ADD_C)
 605 QIR_NODST_2(TEX_S)
 606 QIR_NODST_2(TEX_T)
 607 QIR_NODST_2(TEX_R)
 608 QIR_NODST_2(TEX_B)
 609 QIR_NODST_2(TEX_DIRECT)
 610 QIR_ALU0(FRAG_X)
 611 QIR_ALU0(FRAG_Y)
 612 QIR_ALU0(FRAG_Z)
 613 QIR_ALU0(FRAG_W)
 614 QIR_ALU0(FRAG_REV_FLAG)
 615 QIR_ALU0(TEX_RESULT)
 616 QIR_ALU0(TLB_COLOR_READ)
 617 QIR_NODST_1(TLB_COLOR_WRITE)
 618 QIR_NODST_1(TLB_Z_WRITE)
 619 QIR_NODST_1(TLB_DISCARD_SETUP)
 620 QIR_NODST_1(TLB_STENCIL_SETUP)
 621
 622 static inline struct qreg
 623 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 624 {
 625         struct qreg t = qir_get_temp(c);
 626         qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
 627         return t;
 628 }
 629
 630 static inline struct qreg
 631 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 632 {
 633         struct qreg t = qir_get_temp(c);
 634         qir_emit(c, qir_inst(QOP_UNPACK_8A_I + i, t, src, c->undef));
 635         return t;
 636 }
 637
 638 static inline struct qreg
 639 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 640 {
 641         struct qreg t = qir_get_temp(c);
 642         qir_emit(c, qir_inst(QOP_UNPACK_16A_F + i, t, src, c->undef));
 643         return t;
 644 }
 645
 646 static inline struct qreg
 647 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 648 {
 649         struct qreg t = qir_get_temp(c);
 650         qir_emit(c, qir_inst(QOP_UNPACK_16A_I + i, t, src, c->undef));
 651         return t;
 652 }
 653
 654 static inline struct qreg
 655 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 656 {
 657         qir_emit(c, qir_inst(QOP_PACK_8A_F + chan, dest, val, c->undef));
 658         if (dest.file == QFILE_TEMP)
 659                 c->defs[dest.index] = NULL;
 660         return dest;
 661 }
 662
 663 static inline struct qreg
 664 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 665 {
 666         return qir_EXP2(c, qir_FMUL(c,
 667                                     y,
 668                                     qir_LOG2(c, x)));
 669 }
 670
 671 static inline void
 672 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 673 {
 674         static const struct qreg vpm = { QFILE_VPM, 0 };
 675         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 676 }
 677
 678 #endif /* VC4_QIR_H */