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 "pipe/p_state.h"
  42
  43 struct nir_builder;
  44
  45 enum qfile {
  46         QFILE_NULL,
  47         QFILE_TEMP,
  48         QFILE_VARY,
  49         QFILE_UNIF,
  50         QFILE_VPM,
  51
  52         /**
  53          * Stores an immediate value in the index field that can be turned
  54          * into a small immediate field by qpu_encode_small_immediate().
  55          */
  56         QFILE_SMALL_IMM,
  57 };
  58
  59 struct qreg {
  60         enum qfile file;
  61         uint32_t index;
  62         int pack;
  63 };
  64
  65 enum qop {
  66         QOP_UNDEF,
  67         QOP_MOV,
  68         QOP_FMOV,
  69         QOP_MMOV,
  70         QOP_FADD,
  71         QOP_FSUB,
  72         QOP_FMUL,
  73         QOP_V8MULD,
  74         QOP_V8MIN,
  75         QOP_V8MAX,
  76         QOP_V8ADDS,
  77         QOP_V8SUBS,
  78         QOP_MUL24,
  79         QOP_FMIN,
  80         QOP_FMAX,
  81         QOP_FMINABS,
  82         QOP_FMAXABS,
  83         QOP_ADD,
  84         QOP_SUB,
  85         QOP_SHL,
  86         QOP_SHR,
  87         QOP_ASR,
  88         QOP_MIN,
  89         QOP_MAX,
  90         QOP_AND,
  91         QOP_OR,
  92         QOP_XOR,
  93         QOP_NOT,
  94
  95         /* Note: Orderings of these compares must be the same as in
  96          * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
  97          * otherwise 0. */
  98         QOP_SEL_X_0_ZS,
  99         QOP_SEL_X_0_ZC,
 100         QOP_SEL_X_0_NS,
 101         QOP_SEL_X_0_NC,
 102         QOP_SEL_X_0_CS,
 103         QOP_SEL_X_0_CC,
 104         /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
 105         QOP_SEL_X_Y_ZS,
 106         QOP_SEL_X_Y_ZC,
 107         QOP_SEL_X_Y_NS,
 108         QOP_SEL_X_Y_NC,
 109         QOP_SEL_X_Y_CS,
 110         QOP_SEL_X_Y_CC,
 111
 112         QOP_FTOI,
 113         QOP_ITOF,
 114         QOP_RCP,
 115         QOP_RSQ,
 116         QOP_EXP2,
 117         QOP_LOG2,
 118         QOP_VW_SETUP,
 119         QOP_VR_SETUP,
 120         QOP_TLB_DISCARD_SETUP,
 121         QOP_TLB_STENCIL_SETUP,
 122         QOP_TLB_Z_WRITE,
 123         QOP_TLB_COLOR_WRITE,
 124         QOP_TLB_COLOR_READ,
 125         QOP_VARY_ADD_C,
 126
 127         QOP_FRAG_X,
 128         QOP_FRAG_Y,
 129         QOP_FRAG_Z,
 130         QOP_FRAG_W,
 131         QOP_FRAG_REV_FLAG,
 132
 133         /** Texture x coordinate parameter write */
 134         QOP_TEX_S,
 135         /** Texture y coordinate parameter write */
 136         QOP_TEX_T,
 137         /** Texture border color parameter or cube map z coordinate write */
 138         QOP_TEX_R,
 139         /** Texture LOD bias parameter write */
 140         QOP_TEX_B,
 141
 142         /**
 143          * Texture-unit 4-byte read with address provided direct in S
 144          * cooordinate.
 145          *
 146          * The first operand is the offset from the start of the UBO, and the
 147          * second is the uniform that has the UBO's base pointer.
 148          */
 149         QOP_TEX_DIRECT,
 150
 151         /**
 152          * Signal of texture read being necessary and then reading r4 into
 153          * the destination
 154          */
 155         QOP_TEX_RESULT,
 156 };
 157
 158 struct queued_qpu_inst {
 159         struct list_head link;
 160         uint64_t inst;
 161 };
 162
 163 struct qinst {
 164         struct list_head link;
 165
 166         enum qop op;
 167         struct qreg dst;
 168         struct qreg *src;
 169         bool sf;
 170 };
 171
 172 enum qstage {
 173         /**
 174          * Coordinate shader, runs during binning, before the VS, and just
 175          * outputs position.
 176          */
 177         QSTAGE_COORD,
 178         QSTAGE_VERT,
 179         QSTAGE_FRAG,
 180 };
 181
 182 enum quniform_contents {
 183         /**
 184          * Indicates that a constant 32-bit value is copied from the program's
 185          * uniform contents.
 186          */
 187         QUNIFORM_CONSTANT,
 188         /**
 189          * Indicates that the program's uniform contents are used as an index
 190          * into the GL uniform storage.
 191          */
 192         QUNIFORM_UNIFORM,
 193
 194         /** @{
 195          * Scaling factors from clip coordinates to relative to the viewport
 196          * center.
 197          *
 198          * This is used by the coordinate and vertex shaders to produce the
 199          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 200          * point offsets from the viewport ccenter.
 201          */
 202         QUNIFORM_VIEWPORT_X_SCALE,
 203         QUNIFORM_VIEWPORT_Y_SCALE,
 204         /** @} */
 205
 206         QUNIFORM_VIEWPORT_Z_OFFSET,
 207         QUNIFORM_VIEWPORT_Z_SCALE,
 208
 209         QUNIFORM_USER_CLIP_PLANE,
 210
 211         /**
 212          * A reference to a texture config parameter 0 uniform.
 213          *
 214          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 215          * defines texture type, miplevels, and such.  It will be found as a
 216          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 217          */
 218         QUNIFORM_TEXTURE_CONFIG_P0,
 219
 220         /**
 221          * A reference to a texture config parameter 1 uniform.
 222          *
 223          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 224          * defines texture width, height, filters, and wrap modes.  It will be
 225          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 226          * sequence.
 227          */
 228         QUNIFORM_TEXTURE_CONFIG_P1,
 229
 230         /** A reference to a texture config parameter 2 cubemap stride uniform */
 231         QUNIFORM_TEXTURE_CONFIG_P2,
 232
 233         QUNIFORM_UBO_ADDR,
 234
 235         QUNIFORM_TEXRECT_SCALE_X,
 236         QUNIFORM_TEXRECT_SCALE_Y,
 237
 238         QUNIFORM_TEXTURE_BORDER_COLOR,
 239
 240         QUNIFORM_BLEND_CONST_COLOR_X,
 241         QUNIFORM_BLEND_CONST_COLOR_Y,
 242         QUNIFORM_BLEND_CONST_COLOR_Z,
 243         QUNIFORM_BLEND_CONST_COLOR_W,
 244         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 245         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 246
 247         QUNIFORM_STENCIL,
 248
 249         QUNIFORM_ALPHA_REF,
 250 };
 251
 252 struct vc4_varying_slot {
 253         uint8_t slot;
 254         uint8_t swizzle;
 255 };
 256
 257 struct vc4_compiler_ubo_range {
 258         /**
 259          * offset in bytes from the start of the ubo where this range is
 260          * uploaded.
 261          *
 262          * Only set once used is set.
 263          */
 264         uint32_t dst_offset;
 265
 266         /**
 267          * offset in bytes from the start of the gallium uniforms where the
 268          * data comes from.
 269          */
 270         uint32_t src_offset;
 271
 272         /** size in bytes of this ubo range */
 273         uint32_t size;
 274
 275         /**
 276          * Set if this range is used by the shader for indirect uniforms
 277          * access.
 278          */
 279         bool used;
 280 };
 281
 282 struct vc4_key {
 283         struct vc4_uncompiled_shader *shader_state;
 284         struct {
 285                 enum pipe_format format;
 286                 unsigned compare_mode:1;
 287                 unsigned compare_func:3;
 288                 unsigned wrap_s:3;
 289                 unsigned wrap_t:3;
 290                 uint8_t swizzle[4];
 291         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 292         uint8_t ucp_enables;
 293 };
 294
 295 struct vc4_fs_key {
 296         struct vc4_key base;
 297         enum pipe_format color_format;
 298         bool depth_enabled;
 299         bool stencil_enabled;
 300         bool stencil_twoside;
 301         bool stencil_full_writemasks;
 302         bool is_points;
 303         bool is_lines;
 304         bool alpha_test;
 305         bool point_coord_upper_left;
 306         bool light_twoside;
 307         uint8_t alpha_test_func;
 308         uint8_t logicop_func;
 309         uint32_t point_sprite_mask;
 310
 311         struct pipe_rt_blend_state blend;
 312 };
 313
 314 struct vc4_vs_key {
 315         struct vc4_key base;
 316
 317         /**
 318          * This is a proxy for the array of FS input semantics, which is
 319          * larger than we would want to put in the key.
 320          */
 321         uint64_t compiled_fs_id;
 322
 323         enum pipe_format attr_formats[8];
 324         bool is_coord;
 325         bool per_vertex_point_size;
 326 };
 327
 328 struct vc4_compile {
 329         struct vc4_context *vc4;
 330         nir_shader *s;
 331         nir_function_impl *impl;
 332         struct exec_list *cf_node_list;
 333
 334         /**
 335          * Mapping from nir_register * or nir_ssa_def * to array of struct
 336          * qreg for the values.
 337          */
 338         struct hash_table *def_ht;
 339
 340         /* For each temp, the instruction generating its value. */
 341         struct qinst **defs;
 342         uint32_t defs_array_size;
 343
 344         /**
 345          * Inputs to the shader, arranged by TGSI declaration order.
 346          *
 347          * Not all fragment shader QFILE_VARY reads are present in this array.
 348          */
 349         struct qreg *inputs;
 350         struct qreg *outputs;
 351         uint32_t inputs_array_size;
 352         uint32_t outputs_array_size;
 353         uint32_t uniforms_array_size;
 354
 355         struct vc4_compiler_ubo_range *ubo_ranges;
 356         uint32_t ubo_ranges_array_size;
 357         /** Number of uniform areas declared in ubo_ranges. */
 358         uint32_t num_uniform_ranges;
 359         /** Number of uniform areas used for indirect addressed loads. */
 360         uint32_t num_ubo_ranges;
 361         uint32_t next_ubo_dst_offset;
 362
 363         struct qreg line_x, point_x, point_y;
 364         struct qreg discard;
 365
 366         uint8_t vattr_sizes[8];
 367
 368         /**
 369          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 370          *
 371          * This includes those that aren't part of the VPM varyings, like
 372          * point/line coordinates.
 373          */
 374         struct vc4_varying_slot *input_slots;
 375         uint32_t num_input_slots;
 376         uint32_t input_slots_array_size;
 377
 378         /**
 379          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 380          * of the output is.  Used to emit from the VS in the order that the
 381          * FS needs.
 382          */
 383         struct vc4_varying_slot *output_slots;
 384
 385         struct pipe_shader_state *shader_state;
 386         struct vc4_key *key;
 387         struct vc4_fs_key *fs_key;
 388         struct vc4_vs_key *vs_key;
 389
 390         uint32_t *uniform_data;
 391         enum quniform_contents *uniform_contents;
 392         uint32_t uniform_array_size;
 393         uint32_t num_uniforms;
 394         uint32_t num_outputs;
 395         uint32_t num_texture_samples;
 396         uint32_t output_position_index;
 397         uint32_t output_color_index;
 398         uint32_t output_point_size_index;
 399
 400         struct qreg undef;
 401         enum qstage stage;
 402         uint32_t num_temps;
 403         struct list_head instructions;
 404         uint32_t immediates[1024];
 405
 406         struct list_head qpu_inst_list;
 407         uint64_t *qpu_insts;
 408         uint32_t qpu_inst_count;
 409         uint32_t qpu_inst_size;
 410         uint32_t num_inputs;
 411
 412         uint32_t program_id;
 413         uint32_t variant_id;
 414 };
 415
 416 /* Special nir_load_input intrinsic index for loading the current TLB
 417  * destination color.
 418  */
 419 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 420
 421 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 422  * state-dependent value.
 423  */
 424 #define VC4_NIR_STATE_UNIFORM_OFFSET            2000000000
 425
 426 struct vc4_compile *qir_compile_init(void);
 427 void qir_compile_destroy(struct vc4_compile *c);
 428 struct qinst *qir_inst(enum qop op, struct qreg dst,
 429                        struct qreg src0, struct qreg src1);
 430 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 431                         struct qreg a,
 432                         struct qreg b,
 433                         struct qreg c,
 434                         struct qreg d);
 435 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 436 struct qreg qir_uniform(struct vc4_compile *c,
 437                         enum quniform_contents contents,
 438                         uint32_t data);
 439 void qir_reorder_uniforms(struct vc4_compile *c);
 440
 441 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 442 static inline void qir_emit_nodef(struct vc4_compile *c, struct qinst *inst)
 443 {
 444         list_addtail(&inst->link, &c->instructions);
 445 }
 446
 447 struct qreg qir_get_temp(struct vc4_compile *c);
 448 int qir_get_op_nsrc(enum qop qop);
 449 bool qir_reg_equals(struct qreg a, struct qreg b);
 450 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 451 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 452 bool qir_is_multi_instruction(struct qinst *inst);
 453 bool qir_is_mul(struct qinst *inst);
 454 bool qir_is_raw_mov(struct qinst *inst);
 455 bool qir_is_tex(struct qinst *inst);
 456 bool qir_is_float_input(struct qinst *inst);
 457 bool qir_depends_on_flags(struct qinst *inst);
 458 bool qir_writes_r4(struct qinst *inst);
 459 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 460
 461 void qir_dump(struct vc4_compile *c);
 462 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 463 const char *qir_get_stage_name(enum qstage stage);
 464
 465 void qir_optimize(struct vc4_compile *c);
 466 bool qir_opt_algebraic(struct vc4_compile *c);
 467 bool qir_opt_constant_folding(struct vc4_compile *c);
 468 bool qir_opt_copy_propagation(struct vc4_compile *c);
 469 bool qir_opt_cse(struct vc4_compile *c);
 470 bool qir_opt_dead_code(struct vc4_compile *c);
 471 bool qir_opt_small_immediates(struct vc4_compile *c);
 472 bool qir_opt_vpm_writes(struct vc4_compile *c);
 473 void vc4_nir_lower_blend(struct vc4_compile *c);
 474 void vc4_nir_lower_io(struct vc4_compile *c);
 475 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 476                                        enum quniform_contents contents);
 477 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 478                                           nir_ssa_def **srcs, int swiz);
 479 void qir_lower_uniforms(struct vc4_compile *c);
 480
 481 void qpu_schedule_instructions(struct vc4_compile *c);
 482
 483 void qir_SF(struct vc4_compile *c, struct qreg src);
 484
 485 static inline struct qreg
 486 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 487 {
 488         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 489 }
 490
 491 static inline struct qreg
 492 qir_uniform_f(struct vc4_compile *c, float f)
 493 {
 494         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 495 }
 496
 497 #define QIR_ALU0(name)                                                   \
 498 static inline struct qreg                                                \
 499 qir_##name(struct vc4_compile *c)                                        \
 500 {                                                                        \
 501         struct qreg t = qir_get_temp(c);                                 \
 502         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 503         return t;                                                        \
 504 }
 505
 506 #define QIR_ALU1(name)                                                   \
 507 static inline struct qreg                                                \
 508 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 509 {                                                                        \
 510         struct qreg t = qir_get_temp(c);                                 \
 511         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 512         return t;                                                        \
 513 }                                                                        \
 514 static inline void                                                       \
 515 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 516                   struct qreg a)                                         \
 517 {                                                                        \
 518         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 519 }
 520
 521 #define QIR_ALU2(name)                                                   \
 522 static inline struct qreg                                                \
 523 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 524 {                                                                        \
 525         struct qreg t = qir_get_temp(c);                                 \
 526         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 527         return t;                                                        \
 528 }                                                                        \
 529 static inline void                                                       \
 530 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 531                   struct qreg a, struct qreg b)                          \
 532 {                                                                        \
 533         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, b));             \
 534 }
 535
 536 #define QIR_NODST_1(name)                                               \
 537 static inline void                                                      \
 538 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 539 {                                                                       \
 540         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 541 }
 542
 543 #define QIR_NODST_2(name)                                               \
 544 static inline void                                                      \
 545 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 546 {                                                                       \
 547         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 548 }
 549
 550 #define QIR_PACK(name)                                                   \
 551 static inline struct qreg                                                \
 552 qir_##name(struct vc4_compile *c, struct qreg dest, struct qreg a)       \
 553 {                                                                        \
 554         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 555         return dest;                                                     \
 556 }
 557
 558 QIR_ALU1(MOV)
 559 QIR_ALU1(FMOV)
 560 QIR_ALU1(MMOV)
 561 QIR_ALU2(FADD)
 562 QIR_ALU2(FSUB)
 563 QIR_ALU2(FMUL)
 564 QIR_ALU2(V8MULD)
 565 QIR_ALU2(V8MIN)
 566 QIR_ALU2(V8MAX)
 567 QIR_ALU2(V8ADDS)
 568 QIR_ALU2(V8SUBS)
 569 QIR_ALU2(MUL24)
 570 QIR_ALU1(SEL_X_0_ZS)
 571 QIR_ALU1(SEL_X_0_ZC)
 572 QIR_ALU1(SEL_X_0_NS)
 573 QIR_ALU1(SEL_X_0_NC)
 574 QIR_ALU1(SEL_X_0_CS)
 575 QIR_ALU1(SEL_X_0_CC)
 576 QIR_ALU2(SEL_X_Y_ZS)
 577 QIR_ALU2(SEL_X_Y_ZC)
 578 QIR_ALU2(SEL_X_Y_NS)
 579 QIR_ALU2(SEL_X_Y_NC)
 580 QIR_ALU2(SEL_X_Y_CS)
 581 QIR_ALU2(SEL_X_Y_CC)
 582 QIR_ALU2(FMIN)
 583 QIR_ALU2(FMAX)
 584 QIR_ALU2(FMINABS)
 585 QIR_ALU2(FMAXABS)
 586 QIR_ALU1(FTOI)
 587 QIR_ALU1(ITOF)
 588
 589 QIR_ALU2(ADD)
 590 QIR_ALU2(SUB)
 591 QIR_ALU2(SHL)
 592 QIR_ALU2(SHR)
 593 QIR_ALU2(ASR)
 594 QIR_ALU2(MIN)
 595 QIR_ALU2(MAX)
 596 QIR_ALU2(AND)
 597 QIR_ALU2(OR)
 598 QIR_ALU2(XOR)
 599 QIR_ALU1(NOT)
 600
 601 QIR_ALU1(RCP)
 602 QIR_ALU1(RSQ)
 603 QIR_ALU1(EXP2)
 604 QIR_ALU1(LOG2)
 605 QIR_ALU1(VARY_ADD_C)
 606 QIR_NODST_2(TEX_S)
 607 QIR_NODST_2(TEX_T)
 608 QIR_NODST_2(TEX_R)
 609 QIR_NODST_2(TEX_B)
 610 QIR_NODST_2(TEX_DIRECT)
 611 QIR_ALU0(FRAG_X)
 612 QIR_ALU0(FRAG_Y)
 613 QIR_ALU0(FRAG_Z)
 614 QIR_ALU0(FRAG_W)
 615 QIR_ALU0(FRAG_REV_FLAG)
 616 QIR_ALU0(TEX_RESULT)
 617 QIR_ALU0(TLB_COLOR_READ)
 618 QIR_NODST_1(TLB_COLOR_WRITE)
 619 QIR_NODST_1(TLB_Z_WRITE)
 620 QIR_NODST_1(TLB_DISCARD_SETUP)
 621 QIR_NODST_1(TLB_STENCIL_SETUP)
 622
 623 static inline struct qreg
 624 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 625 {
 626         struct qreg t = qir_FMOV(c, src);
 627         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 628         return t;
 629 }
 630
 631 static inline struct qreg
 632 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 633 {
 634         struct qreg t = qir_MOV(c, src);
 635         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 636         return t;
 637 }
 638
 639 static inline struct qreg
 640 qir_UNPACK_16_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_16A + i;
 644         return t;
 645 }
 646
 647 static inline struct qreg
 648 qir_UNPACK_16_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_16A + i;
 652         return t;
 653 }
 654
 655 static inline void
 656 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 657 {
 658         assert(!dest.pack);
 659         dest.pack = QPU_PACK_MUL_8A + chan;
 660         qir_emit(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 661         if (dest.file == QFILE_TEMP)
 662                 c->defs[dest.index] = NULL;
 663 }
 664
 665 static inline struct qreg
 666 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 667 {
 668         struct qreg dest = qir_MMOV(c, val);
 669         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 670         return dest;
 671 }
 672
 673 static inline struct qreg
 674 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 675 {
 676         return qir_EXP2(c, qir_FMUL(c,
 677                                     y,
 678                                     qir_LOG2(c, x)));
 679 }
 680
 681 static inline void
 682 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 683 {
 684         static const struct qreg vpm = { QFILE_VPM, 0 };
 685         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 686 }
 687
 688 #endif /* VC4_QIR_H */