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 "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 "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         QFILE_TLB_COLOR_WRITE,
  53         QFILE_TLB_COLOR_WRITE_MS,
  54         QFILE_TLB_Z_WRITE,
  55         QFILE_TLB_STENCIL_SETUP,
  56
  57         /**
  58          * Stores an immediate value in the index field that can be turned
  59          * into a small immediate field by qpu_encode_small_immediate().
  60          */
  61         QFILE_SMALL_IMM,
  62 };
  63
  64 struct qreg {
  65         enum qfile file;
  66         uint32_t index;
  67         int pack;
  68 };
  69
  70 static inline struct qreg qir_reg(enum qfile file, uint32_t index)
  71 {
  72         return (struct qreg){file, index};
  73 }
  74
  75 enum qop {
  76         QOP_UNDEF,
  77         QOP_MOV,
  78         QOP_FMOV,
  79         QOP_MMOV,
  80         QOP_FADD,
  81         QOP_FSUB,
  82         QOP_FMUL,
  83         QOP_V8MULD,
  84         QOP_V8MIN,
  85         QOP_V8MAX,
  86         QOP_V8ADDS,
  87         QOP_V8SUBS,
  88         QOP_MUL24,
  89         QOP_FMIN,
  90         QOP_FMAX,
  91         QOP_FMINABS,
  92         QOP_FMAXABS,
  93         QOP_ADD,
  94         QOP_SUB,
  95         QOP_SHL,
  96         QOP_SHR,
  97         QOP_ASR,
  98         QOP_MIN,
  99         QOP_MAX,
 100         QOP_AND,
 101         QOP_OR,
 102         QOP_XOR,
 103         QOP_NOT,
 104
 105         QOP_FTOI,
 106         QOP_ITOF,
 107         QOP_RCP,
 108         QOP_RSQ,
 109         QOP_EXP2,
 110         QOP_LOG2,
 111         QOP_VW_SETUP,
 112         QOP_VR_SETUP,
 113         QOP_TLB_COLOR_READ,
 114         QOP_MS_MASK,
 115         QOP_VARY_ADD_C,
 116
 117         QOP_FRAG_X,
 118         QOP_FRAG_Y,
 119         QOP_FRAG_Z,
 120         QOP_FRAG_W,
 121         QOP_FRAG_REV_FLAG,
 122
 123         /** Texture x coordinate parameter write */
 124         QOP_TEX_S,
 125         /** Texture y coordinate parameter write */
 126         QOP_TEX_T,
 127         /** Texture border color parameter or cube map z coordinate write */
 128         QOP_TEX_R,
 129         /** Texture LOD bias parameter write */
 130         QOP_TEX_B,
 131
 132         /**
 133          * Texture-unit 4-byte read with address provided direct in S
 134          * cooordinate.
 135          *
 136          * The first operand is the offset from the start of the UBO, and the
 137          * second is the uniform that has the UBO's base pointer.
 138          */
 139         QOP_TEX_DIRECT,
 140
 141         /**
 142          * Signal of texture read being necessary and then reading r4 into
 143          * the destination
 144          */
 145         QOP_TEX_RESULT,
 146 };
 147
 148 struct queued_qpu_inst {
 149         struct list_head link;
 150         uint64_t inst;
 151 };
 152
 153 struct qinst {
 154         struct list_head link;
 155
 156         enum qop op;
 157         struct qreg dst;
 158         struct qreg *src;
 159         bool sf;
 160         uint8_t cond;
 161 };
 162
 163 enum qstage {
 164         /**
 165          * Coordinate shader, runs during binning, before the VS, and just
 166          * outputs position.
 167          */
 168         QSTAGE_COORD,
 169         QSTAGE_VERT,
 170         QSTAGE_FRAG,
 171 };
 172
 173 enum quniform_contents {
 174         /**
 175          * Indicates that a constant 32-bit value is copied from the program's
 176          * uniform contents.
 177          */
 178         QUNIFORM_CONSTANT,
 179         /**
 180          * Indicates that the program's uniform contents are used as an index
 181          * into the GL uniform storage.
 182          */
 183         QUNIFORM_UNIFORM,
 184
 185         /** @{
 186          * Scaling factors from clip coordinates to relative to the viewport
 187          * center.
 188          *
 189          * This is used by the coordinate and vertex shaders to produce the
 190          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 191          * point offsets from the viewport ccenter.
 192          */
 193         QUNIFORM_VIEWPORT_X_SCALE,
 194         QUNIFORM_VIEWPORT_Y_SCALE,
 195         /** @} */
 196
 197         QUNIFORM_VIEWPORT_Z_OFFSET,
 198         QUNIFORM_VIEWPORT_Z_SCALE,
 199
 200         QUNIFORM_USER_CLIP_PLANE,
 201
 202         /**
 203          * A reference to a texture config parameter 0 uniform.
 204          *
 205          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 206          * defines texture type, miplevels, and such.  It will be found as a
 207          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 208          */
 209         QUNIFORM_TEXTURE_CONFIG_P0,
 210
 211         /**
 212          * A reference to a texture config parameter 1 uniform.
 213          *
 214          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 215          * defines texture width, height, filters, and wrap modes.  It will be
 216          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 217          * sequence.
 218          */
 219         QUNIFORM_TEXTURE_CONFIG_P1,
 220
 221         /** A reference to a texture config parameter 2 cubemap stride uniform */
 222         QUNIFORM_TEXTURE_CONFIG_P2,
 223
 224         QUNIFORM_TEXTURE_MSAA_ADDR,
 225
 226         QUNIFORM_UBO_ADDR,
 227
 228         QUNIFORM_TEXRECT_SCALE_X,
 229         QUNIFORM_TEXRECT_SCALE_Y,
 230
 231         QUNIFORM_TEXTURE_BORDER_COLOR,
 232
 233         QUNIFORM_BLEND_CONST_COLOR_X,
 234         QUNIFORM_BLEND_CONST_COLOR_Y,
 235         QUNIFORM_BLEND_CONST_COLOR_Z,
 236         QUNIFORM_BLEND_CONST_COLOR_W,
 237         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 238         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 239
 240         QUNIFORM_STENCIL,
 241
 242         QUNIFORM_ALPHA_REF,
 243         QUNIFORM_SAMPLE_MASK,
 244 };
 245
 246 struct vc4_varying_slot {
 247         uint8_t slot;
 248         uint8_t swizzle;
 249 };
 250
 251 struct vc4_compiler_ubo_range {
 252         /**
 253          * offset in bytes from the start of the ubo where this range is
 254          * uploaded.
 255          *
 256          * Only set once used is set.
 257          */
 258         uint32_t dst_offset;
 259
 260         /**
 261          * offset in bytes from the start of the gallium uniforms where the
 262          * data comes from.
 263          */
 264         uint32_t src_offset;
 265
 266         /** size in bytes of this ubo range */
 267         uint32_t size;
 268
 269         /**
 270          * Set if this range is used by the shader for indirect uniforms
 271          * access.
 272          */
 273         bool used;
 274 };
 275
 276 struct vc4_key {
 277         struct vc4_uncompiled_shader *shader_state;
 278         struct {
 279                 enum pipe_format format;
 280                 uint8_t swizzle[4];
 281                 union {
 282                         struct {
 283                                 unsigned compare_mode:1;
 284                                 unsigned compare_func:3;
 285                                 unsigned wrap_s:3;
 286                                 unsigned wrap_t:3;
 287                         };
 288                         struct {
 289                                 uint16_t msaa_width, msaa_height;
 290                         };
 291                 };
 292         } tex[VC4_MAX_TEXTURE_SAMPLERS];
 293         uint8_t ucp_enables;
 294 };
 295
 296 struct vc4_fs_key {
 297         struct vc4_key base;
 298         enum pipe_format color_format;
 299         bool depth_enabled;
 300         bool stencil_enabled;
 301         bool stencil_twoside;
 302         bool stencil_full_writemasks;
 303         bool is_points;
 304         bool is_lines;
 305         bool alpha_test;
 306         bool point_coord_upper_left;
 307         bool light_twoside;
 308         bool msaa;
 309         bool sample_coverage;
 310         bool sample_alpha_to_coverage;
 311         bool sample_alpha_to_one;
 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         bool msaa_per_sample_output;
 357         struct qreg color_reads[VC4_MAX_SAMPLES];
 358         struct qreg sample_colors[VC4_MAX_SAMPLES];
 359         uint32_t inputs_array_size;
 360         uint32_t outputs_array_size;
 361         uint32_t uniforms_array_size;
 362
 363         struct vc4_compiler_ubo_range *ubo_ranges;
 364         uint32_t ubo_ranges_array_size;
 365         /** Number of uniform areas declared in ubo_ranges. */
 366         uint32_t num_uniform_ranges;
 367         /** Number of uniform areas used for indirect addressed loads. */
 368         uint32_t num_ubo_ranges;
 369         uint32_t next_ubo_dst_offset;
 370
 371         struct qreg line_x, point_x, point_y;
 372         struct qreg discard;
 373
 374         uint8_t vattr_sizes[8];
 375
 376         /* Bitfield for whether a given channel of a sampler needs sRGB
 377          * decode.
 378          */
 379         uint8_t tex_srgb_decode[VC4_MAX_TEXTURE_SAMPLERS];
 380
 381         /**
 382          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 383          *
 384          * This includes those that aren't part of the VPM varyings, like
 385          * point/line coordinates.
 386          */
 387         struct vc4_varying_slot *input_slots;
 388         uint32_t num_input_slots;
 389         uint32_t input_slots_array_size;
 390
 391         /**
 392          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 393          * of the output is.  Used to emit from the VS in the order that the
 394          * FS needs.
 395          */
 396         struct vc4_varying_slot *output_slots;
 397
 398         struct pipe_shader_state *shader_state;
 399         struct vc4_key *key;
 400         struct vc4_fs_key *fs_key;
 401         struct vc4_vs_key *vs_key;
 402
 403         uint32_t *uniform_data;
 404         enum quniform_contents *uniform_contents;
 405         uint32_t uniform_array_size;
 406         uint32_t num_uniforms;
 407         uint32_t num_outputs;
 408         uint32_t num_texture_samples;
 409         uint32_t output_position_index;
 410         uint32_t output_color_index;
 411         uint32_t output_point_size_index;
 412         uint32_t output_sample_mask_index;
 413
 414         struct qreg undef;
 415         enum qstage stage;
 416         uint32_t num_temps;
 417         struct list_head instructions;
 418         uint32_t immediates[1024];
 419
 420         struct list_head qpu_inst_list;
 421         uint64_t *qpu_insts;
 422         uint32_t qpu_inst_count;
 423         uint32_t qpu_inst_size;
 424         uint32_t num_inputs;
 425
 426         uint32_t program_id;
 427         uint32_t variant_id;
 428 };
 429
 430 /* Special nir_load_input intrinsic index for loading the current TLB
 431  * destination color.
 432  */
 433 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 434
 435 #define VC4_NIR_MS_MASK_OUTPUT                  2000000000
 436
 437 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 438  * state-dependent value.
 439  */
 440 #define VC4_NIR_STATE_UNIFORM_OFFSET            1000000000
 441
 442 struct vc4_compile *qir_compile_init(void);
 443 void qir_compile_destroy(struct vc4_compile *c);
 444 struct qinst *qir_inst(enum qop op, struct qreg dst,
 445                        struct qreg src0, struct qreg src1);
 446 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 447                         struct qreg a,
 448                         struct qreg b,
 449                         struct qreg c,
 450                         struct qreg d);
 451 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 452 struct qreg qir_uniform(struct vc4_compile *c,
 453                         enum quniform_contents contents,
 454                         uint32_t data);
 455 void qir_schedule_instructions(struct vc4_compile *c);
 456 void qir_reorder_uniforms(struct vc4_compile *c);
 457
 458 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 459 static inline struct qinst *
 460 qir_emit_nodef(struct vc4_compile *c, struct qinst *inst)
 461 {
 462         list_addtail(&inst->link, &c->instructions);
 463         return inst;
 464 }
 465
 466 struct qreg qir_get_temp(struct vc4_compile *c);
 467 int qir_get_op_nsrc(enum qop qop);
 468 bool qir_reg_equals(struct qreg a, struct qreg b);
 469 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 470 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 471 bool qir_is_mul(struct qinst *inst);
 472 bool qir_is_raw_mov(struct qinst *inst);
 473 bool qir_is_tex(struct qinst *inst);
 474 bool qir_is_float_input(struct qinst *inst);
 475 bool qir_depends_on_flags(struct qinst *inst);
 476 bool qir_writes_r4(struct qinst *inst);
 477 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 478
 479 void qir_dump(struct vc4_compile *c);
 480 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 481 const char *qir_get_stage_name(enum qstage stage);
 482
 483 void qir_optimize(struct vc4_compile *c);
 484 bool qir_opt_algebraic(struct vc4_compile *c);
 485 bool qir_opt_constant_folding(struct vc4_compile *c);
 486 bool qir_opt_copy_propagation(struct vc4_compile *c);
 487 bool qir_opt_cse(struct vc4_compile *c);
 488 bool qir_opt_dead_code(struct vc4_compile *c);
 489 bool qir_opt_small_immediates(struct vc4_compile *c);
 490 bool qir_opt_vpm(struct vc4_compile *c);
 491 void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c);
 492 void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c);
 493 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 494                                        enum quniform_contents contents);
 495 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 496                                           nir_ssa_def **srcs, int swiz);
 497 void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c);
 498 void qir_lower_uniforms(struct vc4_compile *c);
 499
 500 uint32_t qpu_schedule_instructions(struct vc4_compile *c);
 501
 502 void qir_SF(struct vc4_compile *c, struct qreg src);
 503
 504 static inline struct qreg
 505 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 506 {
 507         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 508 }
 509
 510 static inline struct qreg
 511 qir_uniform_f(struct vc4_compile *c, float f)
 512 {
 513         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 514 }
 515
 516 #define QIR_ALU0(name)                                                   \
 517 static inline struct qreg                                                \
 518 qir_##name(struct vc4_compile *c)                                        \
 519 {                                                                        \
 520         struct qreg t = qir_get_temp(c);                                 \
 521         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 522         return t;                                                        \
 523 }
 524
 525 #define QIR_ALU1(name)                                                   \
 526 static inline struct qreg                                                \
 527 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 528 {                                                                        \
 529         struct qreg t = qir_get_temp(c);                                 \
 530         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 531         return t;                                                        \
 532 }                                                                        \
 533 static inline struct qinst *                                             \
 534 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 535                   struct qreg a)                                         \
 536 {                                                                        \
 537         return qir_emit_nodef(c, qir_inst(QOP_##name, dest, a,           \
 538                                           c->undef));                    \
 539 }
 540
 541 #define QIR_ALU2(name)                                                   \
 542 static inline struct qreg                                                \
 543 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 544 {                                                                        \
 545         struct qreg t = qir_get_temp(c);                                 \
 546         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 547         return t;                                                        \
 548 }                                                                        \
 549 static inline void                                                       \
 550 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 551                   struct qreg a, struct qreg b)                          \
 552 {                                                                        \
 553         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, b));             \
 554 }
 555
 556 #define QIR_NODST_1(name)                                               \
 557 static inline struct qinst *                                            \
 558 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 559 {                                                                       \
 560         struct qinst *inst = qir_inst(QOP_##name, c->undef,             \
 561                                       a, c->undef);                     \
 562         qir_emit(c, inst);                                              \
 563         return inst;                                                    \
 564 }
 565
 566 #define QIR_NODST_2(name)                                               \
 567 static inline struct qinst *                                            \
 568 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 569 {                                                                       \
 570         struct qinst *inst = qir_inst(QOP_##name, c->undef,             \
 571                                       a, b);                            \
 572         qir_emit(c, inst);                                              \
 573         return inst;                                                    \
 574 }
 575
 576 #define QIR_PACK(name)                                                   \
 577 static inline struct qreg                                                \
 578 qir_##name(struct vc4_compile *c, struct qreg dest, struct qreg a)       \
 579 {                                                                        \
 580         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 581         return dest;                                                     \
 582 }
 583
 584 QIR_ALU1(MOV)
 585 QIR_ALU1(FMOV)
 586 QIR_ALU1(MMOV)
 587 QIR_ALU2(FADD)
 588 QIR_ALU2(FSUB)
 589 QIR_ALU2(FMUL)
 590 QIR_ALU2(V8MULD)
 591 QIR_ALU2(V8MIN)
 592 QIR_ALU2(V8MAX)
 593 QIR_ALU2(V8ADDS)
 594 QIR_ALU2(V8SUBS)
 595 QIR_ALU2(MUL24)
 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(MS_MASK)
 633
 634 static inline struct qreg
 635 qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1)
 636 {
 637         struct qreg t = qir_get_temp(c);
 638         struct qinst *a = qir_MOV_dest(c, t, src0);
 639         struct qinst *b = qir_MOV_dest(c, t, src1);
 640         a->cond = cond;
 641         b->cond = cond ^ 1;
 642         return t;
 643 }
 644
 645 static inline struct qreg
 646 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 647 {
 648         struct qreg t = qir_FMOV(c, src);
 649         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 650         return t;
 651 }
 652
 653 static inline struct qreg
 654 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 655 {
 656         struct qreg t = qir_MOV(c, src);
 657         c->defs[t.index]->src[0].pack = QPU_UNPACK_8A + i;
 658         return t;
 659 }
 660
 661 static inline struct qreg
 662 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 663 {
 664         struct qreg t = qir_FMOV(c, src);
 665         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 666         return t;
 667 }
 668
 669 static inline struct qreg
 670 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 671 {
 672         struct qreg t = qir_MOV(c, src);
 673         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 674         return t;
 675 }
 676
 677 static inline void
 678 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 679 {
 680         assert(!dest.pack);
 681         dest.pack = QPU_PACK_MUL_8A + chan;
 682         qir_emit(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 683         if (dest.file == QFILE_TEMP)
 684                 c->defs[dest.index] = NULL;
 685 }
 686
 687 static inline struct qreg
 688 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 689 {
 690         struct qreg dest = qir_MMOV(c, val);
 691         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 692         return dest;
 693 }
 694
 695 static inline struct qreg
 696 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 697 {
 698         return qir_EXP2(c, qir_FMUL(c,
 699                                     y,
 700                                     qir_LOG2(c, x)));
 701 }
 702
 703 static inline void
 704 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 705 {
 706         qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val);
 707 }
 708
 709 #endif /* VC4_QIR_H */