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         /* Payload registers that aren't in the physical register file, so we
  58          * can just use the corresponding qpu_reg at qpu_emit time.
  59          */
  60         QFILE_FRAG_X,
  61         QFILE_FRAG_Y,
  62         QFILE_FRAG_REV_FLAG,
  63
  64         /**
  65          * Stores an immediate value in the index field that can be turned
  66          * into a small immediate field by qpu_encode_small_immediate().
  67          */
  68         QFILE_SMALL_IMM,
  69 };
  70
  71 struct qreg {
  72         enum qfile file;
  73         uint32_t index;
  74         int pack;
  75 };
  76
  77 static inline struct qreg qir_reg(enum qfile file, uint32_t index)
  78 {
  79         return (struct qreg){file, index};
  80 }
  81
  82 enum qop {
  83         QOP_UNDEF,
  84         QOP_MOV,
  85         QOP_FMOV,
  86         QOP_MMOV,
  87         QOP_FADD,
  88         QOP_FSUB,
  89         QOP_FMUL,
  90         QOP_V8MULD,
  91         QOP_V8MIN,
  92         QOP_V8MAX,
  93         QOP_V8ADDS,
  94         QOP_V8SUBS,
  95         QOP_MUL24,
  96         QOP_FMIN,
  97         QOP_FMAX,
  98         QOP_FMINABS,
  99         QOP_FMAXABS,
 100         QOP_ADD,
 101         QOP_SUB,
 102         QOP_SHL,
 103         QOP_SHR,
 104         QOP_ASR,
 105         QOP_MIN,
 106         QOP_MAX,
 107         QOP_AND,
 108         QOP_OR,
 109         QOP_XOR,
 110         QOP_NOT,
 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_COLOR_READ,
 121         QOP_MS_MASK,
 122         QOP_VARY_ADD_C,
 123
 124         QOP_FRAG_Z,
 125         QOP_FRAG_W,
 126
 127         /** Texture x coordinate parameter write */
 128         QOP_TEX_S,
 129         /** Texture y coordinate parameter write */
 130         QOP_TEX_T,
 131         /** Texture border color parameter or cube map z coordinate write */
 132         QOP_TEX_R,
 133         /** Texture LOD bias parameter write */
 134         QOP_TEX_B,
 135
 136         /**
 137          * Texture-unit 4-byte read with address provided direct in S
 138          * cooordinate.
 139          *
 140          * The first operand is the offset from the start of the UBO, and the
 141          * second is the uniform that has the UBO's base pointer.
 142          */
 143         QOP_TEX_DIRECT,
 144
 145         /**
 146          * Signal of texture read being necessary and then reading r4 into
 147          * the destination
 148          */
 149         QOP_TEX_RESULT,
 150 };
 151
 152 struct queued_qpu_inst {
 153         struct list_head link;
 154         uint64_t inst;
 155 };
 156
 157 struct qinst {
 158         struct list_head link;
 159
 160         enum qop op;
 161         struct qreg dst;
 162         struct qreg *src;
 163         bool sf;
 164         uint8_t cond;
 165 };
 166
 167 enum qstage {
 168         /**
 169          * Coordinate shader, runs during binning, before the VS, and just
 170          * outputs position.
 171          */
 172         QSTAGE_COORD,
 173         QSTAGE_VERT,
 174         QSTAGE_FRAG,
 175 };
 176
 177 enum quniform_contents {
 178         /**
 179          * Indicates that a constant 32-bit value is copied from the program's
 180          * uniform contents.
 181          */
 182         QUNIFORM_CONSTANT,
 183         /**
 184          * Indicates that the program's uniform contents are used as an index
 185          * into the GL uniform storage.
 186          */
 187         QUNIFORM_UNIFORM,
 188
 189         /** @{
 190          * Scaling factors from clip coordinates to relative to the viewport
 191          * center.
 192          *
 193          * This is used by the coordinate and vertex shaders to produce the
 194          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 195          * point offsets from the viewport ccenter.
 196          */
 197         QUNIFORM_VIEWPORT_X_SCALE,
 198         QUNIFORM_VIEWPORT_Y_SCALE,
 199         /** @} */
 200
 201         QUNIFORM_VIEWPORT_Z_OFFSET,
 202         QUNIFORM_VIEWPORT_Z_SCALE,
 203
 204         QUNIFORM_USER_CLIP_PLANE,
 205
 206         /**
 207          * A reference to a texture config parameter 0 uniform.
 208          *
 209          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 210          * defines texture type, miplevels, and such.  It will be found as a
 211          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 212          */
 213         QUNIFORM_TEXTURE_CONFIG_P0,
 214
 215         /**
 216          * A reference to a texture config parameter 1 uniform.
 217          *
 218          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 219          * defines texture width, height, filters, and wrap modes.  It will be
 220          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 221          * sequence.
 222          */
 223         QUNIFORM_TEXTURE_CONFIG_P1,
 224
 225         /** A reference to a texture config parameter 2 cubemap stride uniform */
 226         QUNIFORM_TEXTURE_CONFIG_P2,
 227
 228         QUNIFORM_TEXTURE_MSAA_ADDR,
 229
 230         QUNIFORM_UBO_ADDR,
 231
 232         QUNIFORM_TEXRECT_SCALE_X,
 233         QUNIFORM_TEXRECT_SCALE_Y,
 234
 235         QUNIFORM_TEXTURE_BORDER_COLOR,
 236
 237         QUNIFORM_BLEND_CONST_COLOR_X,
 238         QUNIFORM_BLEND_CONST_COLOR_Y,
 239         QUNIFORM_BLEND_CONST_COLOR_Z,
 240         QUNIFORM_BLEND_CONST_COLOR_W,
 241         QUNIFORM_BLEND_CONST_COLOR_RGBA,
 242         QUNIFORM_BLEND_CONST_COLOR_AAAA,
 243
 244         QUNIFORM_STENCIL,
 245
 246         QUNIFORM_ALPHA_REF,
 247         QUNIFORM_SAMPLE_MASK,
 248 };
 249
 250 struct vc4_varying_slot {
 251         uint8_t slot;
 252         uint8_t swizzle;
 253 };
 254
 255 struct vc4_compiler_ubo_range {
 256         /**
 257          * offset in bytes from the start of the ubo where this range is
 258          * uploaded.
 259          *
 260          * Only set once used is set.
 261          */
 262         uint32_t dst_offset;
 263
 264         /**
 265          * offset in bytes from the start of the gallium uniforms where the
 266          * data comes from.
 267          */
 268         uint32_t src_offset;
 269
 270         /** size in bytes of this ubo range */
 271         uint32_t size;
 272
 273         /**
 274          * Set if this range is used by the shader for indirect uniforms
 275          * access.
 276          */
 277         bool used;
 278 };
 279
 280 struct vc4_key {
 281         struct vc4_uncompiled_shader *shader_state;
 282         struct {
 283                 enum pipe_format format;
 284                 uint8_t swizzle[4];
 285                 union {
 286                         struct {
 287                                 unsigned compare_mode:1;
 288                                 unsigned compare_func:3;
 289                                 unsigned wrap_s:3;
 290                                 unsigned wrap_t:3;
 291                         };
 292                         struct {
 293                                 uint16_t msaa_width, msaa_height;
 294                         };
 295                 };
 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         bool msaa;
 313         bool sample_coverage;
 314         bool sample_alpha_to_coverage;
 315         bool sample_alpha_to_one;
 316         uint8_t alpha_test_func;
 317         uint8_t logicop_func;
 318         uint32_t point_sprite_mask;
 319
 320         struct pipe_rt_blend_state blend;
 321 };
 322
 323 struct vc4_vs_key {
 324         struct vc4_key base;
 325
 326         /**
 327          * This is a proxy for the array of FS input semantics, which is
 328          * larger than we would want to put in the key.
 329          */
 330         uint64_t compiled_fs_id;
 331
 332         enum pipe_format attr_formats[8];
 333         bool is_coord;
 334         bool per_vertex_point_size;
 335 };
 336
 337 struct vc4_compile {
 338         struct vc4_context *vc4;
 339         nir_shader *s;
 340         nir_function_impl *impl;
 341         struct exec_list *cf_node_list;
 342
 343         /**
 344          * Mapping from nir_register * or nir_ssa_def * to array of struct
 345          * qreg for the values.
 346          */
 347         struct hash_table *def_ht;
 348
 349         /* For each temp, the instruction generating its value. */
 350         struct qinst **defs;
 351         uint32_t defs_array_size;
 352
 353         /**
 354          * Inputs to the shader, arranged by TGSI declaration order.
 355          *
 356          * Not all fragment shader QFILE_VARY reads are present in this array.
 357          */
 358         struct qreg *inputs;
 359         struct qreg *outputs;
 360         bool msaa_per_sample_output;
 361         struct qreg color_reads[VC4_MAX_SAMPLES];
 362         struct qreg sample_colors[VC4_MAX_SAMPLES];
 363         uint32_t inputs_array_size;
 364         uint32_t outputs_array_size;
 365         uint32_t uniforms_array_size;
 366
 367         struct vc4_compiler_ubo_range *ubo_ranges;
 368         uint32_t ubo_ranges_array_size;
 369         /** Number of uniform areas declared in ubo_ranges. */
 370         uint32_t num_uniform_ranges;
 371         /** Number of uniform areas used for indirect addressed loads. */
 372         uint32_t num_ubo_ranges;
 373         uint32_t next_ubo_dst_offset;
 374
 375         struct qreg line_x, point_x, point_y;
 376         struct qreg discard;
 377         struct qreg payload_FRAG_Z;
 378         struct qreg payload_FRAG_W;
 379
 380         uint8_t vattr_sizes[8];
 381
 382         /* Bitfield for whether a given channel of a sampler needs sRGB
 383          * decode.
 384          */
 385         uint8_t tex_srgb_decode[VC4_MAX_TEXTURE_SAMPLERS];
 386
 387         /**
 388          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 389          *
 390          * This includes those that aren't part of the VPM varyings, like
 391          * point/line coordinates.
 392          */
 393         struct vc4_varying_slot *input_slots;
 394         uint32_t num_input_slots;
 395         uint32_t input_slots_array_size;
 396
 397         /**
 398          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 399          * of the output is.  Used to emit from the VS in the order that the
 400          * FS needs.
 401          */
 402         struct vc4_varying_slot *output_slots;
 403
 404         struct pipe_shader_state *shader_state;
 405         struct vc4_key *key;
 406         struct vc4_fs_key *fs_key;
 407         struct vc4_vs_key *vs_key;
 408
 409         uint32_t *uniform_data;
 410         enum quniform_contents *uniform_contents;
 411         uint32_t uniform_array_size;
 412         uint32_t num_uniforms;
 413         uint32_t num_outputs;
 414         uint32_t num_texture_samples;
 415         uint32_t output_position_index;
 416         uint32_t output_color_index;
 417         uint32_t output_point_size_index;
 418         uint32_t output_sample_mask_index;
 419
 420         struct qreg undef;
 421         enum qstage stage;
 422         uint32_t num_temps;
 423         struct list_head instructions;
 424
 425         struct list_head qpu_inst_list;
 426         uint64_t *qpu_insts;
 427         uint32_t qpu_inst_count;
 428         uint32_t qpu_inst_size;
 429         uint32_t num_inputs;
 430
 431         uint32_t program_id;
 432         uint32_t variant_id;
 433 };
 434
 435 /* Special nir_load_input intrinsic index for loading the current TLB
 436  * destination color.
 437  */
 438 #define VC4_NIR_TLB_COLOR_READ_INPUT            2000000000
 439
 440 #define VC4_NIR_MS_MASK_OUTPUT                  2000000000
 441
 442 /* Special offset for nir_load_uniform values to get a QUNIFORM_*
 443  * state-dependent value.
 444  */
 445 #define VC4_NIR_STATE_UNIFORM_OFFSET            1000000000
 446
 447 struct vc4_compile *qir_compile_init(void);
 448 void qir_compile_destroy(struct vc4_compile *c);
 449 struct qinst *qir_inst(enum qop op, struct qreg dst,
 450                        struct qreg src0, struct qreg src1);
 451 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 452                         struct qreg a,
 453                         struct qreg b,
 454                         struct qreg c,
 455                         struct qreg d);
 456 void qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst);
 457 struct qreg qir_uniform(struct vc4_compile *c,
 458                         enum quniform_contents contents,
 459                         uint32_t data);
 460 void qir_schedule_instructions(struct vc4_compile *c);
 461 void qir_reorder_uniforms(struct vc4_compile *c);
 462
 463 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 464 static inline struct qinst *
 465 qir_emit_nodef(struct vc4_compile *c, struct qinst *inst)
 466 {
 467         list_addtail(&inst->link, &c->instructions);
 468         return inst;
 469 }
 470
 471 struct qreg qir_get_temp(struct vc4_compile *c);
 472 int qir_get_op_nsrc(enum qop qop);
 473 bool qir_reg_equals(struct qreg a, struct qreg b);
 474 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 475 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 476 bool qir_is_mul(struct qinst *inst);
 477 bool qir_is_raw_mov(struct qinst *inst);
 478 bool qir_is_tex(struct qinst *inst);
 479 bool qir_is_float_input(struct qinst *inst);
 480 bool qir_depends_on_flags(struct qinst *inst);
 481 bool qir_writes_r4(struct qinst *inst);
 482 struct qreg qir_follow_movs(struct vc4_compile *c, struct qreg reg);
 483
 484 void qir_dump(struct vc4_compile *c);
 485 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 486 const char *qir_get_stage_name(enum qstage stage);
 487
 488 void qir_optimize(struct vc4_compile *c);
 489 bool qir_opt_algebraic(struct vc4_compile *c);
 490 bool qir_opt_constant_folding(struct vc4_compile *c);
 491 bool qir_opt_copy_propagation(struct vc4_compile *c);
 492 bool qir_opt_dead_code(struct vc4_compile *c);
 493 bool qir_opt_small_immediates(struct vc4_compile *c);
 494 bool qir_opt_vpm(struct vc4_compile *c);
 495 void vc4_nir_lower_blend(nir_shader *s, struct vc4_compile *c);
 496 void vc4_nir_lower_io(nir_shader *s, struct vc4_compile *c);
 497 nir_ssa_def *vc4_nir_get_state_uniform(struct nir_builder *b,
 498                                        enum quniform_contents contents);
 499 nir_ssa_def *vc4_nir_get_swizzled_channel(struct nir_builder *b,
 500                                           nir_ssa_def **srcs, int swiz);
 501 void vc4_nir_lower_txf_ms(nir_shader *s, struct vc4_compile *c);
 502 void qir_lower_uniforms(struct vc4_compile *c);
 503
 504 uint32_t qpu_schedule_instructions(struct vc4_compile *c);
 505
 506 void qir_SF(struct vc4_compile *c, struct qreg src);
 507
 508 static inline struct qreg
 509 qir_uniform_ui(struct vc4_compile *c, uint32_t ui)
 510 {
 511         return qir_uniform(c, QUNIFORM_CONSTANT, ui);
 512 }
 513
 514 static inline struct qreg
 515 qir_uniform_f(struct vc4_compile *c, float f)
 516 {
 517         return qir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 518 }
 519
 520 #define QIR_ALU0(name)                                                   \
 521 static inline struct qreg                                                \
 522 qir_##name(struct vc4_compile *c)                                        \
 523 {                                                                        \
 524         struct qreg t = qir_get_temp(c);                                 \
 525         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 526         return t;                                                        \
 527 }
 528
 529 #define QIR_ALU1(name)                                                   \
 530 static inline struct qreg                                                \
 531 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 532 {                                                                        \
 533         struct qreg t = qir_get_temp(c);                                 \
 534         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 535         return t;                                                        \
 536 }                                                                        \
 537 static inline struct qinst *                                             \
 538 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 539                   struct qreg a)                                         \
 540 {                                                                        \
 541         return qir_emit_nodef(c, qir_inst(QOP_##name, dest, a,           \
 542                                           c->undef));                    \
 543 }
 544
 545 #define QIR_ALU2(name)                                                   \
 546 static inline struct qreg                                                \
 547 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 548 {                                                                        \
 549         struct qreg t = qir_get_temp(c);                                 \
 550         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 551         return t;                                                        \
 552 }                                                                        \
 553 static inline void                                                       \
 554 qir_##name##_dest(struct vc4_compile *c, struct qreg dest,               \
 555                   struct qreg a, struct qreg b)                          \
 556 {                                                                        \
 557         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, b));             \
 558 }
 559
 560 #define QIR_NODST_1(name)                                               \
 561 static inline struct qinst *                                            \
 562 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 563 {                                                                       \
 564         struct qinst *inst = qir_inst(QOP_##name, c->undef,             \
 565                                       a, c->undef);                     \
 566         qir_emit(c, inst);                                              \
 567         return inst;                                                    \
 568 }
 569
 570 #define QIR_NODST_2(name)                                               \
 571 static inline struct qinst *                                            \
 572 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 573 {                                                                       \
 574         struct qinst *inst = qir_inst(QOP_##name, c->undef,             \
 575                                       a, b);                            \
 576         qir_emit(c, inst);                                              \
 577         return inst;                                                    \
 578 }
 579
 580 #define QIR_PACK(name)                                                   \
 581 static inline struct qreg                                                \
 582 qir_##name(struct vc4_compile *c, struct qreg dest, struct qreg a)       \
 583 {                                                                        \
 584         qir_emit_nodef(c, qir_inst(QOP_##name, dest, a, c->undef));      \
 585         return dest;                                                     \
 586 }
 587
 588 #define QIR_PAYLOAD(name)                                                \
 589 static inline struct qreg                                                \
 590 qir_##name(struct vc4_compile *c)                                        \
 591 {                                                                        \
 592         struct qreg *payload = &c->payload_##name;                       \
 593         if (payload->file != QFILE_NULL)                                 \
 594                 return *payload;                                         \
 595         *payload = qir_get_temp(c);                                      \
 596         struct qinst *inst = qir_inst(QOP_##name, *payload,              \
 597                                       c->undef, c->undef);               \
 598         list_add(&inst->link, &c->instructions);                         \
 599         c->defs[payload->index] = inst;                                  \
 600         return *payload;                                                 \
 601 }
 602
 603 QIR_ALU1(MOV)
 604 QIR_ALU1(FMOV)
 605 QIR_ALU1(MMOV)
 606 QIR_ALU2(FADD)
 607 QIR_ALU2(FSUB)
 608 QIR_ALU2(FMUL)
 609 QIR_ALU2(V8MULD)
 610 QIR_ALU2(V8MIN)
 611 QIR_ALU2(V8MAX)
 612 QIR_ALU2(V8ADDS)
 613 QIR_ALU2(V8SUBS)
 614 QIR_ALU2(MUL24)
 615 QIR_ALU2(FMIN)
 616 QIR_ALU2(FMAX)
 617 QIR_ALU2(FMINABS)
 618 QIR_ALU2(FMAXABS)
 619 QIR_ALU1(FTOI)
 620 QIR_ALU1(ITOF)
 621
 622 QIR_ALU2(ADD)
 623 QIR_ALU2(SUB)
 624 QIR_ALU2(SHL)
 625 QIR_ALU2(SHR)
 626 QIR_ALU2(ASR)
 627 QIR_ALU2(MIN)
 628 QIR_ALU2(MAX)
 629 QIR_ALU2(AND)
 630 QIR_ALU2(OR)
 631 QIR_ALU2(XOR)
 632 QIR_ALU1(NOT)
 633
 634 QIR_ALU1(RCP)
 635 QIR_ALU1(RSQ)
 636 QIR_ALU1(EXP2)
 637 QIR_ALU1(LOG2)
 638 QIR_ALU1(VARY_ADD_C)
 639 QIR_NODST_2(TEX_S)
 640 QIR_NODST_2(TEX_T)
 641 QIR_NODST_2(TEX_R)
 642 QIR_NODST_2(TEX_B)
 643 QIR_NODST_2(TEX_DIRECT)
 644 QIR_PAYLOAD(FRAG_Z)
 645 QIR_PAYLOAD(FRAG_W)
 646 QIR_ALU0(TEX_RESULT)
 647 QIR_ALU0(TLB_COLOR_READ)
 648 QIR_NODST_1(MS_MASK)
 649
 650 static inline struct qreg
 651 qir_SEL(struct vc4_compile *c, uint8_t cond, struct qreg src0, struct qreg src1)
 652 {
 653         struct qreg t = qir_get_temp(c);
 654         struct qinst *a = qir_MOV_dest(c, t, src0);
 655         struct qinst *b = qir_MOV_dest(c, t, src1);
 656         a->cond = cond;
 657         b->cond = cond ^ 1;
 658         return t;
 659 }
 660
 661 static inline struct qreg
 662 qir_UNPACK_8_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_8A + i;
 666         return t;
 667 }
 668
 669 static inline struct qreg
 670 qir_UNPACK_8_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_8A + i;
 674         return t;
 675 }
 676
 677 static inline struct qreg
 678 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 679 {
 680         struct qreg t = qir_FMOV(c, src);
 681         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 682         return t;
 683 }
 684
 685 static inline struct qreg
 686 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 687 {
 688         struct qreg t = qir_MOV(c, src);
 689         c->defs[t.index]->src[0].pack = QPU_UNPACK_16A + i;
 690         return t;
 691 }
 692
 693 static inline void
 694 qir_PACK_8_F(struct vc4_compile *c, struct qreg dest, struct qreg val, int chan)
 695 {
 696         assert(!dest.pack);
 697         dest.pack = QPU_PACK_MUL_8A + chan;
 698         qir_emit(c, qir_inst(QOP_MMOV, dest, val, c->undef));
 699         if (dest.file == QFILE_TEMP)
 700                 c->defs[dest.index] = NULL;
 701 }
 702
 703 static inline struct qreg
 704 qir_PACK_8888_F(struct vc4_compile *c, struct qreg val)
 705 {
 706         struct qreg dest = qir_MMOV(c, val);
 707         c->defs[dest.index]->dst.pack = QPU_PACK_MUL_8888;
 708         return dest;
 709 }
 710
 711 static inline struct qreg
 712 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 713 {
 714         return qir_EXP2(c, qir_FMUL(c,
 715                                     y,
 716                                     qir_LOG2(c, x)));
 717 }
 718
 719 static inline void
 720 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 721 {
 722         qir_MOV_dest(c, qir_reg(QFILE_VPM, 0), val);
 723 }
 724
 725 #endif /* VC4_QIR_H */