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