src/broadcom/compiler/v3d_compiler.h

   1 /*
   2  * Copyright © 2016 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 V3D_COMPILER_H
  25 #define V3D_COMPILER_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 "common/v3d_debug.h"
  36 #include "common/v3d_device_info.h"
  37 #include "compiler/nir/nir.h"
  38 #include "util/list.h"
  39 #include "util/u_math.h"
  40
  41 #include "qpu/qpu_instr.h"
  42 #include "pipe/p_state.h"
  43
  44 #define V3D_MAX_TEXTURE_SAMPLERS 32
  45 #define V3D_MAX_SAMPLES 4
  46 #define V3D_MAX_FS_INPUTS 64
  47 #define V3D_MAX_VS_INPUTS 64
  48
  49 struct nir_builder;
  50
  51 struct v3d_fs_inputs {
  52         /**
  53          * Array of the meanings of the VPM inputs this shader needs.
  54          *
  55          * It doesn't include those that aren't part of the VPM, like
  56          * point/line coordinates.
  57          */
  58         struct v3d_varying_slot *input_slots;
  59         uint32_t num_inputs;
  60 };
  61
  62 enum qfile {
  63         /** An unused source or destination register. */
  64         QFILE_NULL,
  65
  66         /** A physical register, such as the W coordinate payload. */
  67         QFILE_REG,
  68         /** One of the regsiters for fixed function interactions. */
  69         QFILE_MAGIC,
  70
  71         /**
  72          *  A virtual register, that will be allocated to actual accumulator
  73          * or physical registers later.
  74          */
  75         QFILE_TEMP,
  76         QFILE_UNIF,
  77         QFILE_TLB,
  78         QFILE_TLBU,
  79
  80         /**
  81          * VPM reads use this with an index value to say what part of the VPM
  82          * is being read.
  83          */
  84         QFILE_VPM,
  85
  86         /**
  87          * Stores an immediate value in the index field that will be used
  88          * directly by qpu_load_imm().
  89          */
  90         QFILE_LOAD_IMM,
  91
  92         /**
  93          * Stores an immediate value in the index field that can be turned
  94          * into a small immediate field by qpu_encode_small_immediate().
  95          */
  96         QFILE_SMALL_IMM,
  97 };
  98
  99 /**
 100  * A reference to a QPU register or a virtual temp register.
 101  */
 102 struct qreg {
 103         enum qfile file;
 104         uint32_t index;
 105 };
 106
 107 static inline struct qreg vir_reg(enum qfile file, uint32_t index)
 108 {
 109         return (struct qreg){file, index};
 110 }
 111
 112 /**
 113  * A reference to an actual register at the QPU level, for register
 114  * allocation.
 115  */
 116 struct qpu_reg {
 117         bool magic;
 118         int index;
 119 };
 120
 121 struct qinst {
 122         /** Entry in qblock->instructions */
 123         struct list_head link;
 124
 125         /**
 126          * The instruction being wrapped.  Its condition codes, pack flags,
 127          * signals, etc. will all be used, with just the register references
 128          * being replaced by the contents of qinst->dst and qinst->src[].
 129          */
 130         struct v3d_qpu_instr qpu;
 131
 132         /* Pre-register-allocation references to src/dst registers */
 133         struct qreg dst;
 134         struct qreg src[3];
 135         bool cond_is_exec_mask;
 136         bool has_implicit_uniform;
 137         bool is_last_thrsw;
 138
 139         /* After vir_to_qpu.c: If instr reads a uniform, which uniform from
 140          * the uncompiled stream it is.
 141          */
 142         int uniform;
 143 };
 144
 145 enum quniform_contents {
 146         /**
 147          * Indicates that a constant 32-bit value is copied from the program's
 148          * uniform contents.
 149          */
 150         QUNIFORM_CONSTANT,
 151         /**
 152          * Indicates that the program's uniform contents are used as an index
 153          * into the GL uniform storage.
 154          */
 155         QUNIFORM_UNIFORM,
 156
 157         /** @{
 158          * Scaling factors from clip coordinates to relative to the viewport
 159          * center.
 160          *
 161          * This is used by the coordinate and vertex shaders to produce the
 162          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 163          * point offsets from the viewport ccenter.
 164          */
 165         QUNIFORM_VIEWPORT_X_SCALE,
 166         QUNIFORM_VIEWPORT_Y_SCALE,
 167         /** @} */
 168
 169         QUNIFORM_VIEWPORT_Z_OFFSET,
 170         QUNIFORM_VIEWPORT_Z_SCALE,
 171
 172         QUNIFORM_USER_CLIP_PLANE,
 173
 174         /**
 175          * A reference to a V3D 3.x texture config parameter 0 uniform.
 176          *
 177          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 178          * defines texture type, miplevels, and such.  It will be found as a
 179          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 180          */
 181         QUNIFORM_TEXTURE_CONFIG_P0_0,
 182         QUNIFORM_TEXTURE_CONFIG_P0_1,
 183         QUNIFORM_TEXTURE_CONFIG_P0_2,
 184         QUNIFORM_TEXTURE_CONFIG_P0_3,
 185         QUNIFORM_TEXTURE_CONFIG_P0_4,
 186         QUNIFORM_TEXTURE_CONFIG_P0_5,
 187         QUNIFORM_TEXTURE_CONFIG_P0_6,
 188         QUNIFORM_TEXTURE_CONFIG_P0_7,
 189         QUNIFORM_TEXTURE_CONFIG_P0_8,
 190         QUNIFORM_TEXTURE_CONFIG_P0_9,
 191         QUNIFORM_TEXTURE_CONFIG_P0_10,
 192         QUNIFORM_TEXTURE_CONFIG_P0_11,
 193         QUNIFORM_TEXTURE_CONFIG_P0_12,
 194         QUNIFORM_TEXTURE_CONFIG_P0_13,
 195         QUNIFORM_TEXTURE_CONFIG_P0_14,
 196         QUNIFORM_TEXTURE_CONFIG_P0_15,
 197         QUNIFORM_TEXTURE_CONFIG_P0_16,
 198         QUNIFORM_TEXTURE_CONFIG_P0_17,
 199         QUNIFORM_TEXTURE_CONFIG_P0_18,
 200         QUNIFORM_TEXTURE_CONFIG_P0_19,
 201         QUNIFORM_TEXTURE_CONFIG_P0_20,
 202         QUNIFORM_TEXTURE_CONFIG_P0_21,
 203         QUNIFORM_TEXTURE_CONFIG_P0_22,
 204         QUNIFORM_TEXTURE_CONFIG_P0_23,
 205         QUNIFORM_TEXTURE_CONFIG_P0_24,
 206         QUNIFORM_TEXTURE_CONFIG_P0_25,
 207         QUNIFORM_TEXTURE_CONFIG_P0_26,
 208         QUNIFORM_TEXTURE_CONFIG_P0_27,
 209         QUNIFORM_TEXTURE_CONFIG_P0_28,
 210         QUNIFORM_TEXTURE_CONFIG_P0_29,
 211         QUNIFORM_TEXTURE_CONFIG_P0_30,
 212         QUNIFORM_TEXTURE_CONFIG_P0_31,
 213         QUNIFORM_TEXTURE_CONFIG_P0_32,
 214
 215         /**
 216          * A reference to a V3D 3.x texture config parameter 1 uniform.
 217          *
 218          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 219          * has the pointer to the indirect texture state.  Our data[] field
 220          * will have a packed p1 value, but the address field will be just
 221          * which texture unit's texture should be referenced.
 222          */
 223         QUNIFORM_TEXTURE_CONFIG_P1,
 224
 225         /* A a V3D 4.x texture config parameter.  The high 8 bits will be
 226          * which texture or sampler is being sampled, and the driver must
 227          * replace the address field with the appropriate address.
 228          */
 229         QUNIFORM_TMU_CONFIG_P0,
 230         QUNIFORM_TMU_CONFIG_P1,
 231
 232         QUNIFORM_TEXTURE_FIRST_LEVEL,
 233
 234         QUNIFORM_TEXTURE_WIDTH,
 235         QUNIFORM_TEXTURE_HEIGHT,
 236         QUNIFORM_TEXTURE_DEPTH,
 237         QUNIFORM_TEXTURE_ARRAY_SIZE,
 238         QUNIFORM_TEXTURE_LEVELS,
 239
 240         QUNIFORM_UBO_ADDR,
 241
 242         QUNIFORM_TEXRECT_SCALE_X,
 243         QUNIFORM_TEXRECT_SCALE_Y,
 244
 245         QUNIFORM_TEXTURE_BORDER_COLOR,
 246
 247         QUNIFORM_STENCIL,
 248
 249         QUNIFORM_ALPHA_REF,
 250         QUNIFORM_SAMPLE_MASK,
 251
 252         /**
 253          * Returns the the offset of the scratch buffer for register spilling.
 254          */
 255         QUNIFORM_SPILL_OFFSET,
 256         QUNIFORM_SPILL_SIZE_PER_THREAD,
 257 };
 258
 259 struct v3d_varying_slot {
 260         uint8_t slot_and_component;
 261 };
 262
 263 static inline struct v3d_varying_slot
 264 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
 265 {
 266         assert(slot < 255 / 4);
 267         return (struct v3d_varying_slot){ (slot << 2) + component };
 268 }
 269
 270 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
 271 {
 272         return slot.slot_and_component >> 2;
 273 }
 274
 275 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
 276 {
 277         return slot.slot_and_component & 3;
 278 }
 279
 280 struct v3d_ubo_range {
 281         /**
 282          * offset in bytes from the start of the ubo where this range is
 283          * uploaded.
 284          *
 285          * Only set once used is set.
 286          */
 287         uint32_t dst_offset;
 288
 289         /**
 290          * offset in bytes from the start of the gallium uniforms where the
 291          * data comes from.
 292          */
 293         uint32_t src_offset;
 294
 295         /** size in bytes of this ubo range */
 296         uint32_t size;
 297 };
 298
 299 struct v3d_key {
 300         void *shader_state;
 301         struct {
 302                 uint8_t swizzle[4];
 303                 uint8_t return_size;
 304                 uint8_t return_channels;
 305                 union {
 306                         struct {
 307                                 unsigned compare_mode:1;
 308                                 unsigned compare_func:3;
 309                                 bool clamp_s:1;
 310                                 bool clamp_t:1;
 311                                 bool clamp_r:1;
 312                         };
 313                         struct {
 314                                 uint16_t msaa_width, msaa_height;
 315                         };
 316                 };
 317         } tex[V3D_MAX_TEXTURE_SAMPLERS];
 318         uint8_t ucp_enables;
 319 };
 320
 321 struct v3d_fs_key {
 322         struct v3d_key base;
 323         bool depth_enabled;
 324         bool is_points;
 325         bool is_lines;
 326         bool alpha_test;
 327         bool point_coord_upper_left;
 328         bool light_twoside;
 329         bool msaa;
 330         bool sample_coverage;
 331         bool sample_alpha_to_coverage;
 332         bool sample_alpha_to_one;
 333         bool clamp_color;
 334         bool shade_model_flat;
 335         uint8_t nr_cbufs;
 336         uint8_t swap_color_rb;
 337         /* Mask of which render targets need to be written as 32-bit floats */
 338         uint8_t f32_color_rb;
 339         uint8_t alpha_test_func;
 340         uint8_t logicop_func;
 341         uint32_t point_sprite_mask;
 342
 343         struct pipe_rt_blend_state blend;
 344 };
 345
 346 struct v3d_vs_key {
 347         struct v3d_key base;
 348
 349         struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
 350         uint8_t num_fs_inputs;
 351
 352         bool is_coord;
 353         bool per_vertex_point_size;
 354         bool clamp_color;
 355 };
 356
 357 /** A basic block of VIR intructions. */
 358 struct qblock {
 359         struct list_head link;
 360
 361         struct list_head instructions;
 362
 363         struct set *predecessors;
 364         struct qblock *successors[2];
 365
 366         int index;
 367
 368         /* Instruction IPs for the first and last instruction of the block.
 369          * Set by qpu_schedule.c.
 370          */
 371         uint32_t start_qpu_ip;
 372         uint32_t end_qpu_ip;
 373
 374         /* Instruction IP for the branch instruction of the block.  Set by
 375          * qpu_schedule.c.
 376          */
 377         uint32_t branch_qpu_ip;
 378
 379         /** Offset within the uniform stream at the start of the block. */
 380         uint32_t start_uniform;
 381         /** Offset within the uniform stream of the branch instruction */
 382         uint32_t branch_uniform;
 383
 384         /** @{ used by v3d_vir_live_variables.c */
 385         BITSET_WORD *def;
 386         BITSET_WORD *use;
 387         BITSET_WORD *live_in;
 388         BITSET_WORD *live_out;
 389         int start_ip, end_ip;
 390         /** @} */
 391 };
 392
 393 /** Which util/list.h add mode we should use when inserting an instruction. */
 394 enum vir_cursor_mode {
 395         vir_cursor_add,
 396         vir_cursor_addtail,
 397 };
 398
 399 /**
 400  * Tracking structure for where new instructions should be inserted.  Create
 401  * with one of the vir_after_inst()-style helper functions.
 402  *
 403  * This does not protect against removal of the block or instruction, so we
 404  * have an assert in instruction removal to try to catch it.
 405  */
 406 struct vir_cursor {
 407         enum vir_cursor_mode mode;
 408         struct list_head *link;
 409 };
 410
 411 static inline struct vir_cursor
 412 vir_before_inst(struct qinst *inst)
 413 {
 414         return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
 415 }
 416
 417 static inline struct vir_cursor
 418 vir_after_inst(struct qinst *inst)
 419 {
 420         return (struct vir_cursor){ vir_cursor_add, &inst->link };
 421 }
 422
 423 static inline struct vir_cursor
 424 vir_before_block(struct qblock *block)
 425 {
 426         return (struct vir_cursor){ vir_cursor_add, &block->instructions };
 427 }
 428
 429 static inline struct vir_cursor
 430 vir_after_block(struct qblock *block)
 431 {
 432         return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
 433 }
 434
 435 /**
 436  * Compiler state saved across compiler invocations, for any expensive global
 437  * setup.
 438  */
 439 struct v3d_compiler {
 440         const struct v3d_device_info *devinfo;
 441         struct ra_regs *regs;
 442         unsigned int reg_class_phys[3];
 443         unsigned int reg_class_phys_or_acc[3];
 444 };
 445
 446 struct v3d_compile {
 447         const struct v3d_device_info *devinfo;
 448         nir_shader *s;
 449         nir_function_impl *impl;
 450         struct exec_list *cf_node_list;
 451         const struct v3d_compiler *compiler;
 452
 453         /**
 454          * Mapping from nir_register * or nir_ssa_def * to array of struct
 455          * qreg for the values.
 456          */
 457         struct hash_table *def_ht;
 458
 459         /* For each temp, the instruction generating its value. */
 460         struct qinst **defs;
 461         uint32_t defs_array_size;
 462
 463         /**
 464          * Inputs to the shader, arranged by TGSI declaration order.
 465          *
 466          * Not all fragment shader QFILE_VARY reads are present in this array.
 467          */
 468         struct qreg *inputs;
 469         struct qreg *outputs;
 470         bool msaa_per_sample_output;
 471         struct qreg color_reads[V3D_MAX_SAMPLES];
 472         struct qreg sample_colors[V3D_MAX_SAMPLES];
 473         uint32_t inputs_array_size;
 474         uint32_t outputs_array_size;
 475         uint32_t uniforms_array_size;
 476
 477         /* Booleans for whether the corresponding QFILE_VARY[i] is
 478          * flat-shaded.  This includes gl_FragColor flat-shading, which is
 479          * customized based on the shademodel_flat shader key.
 480          */
 481         uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
 482
 483         struct v3d_ubo_range *ubo_ranges;
 484         bool *ubo_range_used;
 485         uint32_t ubo_ranges_array_size;
 486         /** Number of uniform areas tracked in ubo_ranges. */
 487         uint32_t num_ubo_ranges;
 488         uint32_t next_ubo_dst_offset;
 489
 490         /* State for whether we're executing on each channel currently.  0 if
 491          * yes, otherwise a block number + 1 that the channel jumped to.
 492          */
 493         struct qreg execute;
 494
 495         struct qreg line_x, point_x, point_y;
 496
 497         /**
 498          * Instance ID, which comes in before the vertex attribute payload if
 499          * the shader record requests it.
 500          */
 501         struct qreg iid;
 502
 503         /**
 504          * Vertex ID, which comes in before the vertex attribute payload
 505          * (after Instance ID) if the shader record requests it.
 506          */
 507         struct qreg vid;
 508
 509         /* Fragment shader payload regs. */
 510         struct qreg payload_w, payload_w_centroid, payload_z;
 511
 512         uint8_t vattr_sizes[V3D_MAX_VS_INPUTS];
 513         uint32_t num_vpm_writes;
 514
 515         /* Size in bytes of registers that have been spilled. This is how much
 516          * space needs to be available in the spill BO per thread per QPU.
 517          */
 518         uint32_t spill_size;
 519         /* Shader-db stats for register spilling. */
 520         uint32_t spills, fills;
 521         /**
 522          * Register spilling's per-thread base address, shared between each
 523          * spill/fill's addressing calculations.
 524          */
 525         struct qreg spill_base;
 526         /* Bit vector of which temps may be spilled */
 527         BITSET_WORD *spillable;
 528
 529         /**
 530          * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
 531          *
 532          * This includes those that aren't part of the VPM varyings, like
 533          * point/line coordinates.
 534          */
 535         struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
 536
 537         /**
 538          * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
 539          * of the output is.  Used to emit from the VS in the order that the
 540          * FS needs.
 541          */
 542         struct v3d_varying_slot *output_slots;
 543
 544         struct pipe_shader_state *shader_state;
 545         struct v3d_key *key;
 546         struct v3d_fs_key *fs_key;
 547         struct v3d_vs_key *vs_key;
 548
 549         /* Live ranges of temps. */
 550         int *temp_start, *temp_end;
 551
 552         uint32_t *uniform_data;
 553         enum quniform_contents *uniform_contents;
 554         uint32_t uniform_array_size;
 555         uint32_t num_uniforms;
 556         uint32_t num_outputs;
 557         uint32_t output_position_index;
 558         nir_variable *output_color_var[4];
 559         uint32_t output_point_size_index;
 560         uint32_t output_sample_mask_index;
 561
 562         struct qreg undef;
 563         uint32_t num_temps;
 564
 565         struct vir_cursor cursor;
 566         struct list_head blocks;
 567         int next_block_index;
 568         struct qblock *cur_block;
 569         struct qblock *loop_cont_block;
 570         struct qblock *loop_break_block;
 571
 572         uint64_t *qpu_insts;
 573         uint32_t qpu_inst_count;
 574         uint32_t qpu_inst_size;
 575
 576         /* For the FS, the number of varying inputs not counting the
 577          * point/line varyings payload
 578          */
 579         uint32_t num_inputs;
 580
 581         /**
 582          * Number of inputs from num_inputs remaining to be queued to the read
 583          * FIFO in the VS/CS.
 584          */
 585         uint32_t num_inputs_remaining;
 586
 587         /* Number of inputs currently in the read FIFO for the VS/CS */
 588         uint32_t num_inputs_in_fifo;
 589
 590         /** Next offset in the VPM to read from in the VS/CS */
 591         uint32_t vpm_read_offset;
 592
 593         uint32_t program_id;
 594         uint32_t variant_id;
 595
 596         /* Set to compile program in in 1x, 2x, or 4x threaded mode, where
 597          * SIG_THREAD_SWITCH is used to hide texturing latency at the cost of
 598          * limiting ourselves to the part of the physical reg space.
 599          *
 600          * On V3D 3.x, 2x or 4x divide the physical reg space by 2x or 4x.  On
 601          * V3D 4.x, all shaders are 2x threaded, and 4x only divides the
 602          * physical reg space in half.
 603          */
 604         uint8_t threads;
 605         struct qinst *last_thrsw;
 606         bool last_thrsw_at_top_level;
 607
 608         bool failed;
 609 };
 610
 611 struct v3d_uniform_list {
 612         enum quniform_contents *contents;
 613         uint32_t *data;
 614         uint32_t count;
 615 };
 616
 617 struct v3d_prog_data {
 618         struct v3d_uniform_list uniforms;
 619
 620         struct v3d_ubo_range *ubo_ranges;
 621         uint32_t num_ubo_ranges;
 622         uint32_t ubo_size;
 623         uint32_t spill_size;
 624
 625         uint8_t num_inputs;
 626         uint8_t threads;
 627
 628         /* For threads > 1, whether the program should be dispatched in the
 629          * after-final-THRSW state.
 630          */
 631         bool single_seg;
 632 };
 633
 634 struct v3d_vs_prog_data {
 635         struct v3d_prog_data base;
 636
 637         bool uses_iid, uses_vid;
 638
 639         /* Number of components read from each vertex attribute. */
 640         uint8_t vattr_sizes[32];
 641
 642         /* Total number of components read, for the shader state record. */
 643         uint32_t vpm_input_size;
 644
 645         /* Total number of components written, for the shader state record. */
 646         uint32_t vpm_output_size;
 647 };
 648
 649 struct v3d_fs_prog_data {
 650         struct v3d_prog_data base;
 651
 652         struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
 653
 654         /* Array of flat shade flags.
 655          *
 656          * Each entry is only 24 bits (high 8 bits 0), to match the hardware
 657          * packet layout.
 658          */
 659         uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
 660
 661         bool writes_z;
 662         bool discard;
 663 };
 664
 665 /* Special nir_load_input intrinsic index for loading the current TLB
 666  * destination color.
 667  */
 668 #define V3D_NIR_TLB_COLOR_READ_INPUT            2000000000
 669
 670 #define V3D_NIR_MS_MASK_OUTPUT                  2000000000
 671
 672 extern const nir_shader_compiler_options v3d_nir_options;
 673
 674 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
 675 void v3d_compiler_free(const struct v3d_compiler *compiler);
 676 void v3d_optimize_nir(struct nir_shader *s);
 677
 678 uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
 679                          struct v3d_vs_key *key,
 680                          struct v3d_vs_prog_data *prog_data,
 681                          nir_shader *s,
 682                          int program_id, int variant_id,
 683                          uint32_t *final_assembly_size);
 684
 685 uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
 686                          struct v3d_fs_key *key,
 687                          struct v3d_fs_prog_data *prog_data,
 688                          nir_shader *s,
 689                          int program_id, int variant_id,
 690                          uint32_t *final_assembly_size);
 691
 692 void v3d_nir_to_vir(struct v3d_compile *c);
 693
 694 void vir_compile_destroy(struct v3d_compile *c);
 695 const char *vir_get_stage_name(struct v3d_compile *c);
 696 struct qblock *vir_new_block(struct v3d_compile *c);
 697 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
 698 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
 699 struct qblock *vir_entry_block(struct v3d_compile *c);
 700 struct qblock *vir_exit_block(struct v3d_compile *c);
 701 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
 702                            struct qreg src0, struct qreg src1);
 703 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
 704                            struct qreg src0, struct qreg src1);
 705 struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
 706 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
 707 struct qreg vir_uniform(struct v3d_compile *c,
 708                         enum quniform_contents contents,
 709                         uint32_t data);
 710 void vir_schedule_instructions(struct v3d_compile *c);
 711 struct v3d_qpu_instr v3d_qpu_nop(void);
 712
 713 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
 714 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
 715 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
 716 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
 717 void vir_set_unpack(struct qinst *inst, int src,
 718                     enum v3d_qpu_input_unpack unpack);
 719
 720 struct qreg vir_get_temp(struct v3d_compile *c);
 721 void vir_emit_last_thrsw(struct v3d_compile *c);
 722 void vir_calculate_live_intervals(struct v3d_compile *c);
 723 bool vir_has_implicit_uniform(struct qinst *inst);
 724 int vir_get_implicit_uniform_src(struct qinst *inst);
 725 int vir_get_non_sideband_nsrc(struct qinst *inst);
 726 int vir_get_nsrc(struct qinst *inst);
 727 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
 728 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
 729 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
 730 bool vir_is_raw_mov(struct qinst *inst);
 731 bool vir_is_tex(struct qinst *inst);
 732 bool vir_is_add(struct qinst *inst);
 733 bool vir_is_mul(struct qinst *inst);
 734 bool vir_is_float_input(struct qinst *inst);
 735 bool vir_depends_on_flags(struct qinst *inst);
 736 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
 737 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
 738 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
 739 uint8_t vir_channels_written(struct qinst *inst);
 740 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
 741 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
 742                     struct qreg result);
 743 void vir_emit_thrsw(struct v3d_compile *c);
 744
 745 void vir_dump(struct v3d_compile *c);
 746 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
 747
 748 void vir_validate(struct v3d_compile *c);
 749
 750 void vir_optimize(struct v3d_compile *c);
 751 bool vir_opt_algebraic(struct v3d_compile *c);
 752 bool vir_opt_constant_folding(struct v3d_compile *c);
 753 bool vir_opt_copy_propagate(struct v3d_compile *c);
 754 bool vir_opt_dead_code(struct v3d_compile *c);
 755 bool vir_opt_peephole_sf(struct v3d_compile *c);
 756 bool vir_opt_small_immediates(struct v3d_compile *c);
 757 bool vir_opt_vpm(struct v3d_compile *c);
 758 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
 759 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
 760 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
 761 void vir_lower_uniforms(struct v3d_compile *c);
 762
 763 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
 764 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
 765 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
 766 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
 767
 768 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
 769 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
 770 void qpu_validate(struct v3d_compile *c);
 771 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
 772 bool vir_init_reg_sets(struct v3d_compiler *compiler);
 773
 774 void vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf);
 775
 776 static inline bool
 777 quniform_contents_is_texture_p0(enum quniform_contents contents)
 778 {
 779         return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
 780                 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
 781                             V3D_MAX_TEXTURE_SAMPLERS));
 782 }
 783
 784 static inline struct qreg
 785 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
 786 {
 787         return vir_uniform(c, QUNIFORM_CONSTANT, ui);
 788 }
 789
 790 static inline struct qreg
 791 vir_uniform_f(struct v3d_compile *c, float f)
 792 {
 793         return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
 794 }
 795
 796 #define VIR_ALU0(name, vir_inst, op)                                     \
 797 static inline struct qreg                                                \
 798 vir_##name(struct v3d_compile *c)                                        \
 799 {                                                                        \
 800         return vir_emit_def(c, vir_inst(op, c->undef,                    \
 801                                         c->undef, c->undef));            \
 802 }                                                                        \
 803 static inline struct qinst *                                             \
 804 vir_##name##_dest(struct v3d_compile *c, struct qreg dest)               \
 805 {                                                                        \
 806         return vir_emit_nondef(c, vir_inst(op, dest,                     \
 807                                            c->undef, c->undef));         \
 808 }
 809
 810 #define VIR_ALU1(name, vir_inst, op)                                     \
 811 static inline struct qreg                                                \
 812 vir_##name(struct v3d_compile *c, struct qreg a)                         \
 813 {                                                                        \
 814         return vir_emit_def(c, vir_inst(op, c->undef,                    \
 815                                         a, c->undef));                   \
 816 }                                                                        \
 817 static inline struct qinst *                                             \
 818 vir_##name##_dest(struct v3d_compile *c, struct qreg dest,               \
 819                   struct qreg a)                                         \
 820 {                                                                        \
 821         return vir_emit_nondef(c, vir_inst(op, dest, a,          \
 822                                            c->undef));                   \
 823 }
 824
 825 #define VIR_ALU2(name, vir_inst, op)                                       \
 826 static inline struct qreg                                                \
 827 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b)          \
 828 {                                                                        \
 829         return vir_emit_def(c, vir_inst(op, c->undef, a, b));    \
 830 }                                                                        \
 831 static inline struct qinst *                                             \
 832 vir_##name##_dest(struct v3d_compile *c, struct qreg dest,               \
 833                   struct qreg a, struct qreg b)                          \
 834 {                                                                        \
 835         return vir_emit_nondef(c, vir_inst(op, dest, a, b));     \
 836 }
 837
 838 #define VIR_NODST_0(name, vir_inst, op)                                 \
 839 static inline struct qinst *                                            \
 840 vir_##name(struct v3d_compile *c)                                       \
 841 {                                                                       \
 842         return vir_emit_nondef(c, vir_inst(op, c->undef,                \
 843                                            c->undef, c->undef));        \
 844 }
 845
 846 #define VIR_NODST_1(name, vir_inst, op)                                               \
 847 static inline struct qinst *                                            \
 848 vir_##name(struct v3d_compile *c, struct qreg a)                        \
 849 {                                                                       \
 850         return vir_emit_nondef(c, vir_inst(op, c->undef,        \
 851                                            a, c->undef));               \
 852 }
 853
 854 #define VIR_NODST_2(name, vir_inst, op)                                               \
 855 static inline struct qinst *                                            \
 856 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b)         \
 857 {                                                                       \
 858         return vir_emit_nondef(c, vir_inst(op, c->undef,                \
 859                                            a, b));                      \
 860 }
 861
 862 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
 863 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
 864 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
 865 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
 866 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
 867 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
 868 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
 869 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
 870 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
 871 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
 872 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
 873
 874 VIR_A_ALU2(FADD)
 875 VIR_A_ALU2(VFPACK)
 876 VIR_A_ALU2(FSUB)
 877 VIR_A_ALU2(FMIN)
 878 VIR_A_ALU2(FMAX)
 879
 880 VIR_A_ALU2(ADD)
 881 VIR_A_ALU2(SUB)
 882 VIR_A_ALU2(SHL)
 883 VIR_A_ALU2(SHR)
 884 VIR_A_ALU2(ASR)
 885 VIR_A_ALU2(ROR)
 886 VIR_A_ALU2(MIN)
 887 VIR_A_ALU2(MAX)
 888 VIR_A_ALU2(UMIN)
 889 VIR_A_ALU2(UMAX)
 890 VIR_A_ALU2(AND)
 891 VIR_A_ALU2(OR)
 892 VIR_A_ALU2(XOR)
 893 VIR_A_ALU2(VADD)
 894 VIR_A_ALU2(VSUB)
 895 VIR_A_ALU2(STVPMV)
 896 VIR_A_ALU1(NOT)
 897 VIR_A_ALU1(NEG)
 898 VIR_A_ALU1(FLAPUSH)
 899 VIR_A_ALU1(FLBPUSH)
 900 VIR_A_ALU1(FLBPOP)
 901 VIR_A_ALU1(SETMSF)
 902 VIR_A_ALU1(SETREVF)
 903 VIR_A_ALU0(TIDX)
 904 VIR_A_ALU0(EIDX)
 905 VIR_A_ALU1(LDVPMV_IN)
 906 VIR_A_ALU1(LDVPMV_OUT)
 907
 908 VIR_A_ALU0(FXCD)
 909 VIR_A_ALU0(XCD)
 910 VIR_A_ALU0(FYCD)
 911 VIR_A_ALU0(YCD)
 912 VIR_A_ALU0(MSF)
 913 VIR_A_ALU0(REVF)
 914 VIR_A_NODST_1(VPMSETUP)
 915 VIR_A_NODST_0(VPMWT)
 916 VIR_A_ALU2(FCMP)
 917 VIR_A_ALU2(VFMAX)
 918
 919 VIR_A_ALU1(FROUND)
 920 VIR_A_ALU1(FTOIN)
 921 VIR_A_ALU1(FTRUNC)
 922 VIR_A_ALU1(FTOIZ)
 923 VIR_A_ALU1(FFLOOR)
 924 VIR_A_ALU1(FTOUZ)
 925 VIR_A_ALU1(FCEIL)
 926 VIR_A_ALU1(FTOC)
 927
 928 VIR_A_ALU1(FDX)
 929 VIR_A_ALU1(FDY)
 930
 931 VIR_A_ALU1(ITOF)
 932 VIR_A_ALU1(CLZ)
 933 VIR_A_ALU1(UTOF)
 934
 935 VIR_M_ALU2(UMUL24)
 936 VIR_M_ALU2(FMUL)
 937 VIR_M_ALU2(SMUL24)
 938 VIR_M_NODST_2(MULTOP)
 939
 940 VIR_M_ALU1(MOV)
 941 VIR_M_ALU1(FMOV)
 942
 943 static inline struct qinst *
 944 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
 945              struct qreg dest, struct qreg src)
 946 {
 947         struct qinst *mov = vir_MOV_dest(c, dest, src);
 948         vir_set_cond(mov, cond);
 949         return mov;
 950 }
 951
 952 static inline struct qreg
 953 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
 954         struct qreg src0, struct qreg src1)
 955 {
 956         struct qreg t = vir_get_temp(c);
 957         vir_MOV_dest(c, t, src1);
 958         vir_MOV_cond(c, cond, t, src0);
 959         return t;
 960 }
 961
 962 static inline struct qinst *
 963 vir_NOP(struct v3d_compile *c)
 964 {
 965         return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
 966                                                c->undef, c->undef, c->undef));
 967 }
 968
 969 static inline struct qreg
 970 vir_LDTMU(struct v3d_compile *c)
 971 {
 972         if (c->devinfo->ver >= 41) {
 973                 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
 974                                                    c->undef, c->undef);
 975                 ldtmu->qpu.sig.ldtmu = true;
 976
 977                 return vir_emit_def(c, ldtmu);
 978         } else {
 979                 vir_NOP(c)->qpu.sig.ldtmu = true;
 980                 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
 981         }
 982 }
 983
 984 static inline struct qreg
 985 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
 986 {
 987         vir_MULTOP(c, src0, src1);
 988         return vir_UMUL24(c, src0, src1);
 989 }
 990
 991 /*
 992 static inline struct qreg
 993 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
 994 {
 995         return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
 996                                         vir_reg(QFILE_LOAD_IMM, val), c->undef));
 997 }
 998
 999 static inline struct qreg
1000 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1001 {
1002         return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1003                                         vir_reg(QFILE_LOAD_IMM, val),
1004                                         c->undef));
1005 }
1006 static inline struct qreg
1007 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1008 {
1009         return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1010                                         vir_reg(QFILE_LOAD_IMM, val),
1011                                         c->undef));
1012 }
1013 */
1014
1015 static inline struct qinst *
1016 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_cond cond)
1017 {
1018         /* The actual uniform_data value will be set at scheduling time */
1019         return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
1020 }
1021
1022 #define vir_for_each_block(block, c)                                    \
1023         list_for_each_entry(struct qblock, block, &c->blocks, link)
1024
1025 #define vir_for_each_block_rev(block, c)                                \
1026         list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1027
1028 /* Loop over the non-NULL members of the successors array. */
1029 #define vir_for_each_successor(succ, block)                             \
1030         for (struct qblock *succ = block->successors[0];                \
1031              succ != NULL;                                              \
1032              succ = (succ == block->successors[1] ? NULL :              \
1033                      block->successors[1]))
1034
1035 #define vir_for_each_inst(inst, block)                                  \
1036         list_for_each_entry(struct qinst, inst, &block->instructions, link)
1037
1038 #define vir_for_each_inst_rev(inst, block)                                  \
1039         list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1040
1041 #define vir_for_each_inst_safe(inst, block)                             \
1042         list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1043
1044 #define vir_for_each_inst_inorder(inst, c)                              \
1045         vir_for_each_block(_block, c)                                   \
1046                 vir_for_each_inst(inst, _block)
1047
1048 #endif /* V3D_COMPILER_H */