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 <stdio.h>
  28 #include <stdlib.h>
  29 #include <stdbool.h>
  30 #include <stdint.h>
  31 #include <string.h>
  32
  33 #include "util/macros.h"
  34 #include "util/u_simple_list.h"
  35 #include "tgsi/tgsi_parse.h"
  36
  37 enum qfile {
  38         QFILE_NULL,
  39         QFILE_TEMP,
  40         QFILE_VARY,
  41         QFILE_UNIF,
  42         QFILE_VPM,
  43
  44         /**
  45          * Stores an immediate value in the index field that can be turned
  46          * into a small immediate field by qpu_encode_small_immediate().
  47          */
  48         QFILE_SMALL_IMM,
  49 };
  50
  51 struct qreg {
  52         enum qfile file;
  53         uint32_t index;
  54 };
  55
  56 enum qop {
  57         QOP_UNDEF,
  58         QOP_MOV,
  59         QOP_FADD,
  60         QOP_FSUB,
  61         QOP_FMUL,
  62         QOP_MUL24,
  63         QOP_FMIN,
  64         QOP_FMAX,
  65         QOP_FMINABS,
  66         QOP_FMAXABS,
  67         QOP_ADD,
  68         QOP_SUB,
  69         QOP_SHL,
  70         QOP_SHR,
  71         QOP_ASR,
  72         QOP_MIN,
  73         QOP_MAX,
  74         QOP_AND,
  75         QOP_OR,
  76         QOP_XOR,
  77         QOP_NOT,
  78
  79         /* Sets the flag register according to src. */
  80         QOP_SF,
  81
  82         /* Note: Orderings of these compares must be the same as in
  83          * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
  84          * otherwise 0. */
  85         QOP_SEL_X_0_ZS,
  86         QOP_SEL_X_0_ZC,
  87         QOP_SEL_X_0_NS,
  88         QOP_SEL_X_0_NC,
  89         /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
  90         QOP_SEL_X_Y_ZS,
  91         QOP_SEL_X_Y_ZC,
  92         QOP_SEL_X_Y_NS,
  93         QOP_SEL_X_Y_NC,
  94
  95         QOP_FTOI,
  96         QOP_ITOF,
  97         QOP_RCP,
  98         QOP_RSQ,
  99         QOP_EXP2,
 100         QOP_LOG2,
 101         QOP_VW_SETUP,
 102         QOP_VR_SETUP,
 103         QOP_PACK_SCALED,
 104         QOP_PACK_8888_F,
 105         QOP_PACK_8A_F,
 106         QOP_PACK_8B_F,
 107         QOP_PACK_8C_F,
 108         QOP_PACK_8D_F,
 109         QOP_TLB_DISCARD_SETUP,
 110         QOP_TLB_STENCIL_SETUP,
 111         QOP_TLB_Z_WRITE,
 112         QOP_TLB_COLOR_WRITE,
 113         QOP_TLB_COLOR_READ,
 114         QOP_VARY_ADD_C,
 115
 116         QOP_FRAG_X,
 117         QOP_FRAG_Y,
 118         QOP_FRAG_Z,
 119         QOP_FRAG_W,
 120         QOP_FRAG_REV_FLAG,
 121
 122         QOP_UNPACK_8A_F,
 123         QOP_UNPACK_8B_F,
 124         QOP_UNPACK_8C_F,
 125         QOP_UNPACK_8D_F,
 126         QOP_UNPACK_16A_F,
 127         QOP_UNPACK_16B_F,
 128
 129         QOP_UNPACK_8A_I,
 130         QOP_UNPACK_8B_I,
 131         QOP_UNPACK_8C_I,
 132         QOP_UNPACK_8D_I,
 133         QOP_UNPACK_16A_I,
 134         QOP_UNPACK_16B_I,
 135
 136         /** Texture x coordinate parameter write */
 137         QOP_TEX_S,
 138         /** Texture y coordinate parameter write */
 139         QOP_TEX_T,
 140         /** Texture border color parameter or cube map z coordinate write */
 141         QOP_TEX_R,
 142         /** Texture LOD bias parameter write */
 143         QOP_TEX_B,
 144
 145         /**
 146          * Texture-unit 4-byte read with address provided direct in S
 147          * cooordinate.
 148          *
 149          * The first operand is the offset from the start of the UBO, and the
 150          * second is the uniform that has the UBO's base pointer.
 151          */
 152         QOP_TEX_DIRECT,
 153
 154         /**
 155          * Signal of texture read being necessary and then reading r4 into
 156          * the destination
 157          */
 158         QOP_TEX_RESULT,
 159         QOP_R4_UNPACK_A,
 160         QOP_R4_UNPACK_B,
 161         QOP_R4_UNPACK_C,
 162         QOP_R4_UNPACK_D
 163 };
 164
 165 struct simple_node {
 166         struct simple_node *next;
 167         struct simple_node *prev;
 168 };
 169
 170 struct queued_qpu_inst {
 171         struct simple_node link;
 172         uint64_t inst;
 173 };
 174
 175 struct qinst {
 176         struct simple_node link;
 177
 178         enum qop op;
 179         struct qreg dst;
 180         struct qreg *src;
 181 };
 182
 183 enum qstage {
 184         /**
 185          * Coordinate shader, runs during binning, before the VS, and just
 186          * outputs position.
 187          */
 188         QSTAGE_COORD,
 189         QSTAGE_VERT,
 190         QSTAGE_FRAG,
 191 };
 192
 193 enum quniform_contents {
 194         /**
 195          * Indicates that a constant 32-bit value is copied from the program's
 196          * uniform contents.
 197          */
 198         QUNIFORM_CONSTANT,
 199         /**
 200          * Indicates that the program's uniform contents are used as an index
 201          * into the GL uniform storage.
 202          */
 203         QUNIFORM_UNIFORM,
 204
 205         /** @{
 206          * Scaling factors from clip coordinates to relative to the viewport
 207          * center.
 208          *
 209          * This is used by the coordinate and vertex shaders to produce the
 210          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 211          * point offsets from the viewport ccenter.
 212          */
 213         QUNIFORM_VIEWPORT_X_SCALE,
 214         QUNIFORM_VIEWPORT_Y_SCALE,
 215         /** @} */
 216
 217         QUNIFORM_VIEWPORT_Z_OFFSET,
 218         QUNIFORM_VIEWPORT_Z_SCALE,
 219
 220         QUNIFORM_USER_CLIP_PLANE,
 221
 222         /**
 223          * A reference to a texture config parameter 0 uniform.
 224          *
 225          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 226          * defines texture type, miplevels, and such.  It will be found as a
 227          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 228          */
 229         QUNIFORM_TEXTURE_CONFIG_P0,
 230
 231         /**
 232          * A reference to a texture config parameter 1 uniform.
 233          *
 234          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 235          * defines texture width, height, filters, and wrap modes.  It will be
 236          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 237          * sequence.
 238          */
 239         QUNIFORM_TEXTURE_CONFIG_P1,
 240
 241         /** A reference to a texture config parameter 2 cubemap stride uniform */
 242         QUNIFORM_TEXTURE_CONFIG_P2,
 243
 244         QUNIFORM_UBO_ADDR,
 245
 246         QUNIFORM_TEXRECT_SCALE_X,
 247         QUNIFORM_TEXRECT_SCALE_Y,
 248
 249         QUNIFORM_TEXTURE_BORDER_COLOR,
 250
 251         QUNIFORM_BLEND_CONST_COLOR,
 252         QUNIFORM_STENCIL,
 253
 254         QUNIFORM_ALPHA_REF,
 255 };
 256
 257 struct vc4_varying_semantic {
 258         uint8_t semantic;
 259         uint8_t index;
 260         uint8_t swizzle;
 261 };
 262
 263 struct vc4_compiler_ubo_range {
 264         /**
 265          * offset in bytes from the start of the ubo where this range is
 266          * uploaded.
 267          *
 268          * Only set once used is set.
 269          */
 270         uint32_t dst_offset;
 271
 272         /**
 273          * offset in bytes from the start of the gallium uniforms where the
 274          * data comes from.
 275          */
 276         uint32_t src_offset;
 277
 278         /** size in bytes of this ubo range */
 279         uint32_t size;
 280
 281         /**
 282          * Set if this range is used by the shader for indirect uniforms
 283          * access.
 284          */
 285         bool used;
 286 };
 287
 288 struct vc4_compile {
 289         struct vc4_context *vc4;
 290         struct tgsi_parse_context parser;
 291         struct qreg *temps;
 292         /**
 293          * Inputs to the shader, arranged by TGSI declaration order.
 294          *
 295          * Not all fragment shader QFILE_VARY reads are present in this array.
 296          */
 297         struct qreg *inputs;
 298         struct qreg *outputs;
 299         struct qreg *consts;
 300         struct qreg addr[4]; /* TGSI ARL destination. */
 301         uint32_t temps_array_size;
 302         uint32_t inputs_array_size;
 303         uint32_t outputs_array_size;
 304         uint32_t uniforms_array_size;
 305         uint32_t consts_array_size;
 306         uint32_t num_consts;
 307
 308         struct vc4_compiler_ubo_range *ubo_ranges;
 309         uint32_t ubo_ranges_array_size;
 310         uint32_t num_ubo_ranges;
 311         uint32_t next_ubo_dst_offset;
 312
 313         struct qreg line_x, point_x, point_y;
 314         struct qreg discard;
 315
 316         uint8_t vattr_sizes[8];
 317
 318         /**
 319          * Array of the TGSI semantics of all FS QFILE_VARY reads.
 320          *
 321          * This includes those that aren't part of the VPM varyings, like
 322          * point/line coordinates.
 323          */
 324         struct vc4_varying_semantic *input_semantics;
 325         uint32_t num_input_semantics;
 326         uint32_t input_semantics_array_size;
 327
 328         /**
 329          * An entry per outputs[] in the VS indicating what the semantic of
 330          * the output is.  Used to emit from the VS in the order that the FS
 331          * needs.
 332          */
 333         struct vc4_varying_semantic *output_semantics;
 334
 335         struct pipe_shader_state *shader_state;
 336         struct vc4_key *key;
 337         struct vc4_fs_key *fs_key;
 338         struct vc4_vs_key *vs_key;
 339
 340         uint32_t *uniform_data;
 341         enum quniform_contents *uniform_contents;
 342         uint32_t uniform_array_size;
 343         uint32_t num_uniforms;
 344         uint32_t num_outputs;
 345         uint32_t num_texture_samples;
 346         uint32_t output_position_index;
 347         uint32_t output_clipvertex_index;
 348         uint32_t output_color_index;
 349         uint32_t output_point_size_index;
 350
 351         struct qreg undef;
 352         enum qstage stage;
 353         uint32_t num_temps;
 354         struct simple_node instructions;
 355         uint32_t immediates[1024];
 356
 357         struct simple_node qpu_inst_list;
 358         uint64_t *qpu_insts;
 359         uint32_t qpu_inst_count;
 360         uint32_t qpu_inst_size;
 361         uint32_t num_inputs;
 362
 363         uint32_t program_id;
 364         uint32_t variant_id;
 365 };
 366
 367 struct vc4_compile *qir_compile_init(void);
 368 void qir_compile_destroy(struct vc4_compile *c);
 369 struct qinst *qir_inst(enum qop op, struct qreg dst,
 370                        struct qreg src0, struct qreg src1);
 371 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 372                         struct qreg a,
 373                         struct qreg b,
 374                         struct qreg c,
 375                         struct qreg d);
 376 void qir_remove_instruction(struct qinst *qinst);
 377 void qir_reorder_uniforms(struct vc4_compile *c);
 378 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 379 struct qreg qir_get_temp(struct vc4_compile *c);
 380 int qir_get_op_nsrc(enum qop qop);
 381 bool qir_reg_equals(struct qreg a, struct qreg b);
 382 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 383 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 384 bool qir_is_multi_instruction(struct qinst *inst);
 385 bool qir_depends_on_flags(struct qinst *inst);
 386 bool qir_writes_r4(struct qinst *inst);
 387 bool qir_reads_r4(struct qinst *inst);
 388 struct qreg qir_follow_movs(struct qinst **defs, struct qreg reg);
 389
 390 void qir_dump(struct vc4_compile *c);
 391 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 392 const char *qir_get_stage_name(enum qstage stage);
 393
 394 void qir_optimize(struct vc4_compile *c);
 395 bool qir_opt_algebraic(struct vc4_compile *c);
 396 bool qir_opt_copy_propagation(struct vc4_compile *c);
 397 bool qir_opt_cse(struct vc4_compile *c);
 398 bool qir_opt_dead_code(struct vc4_compile *c);
 399 bool qir_opt_small_immediates(struct vc4_compile *c);
 400 bool qir_opt_vpm_writes(struct vc4_compile *c);
 401
 402 void qpu_schedule_instructions(struct vc4_compile *c);
 403
 404 #define QIR_ALU0(name)                                                   \
 405 static inline struct qreg                                                \
 406 qir_##name(struct vc4_compile *c)                                        \
 407 {                                                                        \
 408         struct qreg t = qir_get_temp(c);                                 \
 409         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 410         return t;                                                        \
 411 }
 412
 413 #define QIR_ALU1(name)                                                   \
 414 static inline struct qreg                                                \
 415 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 416 {                                                                        \
 417         struct qreg t = qir_get_temp(c);                                 \
 418         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 419         return t;                                                        \
 420 }
 421
 422 #define QIR_ALU2(name)                                                   \
 423 static inline struct qreg                                                \
 424 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 425 {                                                                        \
 426         struct qreg t = qir_get_temp(c);                                 \
 427         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 428         return t;                                                        \
 429 }
 430
 431 #define QIR_NODST_1(name)                                               \
 432 static inline void                                                      \
 433 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 434 {                                                                       \
 435         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 436 }
 437
 438 #define QIR_NODST_2(name)                                               \
 439 static inline void                                                      \
 440 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 441 {                                                                       \
 442         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 443 }
 444
 445 QIR_ALU1(MOV)
 446 QIR_ALU2(FADD)
 447 QIR_ALU2(FSUB)
 448 QIR_ALU2(FMUL)
 449 QIR_ALU2(MUL24)
 450 QIR_NODST_1(SF)
 451 QIR_ALU1(SEL_X_0_ZS)
 452 QIR_ALU1(SEL_X_0_ZC)
 453 QIR_ALU1(SEL_X_0_NS)
 454 QIR_ALU1(SEL_X_0_NC)
 455 QIR_ALU2(SEL_X_Y_ZS)
 456 QIR_ALU2(SEL_X_Y_ZC)
 457 QIR_ALU2(SEL_X_Y_NS)
 458 QIR_ALU2(SEL_X_Y_NC)
 459 QIR_ALU2(FMIN)
 460 QIR_ALU2(FMAX)
 461 QIR_ALU2(FMINABS)
 462 QIR_ALU2(FMAXABS)
 463 QIR_ALU1(FTOI)
 464 QIR_ALU1(ITOF)
 465
 466 QIR_ALU2(ADD)
 467 QIR_ALU2(SUB)
 468 QIR_ALU2(SHL)
 469 QIR_ALU2(SHR)
 470 QIR_ALU2(ASR)
 471 QIR_ALU2(MIN)
 472 QIR_ALU2(MAX)
 473 QIR_ALU2(AND)
 474 QIR_ALU2(OR)
 475 QIR_ALU2(XOR)
 476 QIR_ALU1(NOT)
 477
 478 QIR_ALU1(RCP)
 479 QIR_ALU1(RSQ)
 480 QIR_ALU1(EXP2)
 481 QIR_ALU1(LOG2)
 482 QIR_ALU2(PACK_SCALED)
 483 QIR_ALU1(PACK_8888_F)
 484 QIR_ALU2(PACK_8A_F)
 485 QIR_ALU2(PACK_8B_F)
 486 QIR_ALU2(PACK_8C_F)
 487 QIR_ALU2(PACK_8D_F)
 488 QIR_ALU1(VARY_ADD_C)
 489 QIR_NODST_2(TEX_S)
 490 QIR_NODST_2(TEX_T)
 491 QIR_NODST_2(TEX_R)
 492 QIR_NODST_2(TEX_B)
 493 QIR_NODST_2(TEX_DIRECT)
 494 QIR_ALU0(FRAG_X)
 495 QIR_ALU0(FRAG_Y)
 496 QIR_ALU0(FRAG_Z)
 497 QIR_ALU0(FRAG_W)
 498 QIR_ALU0(FRAG_REV_FLAG)
 499 QIR_ALU0(TEX_RESULT)
 500 QIR_ALU0(TLB_COLOR_READ)
 501 QIR_NODST_1(TLB_Z_WRITE)
 502 QIR_NODST_1(TLB_DISCARD_SETUP)
 503 QIR_NODST_1(TLB_STENCIL_SETUP)
 504
 505 static inline struct qreg
 506 qir_R4_UNPACK(struct vc4_compile *c, struct qreg r4, int i)
 507 {
 508         struct qreg t = qir_get_temp(c);
 509         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, r4, c->undef));
 510         return t;
 511 }
 512
 513 static inline struct qreg
 514 qir_SEL_X_0_COND(struct vc4_compile *c, int i)
 515 {
 516         struct qreg t = qir_get_temp(c);
 517         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, c->undef, c->undef));
 518         return t;
 519 }
 520
 521 static inline struct qreg
 522 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 523 {
 524         struct qreg t = qir_get_temp(c);
 525         qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
 526         return t;
 527 }
 528
 529 static inline struct qreg
 530 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 531 {
 532         struct qreg t = qir_get_temp(c);
 533         qir_emit(c, qir_inst(QOP_UNPACK_8A_I + i, t, src, c->undef));
 534         return t;
 535 }
 536
 537 static inline struct qreg
 538 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 539 {
 540         struct qreg t = qir_get_temp(c);
 541         qir_emit(c, qir_inst(QOP_UNPACK_16A_F + i, t, src, c->undef));
 542         return t;
 543 }
 544
 545 static inline struct qreg
 546 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 547 {
 548         struct qreg t = qir_get_temp(c);
 549         qir_emit(c, qir_inst(QOP_UNPACK_16A_I + i, t, src, c->undef));
 550         return t;
 551 }
 552
 553 static inline struct qreg
 554 qir_PACK_8_F(struct vc4_compile *c, struct qreg rest, struct qreg val, int chan)
 555 {
 556         struct qreg t = qir_get_temp(c);
 557         qir_emit(c, qir_inst(QOP_PACK_8A_F + chan, t, rest, val));
 558         return t;
 559 }
 560
 561 static inline struct qreg
 562 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 563 {
 564         return qir_EXP2(c, qir_FMUL(c,
 565                                     y,
 566                                     qir_LOG2(c, x)));
 567 }
 568
 569 static inline void
 570 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 571 {
 572         static const struct qreg vpm = { QFILE_VPM, 0 };
 573         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 574 }
 575
 576 #endif /* VC4_QIR_H */