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/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 queued_qpu_inst {
 166         struct simple_node link;
 167         uint64_t inst;
 168 };
 169
 170 struct qinst {
 171         struct simple_node link;
 172
 173         enum qop op;
 174         struct qreg dst;
 175         struct qreg *src;
 176 };
 177
 178 enum qstage {
 179         /**
 180          * Coordinate shader, runs during binning, before the VS, and just
 181          * outputs position.
 182          */
 183         QSTAGE_COORD,
 184         QSTAGE_VERT,
 185         QSTAGE_FRAG,
 186 };
 187
 188 enum quniform_contents {
 189         /**
 190          * Indicates that a constant 32-bit value is copied from the program's
 191          * uniform contents.
 192          */
 193         QUNIFORM_CONSTANT,
 194         /**
 195          * Indicates that the program's uniform contents are used as an index
 196          * into the GL uniform storage.
 197          */
 198         QUNIFORM_UNIFORM,
 199
 200         /** @{
 201          * Scaling factors from clip coordinates to relative to the viewport
 202          * center.
 203          *
 204          * This is used by the coordinate and vertex shaders to produce the
 205          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 206          * point offsets from the viewport ccenter.
 207          */
 208         QUNIFORM_VIEWPORT_X_SCALE,
 209         QUNIFORM_VIEWPORT_Y_SCALE,
 210         /** @} */
 211
 212         QUNIFORM_VIEWPORT_Z_OFFSET,
 213         QUNIFORM_VIEWPORT_Z_SCALE,
 214
 215         QUNIFORM_USER_CLIP_PLANE,
 216
 217         /**
 218          * A reference to a texture config parameter 0 uniform.
 219          *
 220          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 221          * defines texture type, miplevels, and such.  It will be found as a
 222          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 223          */
 224         QUNIFORM_TEXTURE_CONFIG_P0,
 225
 226         /**
 227          * A reference to a texture config parameter 1 uniform.
 228          *
 229          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 230          * defines texture width, height, filters, and wrap modes.  It will be
 231          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 232          * sequence.
 233          */
 234         QUNIFORM_TEXTURE_CONFIG_P1,
 235
 236         /** A reference to a texture config parameter 2 cubemap stride uniform */
 237         QUNIFORM_TEXTURE_CONFIG_P2,
 238
 239         QUNIFORM_UBO_ADDR,
 240
 241         QUNIFORM_TEXRECT_SCALE_X,
 242         QUNIFORM_TEXRECT_SCALE_Y,
 243
 244         QUNIFORM_TEXTURE_BORDER_COLOR,
 245
 246         QUNIFORM_BLEND_CONST_COLOR,
 247         QUNIFORM_STENCIL,
 248
 249         QUNIFORM_ALPHA_REF,
 250 };
 251
 252 struct vc4_varying_semantic {
 253         uint8_t semantic;
 254         uint8_t index;
 255         uint8_t swizzle;
 256 };
 257
 258 struct vc4_compiler_ubo_range {
 259         /**
 260          * offset in bytes from the start of the ubo where this range is
 261          * uploaded.
 262          *
 263          * Only set once used is set.
 264          */
 265         uint32_t dst_offset;
 266
 267         /**
 268          * offset in bytes from the start of the gallium uniforms where the
 269          * data comes from.
 270          */
 271         uint32_t src_offset;
 272
 273         /** size in bytes of this ubo range */
 274         uint32_t size;
 275
 276         /**
 277          * Set if this range is used by the shader for indirect uniforms
 278          * access.
 279          */
 280         bool used;
 281 };
 282
 283 struct vc4_compile {
 284         struct vc4_context *vc4;
 285         struct tgsi_parse_context parser;
 286         struct qreg *temps;
 287         /**
 288          * Inputs to the shader, arranged by TGSI declaration order.
 289          *
 290          * Not all fragment shader QFILE_VARY reads are present in this array.
 291          */
 292         struct qreg *inputs;
 293         struct qreg *outputs;
 294         struct qreg *consts;
 295         struct qreg addr[4]; /* TGSI ARL destination. */
 296         uint32_t temps_array_size;
 297         uint32_t inputs_array_size;
 298         uint32_t outputs_array_size;
 299         uint32_t uniforms_array_size;
 300         uint32_t consts_array_size;
 301         uint32_t num_consts;
 302
 303         struct vc4_compiler_ubo_range *ubo_ranges;
 304         uint32_t ubo_ranges_array_size;
 305         uint32_t num_ubo_ranges;
 306         uint32_t next_ubo_dst_offset;
 307
 308         struct qreg line_x, point_x, point_y;
 309         struct qreg discard;
 310
 311         uint8_t vattr_sizes[8];
 312
 313         /**
 314          * Array of the TGSI semantics of all FS QFILE_VARY reads.
 315          *
 316          * This includes those that aren't part of the VPM varyings, like
 317          * point/line coordinates.
 318          */
 319         struct vc4_varying_semantic *input_semantics;
 320         uint32_t num_input_semantics;
 321         uint32_t input_semantics_array_size;
 322
 323         /**
 324          * An entry per outputs[] in the VS indicating what the semantic of
 325          * the output is.  Used to emit from the VS in the order that the FS
 326          * needs.
 327          */
 328         struct vc4_varying_semantic *output_semantics;
 329
 330         struct pipe_shader_state *shader_state;
 331         struct vc4_key *key;
 332         struct vc4_fs_key *fs_key;
 333         struct vc4_vs_key *vs_key;
 334
 335         uint32_t *uniform_data;
 336         enum quniform_contents *uniform_contents;
 337         uint32_t uniform_array_size;
 338         uint32_t num_uniforms;
 339         uint32_t num_outputs;
 340         uint32_t num_texture_samples;
 341         uint32_t output_position_index;
 342         uint32_t output_clipvertex_index;
 343         uint32_t output_color_index;
 344         uint32_t output_point_size_index;
 345
 346         struct qreg undef;
 347         enum qstage stage;
 348         uint32_t num_temps;
 349         struct simple_node instructions;
 350         uint32_t immediates[1024];
 351
 352         struct simple_node qpu_inst_list;
 353         uint64_t *qpu_insts;
 354         uint32_t qpu_inst_count;
 355         uint32_t qpu_inst_size;
 356         uint32_t num_inputs;
 357
 358         uint32_t program_id;
 359         uint32_t variant_id;
 360 };
 361
 362 struct vc4_compile *qir_compile_init(void);
 363 void qir_compile_destroy(struct vc4_compile *c);
 364 struct qinst *qir_inst(enum qop op, struct qreg dst,
 365                        struct qreg src0, struct qreg src1);
 366 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 367                         struct qreg a,
 368                         struct qreg b,
 369                         struct qreg c,
 370                         struct qreg d);
 371 void qir_remove_instruction(struct qinst *qinst);
 372 void qir_reorder_uniforms(struct vc4_compile *c);
 373 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 374 struct qreg qir_get_temp(struct vc4_compile *c);
 375 int qir_get_op_nsrc(enum qop qop);
 376 bool qir_reg_equals(struct qreg a, struct qreg b);
 377 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 378 bool qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst);
 379 bool qir_is_multi_instruction(struct qinst *inst);
 380 bool qir_depends_on_flags(struct qinst *inst);
 381 bool qir_writes_r4(struct qinst *inst);
 382 bool qir_reads_r4(struct qinst *inst);
 383 bool qir_src_needs_a_file(struct qinst *inst);
 384 struct qreg qir_follow_movs(struct qinst **defs, struct qreg reg);
 385
 386 void qir_dump(struct vc4_compile *c);
 387 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 388 const char *qir_get_stage_name(enum qstage stage);
 389
 390 void qir_optimize(struct vc4_compile *c);
 391 bool qir_opt_algebraic(struct vc4_compile *c);
 392 bool qir_opt_copy_propagation(struct vc4_compile *c);
 393 bool qir_opt_cse(struct vc4_compile *c);
 394 bool qir_opt_dead_code(struct vc4_compile *c);
 395 bool qir_opt_small_immediates(struct vc4_compile *c);
 396 bool qir_opt_vpm_writes(struct vc4_compile *c);
 397
 398 void qpu_schedule_instructions(struct vc4_compile *c);
 399
 400 #define QIR_ALU0(name)                                                   \
 401 static inline struct qreg                                                \
 402 qir_##name(struct vc4_compile *c)                                        \
 403 {                                                                        \
 404         struct qreg t = qir_get_temp(c);                                 \
 405         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 406         return t;                                                        \
 407 }
 408
 409 #define QIR_ALU1(name)                                                   \
 410 static inline struct qreg                                                \
 411 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 412 {                                                                        \
 413         struct qreg t = qir_get_temp(c);                                 \
 414         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 415         return t;                                                        \
 416 }
 417
 418 #define QIR_ALU2(name)                                                   \
 419 static inline struct qreg                                                \
 420 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 421 {                                                                        \
 422         struct qreg t = qir_get_temp(c);                                 \
 423         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 424         return t;                                                        \
 425 }
 426
 427 #define QIR_NODST_1(name)                                               \
 428 static inline void                                                      \
 429 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 430 {                                                                       \
 431         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 432 }
 433
 434 #define QIR_NODST_2(name)                                               \
 435 static inline void                                                      \
 436 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 437 {                                                                       \
 438         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 439 }
 440
 441 QIR_ALU1(MOV)
 442 QIR_ALU2(FADD)
 443 QIR_ALU2(FSUB)
 444 QIR_ALU2(FMUL)
 445 QIR_ALU2(MUL24)
 446 QIR_NODST_1(SF)
 447 QIR_ALU1(SEL_X_0_ZS)
 448 QIR_ALU1(SEL_X_0_ZC)
 449 QIR_ALU1(SEL_X_0_NS)
 450 QIR_ALU1(SEL_X_0_NC)
 451 QIR_ALU2(SEL_X_Y_ZS)
 452 QIR_ALU2(SEL_X_Y_ZC)
 453 QIR_ALU2(SEL_X_Y_NS)
 454 QIR_ALU2(SEL_X_Y_NC)
 455 QIR_ALU2(FMIN)
 456 QIR_ALU2(FMAX)
 457 QIR_ALU2(FMINABS)
 458 QIR_ALU2(FMAXABS)
 459 QIR_ALU1(FTOI)
 460 QIR_ALU1(ITOF)
 461
 462 QIR_ALU2(ADD)
 463 QIR_ALU2(SUB)
 464 QIR_ALU2(SHL)
 465 QIR_ALU2(SHR)
 466 QIR_ALU2(ASR)
 467 QIR_ALU2(MIN)
 468 QIR_ALU2(MAX)
 469 QIR_ALU2(AND)
 470 QIR_ALU2(OR)
 471 QIR_ALU2(XOR)
 472 QIR_ALU1(NOT)
 473
 474 QIR_ALU1(RCP)
 475 QIR_ALU1(RSQ)
 476 QIR_ALU1(EXP2)
 477 QIR_ALU1(LOG2)
 478 QIR_ALU2(PACK_SCALED)
 479 QIR_ALU1(PACK_8888_F)
 480 QIR_ALU2(PACK_8A_F)
 481 QIR_ALU2(PACK_8B_F)
 482 QIR_ALU2(PACK_8C_F)
 483 QIR_ALU2(PACK_8D_F)
 484 QIR_ALU1(VARY_ADD_C)
 485 QIR_NODST_2(TEX_S)
 486 QIR_NODST_2(TEX_T)
 487 QIR_NODST_2(TEX_R)
 488 QIR_NODST_2(TEX_B)
 489 QIR_NODST_2(TEX_DIRECT)
 490 QIR_ALU0(FRAG_X)
 491 QIR_ALU0(FRAG_Y)
 492 QIR_ALU0(FRAG_Z)
 493 QIR_ALU0(FRAG_W)
 494 QIR_ALU0(FRAG_REV_FLAG)
 495 QIR_ALU0(TEX_RESULT)
 496 QIR_ALU0(TLB_COLOR_READ)
 497 QIR_NODST_1(TLB_Z_WRITE)
 498 QIR_NODST_1(TLB_DISCARD_SETUP)
 499 QIR_NODST_1(TLB_STENCIL_SETUP)
 500
 501 static inline struct qreg
 502 qir_R4_UNPACK(struct vc4_compile *c, struct qreg r4, int i)
 503 {
 504         struct qreg t = qir_get_temp(c);
 505         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, r4, c->undef));
 506         return t;
 507 }
 508
 509 static inline struct qreg
 510 qir_SEL_X_0_COND(struct vc4_compile *c, int i)
 511 {
 512         struct qreg t = qir_get_temp(c);
 513         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, c->undef, c->undef));
 514         return t;
 515 }
 516
 517 static inline struct qreg
 518 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 519 {
 520         struct qreg t = qir_get_temp(c);
 521         qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
 522         return t;
 523 }
 524
 525 static inline struct qreg
 526 qir_UNPACK_8_I(struct vc4_compile *c, struct qreg src, int i)
 527 {
 528         struct qreg t = qir_get_temp(c);
 529         qir_emit(c, qir_inst(QOP_UNPACK_8A_I + i, t, src, c->undef));
 530         return t;
 531 }
 532
 533 static inline struct qreg
 534 qir_UNPACK_16_F(struct vc4_compile *c, struct qreg src, int i)
 535 {
 536         struct qreg t = qir_get_temp(c);
 537         qir_emit(c, qir_inst(QOP_UNPACK_16A_F + i, t, src, c->undef));
 538         return t;
 539 }
 540
 541 static inline struct qreg
 542 qir_UNPACK_16_I(struct vc4_compile *c, struct qreg src, int i)
 543 {
 544         struct qreg t = qir_get_temp(c);
 545         qir_emit(c, qir_inst(QOP_UNPACK_16A_I + i, t, src, c->undef));
 546         return t;
 547 }
 548
 549 static inline struct qreg
 550 qir_PACK_8_F(struct vc4_compile *c, struct qreg rest, struct qreg val, int chan)
 551 {
 552         struct qreg t = qir_get_temp(c);
 553         qir_emit(c, qir_inst(QOP_PACK_8A_F + chan, t, rest, val));
 554         return t;
 555 }
 556
 557 static inline struct qreg
 558 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 559 {
 560         return qir_EXP2(c, qir_FMUL(c,
 561                                     y,
 562                                     qir_LOG2(c, x)));
 563 }
 564
 565 static inline void
 566 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 567 {
 568         static const struct qreg vpm = { QFILE_VPM, 0 };
 569         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 570 }
 571
 572 #endif /* VC4_QIR_H */