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/u_simple_list.h"
  34 #include "tgsi/tgsi_parse.h"
  35
  36 enum qfile {
  37         QFILE_NULL,
  38         QFILE_TEMP,
  39         QFILE_VARY,
  40         QFILE_UNIF,
  41         QFILE_VPM,
  42
  43         /**
  44          * Stores an immediate value in the index field that can be turned
  45          * into a small immediate field by qpu_encode_small_immediate().
  46          */
  47         QFILE_SMALL_IMM,
  48 };
  49
  50 struct qreg {
  51         enum qfile file;
  52         uint32_t index;
  53 };
  54
  55 enum qop {
  56         QOP_UNDEF,
  57         QOP_MOV,
  58         QOP_FADD,
  59         QOP_FSUB,
  60         QOP_FMUL,
  61         QOP_MUL24,
  62         QOP_FMIN,
  63         QOP_FMAX,
  64         QOP_FMINABS,
  65         QOP_FMAXABS,
  66         QOP_ADD,
  67         QOP_SUB,
  68         QOP_SHL,
  69         QOP_SHR,
  70         QOP_ASR,
  71         QOP_MIN,
  72         QOP_MAX,
  73         QOP_AND,
  74         QOP_OR,
  75         QOP_XOR,
  76         QOP_NOT,
  77
  78         /* Sets the flag register according to src. */
  79         QOP_SF,
  80
  81         /* Note: Orderings of these compares must be the same as in
  82          * qpu_defines.h.  Selects the src[0] if the ns flag bit is set,
  83          * otherwise 0. */
  84         QOP_SEL_X_0_ZS,
  85         QOP_SEL_X_0_ZC,
  86         QOP_SEL_X_0_NS,
  87         QOP_SEL_X_0_NC,
  88         /* Selects the src[0] if the ns flag bit is set, otherwise src[1]. */
  89         QOP_SEL_X_Y_ZS,
  90         QOP_SEL_X_Y_ZC,
  91         QOP_SEL_X_Y_NS,
  92         QOP_SEL_X_Y_NC,
  93
  94         QOP_FTOI,
  95         QOP_ITOF,
  96         QOP_RCP,
  97         QOP_RSQ,
  98         QOP_EXP2,
  99         QOP_LOG2,
 100         QOP_VW_SETUP,
 101         QOP_VR_SETUP,
 102         QOP_PACK_SCALED,
 103         QOP_PACK_COLORS,
 104         QOP_VPM_READ,
 105         QOP_TLB_DISCARD_SETUP,
 106         QOP_TLB_STENCIL_SETUP,
 107         QOP_TLB_Z_WRITE,
 108         QOP_TLB_COLOR_WRITE,
 109         QOP_TLB_COLOR_READ,
 110         QOP_VARY_ADD_C,
 111
 112         QOP_FRAG_X,
 113         QOP_FRAG_Y,
 114         QOP_FRAG_Z,
 115         QOP_FRAG_W,
 116         QOP_FRAG_REV_FLAG,
 117
 118         QOP_UNPACK_8A_F,
 119         QOP_UNPACK_8B_F,
 120         QOP_UNPACK_8C_F,
 121         QOP_UNPACK_8D_F,
 122         QOP_UNPACK_16A_F,
 123         QOP_UNPACK_16B_F,
 124
 125         QOP_UNPACK_8A_I,
 126         QOP_UNPACK_8B_I,
 127         QOP_UNPACK_8C_I,
 128         QOP_UNPACK_8D_I,
 129         QOP_UNPACK_16A_I,
 130         QOP_UNPACK_16B_I,
 131
 132         /** Texture x coordinate parameter write */
 133         QOP_TEX_S,
 134         /** Texture y coordinate parameter write */
 135         QOP_TEX_T,
 136         /** Texture border color parameter or cube map z coordinate write */
 137         QOP_TEX_R,
 138         /** Texture LOD bias parameter write */
 139         QOP_TEX_B,
 140
 141         /**
 142          * Texture-unit 4-byte read with address provided direct in S
 143          * cooordinate.
 144          *
 145          * The first operand is the offset from the start of the UBO, and the
 146          * second is the uniform that has the UBO's base pointer.
 147          */
 148         QOP_TEX_DIRECT,
 149
 150         /**
 151          * Signal of texture read being necessary and then reading r4 into
 152          * the destination
 153          */
 154         QOP_TEX_RESULT,
 155         QOP_R4_UNPACK_A,
 156         QOP_R4_UNPACK_B,
 157         QOP_R4_UNPACK_C,
 158         QOP_R4_UNPACK_D
 159 };
 160
 161 struct simple_node {
 162         struct simple_node *next;
 163         struct simple_node *prev;
 164 };
 165
 166 struct queued_qpu_inst {
 167         struct simple_node link;
 168         uint64_t inst;
 169 };
 170
 171 struct qinst {
 172         struct simple_node link;
 173
 174         enum qop op;
 175         struct qreg dst;
 176         struct qreg *src;
 177 };
 178
 179 enum qstage {
 180         /**
 181          * Coordinate shader, runs during binning, before the VS, and just
 182          * outputs position.
 183          */
 184         QSTAGE_COORD,
 185         QSTAGE_VERT,
 186         QSTAGE_FRAG,
 187 };
 188
 189 enum quniform_contents {
 190         /**
 191          * Indicates that a constant 32-bit value is copied from the program's
 192          * uniform contents.
 193          */
 194         QUNIFORM_CONSTANT,
 195         /**
 196          * Indicates that the program's uniform contents are used as an index
 197          * into the GL uniform storage.
 198          */
 199         QUNIFORM_UNIFORM,
 200
 201         /** @{
 202          * Scaling factors from clip coordinates to relative to the viewport
 203          * center.
 204          *
 205          * This is used by the coordinate and vertex shaders to produce the
 206          * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
 207          * point offsets from the viewport ccenter.
 208          */
 209         QUNIFORM_VIEWPORT_X_SCALE,
 210         QUNIFORM_VIEWPORT_Y_SCALE,
 211         /** @} */
 212
 213         QUNIFORM_VIEWPORT_Z_OFFSET,
 214         QUNIFORM_VIEWPORT_Z_SCALE,
 215
 216         QUNIFORM_USER_CLIP_PLANE,
 217
 218         /**
 219          * A reference to a texture config parameter 0 uniform.
 220          *
 221          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 222          * defines texture type, miplevels, and such.  It will be found as a
 223          * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
 224          */
 225         QUNIFORM_TEXTURE_CONFIG_P0,
 226
 227         /**
 228          * A reference to a texture config parameter 1 uniform.
 229          *
 230          * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
 231          * defines texture width, height, filters, and wrap modes.  It will be
 232          * found as a parameter to the second QOP_TEX_[STRB] instruction in a
 233          * sequence.
 234          */
 235         QUNIFORM_TEXTURE_CONFIG_P1,
 236
 237         /** A reference to a texture config parameter 2 cubemap stride uniform */
 238         QUNIFORM_TEXTURE_CONFIG_P2,
 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_BLEND_CONST_COLOR,
 248         QUNIFORM_STENCIL,
 249
 250         QUNIFORM_ALPHA_REF,
 251 };
 252
 253 struct vc4_varying_semantic {
 254         uint8_t semantic;
 255         uint8_t index;
 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_compile {
 285         struct vc4_context *vc4;
 286         struct tgsi_parse_context parser;
 287         struct qreg *temps;
 288         /**
 289          * Inputs to the shader, arranged by TGSI declaration order.
 290          *
 291          * Not all fragment shader QFILE_VARY reads are present in this array.
 292          */
 293         struct qreg *inputs;
 294         struct qreg *outputs;
 295         struct qreg *consts;
 296         struct qreg addr[4]; /* TGSI ARL destination. */
 297         uint32_t temps_array_size;
 298         uint32_t inputs_array_size;
 299         uint32_t outputs_array_size;
 300         uint32_t uniforms_array_size;
 301         uint32_t consts_array_size;
 302         uint32_t num_consts;
 303
 304         struct vc4_compiler_ubo_range *ubo_ranges;
 305         uint32_t ubo_ranges_array_size;
 306         uint32_t num_ubo_ranges;
 307         uint32_t next_ubo_dst_offset;
 308
 309         struct qreg line_x, point_x, point_y;
 310         struct qreg discard;
 311
 312         /**
 313          * Array of the TGSI semantics of all FS QFILE_VARY reads.
 314          *
 315          * This includes those that aren't part of the VPM varyings, like
 316          * point/line coordinates.
 317          */
 318         struct vc4_varying_semantic *input_semantics;
 319         uint32_t num_input_semantics;
 320         uint32_t input_semantics_array_size;
 321
 322         /**
 323          * An entry per outputs[] in the VS indicating what the semantic of
 324          * the output is.  Used to emit from the VS in the order that the FS
 325          * needs.
 326          */
 327         struct vc4_varying_semantic *output_semantics;
 328
 329         struct pipe_shader_state *shader_state;
 330         struct vc4_key *key;
 331         struct vc4_fs_key *fs_key;
 332         struct vc4_vs_key *vs_key;
 333
 334         uint32_t *uniform_data;
 335         enum quniform_contents *uniform_contents;
 336         uint32_t uniform_array_size;
 337         uint32_t num_uniforms;
 338         uint32_t num_outputs;
 339         uint32_t num_texture_samples;
 340         uint32_t output_position_index;
 341         uint32_t output_clipvertex_index;
 342         uint32_t output_color_index;
 343         uint32_t output_point_size_index;
 344
 345         struct qreg undef;
 346         enum qstage stage;
 347         uint32_t num_temps;
 348         struct simple_node instructions;
 349         uint32_t immediates[1024];
 350
 351         struct simple_node qpu_inst_list;
 352         uint64_t *qpu_insts;
 353         uint32_t qpu_inst_count;
 354         uint32_t qpu_inst_size;
 355         uint32_t num_inputs;
 356
 357         uint32_t program_id;
 358         uint32_t variant_id;
 359 };
 360
 361 struct vc4_compile *qir_compile_init(void);
 362 void qir_compile_destroy(struct vc4_compile *c);
 363 struct qinst *qir_inst(enum qop op, struct qreg dst,
 364                        struct qreg src0, struct qreg src1);
 365 struct qinst *qir_inst4(enum qop op, struct qreg dst,
 366                         struct qreg a,
 367                         struct qreg b,
 368                         struct qreg c,
 369                         struct qreg d);
 370 void qir_remove_instruction(struct qinst *qinst);
 371 void qir_reorder_uniforms(struct vc4_compile *c);
 372 void qir_emit(struct vc4_compile *c, struct qinst *inst);
 373 struct qreg qir_get_temp(struct vc4_compile *c);
 374 int qir_get_op_nsrc(enum qop qop);
 375 bool qir_reg_equals(struct qreg a, struct qreg b);
 376 bool qir_has_side_effects(struct vc4_compile *c, struct qinst *inst);
 377 bool qir_is_multi_instruction(struct qinst *inst);
 378 bool qir_depends_on_flags(struct qinst *inst);
 379 bool qir_writes_r4(struct qinst *inst);
 380 bool qir_reads_r4(struct qinst *inst);
 381 struct qreg qir_follow_movs(struct qinst **defs, struct qreg reg);
 382
 383 void qir_dump(struct vc4_compile *c);
 384 void qir_dump_inst(struct vc4_compile *c, struct qinst *inst);
 385 const char *qir_get_stage_name(enum qstage stage);
 386
 387 void qir_optimize(struct vc4_compile *c);
 388 bool qir_opt_algebraic(struct vc4_compile *c);
 389 bool qir_opt_copy_propagation(struct vc4_compile *c);
 390 bool qir_opt_cse(struct vc4_compile *c);
 391 bool qir_opt_dead_code(struct vc4_compile *c);
 392 bool qir_opt_small_immediates(struct vc4_compile *c);
 393 bool qir_opt_vpm_writes(struct vc4_compile *c);
 394
 395 void qpu_schedule_instructions(struct vc4_compile *c);
 396
 397 #define QIR_ALU0(name)                                                   \
 398 static inline struct qreg                                                \
 399 qir_##name(struct vc4_compile *c)                                        \
 400 {                                                                        \
 401         struct qreg t = qir_get_temp(c);                                 \
 402         qir_emit(c, qir_inst(QOP_##name, t, c->undef, c->undef));        \
 403         return t;                                                        \
 404 }
 405
 406 #define QIR_ALU1(name)                                                   \
 407 static inline struct qreg                                                \
 408 qir_##name(struct vc4_compile *c, struct qreg a)                         \
 409 {                                                                        \
 410         struct qreg t = qir_get_temp(c);                                 \
 411         qir_emit(c, qir_inst(QOP_##name, t, a, c->undef));               \
 412         return t;                                                        \
 413 }
 414
 415 #define QIR_ALU2(name)                                                   \
 416 static inline struct qreg                                                \
 417 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)          \
 418 {                                                                        \
 419         struct qreg t = qir_get_temp(c);                                 \
 420         qir_emit(c, qir_inst(QOP_##name, t, a, b));                      \
 421         return t;                                                        \
 422 }
 423
 424 #define QIR_NODST_1(name)                                               \
 425 static inline void                                                      \
 426 qir_##name(struct vc4_compile *c, struct qreg a)                        \
 427 {                                                                       \
 428         qir_emit(c, qir_inst(QOP_##name, c->undef, a, c->undef));       \
 429 }
 430
 431 #define QIR_NODST_2(name)                                               \
 432 static inline void                                                      \
 433 qir_##name(struct vc4_compile *c, struct qreg a, struct qreg b)         \
 434 {                                                                       \
 435         qir_emit(c, qir_inst(QOP_##name, c->undef, a, b));       \
 436 }
 437
 438 QIR_ALU1(MOV)
 439 QIR_ALU2(FADD)
 440 QIR_ALU2(FSUB)
 441 QIR_ALU2(FMUL)
 442 QIR_ALU2(MUL24)
 443 QIR_NODST_1(SF)
 444 QIR_ALU1(SEL_X_0_ZS)
 445 QIR_ALU1(SEL_X_0_ZC)
 446 QIR_ALU1(SEL_X_0_NS)
 447 QIR_ALU1(SEL_X_0_NC)
 448 QIR_ALU2(SEL_X_Y_ZS)
 449 QIR_ALU2(SEL_X_Y_ZC)
 450 QIR_ALU2(SEL_X_Y_NS)
 451 QIR_ALU2(SEL_X_Y_NC)
 452 QIR_ALU2(FMIN)
 453 QIR_ALU2(FMAX)
 454 QIR_ALU2(FMINABS)
 455 QIR_ALU2(FMAXABS)
 456 QIR_ALU1(FTOI)
 457 QIR_ALU1(ITOF)
 458
 459 QIR_ALU2(ADD)
 460 QIR_ALU2(SUB)
 461 QIR_ALU2(SHL)
 462 QIR_ALU2(SHR)
 463 QIR_ALU2(ASR)
 464 QIR_ALU2(MIN)
 465 QIR_ALU2(MAX)
 466 QIR_ALU2(AND)
 467 QIR_ALU2(OR)
 468 QIR_ALU2(XOR)
 469 QIR_ALU1(NOT)
 470
 471 QIR_ALU1(RCP)
 472 QIR_ALU1(RSQ)
 473 QIR_ALU1(EXP2)
 474 QIR_ALU1(LOG2)
 475 QIR_ALU2(PACK_SCALED)
 476 QIR_ALU1(VARY_ADD_C)
 477 QIR_NODST_2(TEX_S)
 478 QIR_NODST_2(TEX_T)
 479 QIR_NODST_2(TEX_R)
 480 QIR_NODST_2(TEX_B)
 481 QIR_NODST_2(TEX_DIRECT)
 482 QIR_ALU0(FRAG_X)
 483 QIR_ALU0(FRAG_Y)
 484 QIR_ALU0(FRAG_Z)
 485 QIR_ALU0(FRAG_W)
 486 QIR_ALU0(FRAG_REV_FLAG)
 487 QIR_ALU0(TEX_RESULT)
 488 QIR_ALU0(TLB_COLOR_READ)
 489 QIR_NODST_1(TLB_Z_WRITE)
 490 QIR_NODST_1(TLB_DISCARD_SETUP)
 491 QIR_NODST_1(TLB_STENCIL_SETUP)
 492
 493 static inline struct qreg
 494 qir_R4_UNPACK(struct vc4_compile *c, struct qreg r4, int i)
 495 {
 496         struct qreg t = qir_get_temp(c);
 497         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, r4, c->undef));
 498         return t;
 499 }
 500
 501 static inline struct qreg
 502 qir_SEL_X_0_COND(struct vc4_compile *c, int i)
 503 {
 504         struct qreg t = qir_get_temp(c);
 505         qir_emit(c, qir_inst(QOP_R4_UNPACK_A + i, t, c->undef, c->undef));
 506         return t;
 507 }
 508
 509 static inline struct qreg
 510 qir_UNPACK_8_F(struct vc4_compile *c, struct qreg src, int i)
 511 {
 512         struct qreg t = qir_get_temp(c);
 513         qir_emit(c, qir_inst(QOP_UNPACK_8A_F + i, t, src, c->undef));
 514         return t;
 515 }
 516
 517 static inline struct qreg
 518 qir_UNPACK_8_I(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_I + i, t, src, c->undef));
 522         return t;
 523 }
 524
 525 static inline struct qreg
 526 qir_UNPACK_16_F(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_16A_F + i, t, src, c->undef));
 530         return t;
 531 }
 532
 533 static inline struct qreg
 534 qir_UNPACK_16_I(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_I + i, t, src, c->undef));
 538         return t;
 539 }
 540
 541 static inline struct qreg
 542 qir_POW(struct vc4_compile *c, struct qreg x, struct qreg y)
 543 {
 544         return qir_EXP2(c, qir_FMUL(c,
 545                                     y,
 546                                     qir_LOG2(c, x)));
 547 }
 548
 549 static inline void
 550 qir_VPM_WRITE(struct vc4_compile *c, struct qreg val)
 551 {
 552         static const struct qreg vpm = { QFILE_VPM, 0 };
 553         qir_emit(c, qir_inst(QOP_MOV, vpm, val, c->undef));
 554 }
 555
 556 #endif /* VC4_QIR_H */