src/gallium/drivers/vc4/vc4_qir.c

   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 #include "util/u_memory.h"
  25 #include "util/ralloc.h"
  26
  27 #include "vc4_qir.h"
  28 #include "vc4_qpu.h"
  29
  30 struct qir_op_info {
  31         const char *name;
  32         uint8_t ndst, nsrc;
  33         bool has_side_effects;
  34         bool multi_instruction;
  35 };
  36
  37 static const struct qir_op_info qir_op_info[] = {
  38         [QOP_MOV] = { "mov", 1, 1 },
  39         [QOP_FADD] = { "fadd", 1, 2 },
  40         [QOP_FSUB] = { "fsub", 1, 2 },
  41         [QOP_FMUL] = { "fmul", 1, 2 },
  42         [QOP_MUL24] = { "mul24", 1, 2 },
  43         [QOP_V8MULD] = {"v8muld", 1, 2 },
  44         [QOP_V8MIN] = {"v8min", 1, 2 },
  45         [QOP_V8MAX] = {"v8max", 1, 2 },
  46         [QOP_V8ADDS] = {"v8adds", 1, 2 },
  47         [QOP_V8SUBS] = {"v8subs", 1, 2 },
  48         [QOP_FMIN] = { "fmin", 1, 2 },
  49         [QOP_FMAX] = { "fmax", 1, 2 },
  50         [QOP_FMINABS] = { "fminabs", 1, 2 },
  51         [QOP_FMAXABS] = { "fmaxabs", 1, 2 },
  52         [QOP_FTOI] = { "ftoi", 1, 1 },
  53         [QOP_ITOF] = { "itof", 1, 1 },
  54         [QOP_ADD] = { "add", 1, 2 },
  55         [QOP_SUB] = { "sub", 1, 2 },
  56         [QOP_SHR] = { "shr", 1, 2 },
  57         [QOP_ASR] = { "asr", 1, 2 },
  58         [QOP_SHL] = { "shl", 1, 2 },
  59         [QOP_MIN] = { "min", 1, 2 },
  60         [QOP_MAX] = { "max", 1, 2 },
  61         [QOP_AND] = { "and", 1, 2 },
  62         [QOP_OR] = { "or", 1, 2 },
  63         [QOP_XOR] = { "xor", 1, 2 },
  64         [QOP_NOT] = { "not", 1, 1 },
  65
  66         [QOP_SEL_X_0_NS] = { "fsel_x_0_ns", 1, 1, false, true },
  67         [QOP_SEL_X_0_NC] = { "fsel_x_0_nc", 1, 1, false, true },
  68         [QOP_SEL_X_0_ZS] = { "fsel_x_0_zs", 1, 1, false, true },
  69         [QOP_SEL_X_0_ZC] = { "fsel_x_0_zc", 1, 1, false, true },
  70         [QOP_SEL_X_Y_NS] = { "fsel_x_y_ns", 1, 2, false, true },
  71         [QOP_SEL_X_Y_NC] = { "fsel_x_y_nc", 1, 2, false, true },
  72         [QOP_SEL_X_Y_ZS] = { "fsel_x_y_zs", 1, 2, false, true },
  73         [QOP_SEL_X_Y_ZC] = { "fsel_x_y_zc", 1, 2, false, true },
  74
  75         [QOP_RCP] = { "rcp", 1, 1, false, true },
  76         [QOP_RSQ] = { "rsq", 1, 1, false, true },
  77         [QOP_EXP2] = { "exp2", 1, 2, false, true },
  78         [QOP_LOG2] = { "log2", 1, 2, false, true },
  79         [QOP_PACK_8888_F] = { "pack_8888_f", 1, 1 },
  80         [QOP_PACK_8A_F] = { "pack_8a_f", 1, 1 },
  81         [QOP_PACK_8B_F] = { "pack_8b_f", 1, 1 },
  82         [QOP_PACK_8C_F] = { "pack_8c_f", 1, 1 },
  83         [QOP_PACK_8D_F] = { "pack_8d_f", 1, 1 },
  84         [QOP_TLB_DISCARD_SETUP] = { "discard", 0, 1, true },
  85         [QOP_TLB_STENCIL_SETUP] = { "tlb_stencil_setup", 0, 1, true },
  86         [QOP_TLB_Z_WRITE] = { "tlb_z", 0, 1, true },
  87         [QOP_TLB_COLOR_WRITE] = { "tlb_color", 0, 1, true },
  88         [QOP_TLB_COLOR_READ] = { "tlb_color_read", 1, 0 },
  89         [QOP_VARY_ADD_C] = { "vary_add_c", 1, 1 },
  90
  91         [QOP_FRAG_X] = { "frag_x", 1, 0 },
  92         [QOP_FRAG_Y] = { "frag_y", 1, 0 },
  93         [QOP_FRAG_Z] = { "frag_z", 1, 0 },
  94         [QOP_FRAG_W] = { "frag_w", 1, 0 },
  95         [QOP_FRAG_REV_FLAG] = { "frag_rev_flag", 1, 0 },
  96
  97         [QOP_TEX_S] = { "tex_s", 0, 2 },
  98         [QOP_TEX_T] = { "tex_t", 0, 2 },
  99         [QOP_TEX_R] = { "tex_r", 0, 2 },
 100         [QOP_TEX_B] = { "tex_b", 0, 2 },
 101         [QOP_TEX_DIRECT] = { "tex_direct", 0, 2 },
 102         [QOP_TEX_RESULT] = { "tex_result", 1, 0, true },
 103         [QOP_UNPACK_8A_F] = { "unpack_8a_f", 1, 1 },
 104         [QOP_UNPACK_8B_F] = { "unpack_8b_f", 1, 1 },
 105         [QOP_UNPACK_8C_F] = { "unpack_8c_f", 1, 1 },
 106         [QOP_UNPACK_8D_F] = { "unpack_8d_f", 1, 1 },
 107         [QOP_UNPACK_16A_F] = { "unpack_16a_f", 1, 1 },
 108         [QOP_UNPACK_16B_F] = { "unpack_16b_f", 1, 1 },
 109         [QOP_UNPACK_8A_I] = { "unpack_8a_i", 1, 1 },
 110         [QOP_UNPACK_8B_I] = { "unpack_8b_i", 1, 1 },
 111         [QOP_UNPACK_8C_I] = { "unpack_8c_i", 1, 1 },
 112         [QOP_UNPACK_8D_I] = { "unpack_8d_i", 1, 1 },
 113         [QOP_UNPACK_16A_I] = { "unpack_16a_i", 1, 1 },
 114         [QOP_UNPACK_16B_I] = { "unpack_16b_i", 1, 1 },
 115 };
 116
 117 static const char *
 118 qir_get_op_name(enum qop qop)
 119 {
 120         if (qop < ARRAY_SIZE(qir_op_info) && qir_op_info[qop].name)
 121                 return qir_op_info[qop].name;
 122         else
 123                 return "???";
 124 }
 125
 126 int
 127 qir_get_op_nsrc(enum qop qop)
 128 {
 129         if (qop < ARRAY_SIZE(qir_op_info) && qir_op_info[qop].name)
 130                 return qir_op_info[qop].nsrc;
 131         else
 132                 abort();
 133 }
 134
 135 /**
 136  * Returns whether the instruction has any side effects that must be
 137  * preserved.
 138  */
 139 bool
 140 qir_has_side_effects(struct vc4_compile *c, struct qinst *inst)
 141 {
 142         return qir_op_info[inst->op].has_side_effects;
 143 }
 144
 145 bool
 146 qir_has_side_effect_reads(struct vc4_compile *c, struct qinst *inst)
 147 {
 148         /* We can dead-code eliminate varyings, because we only tell the VS
 149          * about the live ones at the end.  But we have to preserve the
 150          * point/line coordinates reads, because they're generated by
 151          * fixed-function hardware.
 152          */
 153         for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
 154                 if (inst->src[i].file == QFILE_VARY &&
 155                     c->input_slots[inst->src[i].index].slot == 0xff) {
 156                         return true;
 157                 }
 158
 159                 if (inst->src[i].file == QFILE_VPM)
 160                         return true;
 161         }
 162
 163         if (inst->dst.file == QFILE_VPM)
 164                 return true;
 165
 166         return false;
 167 }
 168
 169 bool
 170 qir_is_multi_instruction(struct qinst *inst)
 171 {
 172         return qir_op_info[inst->op].multi_instruction;
 173 }
 174
 175 bool
 176 qir_is_mul(struct qinst *inst)
 177 {
 178         switch (inst->op) {
 179         case QOP_FMUL:
 180         case QOP_MUL24:
 181         case QOP_V8MULD:
 182         case QOP_V8MIN:
 183         case QOP_V8MAX:
 184         case QOP_V8ADDS:
 185         case QOP_V8SUBS:
 186                 return true;
 187         default:
 188                 return false;
 189         }
 190 }
 191
 192 bool
 193 qir_is_tex(struct qinst *inst)
 194 {
 195         return inst->op >= QOP_TEX_S && inst->op <= QOP_TEX_DIRECT;
 196 }
 197
 198 bool
 199 qir_depends_on_flags(struct qinst *inst)
 200 {
 201         switch (inst->op) {
 202         case QOP_SEL_X_0_NS:
 203         case QOP_SEL_X_0_NC:
 204         case QOP_SEL_X_0_ZS:
 205         case QOP_SEL_X_0_ZC:
 206         case QOP_SEL_X_Y_NS:
 207         case QOP_SEL_X_Y_NC:
 208         case QOP_SEL_X_Y_ZS:
 209         case QOP_SEL_X_Y_ZC:
 210                 return true;
 211         default:
 212                 return false;
 213         }
 214 }
 215
 216 bool
 217 qir_src_needs_a_file(struct qinst *inst)
 218 {
 219         switch (inst->op) {
 220         case QOP_UNPACK_8A_F:
 221         case QOP_UNPACK_8B_F:
 222         case QOP_UNPACK_8C_F:
 223         case QOP_UNPACK_8D_F:
 224         case QOP_UNPACK_16A_F:
 225         case QOP_UNPACK_16B_F:
 226         case QOP_UNPACK_8A_I:
 227         case QOP_UNPACK_8B_I:
 228         case QOP_UNPACK_8C_I:
 229         case QOP_UNPACK_8D_I:
 230         case QOP_UNPACK_16A_I:
 231         case QOP_UNPACK_16B_I:
 232                 return true;
 233         default:
 234                 return false;
 235         }
 236 }
 237
 238 bool
 239 qir_writes_r4(struct qinst *inst)
 240 {
 241         switch (inst->op) {
 242         case QOP_TEX_RESULT:
 243         case QOP_TLB_COLOR_READ:
 244         case QOP_RCP:
 245         case QOP_RSQ:
 246         case QOP_EXP2:
 247         case QOP_LOG2:
 248                 return true;
 249         default:
 250                 return false;
 251         }
 252 }
 253
 254 static void
 255 qir_print_reg(struct vc4_compile *c, struct qreg reg, bool write)
 256 {
 257         static const char *files[] = {
 258                 [QFILE_TEMP] = "t",
 259                 [QFILE_VARY] = "v",
 260                 [QFILE_UNIF] = "u",
 261         };
 262
 263         if (reg.file == QFILE_NULL) {
 264                 fprintf(stderr, "null");
 265         } else if (reg.file == QFILE_SMALL_IMM) {
 266                 if ((int)reg.index >= -16 && (int)reg.index <= 15)
 267                         fprintf(stderr, "%d", reg.index);
 268                 else
 269                         fprintf(stderr, "%f", uif(reg.index));
 270         } else if (reg.file == QFILE_VPM) {
 271                 if (write) {
 272                         fprintf(stderr, "vpm");
 273                 } else {
 274                         fprintf(stderr, "vpm%d.%d",
 275                                 reg.index / 4, reg.index % 4);
 276                 }
 277         } else {
 278                 fprintf(stderr, "%s%d", files[reg.file], reg.index);
 279         }
 280
 281         if (reg.file == QFILE_UNIF &&
 282             c->uniform_contents[reg.index] == QUNIFORM_CONSTANT) {
 283                 fprintf(stderr, " (0x%08x / %f)",
 284                         c->uniform_data[reg.index],
 285                         uif(c->uniform_data[reg.index]));
 286         }
 287 }
 288
 289 void
 290 qir_dump_inst(struct vc4_compile *c, struct qinst *inst)
 291 {
 292         fprintf(stderr, "%s%s ",
 293                 qir_get_op_name(inst->op),
 294                 inst->sf ? ".sf" : "");
 295
 296         qir_print_reg(c, inst->dst, true);
 297         if (inst->dst.pack) {
 298                 if (inst->dst.pack) {
 299                         if (qir_is_mul(inst))
 300                                 vc4_qpu_disasm_pack_mul(stderr, inst->dst.pack);
 301                         else
 302                                 vc4_qpu_disasm_pack_a(stderr, inst->dst.pack);
 303                 }
 304         }
 305         for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
 306                 fprintf(stderr, ", ");
 307                 qir_print_reg(c, inst->src[i], false);
 308         }
 309 }
 310
 311 void
 312 qir_dump(struct vc4_compile *c)
 313 {
 314         list_for_each_entry(struct qinst, inst, &c->instructions, link) {
 315                 qir_dump_inst(c, inst);
 316                 fprintf(stderr, "\n");
 317         }
 318 }
 319
 320 struct qreg
 321 qir_get_temp(struct vc4_compile *c)
 322 {
 323         struct qreg reg;
 324
 325         reg.file = QFILE_TEMP;
 326         reg.index = c->num_temps++;
 327         reg.pack = 0;
 328
 329         if (c->num_temps > c->defs_array_size) {
 330                 uint32_t old_size = c->defs_array_size;
 331                 c->defs_array_size = MAX2(old_size * 2, 16);
 332                 c->defs = reralloc(c, c->defs, struct qinst *,
 333                                    c->defs_array_size);
 334                 memset(&c->defs[old_size], 0,
 335                        sizeof(c->defs[0]) * (c->defs_array_size - old_size));
 336         }
 337
 338         return reg;
 339 }
 340
 341 struct qinst *
 342 qir_inst(enum qop op, struct qreg dst, struct qreg src0, struct qreg src1)
 343 {
 344         struct qinst *inst = CALLOC_STRUCT(qinst);
 345
 346         inst->op = op;
 347         inst->dst = dst;
 348         inst->src = calloc(2, sizeof(inst->src[0]));
 349         inst->src[0] = src0;
 350         inst->src[1] = src1;
 351
 352         return inst;
 353 }
 354
 355 struct qinst *
 356 qir_inst4(enum qop op, struct qreg dst,
 357           struct qreg a,
 358           struct qreg b,
 359           struct qreg c,
 360           struct qreg d)
 361 {
 362         struct qinst *inst = CALLOC_STRUCT(qinst);
 363
 364         inst->op = op;
 365         inst->dst = dst;
 366         inst->src = calloc(4, sizeof(*inst->src));
 367         inst->src[0] = a;
 368         inst->src[1] = b;
 369         inst->src[2] = c;
 370         inst->src[3] = d;
 371
 372         return inst;
 373 }
 374
 375 void
 376 qir_emit(struct vc4_compile *c, struct qinst *inst)
 377 {
 378         if (inst->dst.file == QFILE_TEMP)
 379                 c->defs[inst->dst.index] = inst;
 380
 381         qir_emit_nodef(c, inst);
 382 }
 383
 384 bool
 385 qir_reg_equals(struct qreg a, struct qreg b)
 386 {
 387         return a.file == b.file && a.index == b.index;
 388 }
 389
 390 struct vc4_compile *
 391 qir_compile_init(void)
 392 {
 393         struct vc4_compile *c = rzalloc(NULL, struct vc4_compile);
 394
 395         list_inithead(&c->instructions);
 396
 397         c->output_position_index = -1;
 398         c->output_color_index = -1;
 399         c->output_point_size_index = -1;
 400
 401         c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
 402                                             _mesa_key_pointer_equal);
 403
 404         return c;
 405 }
 406
 407 void
 408 qir_remove_instruction(struct vc4_compile *c, struct qinst *qinst)
 409 {
 410         if (qinst->dst.file == QFILE_TEMP)
 411                 c->defs[qinst->dst.index] = NULL;
 412
 413         list_del(&qinst->link);
 414         free(qinst->src);
 415         free(qinst);
 416 }
 417
 418 struct qreg
 419 qir_follow_movs(struct vc4_compile *c, struct qreg reg)
 420 {
 421         while (reg.file == QFILE_TEMP &&
 422                c->defs[reg.index] &&
 423                c->defs[reg.index]->op == QOP_MOV) {
 424                 reg = c->defs[reg.index]->src[0];
 425         }
 426
 427         return reg;
 428 }
 429
 430 void
 431 qir_compile_destroy(struct vc4_compile *c)
 432 {
 433         while (!list_empty(&c->instructions)) {
 434                 struct qinst *qinst =
 435                         (struct qinst *)c->instructions.next;
 436                 qir_remove_instruction(c, qinst);
 437         }
 438
 439         ralloc_free(c);
 440 }
 441
 442 const char *
 443 qir_get_stage_name(enum qstage stage)
 444 {
 445         static const char *names[] = {
 446                 [QSTAGE_FRAG] = "FS",
 447                 [QSTAGE_VERT] = "VS",
 448                 [QSTAGE_COORD] = "CS",
 449         };
 450
 451         return names[stage];
 452 }
 453
 454 struct qreg
 455 qir_uniform(struct vc4_compile *c,
 456             enum quniform_contents contents,
 457             uint32_t data)
 458 {
 459         for (int i = 0; i < c->num_uniforms; i++) {
 460                 if (c->uniform_contents[i] == contents &&
 461                     c->uniform_data[i] == data) {
 462                         return (struct qreg) { QFILE_UNIF, i };
 463                 }
 464         }
 465
 466         uint32_t uniform = c->num_uniforms++;
 467         struct qreg u = { QFILE_UNIF, uniform };
 468
 469         if (uniform >= c->uniform_array_size) {
 470                 c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
 471                                              c->uniform_array_size * 2);
 472
 473                 c->uniform_data = reralloc(c, c->uniform_data,
 474                                            uint32_t,
 475                                            c->uniform_array_size);
 476                 c->uniform_contents = reralloc(c, c->uniform_contents,
 477                                                enum quniform_contents,
 478                                                c->uniform_array_size);
 479         }
 480
 481         c->uniform_contents[uniform] = contents;
 482         c->uniform_data[uniform] = data;
 483
 484         return u;
 485 }
 486
 487 void
 488 qir_SF(struct vc4_compile *c, struct qreg src)
 489 {
 490         struct qinst *last_inst = NULL;
 491         if (!list_empty(&c->instructions))
 492                 last_inst = (struct qinst *)c->instructions.prev;
 493
 494         if (!last_inst ||
 495             last_inst->dst.file != src.file ||
 496             last_inst->dst.index != src.index ||
 497             qir_is_multi_instruction(last_inst)) {
 498                 src = qir_MOV(c, src);
 499                 last_inst = (struct qinst *)c->instructions.prev;
 500         }
 501         last_inst->sf = true;
 502 }
 503
 504 #define OPTPASS(func)                                                   \
 505         do {                                                            \
 506                 bool stage_progress = func(c);                          \
 507                 if (stage_progress) {                                   \
 508                         progress = true;                                \
 509                         if (print_opt_debug) {                          \
 510                                 fprintf(stderr,                         \
 511                                         "QIR opt pass %2d: %s progress\n", \
 512                                         pass, #func);                   \
 513                         }                                               \
 514                 }                                                       \
 515         } while (0)
 516
 517 void
 518 qir_optimize(struct vc4_compile *c)
 519 {
 520         bool print_opt_debug = false;
 521         int pass = 1;
 522
 523         while (true) {
 524                 bool progress = false;
 525
 526                 OPTPASS(qir_opt_algebraic);
 527                 OPTPASS(qir_opt_cse);
 528                 OPTPASS(qir_opt_constant_folding);
 529                 OPTPASS(qir_opt_copy_propagation);
 530                 OPTPASS(qir_opt_dead_code);
 531                 OPTPASS(qir_opt_small_immediates);
 532                 OPTPASS(qir_opt_vpm_writes);
 533
 534                 if (!progress)
 535                         break;
 536
 537                 pass++;
 538         }
 539 }