src/freedreno/ir3/ir3_shader.c

   1 /*
   2  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "util/u_atomic.h"
  28 #include "util/u_string.h"
  29 #include "util/u_math.h"
  30 #include "util/u_memory.h"
  31 #include "util/format/u_format.h"
  32
  33 #include "drm/freedreno_drmif.h"
  34
  35 #include "ir3_shader.h"
  36 #include "ir3_compiler.h"
  37 #include "ir3_nir.h"
  38
  39 int
  40 ir3_glsl_type_size(const struct glsl_type *type, bool bindless)
  41 {
  42         return glsl_count_attribute_slots(type, false);
  43 }
  44
  45 /* for vertex shader, the inputs are loaded into registers before the shader
  46  * is executed, so max_regs from the shader instructions might not properly
  47  * reflect the # of registers actually used, especially in case passthrough
  48  * varyings.
  49  *
  50  * Likewise, for fragment shader, we can have some regs which are passed
  51  * input values but never touched by the resulting shader (ie. as result
  52  * of dead code elimination or simply because we don't know how to turn
  53  * the reg off.
  54  */
  55 static void
  56 fixup_regfootprint(struct ir3_shader_variant *v)
  57 {
  58         unsigned i;
  59
  60         for (i = 0; i < v->inputs_count; i++) {
  61                 /* skip frag inputs fetch via bary.f since their reg's are
  62                  * not written by gpu before shader starts (and in fact the
  63                  * regid's might not even be valid)
  64                  */
  65                 if (v->inputs[i].bary)
  66                         continue;
  67
  68                 /* ignore high regs that are global to all threads in a warp
  69                  * (they exist by default) (a5xx+)
  70                  */
  71                 if (v->inputs[i].regid >= regid(48,0))
  72                         continue;
  73
  74                 if (v->inputs[i].compmask) {
  75                         unsigned n = util_last_bit(v->inputs[i].compmask) - 1;
  76                         int32_t regid = v->inputs[i].regid + n;
  77                         if (v->inputs[i].half) {
  78                                 if (!v->mergedregs) {
  79                                         v->info.max_half_reg = MAX2(v->info.max_half_reg, regid >> 2);
  80                                 } else {
  81                                         v->info.max_reg = MAX2(v->info.max_reg, regid >> 3);
  82                                 }
  83                         } else {
  84                                 v->info.max_reg = MAX2(v->info.max_reg, regid >> 2);
  85                         }
  86                 }
  87         }
  88
  89         for (i = 0; i < v->outputs_count; i++) {
  90                 /* for ex, VS shaders with tess don't have normal varying outs: */
  91                 if (!VALIDREG(v->outputs[i].regid))
  92                         continue;
  93                 int32_t regid = v->outputs[i].regid + 3;
  94                 if (v->outputs[i].half) {
  95                         if (!v->mergedregs) {
  96                                 v->info.max_half_reg = MAX2(v->info.max_half_reg, regid >> 2);
  97                         } else {
  98                                 v->info.max_reg = MAX2(v->info.max_reg, regid >> 3);
  99                         }
 100                 } else {
 101                         v->info.max_reg = MAX2(v->info.max_reg, regid >> 2);
 102                 }
 103         }
 104
 105         for (i = 0; i < v->num_sampler_prefetch; i++) {
 106                 unsigned n = util_last_bit(v->sampler_prefetch[i].wrmask) - 1;
 107                 int32_t regid = v->sampler_prefetch[i].dst + n;
 108                 if (v->sampler_prefetch[i].half_precision) {
 109                         if (!v->mergedregs) {
 110                                 v->info.max_half_reg = MAX2(v->info.max_half_reg, regid >> 2);
 111                         } else {
 112                                 v->info.max_reg = MAX2(v->info.max_reg, regid >> 3);
 113                         }
 114                 } else {
 115                         v->info.max_reg = MAX2(v->info.max_reg, regid >> 2);
 116                 }
 117         }
 118 }
 119
 120 /* wrapper for ir3_assemble() which does some info fixup based on
 121  * shader state.  Non-static since used by ir3_cmdline too.
 122  */
 123 void * ir3_shader_assemble(struct ir3_shader_variant *v)
 124 {
 125         unsigned gpu_id = v->shader->compiler->gpu_id;
 126         void *bin;
 127
 128         bin = ir3_assemble(v);
 129         if (!bin)
 130                 return NULL;
 131
 132         if (gpu_id >= 400) {
 133                 v->instrlen = v->info.sizedwords / (2 * 16);
 134         } else {
 135                 v->instrlen = v->info.sizedwords / (2 * 4);
 136         }
 137
 138         /* NOTE: if relative addressing is used, we set constlen in
 139          * the compiler (to worst-case value) since we don't know in
 140          * the assembler what the max addr reg value can be:
 141          */
 142         v->constlen = MAX2(v->constlen, v->info.max_const + 1);
 143
 144         /* On a4xx and newer, constlen must be a multiple of 16 dwords even though
 145          * uploads are in units of 4 dwords. Round it up here to make calculations
 146          * regarding the shared constlen simpler.
 147          */
 148         if (gpu_id >= 400)
 149                 v->constlen = align(v->constlen, 4);
 150
 151         fixup_regfootprint(v);
 152
 153         return bin;
 154 }
 155
 156 static void
 157 assemble_variant(struct ir3_shader_variant *v)
 158 {
 159         v->bin = ir3_shader_assemble(v);
 160
 161         if (shader_debug_enabled(v->shader->type)) {
 162                 fprintf(stdout, "Native code for unnamed %s shader %s:\n",
 163                         ir3_shader_stage(v), v->shader->nir->info.name);
 164                 if (v->shader->type == MESA_SHADER_FRAGMENT)
 165                         fprintf(stdout, "SIMD0\n");
 166                 ir3_shader_disasm(v, v->bin, stdout);
 167         }
 168
 169         /* no need to keep the ir around beyond this point: */
 170         ir3_destroy(v->ir);
 171         v->ir = NULL;
 172 }
 173
 174 /*
 175  * For creating normal shader variants, 'nonbinning' is NULL.  For
 176  * creating binning pass shader, it is link to corresponding normal
 177  * (non-binning) variant.
 178  */
 179 static struct ir3_shader_variant *
 180 create_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
 181                 struct ir3_shader_variant *nonbinning)
 182 {
 183         struct ir3_shader_variant *v = rzalloc_size(shader, sizeof(*v));
 184         int ret;
 185
 186         if (!v)
 187                 return NULL;
 188
 189         v->id = ++shader->variant_count;
 190         v->shader = shader;
 191         v->binning_pass = !!nonbinning;
 192         v->nonbinning = nonbinning;
 193         v->key = *key;
 194         v->type = shader->type;
 195         v->mergedregs = shader->compiler->gpu_id >= 600;
 196
 197         if (!v->binning_pass)
 198                 v->const_state = rzalloc_size(v, sizeof(*v->const_state));
 199
 200         ret = ir3_compile_shader_nir(shader->compiler, v);
 201         if (ret) {
 202                 debug_error("compile failed!");
 203                 goto fail;
 204         }
 205
 206         assemble_variant(v);
 207         if (!v->bin) {
 208                 debug_error("assemble failed!");
 209                 goto fail;
 210         }
 211
 212         return v;
 213
 214 fail:
 215         ralloc_free(v);
 216         return NULL;
 217 }
 218
 219 static inline struct ir3_shader_variant *
 220 shader_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
 221                 bool *created)
 222 {
 223         struct ir3_shader_variant *v;
 224
 225         *created = false;
 226
 227         for (v = shader->variants; v; v = v->next)
 228                 if (ir3_shader_key_equal(key, &v->key))
 229                         return v;
 230
 231         /* compile new variant if it doesn't exist already: */
 232         v = create_variant(shader, key, NULL);
 233         if (v) {
 234                 v->next = shader->variants;
 235                 shader->variants = v;
 236                 *created = true;
 237         }
 238
 239         return v;
 240 }
 241
 242 struct ir3_shader_variant *
 243 ir3_shader_get_variant(struct ir3_shader *shader, const struct ir3_shader_key *key,
 244                 bool binning_pass, bool *created)
 245 {
 246         mtx_lock(&shader->variants_lock);
 247         struct ir3_shader_variant *v =
 248                         shader_variant(shader, key, created);
 249
 250         if (v && binning_pass) {
 251                 if (!v->binning) {
 252                         v->binning = create_variant(shader, key, v);
 253                         *created = true;
 254                 }
 255                 mtx_unlock(&shader->variants_lock);
 256                 return v->binning;
 257         }
 258         mtx_unlock(&shader->variants_lock);
 259
 260         return v;
 261 }
 262
 263 void
 264 ir3_shader_destroy(struct ir3_shader *shader)
 265 {
 266         ralloc_free(shader->nir);
 267         mtx_destroy(&shader->variants_lock);
 268         ralloc_free(shader);
 269 }
 270
 271 /**
 272  * Creates a bitmask of the used bits of the shader key by this particular
 273  * shader.  Used by the gallium driver to skip state-dependent recompiles when
 274  * possible.
 275  */
 276 static void
 277 ir3_setup_used_key(struct ir3_shader *shader)
 278 {
 279         nir_shader *nir = shader->nir;
 280         struct shader_info *info = &nir->info;
 281         struct ir3_shader_key *key = &shader->key_mask;
 282
 283         /* This key flag is just used to make for a cheaper ir3_shader_key_equal
 284          * check in the common case.
 285          */
 286         key->has_per_samp = true;
 287
 288         if (info->stage == MESA_SHADER_FRAGMENT) {
 289                 key->fsaturate_s = ~0;
 290                 key->fsaturate_t = ~0;
 291                 key->fsaturate_r = ~0;
 292                 key->fastc_srgb = ~0;
 293                 key->fsamples = ~0;
 294
 295                 if (info->inputs_read & VARYING_BITS_COLOR) {
 296                         key->rasterflat = true;
 297                         key->color_two_side = true;
 298                 }
 299
 300                 if ((info->outputs_written & ~(FRAG_RESULT_DEPTH |
 301                                                                 FRAG_RESULT_STENCIL |
 302                                                                 FRAG_RESULT_SAMPLE_MASK)) != 0) {
 303                         key->fclamp_color = true;
 304                 }
 305
 306                 /* Only used for deciding on behavior of
 307                  * nir_intrinsic_load_barycentric_sample
 308                  */
 309                 key->msaa = info->fs.uses_sample_qualifier;
 310         } else {
 311                 key->tessellation = ~0;
 312                 key->has_gs = true;
 313
 314                 if (info->outputs_written & VARYING_BITS_COLOR)
 315                         key->vclamp_color = true;
 316
 317                 if (info->stage == MESA_SHADER_VERTEX) {
 318                         key->vsaturate_s = ~0;
 319                         key->vsaturate_t = ~0;
 320                         key->vsaturate_r = ~0;
 321                         key->vastc_srgb = ~0;
 322                         key->vsamples = ~0;
 323                 }
 324         }
 325 }
 326
 327 struct ir3_shader *
 328 ir3_shader_from_nir(struct ir3_compiler *compiler, nir_shader *nir,
 329                 unsigned reserved_user_consts, struct ir3_stream_output_info *stream_output)
 330 {
 331         struct ir3_shader *shader = rzalloc_size(NULL, sizeof(*shader));
 332
 333         mtx_init(&shader->variants_lock, mtx_plain);
 334         shader->compiler = compiler;
 335         shader->id = p_atomic_inc_return(&shader->compiler->shader_count);
 336         shader->type = nir->info.stage;
 337         if (stream_output)
 338                 memcpy(&shader->stream_output, stream_output, sizeof(shader->stream_output));
 339         shader->num_reserved_user_consts = reserved_user_consts;
 340
 341         if (nir->info.stage == MESA_SHADER_GEOMETRY)
 342                 NIR_PASS_V(nir, ir3_nir_lower_gs);
 343
 344         NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size,
 345                            (nir_lower_io_options)0);
 346
 347         if (compiler->gpu_id >= 600 &&
 348                         nir->info.stage == MESA_SHADER_FRAGMENT &&
 349                         !(ir3_shader_debug & IR3_DBG_NOFP16))
 350                 NIR_PASS_V(nir, nir_lower_mediump_outputs);
 351
 352         if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 353                 /* NOTE: lower load_barycentric_at_sample first, since it
 354                  * produces load_barycentric_at_offset:
 355                  */
 356                 NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_sample);
 357                 NIR_PASS_V(nir, ir3_nir_lower_load_barycentric_at_offset);
 358
 359                 NIR_PASS_V(nir, ir3_nir_move_varying_inputs);
 360         }
 361
 362         NIR_PASS_V(nir, nir_lower_io_arrays_to_elements_no_indirects, false);
 363
 364         NIR_PASS_V(nir, nir_lower_amul, ir3_glsl_type_size);
 365
 366         /* do first pass optimization, ignoring the key: */
 367         ir3_optimize_nir(shader, nir);
 368
 369         shader->nir = nir;
 370         if (ir3_shader_debug & IR3_DBG_DISASM) {
 371                 printf("dump nir%d: type=%d", shader->id, shader->type);
 372                 nir_print_shader(shader->nir, stdout);
 373         }
 374
 375         ir3_setup_used_key(shader);
 376
 377         return shader;
 378 }
 379
 380 static void dump_reg(FILE *out, const char *name, uint32_t r)
 381 {
 382         if (r != regid(63,0)) {
 383                 const char *reg_type = (r & HALF_REG_ID) ? "hr" : "r";
 384                 fprintf(out, "; %s: %s%d.%c\n", name, reg_type,
 385                                 (r & ~HALF_REG_ID) >> 2, "xyzw"[r & 0x3]);
 386         }
 387 }
 388
 389 static void dump_output(FILE *out, struct ir3_shader_variant *so,
 390                 unsigned slot, const char *name)
 391 {
 392         uint32_t regid;
 393         regid = ir3_find_output_regid(so, slot);
 394         dump_reg(out, name, regid);
 395 }
 396
 397 static const char *
 398 input_name(struct ir3_shader_variant *so, int i)
 399 {
 400         if (so->inputs[i].sysval) {
 401                 return gl_system_value_name(so->inputs[i].slot);
 402         } else if (so->type == MESA_SHADER_VERTEX) {
 403                 return gl_vert_attrib_name(so->inputs[i].slot);
 404         } else {
 405                 return gl_varying_slot_name(so->inputs[i].slot);
 406         }
 407 }
 408
 409 static const char *
 410 output_name(struct ir3_shader_variant *so, int i)
 411 {
 412         if (so->type == MESA_SHADER_FRAGMENT) {
 413                 return gl_frag_result_name(so->outputs[i].slot);
 414         } else {
 415                 switch (so->outputs[i].slot) {
 416                 case VARYING_SLOT_GS_HEADER_IR3:
 417                         return "GS_HEADER";
 418                 case VARYING_SLOT_GS_VERTEX_FLAGS_IR3:
 419                         return "GS_VERTEX_FLAGS";
 420                 case VARYING_SLOT_TCS_HEADER_IR3:
 421                         return "TCS_HEADER";
 422                 default:
 423                         return gl_varying_slot_name(so->outputs[i].slot);
 424                 }
 425         }
 426 }
 427
 428 void
 429 ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin, FILE *out)
 430 {
 431         struct ir3 *ir = so->ir;
 432         struct ir3_register *reg;
 433         const char *type = ir3_shader_stage(so);
 434         uint8_t regid;
 435         unsigned i;
 436
 437         foreach_input_n (instr, i, ir) {
 438                 reg = instr->regs[0];
 439                 regid = reg->num;
 440                 fprintf(out, "@in(%sr%d.%c)\tin%d",
 441                                 (reg->flags & IR3_REG_HALF) ? "h" : "",
 442                                 (regid >> 2), "xyzw"[regid & 0x3], i);
 443
 444                 if (reg->wrmask > 0x1)
 445                         fprintf(out, " (wrmask=0x%x)", reg->wrmask);
 446                 fprintf(out, "\n");
 447         }
 448
 449         /* print pre-dispatch texture fetches: */
 450         for (i = 0; i < so->num_sampler_prefetch; i++) {
 451                 const struct ir3_sampler_prefetch *fetch = &so->sampler_prefetch[i];
 452                 fprintf(out, "@tex(%sr%d.%c)\tsrc=%u, samp=%u, tex=%u, wrmask=0x%x, cmd=%u\n",
 453                                 fetch->half_precision ? "h" : "",
 454                                 fetch->dst >> 2, "xyzw"[fetch->dst & 0x3],
 455                                 fetch->src, fetch->samp_id, fetch->tex_id,
 456                                 fetch->wrmask, fetch->cmd);
 457         }
 458
 459         foreach_output_n (instr, i, ir) {
 460                 reg = instr->regs[0];
 461                 regid = reg->num;
 462                 fprintf(out, "@out(%sr%d.%c)\tout%d",
 463                                 (reg->flags & IR3_REG_HALF) ? "h" : "",
 464                                 (regid >> 2), "xyzw"[regid & 0x3], i);
 465                 if (reg->wrmask > 0x1)
 466                         fprintf(out, " (wrmask=0x%x)", reg->wrmask);
 467                 fprintf(out, "\n");
 468         }
 469
 470         const struct ir3_const_state *const_state = ir3_const_state(so);
 471         for (i = 0; i < const_state->immediates_count; i++) {
 472                 fprintf(out, "@const(c%d.x)\t", const_state->offsets.immediate + i);
 473                 fprintf(out, "0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
 474                                 const_state->immediates[i].val[0],
 475                                 const_state->immediates[i].val[1],
 476                                 const_state->immediates[i].val[2],
 477                                 const_state->immediates[i].val[3]);
 478         }
 479
 480         disasm_a3xx(bin, so->info.sizedwords, 0, out, ir->compiler->gpu_id);
 481
 482         fprintf(out, "; %s: outputs:", type);
 483         for (i = 0; i < so->outputs_count; i++) {
 484                 uint8_t regid = so->outputs[i].regid;
 485                 const char *reg_type = so->outputs[i].half ? "hr" : "r";
 486                 fprintf(out, " %s%d.%c (%s)",
 487                                 reg_type, (regid >> 2), "xyzw"[regid & 0x3],
 488                                 output_name(so, i));
 489         }
 490         fprintf(out, "\n");
 491
 492         fprintf(out, "; %s: inputs:", type);
 493         for (i = 0; i < so->inputs_count; i++) {
 494                 uint8_t regid = so->inputs[i].regid;
 495                 fprintf(out, " r%d.%c (%s slot=%d cm=%x,il=%u,b=%u)",
 496                                 (regid >> 2), "xyzw"[regid & 0x3],
 497                                 input_name(so, i),
 498                                 so->inputs[i].slot,
 499                                 so->inputs[i].compmask,
 500                                 so->inputs[i].inloc,
 501                                 so->inputs[i].bary);
 502         }
 503         fprintf(out, "\n");
 504
 505         /* print generic shader info: */
 506         fprintf(out, "; %s prog %d/%d: %u instr, %u nops, %u non-nops, %u mov, %u cov, %u dwords\n",
 507                         type, so->shader->id, so->id,
 508                         so->info.instrs_count,
 509                         so->info.nops_count,
 510                         so->info.instrs_count - so->info.nops_count,
 511                         so->info.mov_count, so->info.cov_count,
 512                         so->info.sizedwords);
 513
 514         fprintf(out, "; %s prog %d/%d: %u last-baryf, %d half, %d full, %u constlen\n",
 515                         type, so->shader->id, so->id,
 516                         so->info.last_baryf,
 517                         so->info.max_half_reg + 1,
 518                         so->info.max_reg + 1,
 519                         so->constlen);
 520
 521         fprintf(out, "; %s prog %d/%d: %u sstall, %u (ss), %u (sy), %d max_sun, %d loops\n",
 522                         type, so->shader->id, so->id,
 523                         so->info.sstall,
 524                         so->info.ss,
 525                         so->info.sy,
 526                         so->max_sun,
 527                         so->loops);
 528
 529         /* print shader type specific info: */
 530         switch (so->type) {
 531         case MESA_SHADER_VERTEX:
 532                 dump_output(out, so, VARYING_SLOT_POS, "pos");
 533                 dump_output(out, so, VARYING_SLOT_PSIZ, "psize");
 534                 break;
 535         case MESA_SHADER_FRAGMENT:
 536                 dump_reg(out, "pos (ij_pixel)",
 537                         ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL));
 538                 dump_reg(out, "pos (ij_centroid)",
 539                         ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID));
 540                 dump_reg(out, "pos (ij_size)",
 541                         ir3_find_sysval_regid(so, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE));
 542                 dump_output(out, so, FRAG_RESULT_DEPTH, "posz");
 543                 if (so->color0_mrt) {
 544                         dump_output(out, so, FRAG_RESULT_COLOR, "color");
 545                 } else {
 546                         dump_output(out, so, FRAG_RESULT_DATA0, "data0");
 547                         dump_output(out, so, FRAG_RESULT_DATA1, "data1");
 548                         dump_output(out, so, FRAG_RESULT_DATA2, "data2");
 549                         dump_output(out, so, FRAG_RESULT_DATA3, "data3");
 550                         dump_output(out, so, FRAG_RESULT_DATA4, "data4");
 551                         dump_output(out, so, FRAG_RESULT_DATA5, "data5");
 552                         dump_output(out, so, FRAG_RESULT_DATA6, "data6");
 553                         dump_output(out, so, FRAG_RESULT_DATA7, "data7");
 554                 }
 555                 dump_reg(out, "fragcoord",
 556                         ir3_find_sysval_regid(so, SYSTEM_VALUE_FRAG_COORD));
 557                 dump_reg(out, "fragface",
 558                         ir3_find_sysval_regid(so, SYSTEM_VALUE_FRONT_FACE));
 559                 break;
 560         default:
 561                 /* TODO */
 562                 break;
 563         }
 564
 565         fprintf(out, "\n");
 566 }
 567
 568 uint64_t
 569 ir3_shader_outputs(const struct ir3_shader *so)
 570 {
 571         return so->nir->info.outputs_written;
 572 }