src/gallium/drivers/freedreno/ir3/ir3_shader.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_string.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_inlines.h"
  33 #include "util/u_format.h"
  34 #include "tgsi/tgsi_dump.h"
  35 #include "tgsi/tgsi_parse.h"
  36
  37 #include "freedreno_context.h"
  38 #include "freedreno_util.h"
  39
  40 #include "ir3_shader.h"
  41 #include "ir3_compiler.h"
  42 #include "ir3_nir.h"
  43
  44 static void
  45 delete_variant(struct ir3_shader_variant *v)
  46 {
  47         if (v->ir)
  48                 ir3_destroy(v->ir);
  49         if (v->bo)
  50                 fd_bo_del(v->bo);
  51         free(v);
  52 }
  53
  54 /* for vertex shader, the inputs are loaded into registers before the shader
  55  * is executed, so max_regs from the shader instructions might not properly
  56  * reflect the # of registers actually used, especially in case passthrough
  57  * varyings.
  58  *
  59  * Likewise, for fragment shader, we can have some regs which are passed
  60  * input values but never touched by the resulting shader (ie. as result
  61  * of dead code elimination or simply because we don't know how to turn
  62  * the reg off.
  63  */
  64 static void
  65 fixup_regfootprint(struct ir3_shader_variant *v)
  66 {
  67         if (v->type == SHADER_VERTEX) {
  68                 unsigned i;
  69                 for (i = 0; i < v->inputs_count; i++) {
  70                         /* skip frag inputs fetch via bary.f since their reg's are
  71                          * not written by gpu before shader starts (and in fact the
  72                          * regid's might not even be valid)
  73                          */
  74                         if (v->inputs[i].bary)
  75                                 continue;
  76
  77                         if (v->inputs[i].compmask) {
  78                                 int32_t regid = (v->inputs[i].regid + 3) >> 2;
  79                                 v->info.max_reg = MAX2(v->info.max_reg, regid);
  80                         }
  81                 }
  82                 for (i = 0; i < v->outputs_count; i++) {
  83                         int32_t regid = (v->outputs[i].regid + 3) >> 2;
  84                         v->info.max_reg = MAX2(v->info.max_reg, regid);
  85                 }
  86         } else if (v->type == SHADER_FRAGMENT) {
  87                 /* NOTE: not sure how to turn pos_regid off..  but this could
  88                  * be, for example, r1.x while max reg used by the shader is
  89                  * r0.*, in which case we need to fixup the reg footprint:
  90                  */
  91                 v->info.max_reg = MAX2(v->info.max_reg, v->pos_regid >> 2);
  92                 if (v->frag_coord)
  93                         debug_assert(v->info.max_reg >= 0); /* hard coded r0.x */
  94                 if (v->frag_face)
  95                         debug_assert(v->info.max_half_reg >= 0); /* hr0.x */
  96         }
  97 }
  98
  99 /* wrapper for ir3_assemble() which does some info fixup based on
 100  * shader state.  Non-static since used by ir3_cmdline too.
 101  */
 102 void * ir3_shader_assemble(struct ir3_shader_variant *v, uint32_t gpu_id)
 103 {
 104         void *bin;
 105
 106         bin = ir3_assemble(v->ir, &v->info, gpu_id);
 107         if (!bin)
 108                 return NULL;
 109
 110         if (gpu_id >= 400) {
 111                 v->instrlen = v->info.sizedwords / (2 * 16);
 112         } else {
 113                 v->instrlen = v->info.sizedwords / (2 * 4);
 114         }
 115
 116         /* NOTE: if relative addressing is used, we set constlen in
 117          * the compiler (to worst-case value) since we don't know in
 118          * the assembler what the max addr reg value can be:
 119          */
 120         v->constlen = MIN2(255, MAX2(v->constlen, v->info.max_const + 1));
 121
 122         fixup_regfootprint(v);
 123
 124         return bin;
 125 }
 126
 127 static void
 128 assemble_variant(struct ir3_shader_variant *v)
 129 {
 130         struct fd_context *ctx = fd_context(v->shader->pctx);
 131         uint32_t gpu_id = v->shader->compiler->gpu_id;
 132         uint32_t sz, *bin;
 133
 134         bin = ir3_shader_assemble(v, gpu_id);
 135         sz = v->info.sizedwords * 4;
 136
 137         v->bo = fd_bo_new(ctx->dev, sz,
 138                         DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
 139                         DRM_FREEDRENO_GEM_TYPE_KMEM);
 140
 141         memcpy(fd_bo_map(v->bo), bin, sz);
 142
 143         if (fd_mesa_debug & FD_DBG_DISASM) {
 144                 struct ir3_shader_key key = v->key;
 145                 DBG("disassemble: type=%d, k={bp=%u,cts=%u,hp=%u}", v->type,
 146                         key.binning_pass, key.color_two_side, key.half_precision);
 147                 ir3_shader_disasm(v, bin);
 148         }
 149
 150         if (fd_mesa_debug & FD_DBG_SHADERDB) {
 151                 /* print generic shader info: */
 152                 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %u instructions, %u dwords\n",
 153                                 ir3_shader_stage(v->shader),
 154                                 v->shader->id, v->id,
 155                                 v->info.instrs_count,
 156                                 v->info.sizedwords);
 157                 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %u half, %u full\n",
 158                                 ir3_shader_stage(v->shader),
 159                                 v->shader->id, v->id,
 160                                 v->info.max_half_reg + 1,
 161                                 v->info.max_reg + 1);
 162                 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %u const, %u constlen\n",
 163                                 ir3_shader_stage(v->shader),
 164                                 v->shader->id, v->id,
 165                                 v->info.max_const + 1,
 166                                 v->constlen);
 167         }
 168
 169         free(bin);
 170
 171         /* no need to keep the ir around beyond this point: */
 172         ir3_destroy(v->ir);
 173         v->ir = NULL;
 174 }
 175
 176 static struct ir3_shader_variant *
 177 create_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 178 {
 179         struct ir3_shader_variant *v = CALLOC_STRUCT(ir3_shader_variant);
 180         int ret;
 181
 182         if (!v)
 183                 return NULL;
 184
 185         v->id = ++shader->variant_count;
 186         v->shader = shader;
 187         v->key = key;
 188         v->type = shader->type;
 189
 190         ret = ir3_compile_shader_nir(shader->compiler, v);
 191         if (ret) {
 192                 debug_error("compile failed!");
 193                 goto fail;
 194         }
 195
 196         assemble_variant(v);
 197         if (!v->bo) {
 198                 debug_error("assemble failed!");
 199                 goto fail;
 200         }
 201
 202         return v;
 203
 204 fail:
 205         delete_variant(v);
 206         return NULL;
 207 }
 208
 209 struct ir3_shader_variant *
 210 ir3_shader_variant(struct ir3_shader *shader, struct ir3_shader_key key)
 211 {
 212         struct ir3_shader_variant *v;
 213
 214         /* some shader key values only apply to vertex or frag shader,
 215          * so normalize the key to avoid constructing multiple identical
 216          * variants:
 217          */
 218         switch (shader->type) {
 219         case SHADER_FRAGMENT:
 220         case SHADER_COMPUTE:
 221                 key.binning_pass = false;
 222                 if (key.has_per_samp) {
 223                         key.vsaturate_s = 0;
 224                         key.vsaturate_t = 0;
 225                         key.vsaturate_r = 0;
 226                 }
 227                 break;
 228         case SHADER_VERTEX:
 229                 key.color_two_side = false;
 230                 key.half_precision = false;
 231                 key.rasterflat = false;
 232                 if (key.has_per_samp) {
 233                         key.fsaturate_s = 0;
 234                         key.fsaturate_t = 0;
 235                         key.fsaturate_r = 0;
 236                 }
 237                 break;
 238         }
 239
 240         for (v = shader->variants; v; v = v->next)
 241                 if (ir3_shader_key_equal(&key, &v->key))
 242                         return v;
 243
 244         /* compile new variant if it doesn't exist already: */
 245         v = create_variant(shader, key);
 246         if (v) {
 247                 v->next = shader->variants;
 248                 shader->variants = v;
 249         }
 250
 251         return v;
 252 }
 253
 254
 255 void
 256 ir3_shader_destroy(struct ir3_shader *shader)
 257 {
 258         struct ir3_shader_variant *v, *t;
 259         for (v = shader->variants; v; ) {
 260                 t = v;
 261                 v = v->next;
 262                 delete_variant(t);
 263         }
 264         ralloc_free(shader->nir);
 265         free(shader);
 266 }
 267
 268 struct ir3_shader *
 269 ir3_shader_create(struct pipe_context *pctx,
 270                 const struct pipe_shader_state *cso,
 271                 enum shader_t type)
 272 {
 273         struct ir3_shader *shader = CALLOC_STRUCT(ir3_shader);
 274         shader->compiler = fd_context(pctx)->screen->compiler;
 275         shader->id = ++shader->compiler->shader_count;
 276         shader->pctx = pctx;
 277         shader->type = type;
 278         if (fd_mesa_debug & FD_DBG_DISASM) {
 279                 DBG("dump tgsi: type=%d", shader->type);
 280                 tgsi_dump(cso->tokens, 0);
 281         }
 282         nir_shader *nir = ir3_tgsi_to_nir(cso->tokens);
 283         /* do first pass optimization, ignoring the key: */
 284         shader->nir = ir3_optimize_nir(shader, nir, NULL);
 285         if (fd_mesa_debug & FD_DBG_DISASM) {
 286                 DBG("dump nir%d: type=%d", shader->id, shader->type);
 287                 nir_print_shader(shader->nir, stdout);
 288         }
 289         shader->stream_output = cso->stream_output;
 290         if (fd_mesa_debug & FD_DBG_SHADERDB) {
 291                 /* if shader-db run, create a standard variant immediately
 292                  * (as otherwise nothing will trigger the shader to be
 293                  * actually compiled)
 294                  */
 295                 static struct ir3_shader_key key = {0};
 296                 ir3_shader_variant(shader, key);
 297         }
 298         return shader;
 299 }
 300
 301 static void dump_reg(const char *name, uint32_t r)
 302 {
 303         if (r != regid(63,0))
 304                 debug_printf("; %s: r%d.%c\n", name, r >> 2, "xyzw"[r & 0x3]);
 305 }
 306
 307 static void dump_output(struct ir3_shader_variant *so,
 308                 unsigned slot, const char *name)
 309 {
 310         uint32_t regid;
 311         regid = ir3_find_output_regid(so, slot);
 312         dump_reg(name, regid);
 313 }
 314
 315 void
 316 ir3_shader_disasm(struct ir3_shader_variant *so, uint32_t *bin)
 317 {
 318         struct ir3 *ir = so->ir;
 319         struct ir3_register *reg;
 320         const char *type = ir3_shader_stage(so->shader);
 321         uint8_t regid;
 322         unsigned i;
 323
 324         for (i = 0; i < ir->ninputs; i++) {
 325                 if (!ir->inputs[i]) {
 326                         debug_printf("; in%d unused\n", i);
 327                         continue;
 328                 }
 329                 reg = ir->inputs[i]->regs[0];
 330                 regid = reg->num;
 331                 debug_printf("@in(%sr%d.%c)\tin%d\n",
 332                                 (reg->flags & IR3_REG_HALF) ? "h" : "",
 333                                 (regid >> 2), "xyzw"[regid & 0x3], i);
 334         }
 335
 336         for (i = 0; i < ir->noutputs; i++) {
 337                 if (!ir->outputs[i]) {
 338                         debug_printf("; out%d unused\n", i);
 339                         continue;
 340                 }
 341                 /* kill shows up as a virtual output.. skip it! */
 342                 if (is_kill(ir->outputs[i]))
 343                         continue;
 344                 reg = ir->outputs[i]->regs[0];
 345                 regid = reg->num;
 346                 debug_printf("@out(%sr%d.%c)\tout%d\n",
 347                                 (reg->flags & IR3_REG_HALF) ? "h" : "",
 348                                 (regid >> 2), "xyzw"[regid & 0x3], i);
 349         }
 350
 351         for (i = 0; i < so->immediates_count; i++) {
 352                 debug_printf("@const(c%d.x)\t", so->first_immediate + i);
 353                 debug_printf("0x%08x, 0x%08x, 0x%08x, 0x%08x\n",
 354                                 so->immediates[i].val[0],
 355                                 so->immediates[i].val[1],
 356                                 so->immediates[i].val[2],
 357                                 so->immediates[i].val[3]);
 358         }
 359
 360         disasm_a3xx(bin, so->info.sizedwords, 0, so->type);
 361
 362         switch (so->type) {
 363         case SHADER_VERTEX:
 364                 debug_printf("; %s: outputs:", type);
 365                 for (i = 0; i < so->outputs_count; i++) {
 366                         uint8_t regid = so->outputs[i].regid;
 367                         debug_printf(" r%d.%c (%s)",
 368                                         (regid >> 2), "xyzw"[regid & 0x3],
 369                                         gl_varying_slot_name(so->outputs[i].slot));
 370                 }
 371                 debug_printf("\n");
 372                 debug_printf("; %s: inputs:", type);
 373                 for (i = 0; i < so->inputs_count; i++) {
 374                         uint8_t regid = so->inputs[i].regid;
 375                         debug_printf(" r%d.%c (cm=%x,il=%u,b=%u)",
 376                                         (regid >> 2), "xyzw"[regid & 0x3],
 377                                         so->inputs[i].compmask,
 378                                         so->inputs[i].inloc,
 379                                         so->inputs[i].bary);
 380                 }
 381                 debug_printf("\n");
 382                 break;
 383         case SHADER_FRAGMENT:
 384                 debug_printf("; %s: outputs:", type);
 385                 for (i = 0; i < so->outputs_count; i++) {
 386                         uint8_t regid = so->outputs[i].regid;
 387                         debug_printf(" r%d.%c (%s)",
 388                                         (regid >> 2), "xyzw"[regid & 0x3],
 389                                         gl_frag_result_name(so->outputs[i].slot));
 390                 }
 391                 debug_printf("\n");
 392                 debug_printf("; %s: inputs:", type);
 393                 for (i = 0; i < so->inputs_count; i++) {
 394                         uint8_t regid = so->inputs[i].regid;
 395                         debug_printf(" r%d.%c (%s,cm=%x,il=%u,b=%u)",
 396                                         (regid >> 2), "xyzw"[regid & 0x3],
 397                                         gl_varying_slot_name(so->inputs[i].slot),
 398                                         so->inputs[i].compmask,
 399                                         so->inputs[i].inloc,
 400                                         so->inputs[i].bary);
 401                 }
 402                 debug_printf("\n");
 403                 break;
 404         case SHADER_COMPUTE:
 405                 break;
 406         }
 407
 408         /* print generic shader info: */
 409         debug_printf("; %s prog %d/%d: %u instructions, %d half, %d full\n",
 410                         type, so->shader->id, so->id,
 411                         so->info.instrs_count,
 412                         so->info.max_half_reg + 1,
 413                         so->info.max_reg + 1);
 414
 415         debug_printf("; %d const, %u constlen\n",
 416                         so->info.max_const + 1,
 417                         so->constlen);
 418
 419         /* print shader type specific info: */
 420         switch (so->type) {
 421         case SHADER_VERTEX:
 422                 dump_output(so, VARYING_SLOT_POS, "pos");
 423                 dump_output(so, VARYING_SLOT_PSIZ, "psize");
 424                 break;
 425         case SHADER_FRAGMENT:
 426                 dump_reg("pos (bary)", so->pos_regid);
 427                 dump_output(so, FRAG_RESULT_DEPTH, "posz");
 428                 if (so->color0_mrt) {
 429                         dump_output(so, FRAG_RESULT_COLOR, "color");
 430                 } else {
 431                         dump_output(so, FRAG_RESULT_DATA0, "data0");
 432                         dump_output(so, FRAG_RESULT_DATA1, "data1");
 433                         dump_output(so, FRAG_RESULT_DATA2, "data2");
 434                         dump_output(so, FRAG_RESULT_DATA3, "data3");
 435                         dump_output(so, FRAG_RESULT_DATA4, "data4");
 436                         dump_output(so, FRAG_RESULT_DATA5, "data5");
 437                         dump_output(so, FRAG_RESULT_DATA6, "data6");
 438                         dump_output(so, FRAG_RESULT_DATA7, "data7");
 439                 }
 440                 /* these two are hard-coded since we don't know how to
 441                  * program them to anything but all 0's...
 442                  */
 443                 if (so->frag_coord)
 444                         debug_printf("; fragcoord: r0.x\n");
 445                 if (so->frag_face)
 446                         debug_printf("; fragface: hr0.x\n");
 447                 break;
 448         case SHADER_COMPUTE:
 449                 break;
 450         }
 451
 452         debug_printf("\n");
 453 }
 454
 455 /* This has to reach into the fd_context a bit more than the rest of
 456  * ir3, but it needs to be aligned with the compiler, so both agree
 457  * on which const regs hold what.  And the logic is identical between
 458  * a3xx/a4xx, the only difference is small details in the actual
 459  * CP_LOAD_STATE packets (which is handled inside the generation
 460  * specific ctx->emit_const(_bo)() fxns)
 461  */
 462
 463 #include "freedreno_resource.h"
 464
 465 static void
 466 emit_user_consts(struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
 467                 struct fd_constbuf_stateobj *constbuf)
 468 {
 469         struct fd_context *ctx = fd_context(v->shader->pctx);
 470         const unsigned index = 0;     /* user consts are index 0 */
 471         /* TODO save/restore dirty_mask for binning pass instead: */
 472         uint32_t dirty_mask = constbuf->enabled_mask;
 473
 474         if (dirty_mask & (1 << index)) {
 475                 struct pipe_constant_buffer *cb = &constbuf->cb[index];
 476                 unsigned size = align(cb->buffer_size, 4) / 4; /* size in dwords */
 477
 478                 /* in particular, with binning shader we may end up with
 479                  * unused consts, ie. we could end up w/ constlen that is
 480                  * smaller than first_driver_param.  In that case truncate
 481                  * the user consts early to avoid HLSQ lockup caused by
 482                  * writing too many consts
 483                  */
 484                 uint32_t max_const = MIN2(v->first_driver_param, v->constlen);
 485
 486                 // I expect that size should be a multiple of vec4's:
 487                 assert(size == align(size, 4));
 488
 489                 /* and even if the start of the const buffer is before
 490                  * first_immediate, the end may not be:
 491                  */
 492                 size = MIN2(size, 4 * max_const);
 493
 494                 if (size > 0) {
 495                         fd_wfi(ctx, ring);
 496                         ctx->emit_const(ring, v->type, 0,
 497                                         cb->buffer_offset, size,
 498                                         cb->user_buffer, cb->buffer);
 499                         constbuf->dirty_mask &= ~(1 << index);
 500                 }
 501         }
 502 }
 503
 504 static void
 505 emit_ubos(struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
 506                 struct fd_constbuf_stateobj *constbuf)
 507 {
 508         uint32_t offset = v->first_driver_param + IR3_UBOS_OFF;
 509         if (v->constlen > offset) {
 510                 struct fd_context *ctx = fd_context(v->shader->pctx);
 511                 uint32_t params = MIN2(4, v->constlen - offset) * 4;
 512                 uint32_t offsets[params];
 513                 struct fd_bo *bos[params];
 514
 515                 for (uint32_t i = 0; i < params; i++) {
 516                         const uint32_t index = i + 1;   /* UBOs start at index 1 */
 517                         struct pipe_constant_buffer *cb = &constbuf->cb[index];
 518                         assert(!cb->user_buffer);
 519
 520                         if ((constbuf->enabled_mask & (1 << index)) && cb->buffer) {
 521                                 offsets[i] = cb->buffer_offset;
 522                                 bos[i] = fd_resource(cb->buffer)->bo;
 523                         } else {
 524                                 offsets[i] = 0;
 525                                 bos[i] = NULL;
 526                         }
 527                 }
 528
 529                 fd_wfi(ctx, ring);
 530                 ctx->emit_const_bo(ring, v->type, false, offset * 4, params, bos, offsets);
 531         }
 532 }
 533
 534 static void
 535 emit_immediates(struct ir3_shader_variant *v, struct fd_ringbuffer *ring)
 536 {
 537         struct fd_context *ctx = fd_context(v->shader->pctx);
 538         int size = v->immediates_count;
 539         uint32_t base = v->first_immediate;
 540
 541         /* truncate size to avoid writing constants that shader
 542          * does not use:
 543          */
 544         size = MIN2(size + base, v->constlen) - base;
 545
 546         /* convert out of vec4: */
 547         base *= 4;
 548         size *= 4;
 549
 550         if (size > 0) {
 551                 fd_wfi(ctx, ring);
 552                 ctx->emit_const(ring, v->type, base,
 553                         0, size, v->immediates[0].val, NULL);
 554         }
 555 }
 556
 557 /* emit stream-out buffers: */
 558 static void
 559 emit_tfbos(struct ir3_shader_variant *v, struct fd_ringbuffer *ring)
 560 {
 561         /* streamout addresses after driver-params: */
 562         uint32_t offset = v->first_driver_param + IR3_TFBOS_OFF;
 563         if (v->constlen > offset) {
 564                 struct fd_context *ctx = fd_context(v->shader->pctx);
 565                 struct fd_streamout_stateobj *so = &ctx->streamout;
 566                 struct pipe_stream_output_info *info = &v->shader->stream_output;
 567                 uint32_t params = 4;
 568                 uint32_t offsets[params];
 569                 struct fd_bo *bos[params];
 570
 571                 for (uint32_t i = 0; i < params; i++) {
 572                         struct pipe_stream_output_target *target = so->targets[i];
 573
 574                         if (target) {
 575                                 offsets[i] = (so->offsets[i] * info->stride[i] * 4) +
 576                                                 target->buffer_offset;
 577                                 bos[i] = fd_resource(target->buffer)->bo;
 578                         } else {
 579                                 offsets[i] = 0;
 580                                 bos[i] = NULL;
 581                         }
 582                 }
 583
 584                 fd_wfi(ctx, ring);
 585                 ctx->emit_const_bo(ring, v->type, true, offset * 4, params, bos, offsets);
 586         }
 587 }
 588
 589 static uint32_t
 590 max_tf_vtx(struct ir3_shader_variant *v)
 591 {
 592         struct fd_context *ctx = fd_context(v->shader->pctx);
 593         struct fd_streamout_stateobj *so = &ctx->streamout;
 594         struct pipe_stream_output_info *info = &v->shader->stream_output;
 595         uint32_t maxvtxcnt = 0x7fffffff;
 596
 597         if (v->key.binning_pass)
 598                 return 0;
 599         if (v->shader->stream_output.num_outputs == 0)
 600                 return 0;
 601         if (so->num_targets == 0)
 602                 return 0;
 603
 604         /* offset to write to is:
 605          *
 606          *   total_vtxcnt = vtxcnt + offsets[i]
 607          *   offset = total_vtxcnt * stride[i]
 608          *
 609          *   offset =   vtxcnt * stride[i]       ; calculated in shader
 610          *            + offsets[i] * stride[i]   ; calculated at emit_tfbos()
 611          *
 612          * assuming for each vtx, each target buffer will have data written
 613          * up to 'offset + stride[i]', that leaves maxvtxcnt as:
 614          *
 615          *   buffer_size = (maxvtxcnt * stride[i]) + stride[i]
 616          *   maxvtxcnt   = (buffer_size - stride[i]) / stride[i]
 617          *
 618          * but shader is actually doing a less-than (rather than less-than-
 619          * equal) check, so we can drop the -stride[i].
 620          *
 621          * TODO is assumption about `offset + stride[i]` legit?
 622          */
 623         for (unsigned i = 0; i < so->num_targets; i++) {
 624                 struct pipe_stream_output_target *target = so->targets[i];
 625                 unsigned stride = info->stride[i] * 4;   /* convert dwords->bytes */
 626                 if (target) {
 627                         uint32_t max = target->buffer_size / stride;
 628                         maxvtxcnt = MIN2(maxvtxcnt, max);
 629                 }
 630         }
 631
 632         return maxvtxcnt;
 633 }
 634
 635 void
 636 ir3_emit_consts(struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
 637                 const struct pipe_draw_info *info, uint32_t dirty)
 638 {
 639         struct fd_context *ctx = fd_context(v->shader->pctx);
 640
 641         if (dirty & (FD_DIRTY_PROG | FD_DIRTY_CONSTBUF)) {
 642                 struct fd_constbuf_stateobj *constbuf;
 643                 bool shader_dirty;
 644
 645                 if (v->type == SHADER_VERTEX) {
 646                         constbuf = &ctx->constbuf[PIPE_SHADER_VERTEX];
 647                         shader_dirty = !!(ctx->prog.dirty & FD_SHADER_DIRTY_VP);
 648                 } else if (v->type == SHADER_FRAGMENT) {
 649                         constbuf = &ctx->constbuf[PIPE_SHADER_FRAGMENT];
 650                         shader_dirty = !!(ctx->prog.dirty & FD_SHADER_DIRTY_FP);
 651                 } else {
 652                         unreachable("bad shader type");
 653                         return;
 654                 }
 655
 656                 emit_user_consts(v, ring, constbuf);
 657                 emit_ubos(v, ring, constbuf);
 658                 if (shader_dirty)
 659                         emit_immediates(v, ring);
 660         }
 661
 662         /* emit driver params every time: */
 663         /* TODO skip emit if shader doesn't use driver params to avoid WFI.. */
 664         if (info && (v->type == SHADER_VERTEX)) {
 665                 uint32_t offset = v->first_driver_param + IR3_DRIVER_PARAM_OFF;
 666                 if (v->constlen >= offset) {
 667                         uint32_t vertex_params[IR3_DP_COUNT] = {
 668                                 [IR3_DP_VTXID_BASE] = info->indexed ?
 669                                                 info->index_bias : info->start,
 670                                 [IR3_DP_VTXCNT_MAX] = max_tf_vtx(v),
 671                         };
 672                         /* if no user-clip-planes, we don't need to emit the
 673                          * entire thing:
 674                          */
 675                         uint32_t vertex_params_size = 4;
 676
 677                         if (v->key.ucp_enables) {
 678                                 struct pipe_clip_state *ucp = &ctx->ucp;
 679                                 unsigned pos = IR3_DP_UCP0_X;
 680                                 for (unsigned i = 0; pos <= IR3_DP_UCP7_W; i++) {
 681                                         for (unsigned j = 0; j < 4; j++) {
 682                                                 vertex_params[pos] = fui(ucp->ucp[i][j]);
 683                                                 pos++;
 684                                         }
 685                                 }
 686                                 vertex_params_size = ARRAY_SIZE(vertex_params);
 687                         }
 688
 689                         fd_wfi(ctx, ring);
 690                         ctx->emit_const(ring, SHADER_VERTEX, offset * 4, 0,
 691                                         vertex_params_size, vertex_params, NULL);
 692
 693                         /* if needed, emit stream-out buffer addresses: */
 694                         if (vertex_params[IR3_DP_VTXCNT_MAX] > 0) {
 695                                 emit_tfbos(v, ring);
 696                         }
 697                 }
 698         }
 699 }