src/gallium/drivers/nv50/nv50_program.c

   1 /*
   2  * Copyright 2010 Chrsitoph Bumiller
   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 shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  */
  22
  23 /* #define NV50_PROGRAM_DEBUG */
  24
  25 #include "nv50_program.h"
  26 #include "nv50_pc.h"
  27 #include "nv50_context.h"
  28
  29 #include "pipe/p_shader_tokens.h"
  30 #include "tgsi/tgsi_parse.h"
  31 #include "tgsi/tgsi_util.h"
  32 #include "tgsi/tgsi_dump.h"
  33
  34 static INLINE unsigned
  35 bitcount4(const uint32_t val)
  36 {
  37    static const unsigned cnt[16]
  38    = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
  39    return cnt[val & 0xf];
  40 }
  41
  42 static unsigned
  43 nv50_tgsi_src_mask(const struct tgsi_full_instruction *inst, int c)
  44 {
  45    unsigned mask = inst->Dst[0].Register.WriteMask;
  46
  47    switch (inst->Instruction.Opcode) {
  48    case TGSI_OPCODE_COS:
  49    case TGSI_OPCODE_SIN:
  50       return (mask & 0x8) | ((mask & 0x7) ? 0x1 : 0x0);
  51    case TGSI_OPCODE_DP3:
  52       return 0x7;
  53    case TGSI_OPCODE_DP4:
  54    case TGSI_OPCODE_DPH:
  55    case TGSI_OPCODE_KIL: /* WriteMask ignored */
  56       return 0xf;
  57    case TGSI_OPCODE_DST:
  58       return mask & (c ? 0xa : 0x6);
  59    case TGSI_OPCODE_EX2:
  60    case TGSI_OPCODE_EXP:
  61    case TGSI_OPCODE_LG2:
  62    case TGSI_OPCODE_LOG:
  63    case TGSI_OPCODE_POW:
  64    case TGSI_OPCODE_RCP:
  65    case TGSI_OPCODE_RSQ:
  66    case TGSI_OPCODE_SCS:
  67       return 0x1;
  68    case TGSI_OPCODE_IF:
  69       return 0x1;
  70    case TGSI_OPCODE_LIT:
  71       return 0xb;
  72    case TGSI_OPCODE_TEX:
  73    case TGSI_OPCODE_TXB:
  74    case TGSI_OPCODE_TXL:
  75    case TGSI_OPCODE_TXP:
  76    {
  77       const struct tgsi_instruction_texture *tex;
  78
  79       assert(inst->Instruction.Texture);
  80       tex = &inst->Texture;
  81
  82       mask = 0x7;
  83       if (inst->Instruction.Opcode != TGSI_OPCODE_TEX &&
  84           inst->Instruction.Opcode != TGSI_OPCODE_TXD)
  85          mask |= 0x8; /* bias, lod or proj */
  86
  87       switch (tex->Texture) {
  88       case TGSI_TEXTURE_1D:
  89          mask &= 0x9;
  90          break;
  91       case TGSI_TEXTURE_SHADOW1D:
  92          mask &= 0x5;
  93          break;
  94       case TGSI_TEXTURE_2D:
  95          mask &= 0xb;
  96          break;
  97       default:
  98          break;
  99       }
 100    }
 101            return mask;
 102    case TGSI_OPCODE_XPD:
 103    {
 104       unsigned x = 0;
 105       if (mask & 1) x |= 0x6;
 106       if (mask & 2) x |= 0x5;
 107       if (mask & 4) x |= 0x3;
 108       return x;
 109    }
 110    default:
 111       break;
 112    }
 113
 114    return mask;
 115 }
 116
 117 static void
 118 nv50_indirect_inputs(struct nv50_translation_info *ti, int id)
 119 {
 120    int i, c;
 121
 122    for (i = 0; i < PIPE_MAX_SHADER_INPUTS; ++i)
 123       for (c = 0; c < 4; ++c)
 124          ti->input_access[i][c] = id;
 125
 126    ti->indirect_inputs = TRUE;
 127 }
 128
 129 static void
 130 nv50_indirect_outputs(struct nv50_translation_info *ti, int id)
 131 {
 132    int i, c;
 133
 134    for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; ++i)
 135       for (c = 0; c < 4; ++c)
 136          ti->output_access[i][c] = id;
 137
 138    ti->indirect_outputs = TRUE;
 139 }
 140
 141 static void
 142 prog_inst(struct nv50_translation_info *ti,
 143           const struct tgsi_full_instruction *inst, int id)
 144 {
 145    const struct tgsi_dst_register *dst;
 146    const struct tgsi_src_register *src;
 147    int s, c, k;
 148    unsigned mask;
 149
 150    if (inst->Dst[0].Register.File == TGSI_FILE_OUTPUT) {
 151       for (c = 0; c < 4; ++c) {
 152          dst = &inst->Dst[0].Register;
 153          if (inst->Dst[0].Register.Indirect)
 154             nv50_indirect_outputs(ti, id);
 155          if (!(dst->WriteMask & (1 << c)))
 156             continue;
 157          ti->output_access[dst->Index][c] = id;
 158       }
 159
 160       if (inst->Instruction.Opcode == TGSI_OPCODE_MOV &&
 161           inst->Src[0].Register.File == TGSI_FILE_INPUT &&
 162           dst->Index == ti->edgeflag_out)
 163          ti->p->vp.edgeflag = inst->Src[0].Register.Index;
 164    }
 165
 166    for (s = 0; s < inst->Instruction.NumSrcRegs; ++s) {
 167       src = &inst->Src[s].Register;
 168       if (src->File != TGSI_FILE_INPUT)
 169          continue;
 170       mask = nv50_tgsi_src_mask(inst, s);
 171
 172       if (inst->Src[s].Register.Indirect)
 173          nv50_indirect_inputs(ti, id);
 174
 175       for (c = 0; c < 4; ++c) {
 176          if (!(mask & (1 << c)))
 177             continue;
 178          k = tgsi_util_get_full_src_register_swizzle(&inst->Src[s], c);
 179          if (k <= TGSI_SWIZZLE_W)
 180             ti->input_access[src->Index][k] = id;
 181       }
 182    }
 183 }
 184
 185 static void
 186 prog_immediate(struct nv50_translation_info *ti,
 187                const struct tgsi_full_immediate *imm)
 188 {
 189    int c;
 190    unsigned n = ti->immd32_nr++;
 191
 192    assert(ti->immd32_nr <= ti->scan.immediate_count);
 193
 194    for (c = 0; c < 4; ++c)
 195       ti->immd32[n * 4 + c] = imm->u[c].Uint;
 196
 197    ti->immd32_ty[n] = imm->Immediate.DataType;
 198 }
 199
 200 static INLINE unsigned
 201 translate_interpolate(const struct tgsi_full_declaration *decl)
 202 {
 203    unsigned mode;
 204
 205    if (decl->Declaration.Interpolate == TGSI_INTERPOLATE_CONSTANT)
 206       mode = NV50_INTERP_FLAT;
 207    else
 208    if (decl->Declaration.Interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
 209       mode = 0;
 210    else
 211       mode = NV50_INTERP_LINEAR;
 212
 213    if (decl->Declaration.Centroid)
 214       mode |= NV50_INTERP_CENTROID;
 215
 216    return mode;
 217 }
 218
 219 static void
 220 prog_decl(struct nv50_translation_info *ti,
 221           const struct tgsi_full_declaration *decl)
 222 {
 223    unsigned i, first, last, sn = 0, si = 0;
 224
 225    first = decl->Range.First;
 226    last = decl->Range.Last;
 227
 228    if (decl->Declaration.Semantic) {
 229       sn = decl->Semantic.Name;
 230       si = decl->Semantic.Index;
 231    }
 232
 233    switch (decl->Declaration.File) {
 234    case TGSI_FILE_INPUT:
 235       for (i = first; i <= last; ++i)
 236          ti->interp_mode[i] = translate_interpolate(decl);
 237
 238       if (!decl->Declaration.Semantic)
 239          break;
 240
 241       for (i = first; i <= last; ++i) {
 242          ti->p->in[i].sn = sn;
 243          ti->p->in[i].si = si;
 244       }
 245
 246       switch (sn) {
 247       case TGSI_SEMANTIC_FACE:
 248          break;
 249       case TGSI_SEMANTIC_COLOR:
 250          if (ti->p->type == PIPE_SHADER_FRAGMENT)
 251             ti->p->vp.bfc[si] = first;
 252          break;
 253       }
 254       break;
 255    case TGSI_FILE_OUTPUT:
 256       if (!decl->Declaration.Semantic)
 257          break;
 258
 259       for (i = first; i <= last; ++i) {
 260          ti->p->out[i].sn = sn;
 261          ti->p->out[i].si = si;
 262       }
 263
 264       switch (sn) {
 265       case TGSI_SEMANTIC_BCOLOR:
 266          ti->p->vp.bfc[si] = first;
 267          break;
 268       case TGSI_SEMANTIC_PSIZE:
 269          ti->p->vp.psiz = first;
 270          break;
 271       case TGSI_SEMANTIC_EDGEFLAG:
 272          ti->edgeflag_out = first;
 273          break;
 274       default:
 275          break;
 276       }
 277       break;
 278    case TGSI_FILE_SYSTEM_VALUE:
 279       switch (decl->Semantic.Name) {
 280       case TGSI_SEMANTIC_FACE:
 281          break;
 282       case TGSI_SEMANTIC_INSTANCEID:
 283          break;
 284       case TGSI_SEMANTIC_PRIMID:
 285          break;
 286          /*
 287       case TGSI_SEMANTIC_PRIMIDIN:
 288          break;
 289       case TGSI_SEMANTIC_VERTEXID:
 290          break;
 291          */
 292       default:
 293          break;
 294       }
 295       break;
 296    case TGSI_FILE_CONSTANT:
 297       ti->p->parm_size = MAX2(ti->p->parm_size, (last + 1) * 16);
 298       break;
 299    case TGSI_FILE_ADDRESS:
 300    case TGSI_FILE_SAMPLER:
 301    case TGSI_FILE_TEMPORARY:
 302       break;
 303    default:
 304       assert(0);
 305       break;
 306    }
 307 }
 308
 309 static int
 310 nv50_vertprog_prepare(struct nv50_translation_info *ti)
 311 {
 312    struct nv50_program *p = ti->p;
 313    int i, c;
 314    unsigned num_inputs = 0;
 315
 316    ti->input_file = NV_FILE_MEM_S;
 317    ti->output_file = NV_FILE_OUT;
 318
 319    for (i = 0; i <= ti->scan.file_max[TGSI_FILE_INPUT]; ++i) {
 320       p->in[i].id = i;
 321       p->in[i].hw = num_inputs;
 322
 323       for (c = 0; c < 4; ++c) {
 324          if (!ti->input_access[i][c])
 325             continue;
 326          ti->input_map[i][c] = num_inputs++;
 327          p->vp.attrs[(4 * i + c) / 32] |= 1 << ((i * 4 + c) % 32);
 328       }
 329    }
 330
 331    for (i = 0; i <= ti->scan.file_max[TGSI_FILE_OUTPUT]; ++i) {
 332       p->out[i].id = i;
 333       p->out[i].hw = p->max_out;
 334
 335       for (c = 0; c < 4; ++c) {
 336          if (!ti->output_access[i][c])
 337             continue;
 338          ti->output_map[i][c] = p->max_out++;
 339          p->out[i].mask |= 1 << c;
 340       }
 341    }
 342
 343    if (p->vp.psiz < 0x40)
 344       p->vp.psiz = p->out[p->vp.psiz].hw;
 345
 346    return 0;
 347 }
 348
 349 static int
 350 nv50_fragprog_prepare(struct nv50_translation_info *ti)
 351 {
 352    struct nv50_program *p = ti->p;
 353    int i, j, c;
 354    unsigned nvary, nintp, depr;
 355    unsigned n = 0, m = 0, skip = 0;
 356    ubyte sn[16], si[16];
 357
 358    /* FP flags */
 359
 360    if (ti->scan.writes_z) {
 361       p->fp.flags[1] = 0x11;
 362       p->fp.flags[0] |= NV50TCL_FP_CONTROL_EXPORTS_Z;
 363    }
 364
 365    if (ti->scan.uses_kill)
 366       p->fp.flags[0] |= NV50TCL_FP_CONTROL_USES_KIL;
 367
 368    /* FP inputs */
 369
 370    ti->input_file = NV_FILE_MEM_V;
 371    ti->output_file = NV_FILE_GPR;
 372
 373    /* count non-flat inputs, save semantic info */
 374    for (i = 0; i < p->in_nr; ++i) {
 375       m += (ti->interp_mode[i] & NV50_INTERP_FLAT) ? 0 : 1;
 376       sn[i] = p->in[i].sn;
 377       si[i] = p->in[i].si;
 378    }
 379
 380    /* reorder p->in[] so that non-flat inputs are first and
 381     * kick out special inputs that don't use VP/GP_RESULT_MAP
 382     */
 383    nintp = 0;
 384    for (i = 0; i < p->in_nr; ++i) {
 385       if (sn[i] == TGSI_SEMANTIC_POSITION) {
 386          for (c = 0; c < 4; ++c) {
 387             ti->input_map[i][c] = nintp;
 388             if (ti->input_access[i][c]) {
 389                p->fp.interp |= 1 << (24 + c);
 390                ++nintp;
 391             }
 392          }
 393          skip++;
 394          continue;
 395       } else
 396       if (sn[i] == TGSI_SEMANTIC_FACE) {
 397          ti->input_map[i][0] = 255;
 398          skip++;
 399          continue;
 400       }
 401
 402       j = (ti->interp_mode[i] & NV50_INTERP_FLAT) ? m++ : n++;
 403
 404       if (sn[i] == TGSI_SEMANTIC_COLOR)
 405          p->vp.bfc[si[i]] = j;
 406
 407       p->in[j].linear = (ti->interp_mode[i] & NV50_INTERP_LINEAR) ? 1 : 0;
 408       p->in[j].id = i;
 409       p->in[j].sn = sn[i];
 410       p->in[j].si = si[i];
 411    }
 412    assert(n <= m);
 413    p->in_nr -= skip;
 414
 415    if (!(p->fp.interp & (8 << 24))) {
 416       p->fp.interp |= (8 << 24);
 417       ++nintp;
 418    }
 419
 420    p->fp.colors = (1 << 24) | 4; /* CLAMP, FFC0_ID = 4 */
 421
 422    for (i = 0; i < p->in_nr; ++i) {
 423       int j = p->in[i].id;
 424       p->in[i].hw = nintp;
 425
 426       for (c = 0; c < 4; ++c) {
 427          if (!ti->input_access[j][c])
 428             continue;
 429          p->in[i].mask |= 1 << c;
 430          ti->input_map[j][c] = nintp++;
 431       }
 432       /* count color inputs */
 433       if (i == p->vp.bfc[0] || i == p->vp.bfc[1])
 434          p->fp.colors += bitcount4(p->in[i].mask) << 16;
 435    }
 436    nintp -= bitcount4(p->fp.interp >> 24); /* subtract position inputs */
 437    nvary = nintp;
 438    if (n < m)
 439       nvary -= p->in[n].hw;
 440
 441    p->fp.interp |= nvary << NV50TCL_FP_INTERPOLANT_CTRL_COUNT_NONFLAT_SHIFT;
 442    p->fp.interp |= nintp << NV50TCL_FP_INTERPOLANT_CTRL_COUNT_SHIFT;
 443
 444    /* FP outputs */
 445
 446    if (p->out_nr > (1 + (ti->scan.writes_z ? 1 : 0)))
 447       p->fp.flags[0] |= NV50TCL_FP_CONTROL_MULTIPLE_RESULTS;
 448
 449    depr = p->out_nr;
 450    for (i = 0; i < p->out_nr; ++i) {
 451       p->out[i].id = i;
 452       if (p->out[i].sn == TGSI_SEMANTIC_POSITION) {
 453          depr = i;
 454          continue;
 455       }
 456       p->out[i].hw = p->max_out;
 457       p->out[i].mask = 0xf;
 458
 459       for (c = 0; c < 4; ++c)
 460          ti->output_map[i][c] = p->max_out++;
 461    }
 462    if (depr < p->out_nr) {
 463       p->out[depr].mask = 0x4;
 464       p->out[depr].hw = ti->output_map[depr][2] = p->max_out++;
 465    }
 466
 467    return 0;
 468 }
 469
 470 static int
 471 nv50_geomprog_prepare(struct nv50_translation_info *ti)
 472 {
 473    ti->input_file = NV_FILE_MEM_S;
 474    ti->output_file = NV_FILE_OUT;
 475
 476    assert(0);
 477    return 1;
 478 }
 479
 480 static int
 481 nv50_prog_scan(struct nv50_translation_info *ti)
 482 {
 483    struct nv50_program *p = ti->p;
 484    struct tgsi_parse_context parse;
 485    int ret;
 486
 487    p->vp.edgeflag = 0x40;
 488    p->vp.psiz = 0x40;
 489    p->vp.bfc[0] = 0x40;
 490    p->vp.bfc[1] = 0x40;
 491    p->gp.primid = 0x80;
 492
 493    tgsi_scan_shader(p->pipe.tokens, &ti->scan);
 494
 495 #ifdef NV50_PROGRAM_DEBUG
 496    tgsi_dump(p->pipe.tokens, 0);
 497 #endif
 498
 499    ti->immd32 = (uint32_t *)MALLOC(ti->scan.immediate_count * 16);
 500    ti->immd32_ty = (ubyte *)MALLOC(ti->scan.immediate_count * sizeof(ubyte));
 501
 502    tgsi_parse_init(&parse, p->pipe.tokens);
 503    while (!tgsi_parse_end_of_tokens(&parse)) {
 504       tgsi_parse_token(&parse);
 505
 506       switch (parse.FullToken.Token.Type) {
 507       case TGSI_TOKEN_TYPE_IMMEDIATE:
 508          prog_immediate(ti, &parse.FullToken.FullImmediate);
 509          break;
 510       case TGSI_TOKEN_TYPE_DECLARATION:
 511          prog_decl(ti, &parse.FullToken.FullDeclaration);
 512          break;
 513       case TGSI_TOKEN_TYPE_INSTRUCTION:
 514          prog_inst(ti, &parse.FullToken.FullInstruction, ++ti->inst_nr);
 515          break;
 516       }
 517    }
 518
 519    p->in_nr = ti->scan.file_max[TGSI_FILE_INPUT] + 1;
 520    p->out_nr = ti->scan.file_max[TGSI_FILE_OUTPUT] + 1;
 521
 522    switch (p->type) {
 523    case PIPE_SHADER_VERTEX:
 524       ret = nv50_vertprog_prepare(ti);
 525       break;
 526    case PIPE_SHADER_FRAGMENT:
 527       ret = nv50_fragprog_prepare(ti);
 528       break;
 529    case PIPE_SHADER_GEOMETRY:
 530       ret = nv50_geomprog_prepare(ti);
 531       break;
 532    default:
 533       assert(!"unsupported program type");
 534       ret = -1;
 535       break;
 536    }
 537
 538    assert(!ret);
 539    return ret;
 540 }
 541
 542 boolean
 543 nv50_program_tx(struct nv50_program *p)
 544 {
 545    struct nv50_translation_info *ti;
 546    int ret;
 547
 548    ti = CALLOC_STRUCT(nv50_translation_info);
 549    ti->p = p;
 550
 551    ti->edgeflag_out = PIPE_MAX_SHADER_OUTPUTS;
 552
 553    ret = nv50_prog_scan(ti);
 554    if (ret) {
 555       NOUVEAU_ERR("unsupported shader program\n");
 556       goto out;
 557    }
 558
 559    ret = nv50_generate_code(ti);
 560    if (ret) {
 561       NOUVEAU_ERR("error during shader translation\n");
 562       goto out;
 563    }
 564
 565 out:
 566    if (ti->immd32)
 567       FREE(ti->immd32);
 568    if (ti->immd32_ty)
 569       FREE(ti->immd32_ty);
 570    FREE(ti);
 571    return ret ? FALSE : TRUE;
 572 }
 573
 574 void
 575 nv50_program_destroy(struct nv50_context *nv50, struct nv50_program *p)
 576 {
 577    nouveau_bo_ref(NULL, &p->bo);
 578
 579    so_ref(NULL, &p->so);
 580
 581    if (p->code)
 582       FREE(p->code);
 583
 584    p->translated = FALSE;
 585 }