src/mesa/drivers/dri/i915/i915_program.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include <strings.h>
  29
  30 #include "main/glheader.h"
  31 #include "main/macros.h"
  32 #include "main/enums.h"
  33
  34 #include "tnl/t_context.h"
  35 #include "intel_batchbuffer.h"
  36
  37 #include "i915_reg.h"
  38 #include "i915_context.h"
  39 #include "i915_program.h"
  40
  41
  42 #define A0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
  43 #define D0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
  44 #define T0_DEST( reg ) (((reg)&UREG_TYPE_NR_MASK)>>UREG_A0_DEST_SHIFT_LEFT)
  45 #define A0_SRC0( reg ) (((reg)&UREG_MASK)>>UREG_A0_SRC0_SHIFT_LEFT)
  46 #define A1_SRC0( reg ) (((reg)&UREG_MASK)<<UREG_A1_SRC0_SHIFT_RIGHT)
  47 #define A1_SRC1( reg ) (((reg)&UREG_MASK)>>UREG_A1_SRC1_SHIFT_LEFT)
  48 #define A2_SRC1( reg ) (((reg)&UREG_MASK)<<UREG_A2_SRC1_SHIFT_RIGHT)
  49 #define A2_SRC2( reg ) (((reg)&UREG_MASK)>>UREG_A2_SRC2_SHIFT_LEFT)
  50
  51 /* These are special, and don't have swizzle/negate bits.
  52  */
  53 #define T0_SAMPLER( reg )     (GET_UREG_NR(reg)<<T0_SAMPLER_NR_SHIFT)
  54 #define T1_ADDRESS_REG( reg ) ((GET_UREG_NR(reg)<<T1_ADDRESS_REG_NR_SHIFT) | \
  55                                (GET_UREG_TYPE(reg)<<T1_ADDRESS_REG_TYPE_SHIFT))
  56
  57
  58 /* Macros for translating UREG's into the various register fields used
  59  * by the I915 programmable unit.
  60  */
  61 #define UREG_A0_DEST_SHIFT_LEFT  (UREG_TYPE_SHIFT - A0_DEST_TYPE_SHIFT)
  62 #define UREG_A0_SRC0_SHIFT_LEFT  (UREG_TYPE_SHIFT - A0_SRC0_TYPE_SHIFT)
  63 #define UREG_A1_SRC0_SHIFT_RIGHT (A1_SRC0_CHANNEL_W_SHIFT - UREG_CHANNEL_W_SHIFT)
  64 #define UREG_A1_SRC1_SHIFT_LEFT  (UREG_TYPE_SHIFT - A1_SRC1_TYPE_SHIFT)
  65 #define UREG_A2_SRC1_SHIFT_RIGHT (A2_SRC1_CHANNEL_W_SHIFT - UREG_CHANNEL_W_SHIFT)
  66 #define UREG_A2_SRC2_SHIFT_LEFT  (UREG_TYPE_SHIFT - A2_SRC2_TYPE_SHIFT)
  67
  68 #define UREG_MASK         0xffffff00
  69 #define UREG_TYPE_NR_MASK ((REG_TYPE_MASK << UREG_TYPE_SHIFT) | \
  70                            (REG_NR_MASK << UREG_NR_SHIFT))
  71
  72
  73 #define I915_CONSTFLAG_PARAM 0x1f
  74
  75 GLuint
  76 i915_get_temp(struct i915_fragment_program *p)
  77 {
  78    int bit = ffs(~p->temp_flag);
  79    if (!bit) {
  80       fprintf(stderr, "%s: out of temporaries\n", __FILE__);
  81       exit(1);
  82    }
  83
  84    p->temp_flag |= 1 << (bit - 1);
  85    return UREG(REG_TYPE_R, (bit - 1));
  86 }
  87
  88
  89 GLuint
  90 i915_get_utemp(struct i915_fragment_program * p)
  91 {
  92    int bit = ffs(~p->utemp_flag);
  93    if (!bit) {
  94       fprintf(stderr, "%s: out of temporaries\n", __FILE__);
  95       exit(1);
  96    }
  97
  98    p->utemp_flag |= 1 << (bit - 1);
  99    return UREG(REG_TYPE_U, (bit - 1));
 100 }
 101
 102 void
 103 i915_release_utemps(struct i915_fragment_program *p)
 104 {
 105    p->utemp_flag = ~0x7;
 106 }
 107
 108
 109 GLuint
 110 i915_emit_decl(struct i915_fragment_program *p,
 111                GLuint type, GLuint nr, GLuint d0_flags)
 112 {
 113    GLuint reg = UREG(type, nr);
 114
 115    if (type == REG_TYPE_T) {
 116       if (p->decl_t & (1 << nr))
 117          return reg;
 118
 119       p->decl_t |= (1 << nr);
 120    }
 121    else if (type == REG_TYPE_S) {
 122       if (p->decl_s & (1 << nr))
 123          return reg;
 124
 125       p->decl_s |= (1 << nr);
 126    }
 127    else
 128       return reg;
 129
 130    *(p->decl++) = (D0_DCL | D0_DEST(reg) | d0_flags);
 131    *(p->decl++) = D1_MBZ;
 132    *(p->decl++) = D2_MBZ;
 133
 134    p->nr_decl_insn++;
 135    return reg;
 136 }
 137
 138 GLuint
 139 i915_emit_arith(struct i915_fragment_program * p,
 140                 GLuint op,
 141                 GLuint dest,
 142                 GLuint mask,
 143                 GLuint saturate, GLuint src0, GLuint src1, GLuint src2)
 144 {
 145    GLuint c[3];
 146    GLuint nr_const = 0;
 147
 148    assert(GET_UREG_TYPE(dest) != REG_TYPE_CONST);
 149    dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest));
 150    assert(dest);
 151
 152    if (GET_UREG_TYPE(src0) == REG_TYPE_CONST)
 153       c[nr_const++] = 0;
 154    if (GET_UREG_TYPE(src1) == REG_TYPE_CONST)
 155       c[nr_const++] = 1;
 156    if (GET_UREG_TYPE(src2) == REG_TYPE_CONST)
 157       c[nr_const++] = 2;
 158
 159    /* Recursively call this function to MOV additional const values
 160     * into temporary registers.  Use utemp registers for this -
 161     * currently shouldn't be possible to run out, but keep an eye on
 162     * this.
 163     */
 164    if (nr_const > 1) {
 165       GLuint s[3], first, i, old_utemp_flag;
 166
 167       s[0] = src0;
 168       s[1] = src1;
 169       s[2] = src2;
 170       old_utemp_flag = p->utemp_flag;
 171
 172       first = GET_UREG_NR(s[c[0]]);
 173       for (i = 1; i < nr_const; i++) {
 174          if (GET_UREG_NR(s[c[i]]) != first) {
 175             GLuint tmp = i915_get_utemp(p);
 176
 177             i915_emit_arith(p, A0_MOV, tmp, A0_DEST_CHANNEL_ALL, 0,
 178                             s[c[i]], 0, 0);
 179             s[c[i]] = tmp;
 180          }
 181       }
 182
 183       src0 = s[0];
 184       src1 = s[1];
 185       src2 = s[2];
 186       p->utemp_flag = old_utemp_flag;   /* restore */
 187    }
 188
 189    *(p->csr++) = (op | A0_DEST(dest) | mask | saturate | A0_SRC0(src0));
 190    *(p->csr++) = (A1_SRC0(src0) | A1_SRC1(src1));
 191    *(p->csr++) = (A2_SRC1(src1) | A2_SRC2(src2));
 192
 193    if (GET_UREG_TYPE(dest) == REG_TYPE_R)
 194       p->register_phases[GET_UREG_NR(dest)] = p->nr_tex_indirect;
 195
 196    p->nr_alu_insn++;
 197    return dest;
 198 }
 199
 200 static GLuint get_free_rreg (struct i915_fragment_program *p,
 201                              GLuint live_regs)
 202 {
 203     int bit = ffs(~live_regs);
 204     if (!bit) {
 205         i915_program_error(p, "Can't find free R reg");
 206         return UREG_BAD;
 207     }
 208     return UREG(REG_TYPE_R, bit - 1);
 209 }
 210
 211 GLuint i915_emit_texld( struct i915_fragment_program *p,
 212                         GLuint live_regs,
 213                         GLuint dest,
 214                         GLuint destmask,
 215                         GLuint sampler,
 216                         GLuint coord,
 217                         GLuint op )
 218 {
 219     if (coord != UREG(GET_UREG_TYPE(coord), GET_UREG_NR(coord))) {
 220         /* With the help of the "needed registers" table created earlier, pick
 221          * a register we can MOV the swizzled TC to (since TEX doesn't support
 222          * swizzled sources) */
 223         GLuint swizCoord = get_free_rreg(p, live_regs);
 224         if (swizCoord == UREG_BAD)
 225             return 0;
 226
 227         i915_emit_arith( p, A0_MOV, swizCoord, A0_DEST_CHANNEL_ALL, 0, coord, 0, 0 );
 228         coord = swizCoord;
 229     }
 230
 231    /* Don't worry about saturate as we only support texture formats
 232     * that are always in the 0..1 range.
 233     */
 234    if (destmask != A0_DEST_CHANNEL_ALL) {
 235       GLuint tmp = i915_get_utemp(p);
 236       i915_emit_texld( p, 0, tmp, A0_DEST_CHANNEL_ALL, sampler, coord, op );
 237       i915_emit_arith( p, A0_MOV, dest, destmask, 0, tmp, 0, 0 );
 238       return dest;
 239    }
 240    else {
 241       assert(GET_UREG_TYPE(dest) != REG_TYPE_CONST);
 242       assert(dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest)));
 243       /* Can't use unsaved temps for coords, as the phase boundary would result
 244        * in the contents becoming undefined.
 245        */
 246       assert(GET_UREG_TYPE(coord) != REG_TYPE_U);
 247
 248       /* Output register being oC or oD defines a phase boundary */
 249       if (GET_UREG_TYPE(dest) == REG_TYPE_OC ||
 250           GET_UREG_TYPE(dest) == REG_TYPE_OD)
 251          p->nr_tex_indirect++;
 252
 253       /* Reading from an r# register whose contents depend on output of the
 254        * current phase defines a phase boundary.
 255        */
 256       if (GET_UREG_TYPE(coord) == REG_TYPE_R &&
 257           p->register_phases[GET_UREG_NR(coord)] == p->nr_tex_indirect)
 258          p->nr_tex_indirect++;
 259
 260       *(p->csr++) = (op |
 261                      T0_DEST( dest ) |
 262                      T0_SAMPLER( sampler ));
 263
 264       *(p->csr++) = T1_ADDRESS_REG( coord );
 265       *(p->csr++) = T2_MBZ;
 266
 267       if (GET_UREG_TYPE(dest) == REG_TYPE_R)
 268          p->register_phases[GET_UREG_NR(dest)] = p->nr_tex_indirect;
 269
 270       p->nr_tex_insn++;
 271       return dest;
 272    }
 273 }
 274
 275
 276 GLuint
 277 i915_emit_const1f(struct i915_fragment_program * p, GLfloat c0)
 278 {
 279    GLint reg, idx;
 280
 281    if (c0 == 0.0)
 282       return swizzle(UREG(REG_TYPE_R, 0), ZERO, ZERO, ZERO, ZERO);
 283    if (c0 == 1.0)
 284       return swizzle(UREG(REG_TYPE_R, 0), ONE, ONE, ONE, ONE);
 285
 286    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 287       if (p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
 288          continue;
 289       for (idx = 0; idx < 4; idx++) {
 290          if (!(p->constant_flags[reg] & (1 << idx)) ||
 291              p->constant[reg][idx] == c0) {
 292             p->constant[reg][idx] = c0;
 293             p->constant_flags[reg] |= 1 << idx;
 294             if (reg + 1 > p->nr_constants)
 295                p->nr_constants = reg + 1;
 296             return swizzle(UREG(REG_TYPE_CONST, reg), idx, ZERO, ZERO, ONE);
 297          }
 298       }
 299    }
 300
 301    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 302    p->error = 1;
 303    return 0;
 304 }
 305
 306 GLuint
 307 i915_emit_const2f(struct i915_fragment_program * p, GLfloat c0, GLfloat c1)
 308 {
 309    GLint reg, idx;
 310
 311    if (c0 == 0.0)
 312       return swizzle(i915_emit_const1f(p, c1), ZERO, X, Z, W);
 313    if (c0 == 1.0)
 314       return swizzle(i915_emit_const1f(p, c1), ONE, X, Z, W);
 315
 316    if (c1 == 0.0)
 317       return swizzle(i915_emit_const1f(p, c0), X, ZERO, Z, W);
 318    if (c1 == 1.0)
 319       return swizzle(i915_emit_const1f(p, c0), X, ONE, Z, W);
 320
 321    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 322       if (p->constant_flags[reg] == 0xf ||
 323           p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
 324          continue;
 325       for (idx = 0; idx < 3; idx++) {
 326          if (!(p->constant_flags[reg] & (3 << idx))) {
 327             p->constant[reg][idx] = c0;
 328             p->constant[reg][idx + 1] = c1;
 329             p->constant_flags[reg] |= 3 << idx;
 330             if (reg + 1 > p->nr_constants)
 331                p->nr_constants = reg + 1;
 332             return swizzle(UREG(REG_TYPE_CONST, reg), idx, idx + 1, ZERO,
 333                            ONE);
 334          }
 335       }
 336    }
 337
 338    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 339    p->error = 1;
 340    return 0;
 341 }
 342
 343
 344
 345 GLuint
 346 i915_emit_const4f(struct i915_fragment_program * p,
 347                   GLfloat c0, GLfloat c1, GLfloat c2, GLfloat c3)
 348 {
 349    GLint reg;
 350
 351    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 352       if (p->constant_flags[reg] == 0xf &&
 353           p->constant[reg][0] == c0 &&
 354           p->constant[reg][1] == c1 &&
 355           p->constant[reg][2] == c2 && p->constant[reg][3] == c3) {
 356          return UREG(REG_TYPE_CONST, reg);
 357       }
 358       else if (p->constant_flags[reg] == 0) {
 359          p->constant[reg][0] = c0;
 360          p->constant[reg][1] = c1;
 361          p->constant[reg][2] = c2;
 362          p->constant[reg][3] = c3;
 363          p->constant_flags[reg] = 0xf;
 364          if (reg + 1 > p->nr_constants)
 365             p->nr_constants = reg + 1;
 366          return UREG(REG_TYPE_CONST, reg);
 367       }
 368    }
 369
 370    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 371    p->error = 1;
 372    return 0;
 373 }
 374
 375
 376 GLuint
 377 i915_emit_const4fv(struct i915_fragment_program * p, const GLfloat * c)
 378 {
 379    return i915_emit_const4f(p, c[0], c[1], c[2], c[3]);
 380 }
 381
 382
 383 GLuint
 384 i915_emit_param4fv(struct i915_fragment_program * p, const GLfloat * values)
 385 {
 386    GLint reg, i;
 387
 388    for (i = 0; i < p->nr_params; i++) {
 389       if (p->param[i].values == values)
 390          return UREG(REG_TYPE_CONST, p->param[i].reg);
 391    }
 392
 393
 394    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 395       if (p->constant_flags[reg] == 0) {
 396          p->constant_flags[reg] = I915_CONSTFLAG_PARAM;
 397          i = p->nr_params++;
 398
 399          p->param[i].values = values;
 400          p->param[i].reg = reg;
 401          p->params_uptodate = 0;
 402
 403          if (reg + 1 > p->nr_constants)
 404             p->nr_constants = reg + 1;
 405          return UREG(REG_TYPE_CONST, reg);
 406       }
 407    }
 408
 409    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 410    p->error = 1;
 411    return 0;
 412 }
 413
 414
 415
 416 void
 417 i915_program_error(struct i915_fragment_program *p, const char *msg)
 418 {
 419    _mesa_problem(NULL, "i915_program_error: %s", msg);
 420    p->error = 1;
 421 }
 422
 423
 424 void
 425 i915_init_program(struct i915_context *i915, struct i915_fragment_program *p)
 426 {
 427    GLcontext *ctx = &i915->intel.ctx;
 428
 429    p->translated = 0;
 430    p->params_uptodate = 0;
 431    p->on_hardware = 0;
 432    p->error = 0;
 433
 434    memset(&p->register_phases, 0, sizeof(p->register_phases));
 435    p->nr_tex_indirect = 1;
 436    p->nr_tex_insn = 0;
 437    p->nr_alu_insn = 0;
 438    p->nr_decl_insn = 0;
 439
 440    p->ctx = ctx;
 441    memset(p->constant_flags, 0, sizeof(p->constant_flags));
 442
 443    p->nr_constants = 0;
 444    p->csr = p->program;
 445    p->decl = p->declarations;
 446    p->decl_s = 0;
 447    p->decl_t = 0;
 448    p->temp_flag = 0xffff000;
 449    p->utemp_flag = ~0x7;
 450    p->wpos_tex = -1;
 451    p->depth_written = 0;
 452    p->nr_params = 0;
 453
 454    *(p->decl++) = _3DSTATE_PIXEL_SHADER_PROGRAM;
 455 }
 456
 457
 458 void
 459 i915_fini_program(struct i915_fragment_program *p)
 460 {
 461    GLuint program_size = p->csr - p->program;
 462    GLuint decl_size = p->decl - p->declarations;
 463
 464    if (p->nr_tex_indirect > I915_MAX_TEX_INDIRECT)
 465       i915_program_error(p, "Exceeded max nr indirect texture lookups");
 466
 467    if (p->nr_tex_insn > I915_MAX_TEX_INSN)
 468       i915_program_error(p, "Exceeded max TEX instructions");
 469
 470    if (p->nr_alu_insn > I915_MAX_ALU_INSN)
 471       i915_program_error(p, "Exceeded max ALU instructions");
 472
 473    if (p->nr_decl_insn > I915_MAX_DECL_INSN)
 474       i915_program_error(p, "Exceeded max DECL instructions");
 475
 476    if (p->error) {
 477       p->FragProg.Base.NumNativeInstructions = 0;
 478       p->FragProg.Base.NumNativeAluInstructions = 0;
 479       p->FragProg.Base.NumNativeTexInstructions = 0;
 480       p->FragProg.Base.NumNativeTexIndirections = 0;
 481    }
 482    else {
 483       p->FragProg.Base.NumNativeInstructions = (p->nr_alu_insn +
 484                                                 p->nr_tex_insn +
 485                                                 p->nr_decl_insn);
 486       p->FragProg.Base.NumNativeAluInstructions = p->nr_alu_insn;
 487       p->FragProg.Base.NumNativeTexInstructions = p->nr_tex_insn;
 488       p->FragProg.Base.NumNativeTexIndirections = p->nr_tex_indirect;
 489    }
 490
 491    p->declarations[0] |= program_size + decl_size - 2;
 492 }
 493
 494 void
 495 i915_upload_program(struct i915_context *i915,
 496                     struct i915_fragment_program *p)
 497 {
 498    GLuint program_size = p->csr - p->program;
 499    GLuint decl_size = p->decl - p->declarations;
 500
 501    FALLBACK(&i915->intel, I915_FALLBACK_PROGRAM, p->error);
 502
 503    /* Could just go straight to the batchbuffer from here:
 504     */
 505    if (i915->state.ProgramSize != (program_size + decl_size) ||
 506        memcmp(i915->state.Program + decl_size, p->program,
 507               program_size * sizeof(int)) != 0) {
 508       I915_STATECHANGE(i915, I915_UPLOAD_PROGRAM);
 509       memcpy(i915->state.Program, p->declarations, decl_size * sizeof(int));
 510       memcpy(i915->state.Program + decl_size, p->program,
 511              program_size * sizeof(int));
 512       i915->state.ProgramSize = decl_size + program_size;
 513    }
 514
 515    /* Always seemed to get a failure if I used memcmp() to
 516     * shortcircuit this state upload.  Needs further investigation?
 517     */
 518    if (p->nr_constants) {
 519       GLuint nr = p->nr_constants;
 520
 521       I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 1);
 522       I915_STATECHANGE(i915, I915_UPLOAD_CONSTANTS);
 523
 524       i915->state.Constant[0] = _3DSTATE_PIXEL_SHADER_CONSTANTS | ((nr) * 4);
 525       i915->state.Constant[1] = (1 << (nr - 1)) | ((1 << (nr - 1)) - 1);
 526
 527       memcpy(&i915->state.Constant[2], p->constant, 4 * sizeof(int) * (nr));
 528       i915->state.ConstantSize = 2 + (nr) * 4;
 529
 530       if (0) {
 531          GLuint i;
 532          for (i = 0; i < nr; i++) {
 533             fprintf(stderr, "const[%d]: %f %f %f %f\n", i,
 534                     p->constant[i][0],
 535                     p->constant[i][1], p->constant[i][2], p->constant[i][3]);
 536          }
 537       }
 538    }
 539    else {
 540       I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 0);
 541    }
 542
 543    p->on_hardware = 1;
 544 }