src/mesa/drivers/dri/i915tex/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 "glheader.h"
  31 #include "macros.h"
  32 #include "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    p->nr_alu_insn++;
 194    return dest;
 195 }
 196
 197 GLuint i915_emit_texld( struct i915_fragment_program *p,
 198                         GLuint dest,
 199                         GLuint destmask,
 200                         GLuint sampler,
 201                         GLuint coord,
 202                         GLuint op )
 203 {
 204    if (coord != UREG(GET_UREG_TYPE(coord), GET_UREG_NR(coord))) {
 205       /* No real way to work around this in the general case - need to
 206        * allocate and declare a new temporary register (a utemp won't
 207        * do).  Will fallback for now.
 208        */
 209       i915_program_error(p, "Can't (yet) swizzle TEX arguments");
 210       return 0;
 211    }
 212
 213    /* Don't worry about saturate as we only support
 214     */
 215    if (destmask != A0_DEST_CHANNEL_ALL) {
 216       GLuint tmp = i915_get_utemp(p);
 217       i915_emit_texld( p, tmp, A0_DEST_CHANNEL_ALL, sampler, coord, op );
 218       i915_emit_arith( p, A0_MOV, dest, destmask, 0, tmp, 0, 0 );
 219       return dest;
 220    }
 221    else {
 222       assert(GET_UREG_TYPE(dest) != REG_TYPE_CONST);
 223       assert(dest = UREG(GET_UREG_TYPE(dest), GET_UREG_NR(dest)));
 224
 225       if (GET_UREG_TYPE(coord) != REG_TYPE_T) {
 226          p->nr_tex_indirect++;
 227       }
 228
 229       *(p->csr++) = (op |
 230                      T0_DEST( dest ) |
 231                      T0_SAMPLER( sampler ));
 232
 233       *(p->csr++) = T1_ADDRESS_REG( coord );
 234       *(p->csr++) = T2_MBZ;
 235
 236       p->nr_tex_insn++;
 237       return dest;
 238    }
 239 }
 240
 241
 242 GLuint
 243 i915_emit_const1f(struct i915_fragment_program * p, GLfloat c0)
 244 {
 245    GLint reg, idx;
 246
 247    if (c0 == 0.0)
 248       return swizzle(UREG(REG_TYPE_R, 0), ZERO, ZERO, ZERO, ZERO);
 249    if (c0 == 1.0)
 250       return swizzle(UREG(REG_TYPE_R, 0), ONE, ONE, ONE, ONE);
 251
 252    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 253       if (p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
 254          continue;
 255       for (idx = 0; idx < 4; idx++) {
 256          if (!(p->constant_flags[reg] & (1 << idx)) ||
 257              p->constant[reg][idx] == c0) {
 258             p->constant[reg][idx] = c0;
 259             p->constant_flags[reg] |= 1 << idx;
 260             if (reg + 1 > p->nr_constants)
 261                p->nr_constants = reg + 1;
 262             return swizzle(UREG(REG_TYPE_CONST, reg), idx, ZERO, ZERO, ONE);
 263          }
 264       }
 265    }
 266
 267    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 268    p->error = 1;
 269    return 0;
 270 }
 271
 272 GLuint
 273 i915_emit_const2f(struct i915_fragment_program * p, GLfloat c0, GLfloat c1)
 274 {
 275    GLint reg, idx;
 276
 277    if (c0 == 0.0)
 278       return swizzle(i915_emit_const1f(p, c1), ZERO, X, Z, W);
 279    if (c0 == 1.0)
 280       return swizzle(i915_emit_const1f(p, c1), ONE, X, Z, W);
 281
 282    if (c1 == 0.0)
 283       return swizzle(i915_emit_const1f(p, c0), X, ZERO, Z, W);
 284    if (c1 == 1.0)
 285       return swizzle(i915_emit_const1f(p, c0), X, ONE, Z, W);
 286
 287    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 288       if (p->constant_flags[reg] == 0xf ||
 289           p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
 290          continue;
 291       for (idx = 0; idx < 3; idx++) {
 292          if (!(p->constant_flags[reg] & (3 << idx))) {
 293             p->constant[reg][idx] = c0;
 294             p->constant[reg][idx + 1] = c1;
 295             p->constant_flags[reg] |= 3 << idx;
 296             if (reg + 1 > p->nr_constants)
 297                p->nr_constants = reg + 1;
 298             return swizzle(UREG(REG_TYPE_CONST, reg), idx, idx + 1, ZERO,
 299                            ONE);
 300          }
 301       }
 302    }
 303
 304    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 305    p->error = 1;
 306    return 0;
 307 }
 308
 309
 310
 311 GLuint
 312 i915_emit_const4f(struct i915_fragment_program * p,
 313                   GLfloat c0, GLfloat c1, GLfloat c2, GLfloat c3)
 314 {
 315    GLint reg;
 316
 317    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 318       if (p->constant_flags[reg] == 0xf &&
 319           p->constant[reg][0] == c0 &&
 320           p->constant[reg][1] == c1 &&
 321           p->constant[reg][2] == c2 && p->constant[reg][3] == c3) {
 322          return UREG(REG_TYPE_CONST, reg);
 323       }
 324       else if (p->constant_flags[reg] == 0) {
 325          p->constant[reg][0] = c0;
 326          p->constant[reg][1] = c1;
 327          p->constant[reg][2] = c2;
 328          p->constant[reg][3] = c3;
 329          p->constant_flags[reg] = 0xf;
 330          if (reg + 1 > p->nr_constants)
 331             p->nr_constants = reg + 1;
 332          return UREG(REG_TYPE_CONST, reg);
 333       }
 334    }
 335
 336    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 337    p->error = 1;
 338    return 0;
 339 }
 340
 341
 342 GLuint
 343 i915_emit_const4fv(struct i915_fragment_program * p, const GLfloat * c)
 344 {
 345    return i915_emit_const4f(p, c[0], c[1], c[2], c[3]);
 346 }
 347
 348
 349 GLuint
 350 i915_emit_param4fv(struct i915_fragment_program * p, const GLfloat * values)
 351 {
 352    GLint reg, i;
 353
 354    for (i = 0; i < p->nr_params; i++) {
 355       if (p->param[i].values == values)
 356          return UREG(REG_TYPE_CONST, p->param[i].reg);
 357    }
 358
 359
 360    for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
 361       if (p->constant_flags[reg] == 0) {
 362          p->constant_flags[reg] = I915_CONSTFLAG_PARAM;
 363          i = p->nr_params++;
 364
 365          p->param[i].values = values;
 366          p->param[i].reg = reg;
 367          p->params_uptodate = 0;
 368
 369          if (reg + 1 > p->nr_constants)
 370             p->nr_constants = reg + 1;
 371          return UREG(REG_TYPE_CONST, reg);
 372       }
 373    }
 374
 375    fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
 376    p->error = 1;
 377    return 0;
 378 }
 379
 380
 381
 382
 383 void
 384 i915_program_error(struct i915_fragment_program *p, const char *msg)
 385 {
 386    /* XXX we shouldn't print anything to stdout, record GL error or
 387     * call _mesa_problem()
 388     */
 389    fprintf(stderr, "%s\n", msg);
 390    p->error = 1;
 391 }
 392
 393 void
 394 i915_init_program(struct i915_context *i915, struct i915_fragment_program *p)
 395 {
 396    GLcontext *ctx = &i915->intel.ctx;
 397    TNLcontext *tnl = TNL_CONTEXT(ctx);
 398
 399    p->translated = 0;
 400    p->params_uptodate = 0;
 401    p->on_hardware = 0;
 402    p->error = 0;
 403
 404    p->nr_tex_indirect = 1;      /* correct? */
 405    p->nr_tex_insn = 0;
 406    p->nr_alu_insn = 0;
 407    p->nr_decl_insn = 0;
 408
 409    p->ctx = ctx;
 410    memset(p->constant_flags, 0, sizeof(p->constant_flags));
 411
 412    p->nr_constants = 0;
 413    p->csr = p->program;
 414    p->decl = p->declarations;
 415    p->decl_s = 0;
 416    p->decl_t = 0;
 417    p->temp_flag = 0xffff000;
 418    p->utemp_flag = ~0x7;
 419    p->wpos_tex = -1;
 420    p->depth_written = 0;
 421    p->nr_params = 0;
 422
 423    p->src_texture = UREG_BAD;
 424    p->src_previous = UREG(REG_TYPE_T, T_DIFFUSE);
 425    p->last_tex_stage = 0;
 426    p->VB = &tnl->vb;
 427
 428    *(p->decl++) = _3DSTATE_PIXEL_SHADER_PROGRAM;
 429 }
 430
 431
 432 void
 433 i915_fini_program(struct i915_fragment_program *p)
 434 {
 435    GLuint program_size = p->csr - p->program;
 436    GLuint decl_size = p->decl - p->declarations;
 437
 438    if (p->nr_tex_indirect > I915_MAX_TEX_INDIRECT)
 439       i915_program_error(p, "Exceeded max nr indirect texture lookups");
 440
 441    if (p->nr_tex_insn > I915_MAX_TEX_INSN)
 442       i915_program_error(p, "Exceeded max TEX instructions");
 443
 444    if (p->nr_alu_insn > I915_MAX_ALU_INSN)
 445       i915_program_error(p, "Exceeded max ALU instructions");
 446
 447    if (p->nr_decl_insn > I915_MAX_DECL_INSN)
 448       i915_program_error(p, "Exceeded max DECL instructions");
 449
 450    if (p->error) {
 451       p->FragProg.Base.NumNativeInstructions = 0;
 452       p->FragProg.NumNativeAluInstructions = 0;
 453       p->FragProg.NumNativeTexInstructions = 0;
 454       p->FragProg.NumNativeTexIndirections = 0;
 455    }
 456    else {
 457       p->FragProg.Base.NumNativeInstructions = (p->nr_alu_insn +
 458                                                 p->nr_tex_insn +
 459                                                 p->nr_decl_insn);
 460       p->FragProg.NumNativeAluInstructions = p->nr_alu_insn;
 461       p->FragProg.NumNativeTexInstructions = p->nr_tex_insn;
 462       p->FragProg.NumNativeTexIndirections = p->nr_tex_indirect;
 463    }
 464
 465    p->declarations[0] |= program_size + decl_size - 2;
 466 }
 467
 468 void
 469 i915_upload_program(struct i915_context *i915,
 470                     struct i915_fragment_program *p)
 471 {
 472    GLuint program_size = p->csr - p->program;
 473    GLuint decl_size = p->decl - p->declarations;
 474
 475    FALLBACK(&i915->intel, I915_FALLBACK_PROGRAM, p->error);
 476
 477    /* Could just go straight to the batchbuffer from here:
 478     */
 479    if (i915->state.ProgramSize != (program_size + decl_size) ||
 480        memcmp(i915->state.Program + decl_size, p->program,
 481               program_size * sizeof(int)) != 0) {
 482       I915_STATECHANGE(i915, I915_UPLOAD_PROGRAM);
 483       memcpy(i915->state.Program, p->declarations, decl_size * sizeof(int));
 484       memcpy(i915->state.Program + decl_size, p->program,
 485              program_size * sizeof(int));
 486       i915->state.ProgramSize = decl_size + program_size;
 487    }
 488
 489    /* Always seemed to get a failure if I used memcmp() to
 490     * shortcircuit this state upload.  Needs further investigation?
 491     */
 492    if (p->nr_constants) {
 493       GLuint nr = p->nr_constants;
 494
 495       I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 1);
 496       I915_STATECHANGE(i915, I915_UPLOAD_CONSTANTS);
 497
 498       i915->state.Constant[0] = _3DSTATE_PIXEL_SHADER_CONSTANTS | ((nr) * 4);
 499       i915->state.Constant[1] = (1 << (nr - 1)) | ((1 << (nr - 1)) - 1);
 500
 501       memcpy(&i915->state.Constant[2], p->constant, 4 * sizeof(int) * (nr));
 502       i915->state.ConstantSize = 2 + (nr) * 4;
 503
 504       if (0) {
 505          GLuint i;
 506          for (i = 0; i < nr; i++) {
 507             fprintf(stderr, "const[%d]: %f %f %f %f\n", i,
 508                     p->constant[i][0],
 509                     p->constant[i][1], p->constant[i][2], p->constant[i][3]);
 510          }
 511       }
 512    }
 513    else {
 514       I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 0);
 515    }
 516
 517    p->on_hardware = 1;
 518 }