src/gallium/auxiliary/util/u_simple_shaders.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2008 VMware, Inc.
   4  * All Rights Reserved.
   5  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Simple vertex/fragment shader generators.
  32  *
  33  * @author Brian Paul
  34            Marek Olšák
  35  */
  36
  37
  38 #include "pipe/p_context.h"
  39 #include "pipe/p_shader_tokens.h"
  40 #include "pipe/p_state.h"
  41 #include "util/u_simple_shaders.h"
  42 #include "util/u_debug.h"
  43 #include "util/u_memory.h"
  44 #include "tgsi/tgsi_dump.h"
  45 #include "tgsi/tgsi_strings.h"
  46 #include "tgsi/tgsi_ureg.h"
  47 #include "tgsi/tgsi_text.h"
  48 #include <stdio.h> /* include last */
  49
  50
  51
  52 /**
  53  * Make simple vertex pass-through shader.
  54  * \param num_attribs  number of attributes to pass through
  55  * \param semantic_names  array of semantic names for each attribute
  56  * \param semantic_indexes  array of semantic indexes for each attribute
  57  */
  58 void *
  59 util_make_vertex_passthrough_shader(struct pipe_context *pipe,
  60                                     uint num_attribs,
  61                                     const uint *semantic_names,
  62                                     const uint *semantic_indexes,
  63                                     bool window_space)
  64 {
  65    return util_make_vertex_passthrough_shader_with_so(pipe, num_attribs,
  66                                                       semantic_names,
  67                                                       semantic_indexes,
  68                                                       window_space, NULL);
  69 }
  70
  71 void *
  72 util_make_vertex_passthrough_shader_with_so(struct pipe_context *pipe,
  73                                     uint num_attribs,
  74                                     const uint *semantic_names,
  75                                     const uint *semantic_indexes,
  76                                     bool window_space,
  77                                     const struct pipe_stream_output_info *so)
  78 {
  79    struct ureg_program *ureg;
  80    uint i;
  81
  82    ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
  83    if (ureg == NULL)
  84       return NULL;
  85
  86    if (window_space)
  87       ureg_property_vs_window_space_position(ureg, TRUE);
  88
  89    for (i = 0; i < num_attribs; i++) {
  90       struct ureg_src src;
  91       struct ureg_dst dst;
  92
  93       src = ureg_DECL_vs_input( ureg, i );
  94
  95       dst = ureg_DECL_output( ureg,
  96                               semantic_names[i],
  97                               semantic_indexes[i]);
  98
  99       ureg_MOV( ureg, dst, src );
 100    }
 101
 102    ureg_END( ureg );
 103
 104    return ureg_create_shader_with_so_and_destroy( ureg, pipe, so );
 105 }
 106
 107
 108 void *util_make_layered_clear_vertex_shader(struct pipe_context *pipe)
 109 {
 110    static const char text[] =
 111          "VERT\n"
 112          "DCL IN[0]\n"
 113          "DCL IN[1]\n"
 114          "DCL SV[0], INSTANCEID\n"
 115          "DCL OUT[0], POSITION\n"
 116          "DCL OUT[1], GENERIC[0]\n"
 117          "DCL OUT[2], LAYER\n"
 118
 119          "MOV OUT[0], IN[0]\n"
 120          "MOV OUT[1], IN[1]\n"
 121          "MOV OUT[2], SV[0]\n"
 122          "END\n";
 123    struct tgsi_token tokens[1000];
 124    struct pipe_shader_state state = {tokens};
 125
 126    if (!tgsi_text_translate(text, tokens, Elements(tokens))) {
 127       assert(0);
 128       return NULL;
 129    }
 130    return pipe->create_vs_state(pipe, &state);
 131 }
 132
 133
 134 /**
 135  * Make simple fragment texture shader:
 136  *  IMM {0,0,0,1}                         // (if writemask != 0xf)
 137  *  MOV OUT[0], IMM[0]                    // (if writemask != 0xf)
 138  *  TEX OUT[0].writemask, IN[0], SAMP[0], 2D;
 139  *  END;
 140  *
 141  * \param tex_target  one of PIPE_TEXTURE_x
 142  * \parma interp_mode  either TGSI_INTERPOLATE_LINEAR or PERSPECTIVE
 143  * \param writemask  mask of TGSI_WRITEMASK_x
 144  */
 145 void *
 146 util_make_fragment_tex_shader_writemask(struct pipe_context *pipe,
 147                                         unsigned tex_target,
 148                                         unsigned interp_mode,
 149                                         unsigned writemask )
 150 {
 151    struct ureg_program *ureg;
 152    struct ureg_src sampler;
 153    struct ureg_src tex;
 154    struct ureg_dst out;
 155
 156    assert(interp_mode == TGSI_INTERPOLATE_LINEAR ||
 157           interp_mode == TGSI_INTERPOLATE_PERSPECTIVE);
 158
 159    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 160    if (ureg == NULL)
 161       return NULL;
 162
 163    sampler = ureg_DECL_sampler( ureg, 0 );
 164
 165    tex = ureg_DECL_fs_input( ureg,
 166                              TGSI_SEMANTIC_GENERIC, 0,
 167                              interp_mode );
 168
 169    out = ureg_DECL_output( ureg,
 170                            TGSI_SEMANTIC_COLOR,
 171                            0 );
 172
 173    if (writemask != TGSI_WRITEMASK_XYZW) {
 174       struct ureg_src imm = ureg_imm4f( ureg, 0, 0, 0, 1 );
 175
 176       ureg_MOV( ureg, out, imm );
 177    }
 178
 179    ureg_TEX( ureg,
 180              ureg_writemask(out, writemask),
 181              tex_target, tex, sampler );
 182    ureg_END( ureg );
 183
 184    return ureg_create_shader_and_destroy( ureg, pipe );
 185 }
 186
 187
 188 /**
 189  * Make a simple fragment shader that sets the output color to a color
 190  * taken from a texture.
 191  * \param tex_target  one of PIPE_TEXTURE_x
 192  */
 193 void *
 194 util_make_fragment_tex_shader(struct pipe_context *pipe, unsigned tex_target,
 195                               unsigned interp_mode)
 196 {
 197    return util_make_fragment_tex_shader_writemask( pipe,
 198                                                    tex_target,
 199                                                    interp_mode,
 200                                                    TGSI_WRITEMASK_XYZW );
 201 }
 202
 203
 204 /**
 205  * Make a simple fragment texture shader which reads an X component from
 206  * a texture and writes it as depth.
 207  */
 208 void *
 209 util_make_fragment_tex_shader_writedepth(struct pipe_context *pipe,
 210                                          unsigned tex_target,
 211                                          unsigned interp_mode)
 212 {
 213    struct ureg_program *ureg;
 214    struct ureg_src sampler;
 215    struct ureg_src tex;
 216    struct ureg_dst out, depth;
 217    struct ureg_src imm;
 218
 219    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 220    if (ureg == NULL)
 221       return NULL;
 222
 223    sampler = ureg_DECL_sampler( ureg, 0 );
 224
 225    tex = ureg_DECL_fs_input( ureg,
 226                              TGSI_SEMANTIC_GENERIC, 0,
 227                              interp_mode );
 228
 229    out = ureg_DECL_output( ureg,
 230                            TGSI_SEMANTIC_COLOR,
 231                            0 );
 232
 233    depth = ureg_DECL_output( ureg,
 234                              TGSI_SEMANTIC_POSITION,
 235                              0 );
 236
 237    imm = ureg_imm4f( ureg, 0, 0, 0, 1 );
 238
 239    ureg_MOV( ureg, out, imm );
 240
 241    ureg_TEX( ureg,
 242              ureg_writemask(depth, TGSI_WRITEMASK_Z),
 243              tex_target, tex, sampler );
 244    ureg_END( ureg );
 245
 246    return ureg_create_shader_and_destroy( ureg, pipe );
 247 }
 248
 249
 250 /**
 251  * Make a simple fragment texture shader which reads the texture unit 0 and 1
 252  * and writes it as depth and stencil, respectively.
 253  */
 254 void *
 255 util_make_fragment_tex_shader_writedepthstencil(struct pipe_context *pipe,
 256                                                 unsigned tex_target,
 257                                                 unsigned interp_mode)
 258 {
 259    struct ureg_program *ureg;
 260    struct ureg_src depth_sampler, stencil_sampler;
 261    struct ureg_src tex;
 262    struct ureg_dst out, depth, stencil;
 263    struct ureg_src imm;
 264
 265    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 266    if (ureg == NULL)
 267       return NULL;
 268
 269    depth_sampler = ureg_DECL_sampler( ureg, 0 );
 270    stencil_sampler = ureg_DECL_sampler( ureg, 1 );
 271
 272    tex = ureg_DECL_fs_input( ureg,
 273                              TGSI_SEMANTIC_GENERIC, 0,
 274                              interp_mode );
 275
 276    out = ureg_DECL_output( ureg,
 277                            TGSI_SEMANTIC_COLOR,
 278                            0 );
 279
 280    depth = ureg_DECL_output( ureg,
 281                              TGSI_SEMANTIC_POSITION,
 282                              0 );
 283
 284    stencil = ureg_DECL_output( ureg,
 285                              TGSI_SEMANTIC_STENCIL,
 286                              0 );
 287
 288    imm = ureg_imm4f( ureg, 0, 0, 0, 1 );
 289
 290    ureg_MOV( ureg, out, imm );
 291
 292    ureg_TEX( ureg,
 293              ureg_writemask(depth, TGSI_WRITEMASK_Z),
 294              tex_target, tex, depth_sampler );
 295    ureg_TEX( ureg,
 296              ureg_writemask(stencil, TGSI_WRITEMASK_Y),
 297              tex_target, tex, stencil_sampler );
 298    ureg_END( ureg );
 299
 300    return ureg_create_shader_and_destroy( ureg, pipe );
 301 }
 302
 303
 304 /**
 305  * Make a simple fragment texture shader which reads a texture and writes it
 306  * as stencil.
 307  */
 308 void *
 309 util_make_fragment_tex_shader_writestencil(struct pipe_context *pipe,
 310                                            unsigned tex_target,
 311                                            unsigned interp_mode)
 312 {
 313    struct ureg_program *ureg;
 314    struct ureg_src stencil_sampler;
 315    struct ureg_src tex;
 316    struct ureg_dst out, stencil;
 317    struct ureg_src imm;
 318
 319    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 320    if (ureg == NULL)
 321       return NULL;
 322
 323    stencil_sampler = ureg_DECL_sampler( ureg, 0 );
 324
 325    tex = ureg_DECL_fs_input( ureg,
 326                              TGSI_SEMANTIC_GENERIC, 0,
 327                              interp_mode );
 328
 329    out = ureg_DECL_output( ureg,
 330                            TGSI_SEMANTIC_COLOR,
 331                            0 );
 332
 333    stencil = ureg_DECL_output( ureg,
 334                              TGSI_SEMANTIC_STENCIL,
 335                              0 );
 336
 337    imm = ureg_imm4f( ureg, 0, 0, 0, 1 );
 338
 339    ureg_MOV( ureg, out, imm );
 340
 341    ureg_TEX( ureg,
 342              ureg_writemask(stencil, TGSI_WRITEMASK_Y),
 343              tex_target, tex, stencil_sampler );
 344    ureg_END( ureg );
 345
 346    return ureg_create_shader_and_destroy( ureg, pipe );
 347 }
 348
 349
 350 /**
 351  * Make simple fragment color pass-through shader that replicates OUT[0]
 352  * to all bound colorbuffers.
 353  */
 354 void *
 355 util_make_fragment_passthrough_shader(struct pipe_context *pipe,
 356                                       int input_semantic,
 357                                       int input_interpolate,
 358                                       boolean write_all_cbufs)
 359 {
 360    static const char shader_templ[] =
 361          "FRAG\n"
 362          "%s"
 363          "DCL IN[0], %s[0], %s\n"
 364          "DCL OUT[0], COLOR[0]\n"
 365
 366          "MOV OUT[0], IN[0]\n"
 367          "END\n";
 368
 369    char text[sizeof(shader_templ)+100];
 370    struct tgsi_token tokens[1000];
 371    struct pipe_shader_state state = {tokens};
 372
 373    sprintf(text, shader_templ,
 374            write_all_cbufs ? "PROPERTY FS_COLOR0_WRITES_ALL_CBUFS 1\n" : "",
 375            tgsi_semantic_names[input_semantic],
 376            tgsi_interpolate_names[input_interpolate]);
 377
 378    if (!tgsi_text_translate(text, tokens, Elements(tokens))) {
 379       assert(0);
 380       return NULL;
 381    }
 382 #if 0
 383    tgsi_dump(state.tokens, 0);
 384 #endif
 385
 386    return pipe->create_fs_state(pipe, &state);
 387 }
 388
 389
 390 void *
 391 util_make_empty_fragment_shader(struct pipe_context *pipe)
 392 {
 393    struct ureg_program *ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 394    if (ureg == NULL)
 395       return NULL;
 396
 397    ureg_END(ureg);
 398    return ureg_create_shader_and_destroy(ureg, pipe);
 399 }
 400
 401
 402 /**
 403  * Make a fragment shader that copies the input color to N output colors.
 404  */
 405 void *
 406 util_make_fragment_cloneinput_shader(struct pipe_context *pipe, int num_cbufs,
 407                                      int input_semantic,
 408                                      int input_interpolate)
 409 {
 410    struct ureg_program *ureg;
 411    struct ureg_src src;
 412    struct ureg_dst dst[PIPE_MAX_COLOR_BUFS];
 413    int i;
 414
 415    assert(num_cbufs <= PIPE_MAX_COLOR_BUFS);
 416
 417    ureg = ureg_create( TGSI_PROCESSOR_FRAGMENT );
 418    if (ureg == NULL)
 419       return NULL;
 420
 421    src = ureg_DECL_fs_input( ureg, input_semantic, 0,
 422                              input_interpolate );
 423
 424    for (i = 0; i < num_cbufs; i++)
 425       dst[i] = ureg_DECL_output( ureg, TGSI_SEMANTIC_COLOR, i );
 426
 427    for (i = 0; i < num_cbufs; i++)
 428       ureg_MOV( ureg, dst[i], src );
 429
 430    ureg_END( ureg );
 431
 432    return ureg_create_shader_and_destroy( ureg, pipe );
 433 }
 434
 435
 436 static void *
 437 util_make_fs_blit_msaa_gen(struct pipe_context *pipe,
 438                            unsigned tgsi_tex,
 439                            const char *output_semantic,
 440                            const char *output_mask)
 441 {
 442    static const char shader_templ[] =
 443          "FRAG\n"
 444          "DCL IN[0], GENERIC[0], LINEAR\n"
 445          "DCL SAMP[0]\n"
 446          "DCL OUT[0], %s\n"
 447          "DCL TEMP[0]\n"
 448
 449          "F2U TEMP[0], IN[0]\n"
 450          "TXF OUT[0]%s, TEMP[0], SAMP[0], %s\n"
 451          "END\n";
 452
 453    const char *type = tgsi_texture_names[tgsi_tex];
 454    char text[sizeof(shader_templ)+100];
 455    struct tgsi_token tokens[1000];
 456    struct pipe_shader_state state = {tokens};
 457
 458    assert(tgsi_tex == TGSI_TEXTURE_2D_MSAA ||
 459           tgsi_tex == TGSI_TEXTURE_2D_ARRAY_MSAA);
 460
 461    sprintf(text, shader_templ, output_semantic, output_mask, type);
 462
 463    if (!tgsi_text_translate(text, tokens, Elements(tokens))) {
 464       puts(text);
 465       assert(0);
 466       return NULL;
 467    }
 468 #if 0
 469    tgsi_dump(state.tokens, 0);
 470 #endif
 471
 472    return pipe->create_fs_state(pipe, &state);
 473 }
 474
 475
 476 /**
 477  * Make a fragment shader that sets the output color to a color
 478  * fetched from a multisample texture.
 479  * \param tex_target  one of PIPE_TEXTURE_x
 480  */
 481 void *
 482 util_make_fs_blit_msaa_color(struct pipe_context *pipe,
 483                              unsigned tgsi_tex)
 484 {
 485    return util_make_fs_blit_msaa_gen(pipe, tgsi_tex,
 486                                      "COLOR[0]", "");
 487 }
 488
 489
 490 /**
 491  * Make a fragment shader that sets the output depth to a depth value
 492  * fetched from a multisample texture.
 493  * \param tex_target  one of PIPE_TEXTURE_x
 494  */
 495 void *
 496 util_make_fs_blit_msaa_depth(struct pipe_context *pipe,
 497                              unsigned tgsi_tex)
 498 {
 499    return util_make_fs_blit_msaa_gen(pipe, tgsi_tex,
 500                                      "POSITION", ".z");
 501 }
 502
 503
 504 /**
 505  * Make a fragment shader that sets the output stencil to a stencil value
 506  * fetched from a multisample texture.
 507  * \param tex_target  one of PIPE_TEXTURE_x
 508  */
 509 void *
 510 util_make_fs_blit_msaa_stencil(struct pipe_context *pipe,
 511                                unsigned tgsi_tex)
 512 {
 513    return util_make_fs_blit_msaa_gen(pipe, tgsi_tex,
 514                                      "STENCIL", ".y");
 515 }
 516
 517
 518 /**
 519  * Make a fragment shader that sets the output depth and stencil to depth
 520  * and stencil values fetched from two multisample textures / samplers.
 521  * The sizes of both textures should match (it should be one depth-stencil
 522  * texture).
 523  * \param tex_target  one of PIPE_TEXTURE_x
 524  */
 525 void *
 526 util_make_fs_blit_msaa_depthstencil(struct pipe_context *pipe,
 527                                     unsigned tgsi_tex)
 528 {
 529    static const char shader_templ[] =
 530          "FRAG\n"
 531          "DCL IN[0], GENERIC[0], LINEAR\n"
 532          "DCL SAMP[0..1]\n"
 533          "DCL OUT[0], POSITION\n"
 534          "DCL OUT[1], STENCIL\n"
 535          "DCL TEMP[0]\n"
 536
 537          "F2U TEMP[0], IN[0]\n"
 538          "TXF OUT[0].z, TEMP[0], SAMP[0], %s\n"
 539          "TXF OUT[1].y, TEMP[0], SAMP[1], %s\n"
 540          "END\n";
 541
 542    const char *type = tgsi_texture_names[tgsi_tex];
 543    char text[sizeof(shader_templ)+100];
 544    struct tgsi_token tokens[1000];
 545    struct pipe_shader_state state = {tokens};
 546
 547    assert(tgsi_tex == TGSI_TEXTURE_2D_MSAA ||
 548           tgsi_tex == TGSI_TEXTURE_2D_ARRAY_MSAA);
 549
 550    sprintf(text, shader_templ, type, type);
 551
 552    if (!tgsi_text_translate(text, tokens, Elements(tokens))) {
 553       assert(0);
 554       return NULL;
 555    }
 556 #if 0
 557    tgsi_dump(state.tokens, 0);
 558 #endif
 559
 560    return pipe->create_fs_state(pipe, &state);
 561 }
 562
 563
 564 void *
 565 util_make_fs_msaa_resolve(struct pipe_context *pipe,
 566                           unsigned tgsi_tex, unsigned nr_samples,
 567                           boolean is_uint, boolean is_sint)
 568 {
 569    struct ureg_program *ureg;
 570    struct ureg_src sampler, coord;
 571    struct ureg_dst out, tmp_sum, tmp_coord, tmp;
 572    int i;
 573
 574    ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 575    if (!ureg)
 576       return NULL;
 577
 578    /* Declarations. */
 579    sampler = ureg_DECL_sampler(ureg, 0);
 580    coord = ureg_DECL_fs_input(ureg, TGSI_SEMANTIC_GENERIC, 0,
 581                               TGSI_INTERPOLATE_LINEAR);
 582    out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 583    tmp_sum = ureg_DECL_temporary(ureg);
 584    tmp_coord = ureg_DECL_temporary(ureg);
 585    tmp = ureg_DECL_temporary(ureg);
 586
 587    /* Instructions. */
 588    ureg_MOV(ureg, tmp_sum, ureg_imm1f(ureg, 0));
 589    ureg_F2U(ureg, tmp_coord, coord);
 590
 591    for (i = 0; i < nr_samples; i++) {
 592       /* Read one sample. */
 593       ureg_MOV(ureg, ureg_writemask(tmp_coord, TGSI_WRITEMASK_W),
 594                ureg_imm1u(ureg, i));
 595       ureg_TXF(ureg, tmp, tgsi_tex, ureg_src(tmp_coord), sampler);
 596
 597       if (is_uint)
 598          ureg_U2F(ureg, tmp, ureg_src(tmp));
 599       else if (is_sint)
 600          ureg_I2F(ureg, tmp, ureg_src(tmp));
 601
 602       /* Add it to the sum.*/
 603       ureg_ADD(ureg, tmp_sum, ureg_src(tmp_sum), ureg_src(tmp));
 604    }
 605
 606    /* Calculate the average and return. */
 607    ureg_MUL(ureg, tmp_sum, ureg_src(tmp_sum),
 608             ureg_imm1f(ureg, 1.0 / nr_samples));
 609
 610    if (is_uint)
 611       ureg_F2U(ureg, out, ureg_src(tmp_sum));
 612    else if (is_sint)
 613       ureg_F2I(ureg, out, ureg_src(tmp_sum));
 614    else
 615       ureg_MOV(ureg, out, ureg_src(tmp_sum));
 616
 617    ureg_END(ureg);
 618
 619    return ureg_create_shader_and_destroy(ureg, pipe);
 620 }
 621
 622
 623 void *
 624 util_make_fs_msaa_resolve_bilinear(struct pipe_context *pipe,
 625                                    unsigned tgsi_tex, unsigned nr_samples,
 626                                    boolean is_uint, boolean is_sint)
 627 {
 628    struct ureg_program *ureg;
 629    struct ureg_src sampler, coord;
 630    struct ureg_dst out, tmp, top, bottom;
 631    struct ureg_dst tmp_coord[4], tmp_sum[4];
 632    int i, c;
 633
 634    ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 635    if (!ureg)
 636       return NULL;
 637
 638    /* Declarations. */
 639    sampler = ureg_DECL_sampler(ureg, 0);
 640    coord = ureg_DECL_fs_input(ureg, TGSI_SEMANTIC_GENERIC, 0,
 641                               TGSI_INTERPOLATE_LINEAR);
 642    out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 643    for (c = 0; c < 4; c++)
 644       tmp_sum[c] = ureg_DECL_temporary(ureg);
 645    for (c = 0; c < 4; c++)
 646       tmp_coord[c] = ureg_DECL_temporary(ureg);
 647    tmp = ureg_DECL_temporary(ureg);
 648    top = ureg_DECL_temporary(ureg);
 649    bottom = ureg_DECL_temporary(ureg);
 650
 651    /* Instructions. */
 652    for (c = 0; c < 4; c++)
 653       ureg_MOV(ureg, tmp_sum[c], ureg_imm1f(ureg, 0));
 654
 655    /* Get 4 texture coordinates for the bilinear filter. */
 656    ureg_F2U(ureg, tmp_coord[0], coord); /* top-left */
 657    ureg_UADD(ureg, tmp_coord[1], ureg_src(tmp_coord[0]),
 658              ureg_imm4u(ureg, 1, 0, 0, 0)); /* top-right */
 659    ureg_UADD(ureg, tmp_coord[2], ureg_src(tmp_coord[0]),
 660              ureg_imm4u(ureg, 0, 1, 0, 0)); /* bottom-left */
 661    ureg_UADD(ureg, tmp_coord[3], ureg_src(tmp_coord[0]),
 662              ureg_imm4u(ureg, 1, 1, 0, 0)); /* bottom-right */
 663
 664    for (i = 0; i < nr_samples; i++) {
 665       for (c = 0; c < 4; c++) {
 666          /* Read one sample. */
 667          ureg_MOV(ureg, ureg_writemask(tmp_coord[c], TGSI_WRITEMASK_W),
 668                   ureg_imm1u(ureg, i));
 669          ureg_TXF(ureg, tmp, tgsi_tex, ureg_src(tmp_coord[c]), sampler);
 670
 671          if (is_uint)
 672             ureg_U2F(ureg, tmp, ureg_src(tmp));
 673          else if (is_sint)
 674             ureg_I2F(ureg, tmp, ureg_src(tmp));
 675
 676          /* Add it to the sum.*/
 677          ureg_ADD(ureg, tmp_sum[c], ureg_src(tmp_sum[c]), ureg_src(tmp));
 678       }
 679    }
 680
 681    /* Calculate the average. */
 682    for (c = 0; c < 4; c++)
 683       ureg_MUL(ureg, tmp_sum[c], ureg_src(tmp_sum[c]),
 684                ureg_imm1f(ureg, 1.0 / nr_samples));
 685
 686    /* Take the 4 average values and apply a standard bilinear filter. */
 687    ureg_FRC(ureg, tmp, coord);
 688
 689    ureg_LRP(ureg, top,
 690             ureg_scalar(ureg_src(tmp), 0),
 691             ureg_src(tmp_sum[1]),
 692             ureg_src(tmp_sum[0]));
 693
 694    ureg_LRP(ureg, bottom,
 695             ureg_scalar(ureg_src(tmp), 0),
 696             ureg_src(tmp_sum[3]),
 697             ureg_src(tmp_sum[2]));
 698
 699    ureg_LRP(ureg, tmp,
 700             ureg_scalar(ureg_src(tmp), 1),
 701             ureg_src(bottom),
 702             ureg_src(top));
 703
 704    /* Convert to the texture format and return. */
 705    if (is_uint)
 706       ureg_F2U(ureg, out, ureg_src(tmp));
 707    else if (is_sint)
 708       ureg_F2I(ureg, out, ureg_src(tmp));
 709    else
 710       ureg_MOV(ureg, out, ureg_src(tmp));
 711
 712    ureg_END(ureg);
 713
 714    return ureg_create_shader_and_destroy(ureg, pipe);
 715 }