src/gallium/auxiliary/gallivm/lp_bld_sample.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   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 VMWARE 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 /**
  29  * @file
  30  * Texture sampling -- common code.
  31  *
  32  * @author Jose Fonseca <jfonseca@vmware.com>
  33  */
  34
  35 #include "pipe/p_defines.h"
  36 #include "pipe/p_state.h"
  37 #include "util/u_format.h"
  38 #include "util/u_math.h"
  39 #include "lp_bld_arit.h"
  40 #include "lp_bld_const.h"
  41 #include "lp_bld_debug.h"
  42 #include "lp_bld_flow.h"
  43 #include "lp_bld_sample.h"
  44 #include "lp_bld_swizzle.h"
  45 #include "lp_bld_type.h"
  46
  47
  48 /**
  49  * Does the given texture wrap mode allow sampling the texture border color?
  50  * XXX maybe move this into gallium util code.
  51  */
  52 boolean
  53 lp_sampler_wrap_mode_uses_border_color(unsigned mode,
  54                                        unsigned min_img_filter,
  55                                        unsigned mag_img_filter)
  56 {
  57    switch (mode) {
  58    case PIPE_TEX_WRAP_REPEAT:
  59    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
  60    case PIPE_TEX_WRAP_MIRROR_REPEAT:
  61    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
  62       return FALSE;
  63    case PIPE_TEX_WRAP_CLAMP:
  64    case PIPE_TEX_WRAP_MIRROR_CLAMP:
  65       if (min_img_filter == PIPE_TEX_FILTER_NEAREST &&
  66           mag_img_filter == PIPE_TEX_FILTER_NEAREST) {
  67          return FALSE;
  68       } else {
  69          return TRUE;
  70       }
  71    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
  72    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
  73       return TRUE;
  74    default:
  75       assert(0 && "unexpected wrap mode");
  76       return FALSE;
  77    }
  78 }
  79
  80
  81 /**
  82  * Initialize lp_sampler_static_state object with the gallium sampler
  83  * and texture state.
  84  * The former is considered to be static and the later dynamic.
  85  */
  86 void
  87 lp_sampler_static_state(struct lp_sampler_static_state *state,
  88                         const struct pipe_sampler_view *view,
  89                         const struct pipe_sampler_state *sampler)
  90 {
  91    const struct pipe_resource *texture = view->texture;
  92
  93    memset(state, 0, sizeof *state);
  94
  95    if(!texture)
  96       return;
  97
  98    if(!sampler)
  99       return;
 100
 101    /*
 102     * We don't copy sampler state over unless it is actually enabled, to avoid
 103     * spurious recompiles, as the sampler static state is part of the shader
 104     * key.
 105     *
 106     * Ideally the state tracker or cso_cache module would make all state
 107     * canonical, but until that happens it's better to be safe than sorry here.
 108     *
 109     * XXX: Actually there's much more than can be done here, especially
 110     * regarding 1D/2D/3D/CUBE textures, wrap modes, etc.
 111     */
 112
 113    state->format            = view->format;
 114    state->swizzle_r         = view->swizzle_r;
 115    state->swizzle_g         = view->swizzle_g;
 116    state->swizzle_b         = view->swizzle_b;
 117    state->swizzle_a         = view->swizzle_a;
 118
 119    state->target            = texture->target;
 120    state->pot_width         = util_is_power_of_two(texture->width0);
 121    state->pot_height        = util_is_power_of_two(texture->height0);
 122    state->pot_depth         = util_is_power_of_two(texture->depth0);
 123
 124    state->wrap_s            = sampler->wrap_s;
 125    state->wrap_t            = sampler->wrap_t;
 126    state->wrap_r            = sampler->wrap_r;
 127    state->min_img_filter    = sampler->min_img_filter;
 128    state->mag_img_filter    = sampler->mag_img_filter;
 129    if (view->last_level) {
 130       state->min_mip_filter = sampler->min_mip_filter;
 131    } else {
 132       state->min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 133    }
 134
 135    /* If min_lod == max_lod we can greatly simplify mipmap selection.
 136     * This is a case that occurs during automatic mipmap generation.
 137     */
 138    if (sampler->min_lod == sampler->max_lod) {
 139       state->min_max_lod_equal = 1;
 140    }
 141
 142    state->compare_mode      = sampler->compare_mode;
 143    if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
 144       state->compare_func   = sampler->compare_func;
 145    }
 146
 147    state->normalized_coords = sampler->normalized_coords;
 148
 149    /*
 150     * FIXME: Handle the remainder of pipe_sampler_view.
 151     */
 152 }
 153
 154
 155 /**
 156  * Generate code to compute texture level of detail (lambda).
 157  * \param ddx  partial derivatives of (s, t, r, q) with respect to X
 158  * \param ddy  partial derivatives of (s, t, r, q) with respect to Y
 159  * \param lod_bias  optional float vector with the shader lod bias
 160  * \param explicit_lod  optional float vector with the explicit lod
 161  * \param width  scalar int texture width
 162  * \param height  scalar int texture height
 163  * \param depth  scalar int texture depth
 164  *
 165  * XXX: The resulting lod is scalar, so ignore all but the first element of
 166  * derivatives, lod_bias, etc that are passed by the shader.
 167  */
 168 LLVMValueRef
 169 lp_build_lod_selector(struct lp_build_sample_context *bld,
 170                       unsigned unit,
 171                       const LLVMValueRef ddx[4],
 172                       const LLVMValueRef ddy[4],
 173                       LLVMValueRef lod_bias, /* optional */
 174                       LLVMValueRef explicit_lod, /* optional */
 175                       LLVMValueRef width,
 176                       LLVMValueRef height,
 177                       LLVMValueRef depth)
 178
 179 {
 180    LLVMValueRef min_lod =
 181       bld->dynamic_state->min_lod(bld->dynamic_state, bld->builder, unit);
 182
 183    if (bld->static_state->min_max_lod_equal) {
 184       /* User is forcing sampling from a particular mipmap level.
 185        * This is hit during mipmap generation.
 186        */
 187       return min_lod;
 188    }
 189    else {
 190       struct lp_build_context *float_bld = &bld->float_bld;
 191       LLVMValueRef sampler_lod_bias =
 192          bld->dynamic_state->lod_bias(bld->dynamic_state, bld->builder, unit);
 193       LLVMValueRef max_lod =
 194          bld->dynamic_state->max_lod(bld->dynamic_state, bld->builder, unit);
 195       LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 196       LLVMValueRef lod;
 197
 198       if (explicit_lod) {
 199          lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
 200                                        index0, "");
 201       }
 202       else {
 203          const int dims = texture_dims(bld->static_state->target);
 204          LLVMValueRef dsdx, dsdy;
 205          LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
 206          LLVMValueRef rho;
 207
 208          dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
 209          dsdx = lp_build_abs(float_bld, dsdx);
 210          dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
 211          dsdy = lp_build_abs(float_bld, dsdy);
 212          if (dims > 1) {
 213             dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
 214             dtdx = lp_build_abs(float_bld, dtdx);
 215             dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
 216             dtdy = lp_build_abs(float_bld, dtdy);
 217             if (dims > 2) {
 218                drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
 219                drdx = lp_build_abs(float_bld, drdx);
 220                drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
 221                drdy = lp_build_abs(float_bld, drdy);
 222             }
 223          }
 224
 225          /* Compute rho = max of all partial derivatives scaled by texture size.
 226           * XXX this could be vectorized somewhat
 227           */
 228          rho = LLVMBuildFMul(bld->builder,
 229                             lp_build_max(float_bld, dsdx, dsdy),
 230                             lp_build_int_to_float(float_bld, width), "");
 231          if (dims > 1) {
 232             LLVMValueRef max;
 233             max = LLVMBuildFMul(bld->builder,
 234                                lp_build_max(float_bld, dtdx, dtdy),
 235                                lp_build_int_to_float(float_bld, height), "");
 236             rho = lp_build_max(float_bld, rho, max);
 237             if (dims > 2) {
 238                max = LLVMBuildFMul(bld->builder,
 239                                   lp_build_max(float_bld, drdx, drdy),
 240                                   lp_build_int_to_float(float_bld, depth), "");
 241                rho = lp_build_max(float_bld, rho, max);
 242             }
 243          }
 244
 245          /* compute lod = log2(rho) */
 246 #if 0
 247          lod = lp_build_log2(float_bld, rho);
 248 #else
 249          lod = lp_build_fast_log2(float_bld, rho);
 250 #endif
 251
 252          /* add shader lod bias */
 253          if (lod_bias) {
 254             lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
 255                                                index0, "");
 256             lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
 257          }
 258       }
 259
 260       /* add sampler lod bias */
 261       lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
 262
 263       /* clamp lod */
 264       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
 265
 266       return lod;
 267    }
 268 }
 269
 270
 271 /**
 272  * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
 273  * mipmap level index.
 274  * Note: this is all scalar code.
 275  * \param lod  scalar float texture level of detail
 276  * \param level_out  returns integer
 277  */
 278 void
 279 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
 280                            unsigned unit,
 281                            LLVMValueRef lod,
 282                            LLVMValueRef *level_out)
 283 {
 284    struct lp_build_context *float_bld = &bld->float_bld;
 285    struct lp_build_context *int_bld = &bld->int_bld;
 286    LLVMValueRef last_level, level;
 287
 288    LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
 289
 290    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 291                                                bld->builder, unit);
 292
 293    /* convert float lod to integer */
 294    level = lp_build_iround(float_bld, lod);
 295
 296    /* clamp level to legal range of levels */
 297    *level_out = lp_build_clamp(int_bld, level, zero, last_level);
 298 }
 299
 300
 301 /**
 302  * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
 303  * two (adjacent) mipmap level indexes.  Later, we'll sample from those
 304  * two mipmap levels and interpolate between them.
 305  */
 306 void
 307 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
 308                            unsigned unit,
 309                            LLVMValueRef lod,
 310                            LLVMValueRef *level0_out,
 311                            LLVMValueRef *level1_out,
 312                            LLVMValueRef *weight_out)
 313 {
 314    struct lp_build_context *float_bld = &bld->float_bld;
 315    struct lp_build_context *int_bld = &bld->int_bld;
 316    LLVMValueRef last_level, level;
 317
 318    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 319                                                bld->builder, unit);
 320
 321    /* convert float lod to integer */
 322    lp_build_ifloor_fract(float_bld, lod, &level, weight_out);
 323
 324    /* compute level 0 and clamp to legal range of levels */
 325    *level0_out = lp_build_clamp(int_bld, level,
 326                                 int_bld->zero,
 327                                 last_level);
 328    /* compute level 1 and clamp to legal range of levels */
 329    level = lp_build_add(int_bld, level, int_bld->one);
 330    *level1_out = lp_build_clamp(int_bld, level,
 331                                 int_bld->zero,
 332                                 last_level);
 333 }
 334
 335
 336 /**
 337  * Return pointer to a single mipmap level.
 338  * \param data_array  array of pointers to mipmap levels
 339  * \param level  integer mipmap level
 340  */
 341 LLVMValueRef
 342 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
 343                           LLVMValueRef data_array, LLVMValueRef level)
 344 {
 345    LLVMValueRef indexes[2], data_ptr;
 346    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 347    indexes[1] = level;
 348    data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
 349    data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
 350    return data_ptr;
 351 }
 352
 353
 354 LLVMValueRef
 355 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
 356                                 LLVMValueRef data_array, int level)
 357 {
 358    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
 359    return lp_build_get_mipmap_level(bld, data_array, lvl);
 360 }
 361
 362
 363 /**
 364  * Codegen equivalent for u_minify().
 365  * Return max(1, base_size >> level);
 366  */
 367 static LLVMValueRef
 368 lp_build_minify(struct lp_build_sample_context *bld,
 369                 LLVMValueRef base_size,
 370                 LLVMValueRef level)
 371 {
 372    if (level == bld->int_coord_bld.zero) {
 373       /* if we're using mipmap level zero, no minification is needed */
 374       return base_size;
 375    }
 376    else {
 377       LLVMValueRef size =
 378          LLVMBuildLShr(bld->builder, base_size, level, "minify");
 379       size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
 380       return size;
 381    }
 382 }
 383
 384
 385 /**
 386  * Dereference stride_array[mipmap_level] array to get a stride.
 387  * Return stride as a vector.
 388  */
 389 static LLVMValueRef
 390 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
 391                               LLVMValueRef stride_array, LLVMValueRef level)
 392 {
 393    LLVMValueRef indexes[2], stride;
 394    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 395    indexes[1] = level;
 396    stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
 397    stride = LLVMBuildLoad(bld->builder, stride, "");
 398    stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
 399    return stride;
 400 }
 401
 402
 403 /**
 404  * When sampling a mipmap, we need to compute the width, height, depth
 405  * of the source levels from the level indexes.  This helper function
 406  * does that.
 407  */
 408 void
 409 lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
 410                             unsigned dims,
 411                             LLVMValueRef width_vec,
 412                             LLVMValueRef height_vec,
 413                             LLVMValueRef depth_vec,
 414                             LLVMValueRef ilevel0,
 415                             LLVMValueRef ilevel1,
 416                             LLVMValueRef row_stride_array,
 417                             LLVMValueRef img_stride_array,
 418                             LLVMValueRef *width0_vec,
 419                             LLVMValueRef *width1_vec,
 420                             LLVMValueRef *height0_vec,
 421                             LLVMValueRef *height1_vec,
 422                             LLVMValueRef *depth0_vec,
 423                             LLVMValueRef *depth1_vec,
 424                             LLVMValueRef *row_stride0_vec,
 425                             LLVMValueRef *row_stride1_vec,
 426                             LLVMValueRef *img_stride0_vec,
 427                             LLVMValueRef *img_stride1_vec)
 428 {
 429    const unsigned mip_filter = bld->static_state->min_mip_filter;
 430    LLVMValueRef ilevel0_vec, ilevel1_vec;
 431
 432    ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
 433    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
 434       ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
 435
 436    /*
 437     * Compute width, height, depth at mipmap level 'ilevel0'
 438     */
 439    *width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
 440    if (dims >= 2) {
 441       *height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
 442       *row_stride0_vec = lp_build_get_level_stride_vec(bld,
 443                                                        row_stride_array,
 444                                                        ilevel0);
 445       if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 446          *img_stride0_vec = lp_build_get_level_stride_vec(bld,
 447                                                           img_stride_array,
 448                                                           ilevel0);
 449          if (dims == 3) {
 450             *depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
 451          }
 452       }
 453    }
 454    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
 455       /* compute width, height, depth for second mipmap level at 'ilevel1' */
 456       *width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
 457       if (dims >= 2) {
 458          *height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
 459          *row_stride1_vec = lp_build_get_level_stride_vec(bld,
 460                                                           row_stride_array,
 461                                                           ilevel1);
 462          if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 463             *img_stride1_vec = lp_build_get_level_stride_vec(bld,
 464                                                              img_stride_array,
 465                                                              ilevel1);
 466             if (dims == 3) {
 467                *depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
 468             }
 469          }
 470       }
 471    }
 472 }
 473
 474
 475
 476 /** Helper used by lp_build_cube_lookup() */
 477 static LLVMValueRef
 478 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
 479 {
 480    /* ima = -0.5 / abs(coord); */
 481    LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
 482    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
 483    LLVMValueRef ima = lp_build_div(coord_bld, negHalf, absCoord);
 484    return ima;
 485 }
 486
 487
 488 /**
 489  * Helper used by lp_build_cube_lookup()
 490  * \param sign  scalar +1 or -1
 491  * \param coord  float vector
 492  * \param ima  float vector
 493  */
 494 static LLVMValueRef
 495 lp_build_cube_coord(struct lp_build_context *coord_bld,
 496                     LLVMValueRef sign, int negate_coord,
 497                     LLVMValueRef coord, LLVMValueRef ima)
 498 {
 499    /* return negate(coord) * ima * sign + 0.5; */
 500    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
 501    LLVMValueRef res;
 502
 503    assert(negate_coord == +1 || negate_coord == -1);
 504
 505    if (negate_coord == -1) {
 506       coord = lp_build_negate(coord_bld, coord);
 507    }
 508
 509    res = lp_build_mul(coord_bld, coord, ima);
 510    if (sign) {
 511       sign = lp_build_broadcast_scalar(coord_bld, sign);
 512       res = lp_build_mul(coord_bld, res, sign);
 513    }
 514    res = lp_build_add(coord_bld, res, half);
 515
 516    return res;
 517 }
 518
 519
 520 /** Helper used by lp_build_cube_lookup()
 521  * Return (major_coord >= 0) ? pos_face : neg_face;
 522  */
 523 static LLVMValueRef
 524 lp_build_cube_face(struct lp_build_sample_context *bld,
 525                    LLVMValueRef major_coord,
 526                    unsigned pos_face, unsigned neg_face)
 527 {
 528    LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
 529                                     major_coord,
 530                                     bld->float_bld.zero, "");
 531    LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
 532    LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
 533    LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
 534    return res;
 535 }
 536
 537
 538
 539 /**
 540  * Generate code to do cube face selection and compute per-face texcoords.
 541  */
 542 void
 543 lp_build_cube_lookup(struct lp_build_sample_context *bld,
 544                      LLVMValueRef s,
 545                      LLVMValueRef t,
 546                      LLVMValueRef r,
 547                      LLVMValueRef *face,
 548                      LLVMValueRef *face_s,
 549                      LLVMValueRef *face_t)
 550 {
 551    struct lp_build_context *float_bld = &bld->float_bld;
 552    struct lp_build_context *coord_bld = &bld->coord_bld;
 553    LLVMValueRef rx, ry, rz;
 554    LLVMValueRef arx, ary, arz;
 555    LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
 556    LLVMValueRef arx_ge_ary, arx_ge_arz;
 557    LLVMValueRef ary_ge_arx, ary_ge_arz;
 558    LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
 559    LLVMValueRef rx_pos, ry_pos, rz_pos;
 560
 561    assert(bld->coord_bld.type.length == 4);
 562
 563    /*
 564     * Use the average of the four pixel's texcoords to choose the face.
 565     */
 566    rx = lp_build_mul(float_bld, c25,
 567                      lp_build_sum_vector(&bld->coord_bld, s));
 568    ry = lp_build_mul(float_bld, c25,
 569                      lp_build_sum_vector(&bld->coord_bld, t));
 570    rz = lp_build_mul(float_bld, c25,
 571                      lp_build_sum_vector(&bld->coord_bld, r));
 572
 573    arx = lp_build_abs(float_bld, rx);
 574    ary = lp_build_abs(float_bld, ry);
 575    arz = lp_build_abs(float_bld, rz);
 576
 577    /*
 578     * Compare sign/magnitude of rx,ry,rz to determine face
 579     */
 580    arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
 581    arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
 582    ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
 583    ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
 584
 585    arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
 586    ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 587
 588    rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
 589    ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
 590    rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
 591
 592    {
 593       struct lp_build_flow_context *flow_ctx;
 594       struct lp_build_if_state if_ctx;
 595
 596       flow_ctx = lp_build_flow_create(bld->builder);
 597       lp_build_flow_scope_begin(flow_ctx);
 598
 599       *face_s = bld->coord_bld.undef;
 600       *face_t = bld->coord_bld.undef;
 601       *face = bld->int_bld.undef;
 602
 603       lp_build_name(*face_s, "face_s");
 604       lp_build_name(*face_t, "face_t");
 605       lp_build_name(*face, "face");
 606
 607       lp_build_flow_scope_declare(flow_ctx, face_s);
 608       lp_build_flow_scope_declare(flow_ctx, face_t);
 609       lp_build_flow_scope_declare(flow_ctx, face);
 610
 611       lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
 612       {
 613          /* +/- X face */
 614          LLVMValueRef sign = lp_build_sgn(float_bld, rx);
 615          LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
 616          *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
 617          *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 618          *face = lp_build_cube_face(bld, rx,
 619                                     PIPE_TEX_FACE_POS_X,
 620                                     PIPE_TEX_FACE_NEG_X);
 621       }
 622       lp_build_else(&if_ctx);
 623       {
 624          struct lp_build_flow_context *flow_ctx2;
 625          struct lp_build_if_state if_ctx2;
 626
 627          LLVMValueRef face_s2 = bld->coord_bld.undef;
 628          LLVMValueRef face_t2 = bld->coord_bld.undef;
 629          LLVMValueRef face2 = bld->int_bld.undef;
 630
 631          flow_ctx2 = lp_build_flow_create(bld->builder);
 632          lp_build_flow_scope_begin(flow_ctx2);
 633          lp_build_flow_scope_declare(flow_ctx2, &face_s2);
 634          lp_build_flow_scope_declare(flow_ctx2, &face_t2);
 635          lp_build_flow_scope_declare(flow_ctx2, &face2);
 636
 637          ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 638
 639          lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
 640          {
 641             /* +/- Y face */
 642             LLVMValueRef sign = lp_build_sgn(float_bld, ry);
 643             LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
 644             face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
 645             face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
 646             face2 = lp_build_cube_face(bld, ry,
 647                                        PIPE_TEX_FACE_POS_Y,
 648                                        PIPE_TEX_FACE_NEG_Y);
 649          }
 650          lp_build_else(&if_ctx2);
 651          {
 652             /* +/- Z face */
 653             LLVMValueRef sign = lp_build_sgn(float_bld, rz);
 654             LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
 655             face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
 656             face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 657             face2 = lp_build_cube_face(bld, rz,
 658                                        PIPE_TEX_FACE_POS_Z,
 659                                        PIPE_TEX_FACE_NEG_Z);
 660          }
 661          lp_build_endif(&if_ctx2);
 662          lp_build_flow_scope_end(flow_ctx2);
 663          lp_build_flow_destroy(flow_ctx2);
 664          *face_s = face_s2;
 665          *face_t = face_t2;
 666          *face = face2;
 667       }
 668
 669       lp_build_endif(&if_ctx);
 670       lp_build_flow_scope_end(flow_ctx);
 671       lp_build_flow_destroy(flow_ctx);
 672    }
 673 }
 674
 675
 676 /**
 677  * Compute the partial offset of a pixel block along an arbitrary axis.
 678  *
 679  * @param coord   coordinate in pixels
 680  * @param stride  number of bytes between rows of successive pixel blocks
 681  * @param block_length  number of pixels in a pixels block along the coordinate
 682  *                      axis
 683  * @param out_offset    resulting relative offset of the pixel block in bytes
 684  * @param out_subcoord  resulting sub-block pixel coordinate
 685  */
 686 void
 687 lp_build_sample_partial_offset(struct lp_build_context *bld,
 688                                unsigned block_length,
 689                                LLVMValueRef coord,
 690                                LLVMValueRef stride,
 691                                LLVMValueRef *out_offset,
 692                                LLVMValueRef *out_subcoord)
 693 {
 694    LLVMValueRef offset;
 695    LLVMValueRef subcoord;
 696
 697    if (block_length == 1) {
 698       subcoord = bld->zero;
 699    }
 700    else {
 701       /*
 702        * Pixel blocks have power of two dimensions. LLVM should convert the
 703        * rem/div to bit arithmetic.
 704        * TODO: Verify this.
 705        * It does indeed BUT it does transform it to scalar (and back) when doing so
 706        * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
 707        * The generated code looks seriously unfunny and is quite expensive.
 708        */
 709 #if 0
 710       LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
 711       subcoord = LLVMBuildURem(bld->builder, coord, block_width, "");
 712       coord    = LLVMBuildUDiv(bld->builder, coord, block_width, "");
 713 #else
 714       unsigned logbase2 = util_unsigned_logbase2(block_length);
 715       LLVMValueRef block_shift = lp_build_const_int_vec(bld->type, logbase2);
 716       LLVMValueRef block_mask = lp_build_const_int_vec(bld->type, block_length - 1);
 717       subcoord = LLVMBuildAnd(bld->builder, coord, block_mask, "");
 718       coord = LLVMBuildLShr(bld->builder, coord, block_shift, "");
 719 #endif
 720    }
 721
 722    offset = lp_build_mul(bld, coord, stride);
 723
 724    assert(out_offset);
 725    assert(out_subcoord);
 726
 727    *out_offset = offset;
 728    *out_subcoord = subcoord;
 729 }
 730
 731
 732 /**
 733  * Compute the offset of a pixel block.
 734  *
 735  * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
 736  *
 737  * Returns the relative offset and i,j sub-block coordinates
 738  */
 739 void
 740 lp_build_sample_offset(struct lp_build_context *bld,
 741                        const struct util_format_description *format_desc,
 742                        LLVMValueRef x,
 743                        LLVMValueRef y,
 744                        LLVMValueRef z,
 745                        LLVMValueRef y_stride,
 746                        LLVMValueRef z_stride,
 747                        LLVMValueRef *out_offset,
 748                        LLVMValueRef *out_i,
 749                        LLVMValueRef *out_j)
 750 {
 751    LLVMValueRef x_stride;
 752    LLVMValueRef offset;
 753
 754    x_stride = lp_build_const_vec(bld->type, format_desc->block.bits/8);
 755
 756    lp_build_sample_partial_offset(bld,
 757                                   format_desc->block.width,
 758                                   x, x_stride,
 759                                   &offset, out_i);
 760
 761    if (y && y_stride) {
 762       LLVMValueRef y_offset;
 763       lp_build_sample_partial_offset(bld,
 764                                      format_desc->block.height,
 765                                      y, y_stride,
 766                                      &y_offset, out_j);
 767       offset = lp_build_add(bld, offset, y_offset);
 768    }
 769    else {
 770       *out_j = bld->zero;
 771    }
 772
 773    if (z && z_stride) {
 774       LLVMValueRef z_offset;
 775       LLVMValueRef k;
 776       lp_build_sample_partial_offset(bld,
 777                                      1, /* pixel blocks are always 2D */
 778                                      z, z_stride,
 779                                      &z_offset, &k);
 780       offset = lp_build_add(bld, offset, z_offset);
 781    }
 782
 783    *out_offset = offset;
 784 }