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 (sampler->lod_bias != 0.0) {
 136       state->lod_bias_non_zero = 1;
 137    }
 138
 139    /* If min_lod == max_lod we can greatly simplify mipmap selection.
 140     * This is a case that occurs during automatic mipmap generation.
 141     */
 142    if (sampler->min_lod == sampler->max_lod) {
 143       state->min_max_lod_equal = 1;
 144    }
 145
 146    state->compare_mode      = sampler->compare_mode;
 147    if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
 148       state->compare_func   = sampler->compare_func;
 149    }
 150
 151    state->normalized_coords = sampler->normalized_coords;
 152
 153    /*
 154     * FIXME: Handle the remainder of pipe_sampler_view.
 155     */
 156 }
 157
 158
 159 /**
 160  * Generate code to compute texture level of detail (lambda).
 161  * \param ddx  partial derivatives of (s, t, r, q) with respect to X
 162  * \param ddy  partial derivatives of (s, t, r, q) with respect to Y
 163  * \param lod_bias  optional float vector with the shader lod bias
 164  * \param explicit_lod  optional float vector with the explicit lod
 165  * \param width  scalar int texture width
 166  * \param height  scalar int texture height
 167  * \param depth  scalar int texture depth
 168  *
 169  * XXX: The resulting lod is scalar, so ignore all but the first element of
 170  * derivatives, lod_bias, etc that are passed by the shader.
 171  */
 172 LLVMValueRef
 173 lp_build_lod_selector(struct lp_build_sample_context *bld,
 174                       unsigned unit,
 175                       const LLVMValueRef ddx[4],
 176                       const LLVMValueRef ddy[4],
 177                       LLVMValueRef lod_bias, /* optional */
 178                       LLVMValueRef explicit_lod, /* optional */
 179                       LLVMValueRef width,
 180                       LLVMValueRef height,
 181                       LLVMValueRef depth)
 182
 183 {
 184    LLVMValueRef min_lod =
 185       bld->dynamic_state->min_lod(bld->dynamic_state, bld->builder, unit);
 186
 187    if (bld->static_state->min_max_lod_equal) {
 188       /* User is forcing sampling from a particular mipmap level.
 189        * This is hit during mipmap generation.
 190        */
 191       return min_lod;
 192    }
 193    else {
 194       struct lp_build_context *float_bld = &bld->float_bld;
 195       LLVMValueRef sampler_lod_bias =
 196          bld->dynamic_state->lod_bias(bld->dynamic_state, bld->builder, unit);
 197       LLVMValueRef max_lod =
 198          bld->dynamic_state->max_lod(bld->dynamic_state, bld->builder, unit);
 199       LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 200       LLVMValueRef lod;
 201
 202       if (explicit_lod) {
 203          lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
 204                                        index0, "");
 205       }
 206       else {
 207          const int dims = texture_dims(bld->static_state->target);
 208          LLVMValueRef dsdx, dsdy;
 209          LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
 210          LLVMValueRef rho;
 211
 212          dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
 213          dsdx = lp_build_abs(float_bld, dsdx);
 214          dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
 215          dsdy = lp_build_abs(float_bld, dsdy);
 216          if (dims > 1) {
 217             dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
 218             dtdx = lp_build_abs(float_bld, dtdx);
 219             dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
 220             dtdy = lp_build_abs(float_bld, dtdy);
 221             if (dims > 2) {
 222                drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
 223                drdx = lp_build_abs(float_bld, drdx);
 224                drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
 225                drdy = lp_build_abs(float_bld, drdy);
 226             }
 227          }
 228
 229          /* Compute rho = max of all partial derivatives scaled by texture size.
 230           * XXX this could be vectorized somewhat
 231           */
 232          rho = LLVMBuildFMul(bld->builder,
 233                             lp_build_max(float_bld, dsdx, dsdy),
 234                             lp_build_int_to_float(float_bld, width), "");
 235          if (dims > 1) {
 236             LLVMValueRef max;
 237             max = LLVMBuildFMul(bld->builder,
 238                                lp_build_max(float_bld, dtdx, dtdy),
 239                                lp_build_int_to_float(float_bld, height), "");
 240             rho = lp_build_max(float_bld, rho, max);
 241             if (dims > 2) {
 242                max = LLVMBuildFMul(bld->builder,
 243                                   lp_build_max(float_bld, drdx, drdy),
 244                                   lp_build_int_to_float(float_bld, depth), "");
 245                rho = lp_build_max(float_bld, rho, max);
 246             }
 247          }
 248
 249          /* compute lod = log2(rho) */
 250 #if 0
 251          lod = lp_build_log2(float_bld, rho);
 252 #else
 253          lod = lp_build_fast_log2(float_bld, rho);
 254 #endif
 255
 256          /* add shader lod bias */
 257          if (lod_bias) {
 258             lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
 259                                                index0, "");
 260             lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
 261          }
 262       }
 263
 264       /* add sampler lod bias */
 265       if (bld->static_state->lod_bias_non_zero)
 266          lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
 267
 268       /* clamp lod */
 269       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
 270
 271       return lod;
 272    }
 273 }
 274
 275
 276 /**
 277  * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
 278  * mipmap level index.
 279  * Note: this is all scalar code.
 280  * \param lod  scalar float texture level of detail
 281  * \param level_out  returns integer
 282  */
 283 void
 284 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
 285                            unsigned unit,
 286                            LLVMValueRef lod,
 287                            LLVMValueRef *level_out)
 288 {
 289    struct lp_build_context *float_bld = &bld->float_bld;
 290    struct lp_build_context *int_bld = &bld->int_bld;
 291    LLVMValueRef last_level, level;
 292
 293    LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
 294
 295    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 296                                                bld->builder, unit);
 297
 298    /* convert float lod to integer */
 299    level = lp_build_iround(float_bld, lod);
 300
 301    /* clamp level to legal range of levels */
 302    *level_out = lp_build_clamp(int_bld, level, zero, last_level);
 303 }
 304
 305
 306 /**
 307  * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
 308  * two (adjacent) mipmap level indexes.  Later, we'll sample from those
 309  * two mipmap levels and interpolate between them.
 310  */
 311 void
 312 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
 313                            unsigned unit,
 314                            LLVMValueRef lod,
 315                            LLVMValueRef *level0_out,
 316                            LLVMValueRef *level1_out,
 317                            LLVMValueRef *weight_out)
 318 {
 319    struct lp_build_context *float_bld = &bld->float_bld;
 320    struct lp_build_context *int_bld = &bld->int_bld;
 321    LLVMValueRef last_level, level;
 322
 323    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 324                                                bld->builder, unit);
 325
 326    /* convert float lod to integer */
 327    lp_build_ifloor_fract(float_bld, lod, &level, weight_out);
 328
 329    /* compute level 0 and clamp to legal range of levels */
 330    *level0_out = lp_build_clamp(int_bld, level,
 331                                 int_bld->zero,
 332                                 last_level);
 333    /* compute level 1 and clamp to legal range of levels */
 334    level = lp_build_add(int_bld, level, int_bld->one);
 335    *level1_out = lp_build_clamp(int_bld, level,
 336                                 int_bld->zero,
 337                                 last_level);
 338 }
 339
 340
 341 /**
 342  * Return pointer to a single mipmap level.
 343  * \param data_array  array of pointers to mipmap levels
 344  * \param level  integer mipmap level
 345  */
 346 LLVMValueRef
 347 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
 348                           LLVMValueRef data_array, LLVMValueRef level)
 349 {
 350    LLVMValueRef indexes[2], data_ptr;
 351    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 352    indexes[1] = level;
 353    data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
 354    data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
 355    return data_ptr;
 356 }
 357
 358
 359 LLVMValueRef
 360 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
 361                                 LLVMValueRef data_array, int level)
 362 {
 363    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
 364    return lp_build_get_mipmap_level(bld, data_array, lvl);
 365 }
 366
 367
 368 /**
 369  * Codegen equivalent for u_minify().
 370  * Return max(1, base_size >> level);
 371  */
 372 static LLVMValueRef
 373 lp_build_minify(struct lp_build_sample_context *bld,
 374                 LLVMValueRef base_size,
 375                 LLVMValueRef level)
 376 {
 377    if (level == bld->int_coord_bld.zero) {
 378       /* if we're using mipmap level zero, no minification is needed */
 379       return base_size;
 380    }
 381    else {
 382       LLVMValueRef size =
 383          LLVMBuildLShr(bld->builder, base_size, level, "minify");
 384       size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
 385       return size;
 386    }
 387 }
 388
 389
 390 /**
 391  * Dereference stride_array[mipmap_level] array to get a stride.
 392  * Return stride as a vector.
 393  */
 394 static LLVMValueRef
 395 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
 396                               LLVMValueRef stride_array, LLVMValueRef level)
 397 {
 398    LLVMValueRef indexes[2], stride;
 399    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 400    indexes[1] = level;
 401    stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
 402    stride = LLVMBuildLoad(bld->builder, stride, "");
 403    stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
 404    return stride;
 405 }
 406
 407
 408 /**
 409  * When sampling a mipmap, we need to compute the width, height, depth
 410  * of the source levels from the level indexes.  This helper function
 411  * does that.
 412  */
 413 void
 414 lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
 415                             unsigned dims,
 416                             LLVMValueRef width_vec,
 417                             LLVMValueRef height_vec,
 418                             LLVMValueRef depth_vec,
 419                             LLVMValueRef ilevel0,
 420                             LLVMValueRef ilevel1,
 421                             LLVMValueRef row_stride_array,
 422                             LLVMValueRef img_stride_array,
 423                             LLVMValueRef *width0_vec,
 424                             LLVMValueRef *width1_vec,
 425                             LLVMValueRef *height0_vec,
 426                             LLVMValueRef *height1_vec,
 427                             LLVMValueRef *depth0_vec,
 428                             LLVMValueRef *depth1_vec,
 429                             LLVMValueRef *row_stride0_vec,
 430                             LLVMValueRef *row_stride1_vec,
 431                             LLVMValueRef *img_stride0_vec,
 432                             LLVMValueRef *img_stride1_vec)
 433 {
 434    const unsigned mip_filter = bld->static_state->min_mip_filter;
 435    LLVMValueRef ilevel0_vec, ilevel1_vec;
 436
 437    ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
 438    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
 439       ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
 440
 441    /*
 442     * Compute width, height, depth at mipmap level 'ilevel0'
 443     */
 444    *width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
 445    if (dims >= 2) {
 446       *height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
 447       *row_stride0_vec = lp_build_get_level_stride_vec(bld,
 448                                                        row_stride_array,
 449                                                        ilevel0);
 450       if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 451          *img_stride0_vec = lp_build_get_level_stride_vec(bld,
 452                                                           img_stride_array,
 453                                                           ilevel0);
 454          if (dims == 3) {
 455             *depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
 456          }
 457       }
 458    }
 459    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
 460       /* compute width, height, depth for second mipmap level at 'ilevel1' */
 461       *width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
 462       if (dims >= 2) {
 463          *height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
 464          *row_stride1_vec = lp_build_get_level_stride_vec(bld,
 465                                                           row_stride_array,
 466                                                           ilevel1);
 467          if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 468             *img_stride1_vec = lp_build_get_level_stride_vec(bld,
 469                                                              img_stride_array,
 470                                                              ilevel1);
 471             if (dims == 3) {
 472                *depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
 473             }
 474          }
 475       }
 476    }
 477 }
 478
 479
 480
 481 /** Helper used by lp_build_cube_lookup() */
 482 static LLVMValueRef
 483 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
 484 {
 485    /* ima = -0.5 / abs(coord); */
 486    LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
 487    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
 488    LLVMValueRef ima = lp_build_div(coord_bld, negHalf, absCoord);
 489    return ima;
 490 }
 491
 492
 493 /**
 494  * Helper used by lp_build_cube_lookup()
 495  * \param sign  scalar +1 or -1
 496  * \param coord  float vector
 497  * \param ima  float vector
 498  */
 499 static LLVMValueRef
 500 lp_build_cube_coord(struct lp_build_context *coord_bld,
 501                     LLVMValueRef sign, int negate_coord,
 502                     LLVMValueRef coord, LLVMValueRef ima)
 503 {
 504    /* return negate(coord) * ima * sign + 0.5; */
 505    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
 506    LLVMValueRef res;
 507
 508    assert(negate_coord == +1 || negate_coord == -1);
 509
 510    if (negate_coord == -1) {
 511       coord = lp_build_negate(coord_bld, coord);
 512    }
 513
 514    res = lp_build_mul(coord_bld, coord, ima);
 515    if (sign) {
 516       sign = lp_build_broadcast_scalar(coord_bld, sign);
 517       res = lp_build_mul(coord_bld, res, sign);
 518    }
 519    res = lp_build_add(coord_bld, res, half);
 520
 521    return res;
 522 }
 523
 524
 525 /** Helper used by lp_build_cube_lookup()
 526  * Return (major_coord >= 0) ? pos_face : neg_face;
 527  */
 528 static LLVMValueRef
 529 lp_build_cube_face(struct lp_build_sample_context *bld,
 530                    LLVMValueRef major_coord,
 531                    unsigned pos_face, unsigned neg_face)
 532 {
 533    LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
 534                                     major_coord,
 535                                     bld->float_bld.zero, "");
 536    LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
 537    LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
 538    LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
 539    return res;
 540 }
 541
 542
 543
 544 /**
 545  * Generate code to do cube face selection and compute per-face texcoords.
 546  */
 547 void
 548 lp_build_cube_lookup(struct lp_build_sample_context *bld,
 549                      LLVMValueRef s,
 550                      LLVMValueRef t,
 551                      LLVMValueRef r,
 552                      LLVMValueRef *face,
 553                      LLVMValueRef *face_s,
 554                      LLVMValueRef *face_t)
 555 {
 556    struct lp_build_context *float_bld = &bld->float_bld;
 557    struct lp_build_context *coord_bld = &bld->coord_bld;
 558    LLVMValueRef rx, ry, rz;
 559    LLVMValueRef arx, ary, arz;
 560    LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
 561    LLVMValueRef arx_ge_ary, arx_ge_arz;
 562    LLVMValueRef ary_ge_arx, ary_ge_arz;
 563    LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
 564    LLVMValueRef rx_pos, ry_pos, rz_pos;
 565
 566    assert(bld->coord_bld.type.length == 4);
 567
 568    /*
 569     * Use the average of the four pixel's texcoords to choose the face.
 570     */
 571    rx = lp_build_mul(float_bld, c25,
 572                      lp_build_sum_vector(&bld->coord_bld, s));
 573    ry = lp_build_mul(float_bld, c25,
 574                      lp_build_sum_vector(&bld->coord_bld, t));
 575    rz = lp_build_mul(float_bld, c25,
 576                      lp_build_sum_vector(&bld->coord_bld, r));
 577
 578    arx = lp_build_abs(float_bld, rx);
 579    ary = lp_build_abs(float_bld, ry);
 580    arz = lp_build_abs(float_bld, rz);
 581
 582    /*
 583     * Compare sign/magnitude of rx,ry,rz to determine face
 584     */
 585    arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
 586    arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
 587    ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
 588    ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
 589
 590    arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
 591    ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 592
 593    rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
 594    ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
 595    rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
 596
 597    {
 598       struct lp_build_flow_context *flow_ctx;
 599       struct lp_build_if_state if_ctx;
 600
 601       flow_ctx = lp_build_flow_create(bld->builder);
 602       lp_build_flow_scope_begin(flow_ctx);
 603
 604       *face_s = bld->coord_bld.undef;
 605       *face_t = bld->coord_bld.undef;
 606       *face = bld->int_bld.undef;
 607
 608       lp_build_name(*face_s, "face_s");
 609       lp_build_name(*face_t, "face_t");
 610       lp_build_name(*face, "face");
 611
 612       lp_build_flow_scope_declare(flow_ctx, face_s);
 613       lp_build_flow_scope_declare(flow_ctx, face_t);
 614       lp_build_flow_scope_declare(flow_ctx, face);
 615
 616       lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
 617       {
 618          /* +/- X face */
 619          LLVMValueRef sign = lp_build_sgn(float_bld, rx);
 620          LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
 621          *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
 622          *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 623          *face = lp_build_cube_face(bld, rx,
 624                                     PIPE_TEX_FACE_POS_X,
 625                                     PIPE_TEX_FACE_NEG_X);
 626       }
 627       lp_build_else(&if_ctx);
 628       {
 629          struct lp_build_flow_context *flow_ctx2;
 630          struct lp_build_if_state if_ctx2;
 631
 632          LLVMValueRef face_s2 = bld->coord_bld.undef;
 633          LLVMValueRef face_t2 = bld->coord_bld.undef;
 634          LLVMValueRef face2 = bld->int_bld.undef;
 635
 636          flow_ctx2 = lp_build_flow_create(bld->builder);
 637          lp_build_flow_scope_begin(flow_ctx2);
 638          lp_build_flow_scope_declare(flow_ctx2, &face_s2);
 639          lp_build_flow_scope_declare(flow_ctx2, &face_t2);
 640          lp_build_flow_scope_declare(flow_ctx2, &face2);
 641
 642          ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 643
 644          lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
 645          {
 646             /* +/- Y face */
 647             LLVMValueRef sign = lp_build_sgn(float_bld, ry);
 648             LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
 649             face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
 650             face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
 651             face2 = lp_build_cube_face(bld, ry,
 652                                        PIPE_TEX_FACE_POS_Y,
 653                                        PIPE_TEX_FACE_NEG_Y);
 654          }
 655          lp_build_else(&if_ctx2);
 656          {
 657             /* +/- Z face */
 658             LLVMValueRef sign = lp_build_sgn(float_bld, rz);
 659             LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
 660             face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
 661             face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 662             face2 = lp_build_cube_face(bld, rz,
 663                                        PIPE_TEX_FACE_POS_Z,
 664                                        PIPE_TEX_FACE_NEG_Z);
 665          }
 666          lp_build_endif(&if_ctx2);
 667          lp_build_flow_scope_end(flow_ctx2);
 668          lp_build_flow_destroy(flow_ctx2);
 669          *face_s = face_s2;
 670          *face_t = face_t2;
 671          *face = face2;
 672       }
 673
 674       lp_build_endif(&if_ctx);
 675       lp_build_flow_scope_end(flow_ctx);
 676       lp_build_flow_destroy(flow_ctx);
 677    }
 678 }
 679
 680
 681 /**
 682  * Compute the partial offset of a pixel block along an arbitrary axis.
 683  *
 684  * @param coord   coordinate in pixels
 685  * @param stride  number of bytes between rows of successive pixel blocks
 686  * @param block_length  number of pixels in a pixels block along the coordinate
 687  *                      axis
 688  * @param out_offset    resulting relative offset of the pixel block in bytes
 689  * @param out_subcoord  resulting sub-block pixel coordinate
 690  */
 691 void
 692 lp_build_sample_partial_offset(struct lp_build_context *bld,
 693                                unsigned block_length,
 694                                LLVMValueRef coord,
 695                                LLVMValueRef stride,
 696                                LLVMValueRef *out_offset,
 697                                LLVMValueRef *out_subcoord)
 698 {
 699    LLVMValueRef offset;
 700    LLVMValueRef subcoord;
 701
 702    if (block_length == 1) {
 703       subcoord = bld->zero;
 704    }
 705    else {
 706       /*
 707        * Pixel blocks have power of two dimensions. LLVM should convert the
 708        * rem/div to bit arithmetic.
 709        * TODO: Verify this.
 710        * It does indeed BUT it does transform it to scalar (and back) when doing so
 711        * (using roughly extract, shift/and, mov, unpack) (llvm 2.7).
 712        * The generated code looks seriously unfunny and is quite expensive.
 713        */
 714 #if 0
 715       LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
 716       subcoord = LLVMBuildURem(bld->builder, coord, block_width, "");
 717       coord    = LLVMBuildUDiv(bld->builder, coord, block_width, "");
 718 #else
 719       unsigned logbase2 = util_unsigned_logbase2(block_length);
 720       LLVMValueRef block_shift = lp_build_const_int_vec(bld->type, logbase2);
 721       LLVMValueRef block_mask = lp_build_const_int_vec(bld->type, block_length - 1);
 722       subcoord = LLVMBuildAnd(bld->builder, coord, block_mask, "");
 723       coord = LLVMBuildLShr(bld->builder, coord, block_shift, "");
 724 #endif
 725    }
 726
 727    offset = lp_build_mul(bld, coord, stride);
 728
 729    assert(out_offset);
 730    assert(out_subcoord);
 731
 732    *out_offset = offset;
 733    *out_subcoord = subcoord;
 734 }
 735
 736
 737 /**
 738  * Compute the offset of a pixel block.
 739  *
 740  * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
 741  *
 742  * Returns the relative offset and i,j sub-block coordinates
 743  */
 744 void
 745 lp_build_sample_offset(struct lp_build_context *bld,
 746                        const struct util_format_description *format_desc,
 747                        LLVMValueRef x,
 748                        LLVMValueRef y,
 749                        LLVMValueRef z,
 750                        LLVMValueRef y_stride,
 751                        LLVMValueRef z_stride,
 752                        LLVMValueRef *out_offset,
 753                        LLVMValueRef *out_i,
 754                        LLVMValueRef *out_j)
 755 {
 756    LLVMValueRef x_stride;
 757    LLVMValueRef offset;
 758
 759    x_stride = lp_build_const_vec(bld->type, format_desc->block.bits/8);
 760
 761    lp_build_sample_partial_offset(bld,
 762                                   format_desc->block.width,
 763                                   x, x_stride,
 764                                   &offset, out_i);
 765
 766    if (y && y_stride) {
 767       LLVMValueRef y_offset;
 768       lp_build_sample_partial_offset(bld,
 769                                      format_desc->block.height,
 770                                      y, y_stride,
 771                                      &y_offset, out_j);
 772       offset = lp_build_add(bld, offset, y_offset);
 773    }
 774    else {
 775       *out_j = bld->zero;
 776    }
 777
 778    if (z && z_stride) {
 779       LLVMValueRef z_offset;
 780       LLVMValueRef k;
 781       lp_build_sample_partial_offset(bld,
 782                                      1, /* pixel blocks are always 2D */
 783                                      z, z_stride,
 784                                      &z_offset, &k);
 785       offset = lp_build_add(bld, offset, z_offset);
 786    }
 787
 788    *out_offset = offset;
 789 }