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