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    state->compare_mode      = sampler->compare_mode;
 103    if (sampler->compare_mode != PIPE_TEX_COMPARE_NONE) {
 104       state->compare_func   = sampler->compare_func;
 105    }
 106
 107    state->normalized_coords = sampler->normalized_coords;
 108    state->lod_bias          = sampler->lod_bias;
 109    if (!view->last_level &&
 110        sampler->min_img_filter == sampler->mag_img_filter) {
 111       state->min_lod        = 0.0f;
 112       state->max_lod        = 0.0f;
 113    } else {
 114       state->min_lod        = MAX2(sampler->min_lod, 0.0f);
 115       state->max_lod        = sampler->max_lod;
 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                       const LLVMValueRef ddx[4],
 144                       const LLVMValueRef ddy[4],
 145                       LLVMValueRef lod_bias, /* optional */
 146                       LLVMValueRef explicit_lod, /* optional */
 147                       LLVMValueRef width,
 148                       LLVMValueRef height,
 149                       LLVMValueRef depth)
 150
 151 {
 152    if (bld->static_state->min_lod == bld->static_state->max_lod) {
 153       /* User is forcing sampling from a particular mipmap level.
 154        * This is hit during mipmap generation.
 155        */
 156       return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
 157    }
 158    else {
 159       struct lp_build_context *float_bld = &bld->float_bld;
 160       LLVMValueRef sampler_lod_bias = LLVMConstReal(LLVMFloatType(),
 161                                                     bld->static_state->lod_bias);
 162       LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
 163                                            bld->static_state->min_lod);
 164       LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
 165                                            bld->static_state->max_lod);
 166       LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 167       LLVMValueRef lod;
 168
 169       if (explicit_lod) {
 170          lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
 171                                        index0, "");
 172       }
 173       else {
 174          const int dims = texture_dims(bld->static_state->target);
 175          LLVMValueRef dsdx, dsdy;
 176          LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
 177          LLVMValueRef rho;
 178
 179          dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
 180          dsdx = lp_build_abs(float_bld, dsdx);
 181          dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
 182          dsdy = lp_build_abs(float_bld, dsdy);
 183          if (dims > 1) {
 184             dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
 185             dtdx = lp_build_abs(float_bld, dtdx);
 186             dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
 187             dtdy = lp_build_abs(float_bld, dtdy);
 188             if (dims > 2) {
 189                drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
 190                drdx = lp_build_abs(float_bld, drdx);
 191                drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
 192                drdy = lp_build_abs(float_bld, drdy);
 193             }
 194          }
 195
 196          /* Compute rho = max of all partial derivatives scaled by texture size.
 197           * XXX this could be vectorized somewhat
 198           */
 199          rho = LLVMBuildFMul(bld->builder,
 200                             lp_build_max(float_bld, dsdx, dsdy),
 201                             lp_build_int_to_float(float_bld, width), "");
 202          if (dims > 1) {
 203             LLVMValueRef max;
 204             max = LLVMBuildFMul(bld->builder,
 205                                lp_build_max(float_bld, dtdx, dtdy),
 206                                lp_build_int_to_float(float_bld, height), "");
 207             rho = lp_build_max(float_bld, rho, max);
 208             if (dims > 2) {
 209                max = LLVMBuildFMul(bld->builder,
 210                                   lp_build_max(float_bld, drdx, drdy),
 211                                   lp_build_int_to_float(float_bld, depth), "");
 212                rho = lp_build_max(float_bld, rho, max);
 213             }
 214          }
 215
 216          /* compute lod = log2(rho) */
 217          lod = lp_build_log2(float_bld, rho);
 218
 219          /* add shader lod bias */
 220          if (lod_bias) {
 221             lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
 222                                                index0, "");
 223             lod = LLVMBuildFAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
 224          }
 225       }
 226
 227       /* add sampler lod bias */
 228       lod = LLVMBuildFAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
 229
 230       /* clamp lod */
 231       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
 232
 233       return lod;
 234    }
 235 }
 236
 237
 238 /**
 239  * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
 240  * mipmap level index.
 241  * Note: this is all scalar code.
 242  * \param lod  scalar float texture level of detail
 243  * \param level_out  returns integer
 244  */
 245 void
 246 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
 247                            unsigned unit,
 248                            LLVMValueRef lod,
 249                            LLVMValueRef *level_out)
 250 {
 251    struct lp_build_context *float_bld = &bld->float_bld;
 252    struct lp_build_context *int_bld = &bld->int_bld;
 253    LLVMValueRef last_level, level;
 254
 255    LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
 256
 257    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 258                                                bld->builder, unit);
 259
 260    /* convert float lod to integer */
 261    level = lp_build_iround(float_bld, lod);
 262
 263    /* clamp level to legal range of levels */
 264    *level_out = lp_build_clamp(int_bld, level, zero, last_level);
 265 }
 266
 267
 268 /**
 269  * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
 270  * two (adjacent) mipmap level indexes.  Later, we'll sample from those
 271  * two mipmap levels and interpolate between them.
 272  */
 273 void
 274 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
 275                            unsigned unit,
 276                            LLVMValueRef lod,
 277                            LLVMValueRef *level0_out,
 278                            LLVMValueRef *level1_out,
 279                            LLVMValueRef *weight_out)
 280 {
 281    struct lp_build_context *float_bld = &bld->float_bld;
 282    struct lp_build_context *int_bld = &bld->int_bld;
 283    LLVMValueRef last_level, level;
 284
 285    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 286                                                bld->builder, unit);
 287
 288    /* convert float lod to integer */
 289    level = lp_build_ifloor(float_bld, lod);
 290
 291    /* compute level 0 and clamp to legal range of levels */
 292    *level0_out = lp_build_clamp(int_bld, level,
 293                                 int_bld->zero,
 294                                 last_level);
 295    /* compute level 1 and clamp to legal range of levels */
 296    level = lp_build_add(int_bld, level, int_bld->one);
 297    *level1_out = lp_build_clamp(int_bld, level,
 298                                 int_bld->zero,
 299                                 last_level);
 300
 301    *weight_out = lp_build_fract(float_bld, lod);
 302 }
 303
 304
 305 LLVMValueRef
 306 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
 307                           LLVMValueRef data_array, LLVMValueRef level)
 308 {
 309    LLVMValueRef indexes[2], data_ptr;
 310    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 311    indexes[1] = level;
 312    data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
 313    data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
 314    return data_ptr;
 315 }
 316
 317
 318 LLVMValueRef
 319 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
 320                                 LLVMValueRef data_array, int level)
 321 {
 322    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
 323    return lp_build_get_mipmap_level(bld, data_array, lvl);
 324 }
 325
 326
 327 /**
 328  * Codegen equivalent for u_minify().
 329  * Return max(1, base_size >> level);
 330  */
 331 static LLVMValueRef
 332 lp_build_minify(struct lp_build_sample_context *bld,
 333                 LLVMValueRef base_size,
 334                 LLVMValueRef level)
 335 {
 336    LLVMValueRef size = LLVMBuildLShr(bld->builder, base_size, level, "minify");
 337    size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
 338    return size;
 339 }
 340
 341
 342 /**
 343  * Dereference stride_array[mipmap_level] array to get a stride.
 344  * Return stride as a vector.
 345  */
 346 static LLVMValueRef
 347 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
 348                               LLVMValueRef stride_array, LLVMValueRef level)
 349 {
 350    LLVMValueRef indexes[2], stride;
 351    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 352    indexes[1] = level;
 353    stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
 354    stride = LLVMBuildLoad(bld->builder, stride, "");
 355    stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
 356    return stride;
 357 }
 358
 359
 360 /**
 361  * When sampling a mipmap, we need to compute the width, height, depth
 362  * of the source levels from the level indexes.  This helper function
 363  * does that.
 364  */
 365 void
 366 lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
 367                             unsigned dims,
 368                             LLVMValueRef width_vec,
 369                             LLVMValueRef height_vec,
 370                             LLVMValueRef depth_vec,
 371                             LLVMValueRef ilevel0,
 372                             LLVMValueRef ilevel1,
 373                             LLVMValueRef row_stride_array,
 374                             LLVMValueRef img_stride_array,
 375                             LLVMValueRef *width0_vec,
 376                             LLVMValueRef *width1_vec,
 377                             LLVMValueRef *height0_vec,
 378                             LLVMValueRef *height1_vec,
 379                             LLVMValueRef *depth0_vec,
 380                             LLVMValueRef *depth1_vec,
 381                             LLVMValueRef *row_stride0_vec,
 382                             LLVMValueRef *row_stride1_vec,
 383                             LLVMValueRef *img_stride0_vec,
 384                             LLVMValueRef *img_stride1_vec)
 385 {
 386    const unsigned mip_filter = bld->static_state->min_mip_filter;
 387    LLVMValueRef ilevel0_vec, ilevel1_vec;
 388
 389    ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
 390    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
 391       ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
 392
 393    /*
 394     * Compute width, height, depth at mipmap level 'ilevel0'
 395     */
 396    *width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
 397    if (dims >= 2) {
 398       *height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
 399       *row_stride0_vec = lp_build_get_level_stride_vec(bld,
 400                                                        row_stride_array,
 401                                                        ilevel0);
 402       if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 403          *img_stride0_vec = lp_build_get_level_stride_vec(bld,
 404                                                           img_stride_array,
 405                                                           ilevel0);
 406          if (dims == 3) {
 407             *depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
 408          }
 409       }
 410    }
 411    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
 412       /* compute width, height, depth for second mipmap level at 'ilevel1' */
 413       *width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
 414       if (dims >= 2) {
 415          *height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
 416          *row_stride1_vec = lp_build_get_level_stride_vec(bld,
 417                                                           row_stride_array,
 418                                                           ilevel1);
 419          if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
 420             *img_stride1_vec = lp_build_get_level_stride_vec(bld,
 421                                                              img_stride_array,
 422                                                              ilevel1);
 423             if (dims == 3) {
 424                *depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
 425             }
 426          }
 427       }
 428    }
 429 }
 430
 431
 432
 433 /** Helper used by lp_build_cube_lookup() */
 434 static LLVMValueRef
 435 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
 436 {
 437    /* ima = -0.5 / abs(coord); */
 438    LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
 439    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
 440    LLVMValueRef ima = lp_build_div(coord_bld, negHalf, absCoord);
 441    return ima;
 442 }
 443
 444
 445 /**
 446  * Helper used by lp_build_cube_lookup()
 447  * \param sign  scalar +1 or -1
 448  * \param coord  float vector
 449  * \param ima  float vector
 450  */
 451 static LLVMValueRef
 452 lp_build_cube_coord(struct lp_build_context *coord_bld,
 453                     LLVMValueRef sign, int negate_coord,
 454                     LLVMValueRef coord, LLVMValueRef ima)
 455 {
 456    /* return negate(coord) * ima * sign + 0.5; */
 457    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
 458    LLVMValueRef res;
 459
 460    assert(negate_coord == +1 || negate_coord == -1);
 461
 462    if (negate_coord == -1) {
 463       coord = lp_build_negate(coord_bld, coord);
 464    }
 465
 466    res = lp_build_mul(coord_bld, coord, ima);
 467    if (sign) {
 468       sign = lp_build_broadcast_scalar(coord_bld, sign);
 469       res = lp_build_mul(coord_bld, res, sign);
 470    }
 471    res = lp_build_add(coord_bld, res, half);
 472
 473    return res;
 474 }
 475
 476
 477 /** Helper used by lp_build_cube_lookup()
 478  * Return (major_coord >= 0) ? pos_face : neg_face;
 479  */
 480 static LLVMValueRef
 481 lp_build_cube_face(struct lp_build_sample_context *bld,
 482                    LLVMValueRef major_coord,
 483                    unsigned pos_face, unsigned neg_face)
 484 {
 485    LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
 486                                     major_coord,
 487                                     bld->float_bld.zero, "");
 488    LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
 489    LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
 490    LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
 491    return res;
 492 }
 493
 494
 495
 496 /**
 497  * Generate code to do cube face selection and compute per-face texcoords.
 498  */
 499 void
 500 lp_build_cube_lookup(struct lp_build_sample_context *bld,
 501                      LLVMValueRef s,
 502                      LLVMValueRef t,
 503                      LLVMValueRef r,
 504                      LLVMValueRef *face,
 505                      LLVMValueRef *face_s,
 506                      LLVMValueRef *face_t)
 507 {
 508    struct lp_build_context *float_bld = &bld->float_bld;
 509    struct lp_build_context *coord_bld = &bld->coord_bld;
 510    LLVMValueRef rx, ry, rz;
 511    LLVMValueRef arx, ary, arz;
 512    LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
 513    LLVMValueRef arx_ge_ary, arx_ge_arz;
 514    LLVMValueRef ary_ge_arx, ary_ge_arz;
 515    LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
 516    LLVMValueRef rx_pos, ry_pos, rz_pos;
 517
 518    assert(bld->coord_bld.type.length == 4);
 519
 520    /*
 521     * Use the average of the four pixel's texcoords to choose the face.
 522     */
 523    rx = lp_build_mul(float_bld, c25,
 524                      lp_build_sum_vector(&bld->coord_bld, s));
 525    ry = lp_build_mul(float_bld, c25,
 526                      lp_build_sum_vector(&bld->coord_bld, t));
 527    rz = lp_build_mul(float_bld, c25,
 528                      lp_build_sum_vector(&bld->coord_bld, r));
 529
 530    arx = lp_build_abs(float_bld, rx);
 531    ary = lp_build_abs(float_bld, ry);
 532    arz = lp_build_abs(float_bld, rz);
 533
 534    /*
 535     * Compare sign/magnitude of rx,ry,rz to determine face
 536     */
 537    arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
 538    arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
 539    ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
 540    ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
 541
 542    arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
 543    ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 544
 545    rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
 546    ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
 547    rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
 548
 549    {
 550       struct lp_build_flow_context *flow_ctx;
 551       struct lp_build_if_state if_ctx;
 552
 553       flow_ctx = lp_build_flow_create(bld->builder);
 554       lp_build_flow_scope_begin(flow_ctx);
 555
 556       *face_s = bld->coord_bld.undef;
 557       *face_t = bld->coord_bld.undef;
 558       *face = bld->int_bld.undef;
 559
 560       lp_build_name(*face_s, "face_s");
 561       lp_build_name(*face_t, "face_t");
 562       lp_build_name(*face, "face");
 563
 564       lp_build_flow_scope_declare(flow_ctx, face_s);
 565       lp_build_flow_scope_declare(flow_ctx, face_t);
 566       lp_build_flow_scope_declare(flow_ctx, face);
 567
 568       lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
 569       {
 570          /* +/- X face */
 571          LLVMValueRef sign = lp_build_sgn(float_bld, rx);
 572          LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
 573          *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
 574          *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 575          *face = lp_build_cube_face(bld, rx,
 576                                     PIPE_TEX_FACE_POS_X,
 577                                     PIPE_TEX_FACE_NEG_X);
 578       }
 579       lp_build_else(&if_ctx);
 580       {
 581          struct lp_build_flow_context *flow_ctx2;
 582          struct lp_build_if_state if_ctx2;
 583
 584          LLVMValueRef face_s2 = bld->coord_bld.undef;
 585          LLVMValueRef face_t2 = bld->coord_bld.undef;
 586          LLVMValueRef face2 = bld->int_bld.undef;
 587
 588          flow_ctx2 = lp_build_flow_create(bld->builder);
 589          lp_build_flow_scope_begin(flow_ctx2);
 590          lp_build_flow_scope_declare(flow_ctx2, &face_s2);
 591          lp_build_flow_scope_declare(flow_ctx2, &face_t2);
 592          lp_build_flow_scope_declare(flow_ctx2, &face2);
 593
 594          ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
 595
 596          lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
 597          {
 598             /* +/- Y face */
 599             LLVMValueRef sign = lp_build_sgn(float_bld, ry);
 600             LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
 601             face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
 602             face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
 603             face2 = lp_build_cube_face(bld, ry,
 604                                        PIPE_TEX_FACE_POS_Y,
 605                                        PIPE_TEX_FACE_NEG_Y);
 606          }
 607          lp_build_else(&if_ctx2);
 608          {
 609             /* +/- Z face */
 610             LLVMValueRef sign = lp_build_sgn(float_bld, rz);
 611             LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
 612             face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
 613             face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
 614             face2 = lp_build_cube_face(bld, rz,
 615                                        PIPE_TEX_FACE_POS_Z,
 616                                        PIPE_TEX_FACE_NEG_Z);
 617          }
 618          lp_build_endif(&if_ctx2);
 619          lp_build_flow_scope_end(flow_ctx2);
 620          lp_build_flow_destroy(flow_ctx2);
 621          *face_s = face_s2;
 622          *face_t = face_t2;
 623          *face = face2;
 624       }
 625
 626       lp_build_endif(&if_ctx);
 627       lp_build_flow_scope_end(flow_ctx);
 628       lp_build_flow_destroy(flow_ctx);
 629    }
 630 }
 631
 632
 633 /**
 634  * Compute the partial offset of a pixel block along an arbitrary axis.
 635  *
 636  * @param coord   coordinate in pixels
 637  * @param stride  number of bytes between rows of successive pixel blocks
 638  * @param block_length  number of pixels in a pixels block along the coordinate
 639  *                      axis
 640  * @param out_offset    resulting relative offset of the pixel block in bytes
 641  * @param out_subcoord  resulting sub-block pixel coordinate
 642  */
 643 void
 644 lp_build_sample_partial_offset(struct lp_build_context *bld,
 645                                unsigned block_length,
 646                                LLVMValueRef coord,
 647                                LLVMValueRef stride,
 648                                LLVMValueRef *out_offset,
 649                                LLVMValueRef *out_subcoord)
 650 {
 651    LLVMValueRef offset;
 652    LLVMValueRef subcoord;
 653
 654    if (block_length == 1) {
 655       subcoord = bld->zero;
 656    }
 657    else {
 658       /*
 659        * Pixel blocks have power of two dimensions. LLVM should convert the
 660        * rem/div to bit arithmetic.
 661        * TODO: Verify this.
 662        */
 663
 664       LLVMValueRef block_width = lp_build_const_int_vec(bld->type, block_length);
 665       subcoord = LLVMBuildURem(bld->builder, coord, block_width, "");
 666       coord    = LLVMBuildUDiv(bld->builder, coord, block_width, "");
 667    }
 668
 669    offset = lp_build_mul(bld, coord, stride);
 670
 671    assert(out_offset);
 672    assert(out_subcoord);
 673
 674    *out_offset = offset;
 675    *out_subcoord = subcoord;
 676 }
 677
 678
 679 /**
 680  * Compute the offset of a pixel block.
 681  *
 682  * x, y, z, y_stride, z_stride are vectors, and they refer to pixels.
 683  *
 684  * Returns the relative offset and i,j sub-block coordinates
 685  */
 686 void
 687 lp_build_sample_offset(struct lp_build_context *bld,
 688                        const struct util_format_description *format_desc,
 689                        LLVMValueRef x,
 690                        LLVMValueRef y,
 691                        LLVMValueRef z,
 692                        LLVMValueRef y_stride,
 693                        LLVMValueRef z_stride,
 694                        LLVMValueRef *out_offset,
 695                        LLVMValueRef *out_i,
 696                        LLVMValueRef *out_j)
 697 {
 698    LLVMValueRef x_stride;
 699    LLVMValueRef offset;
 700
 701    x_stride = lp_build_const_vec(bld->type, format_desc->block.bits/8);
 702
 703    lp_build_sample_partial_offset(bld,
 704                                   format_desc->block.width,
 705                                   x, x_stride,
 706                                   &offset, out_i);
 707
 708    if (y && y_stride) {
 709       LLVMValueRef y_offset;
 710       lp_build_sample_partial_offset(bld,
 711                                      format_desc->block.height,
 712                                      y, y_stride,
 713                                      &y_offset, out_j);
 714       offset = lp_build_add(bld, offset, y_offset);
 715    }
 716    else {
 717       *out_j = bld->zero;
 718    }
 719
 720    if (z && z_stride) {
 721       LLVMValueRef z_offset;
 722       LLVMValueRef k;
 723       lp_build_sample_partial_offset(bld,
 724                                      1, /* pixel blocks are always 2D */
 725                                      z, z_stride,
 726                                      &z_offset, &k);
 727       offset = lp_build_add(bld, offset, z_offset);
 728    }
 729
 730    *out_offset = offset;
 731 }