src/gallium/auxiliary/gallivm/lp_bld_sample_soa.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 -- SoA.
  31  *
  32  * @author Jose Fonseca <jfonseca@vmware.com>
  33  * @author Brian Paul <brianp@vmware.com>
  34  */
  35
  36 #include "pipe/p_defines.h"
  37 #include "pipe/p_state.h"
  38 #include "pipe/p_shader_tokens.h"
  39 #include "util/u_debug.h"
  40 #include "util/u_dump.h"
  41 #include "util/u_memory.h"
  42 #include "util/u_math.h"
  43 #include "util/u_format.h"
  44 #include "util/u_cpu_detect.h"
  45 #include "lp_bld_debug.h"
  46 #include "lp_bld_type.h"
  47 #include "lp_bld_const.h"
  48 #include "lp_bld_conv.h"
  49 #include "lp_bld_arit.h"
  50 #include "lp_bld_bitarit.h"
  51 #include "lp_bld_logic.h"
  52 #include "lp_bld_printf.h"
  53 #include "lp_bld_swizzle.h"
  54 #include "lp_bld_flow.h"
  55 #include "lp_bld_gather.h"
  56 #include "lp_bld_format.h"
  57 #include "lp_bld_sample.h"
  58 #include "lp_bld_sample_aos.h"
  59 #include "lp_bld_struct.h"
  60 #include "lp_bld_quad.h"
  61 #include "lp_bld_pack.h"
  62
  63
  64 /**
  65  * Generate code to fetch a texel from a texture at int coords (x, y, z).
  66  * The computation depends on whether the texture is 1D, 2D or 3D.
  67  * The result, texel, will be float vectors:
  68  *   texel[0] = red values
  69  *   texel[1] = green values
  70  *   texel[2] = blue values
  71  *   texel[3] = alpha values
  72  */
  73 static void
  74 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
  75                           unsigned sampler_unit,
  76                           LLVMValueRef width,
  77                           LLVMValueRef height,
  78                           LLVMValueRef depth,
  79                           LLVMValueRef x,
  80                           LLVMValueRef y,
  81                           LLVMValueRef z,
  82                           LLVMValueRef y_stride,
  83                           LLVMValueRef z_stride,
  84                           LLVMValueRef data_ptr,
  85                           LLVMValueRef mipoffsets,
  86                           LLVMValueRef texel_out[4])
  87 {
  88    const struct lp_static_sampler_state *static_state = bld->static_sampler_state;
  89    const unsigned dims = bld->dims;
  90    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
  91    LLVMBuilderRef builder = bld->gallivm->builder;
  92    LLVMValueRef offset;
  93    LLVMValueRef i, j;
  94    LLVMValueRef use_border = NULL;
  95
  96    /* use_border = x < 0 || x >= width || y < 0 || y >= height */
  97    if (lp_sampler_wrap_mode_uses_border_color(static_state->wrap_s,
  98                                               static_state->min_img_filter,
  99                                               static_state->mag_img_filter)) {
 100       LLVMValueRef b1, b2;
 101       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
 102       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
 103       use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
 104    }
 105
 106    if (dims >= 2 &&
 107        lp_sampler_wrap_mode_uses_border_color(static_state->wrap_t,
 108                                               static_state->min_img_filter,
 109                                               static_state->mag_img_filter)) {
 110       LLVMValueRef b1, b2;
 111       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
 112       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
 113       if (use_border) {
 114          use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
 115          use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
 116       }
 117       else {
 118          use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
 119       }
 120    }
 121
 122    if (dims == 3 &&
 123        lp_sampler_wrap_mode_uses_border_color(static_state->wrap_r,
 124                                               static_state->min_img_filter,
 125                                               static_state->mag_img_filter)) {
 126       LLVMValueRef b1, b2;
 127       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
 128       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
 129       if (use_border) {
 130          use_border = LLVMBuildOr(builder, use_border, b1, "ub_or_b1");
 131          use_border = LLVMBuildOr(builder, use_border, b2, "ub_or_b2");
 132       }
 133       else {
 134          use_border = LLVMBuildOr(builder, b1, b2, "b1_or_b2");
 135       }
 136    }
 137
 138    /* convert x,y,z coords to linear offset from start of texture, in bytes */
 139    lp_build_sample_offset(&bld->int_coord_bld,
 140                           bld->format_desc,
 141                           x, y, z, y_stride, z_stride,
 142                           &offset, &i, &j);
 143    if (mipoffsets) {
 144       offset = lp_build_add(&bld->int_coord_bld, offset, mipoffsets);
 145    }
 146
 147    if (use_border) {
 148       /* If we can sample the border color, it means that texcoords may
 149        * lie outside the bounds of the texture image.  We need to do
 150        * something to prevent reading out of bounds and causing a segfault.
 151        *
 152        * Simply AND the texture coords with !use_border.  This will cause
 153        * coords which are out of bounds to become zero.  Zero's guaranteed
 154        * to be inside the texture image.
 155        */
 156       offset = lp_build_andnot(&bld->int_coord_bld, offset, use_border);
 157    }
 158
 159    lp_build_fetch_rgba_soa(bld->gallivm,
 160                            bld->format_desc,
 161                            bld->texel_type,
 162                            data_ptr, offset,
 163                            i, j,
 164                            texel_out);
 165
 166    /*
 167     * Note: if we find an app which frequently samples the texture border
 168     * we might want to implement a true conditional here to avoid sampling
 169     * the texture whenever possible (since that's quite a bit of code).
 170     * Ex:
 171     *   if (use_border) {
 172     *      texel = border_color;
 173     *   }
 174     *   else {
 175     *      texel = sample_texture(coord);
 176     *   }
 177     * As it is now, we always sample the texture, then selectively replace
 178     * the texel color results with the border color.
 179     */
 180
 181    if (use_border) {
 182       /* select texel color or border color depending on use_border */
 183       LLVMValueRef border_color_ptr =
 184          bld->dynamic_state->border_color(bld->dynamic_state,
 185                                           bld->gallivm, sampler_unit);
 186       int chan;
 187       for (chan = 0; chan < 4; chan++) {
 188          LLVMValueRef border_chan =
 189             lp_build_array_get(bld->gallivm, border_color_ptr,
 190                                lp_build_const_int32(bld->gallivm, chan));
 191          LLVMValueRef border_chan_vec =
 192             lp_build_broadcast_scalar(&bld->float_vec_bld, border_chan);
 193
 194          if (!bld->texel_type.floating) {
 195             border_chan_vec = LLVMBuildBitCast(builder, border_chan_vec,
 196                                                bld->texel_bld.vec_type, "");
 197          }
 198          texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
 199                                            border_chan_vec, texel_out[chan]);
 200       }
 201    }
 202 }
 203
 204
 205 /**
 206  * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
 207  */
 208 static LLVMValueRef
 209 lp_build_coord_mirror(struct lp_build_sample_context *bld,
 210                       LLVMValueRef coord)
 211 {
 212    struct lp_build_context *coord_bld = &bld->coord_bld;
 213    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 214    LLVMValueRef fract, flr, isOdd;
 215
 216    lp_build_ifloor_fract(coord_bld, coord, &flr, &fract);
 217
 218    /* isOdd = flr & 1 */
 219    isOdd = LLVMBuildAnd(bld->gallivm->builder, flr, int_coord_bld->one, "");
 220
 221    /* make coord positive or negative depending on isOdd */
 222    coord = lp_build_set_sign(coord_bld, fract, isOdd);
 223
 224    /* convert isOdd to float */
 225    isOdd = lp_build_int_to_float(coord_bld, isOdd);
 226
 227    /* add isOdd to coord */
 228    coord = lp_build_add(coord_bld, coord, isOdd);
 229
 230    return coord;
 231 }
 232
 233
 234 /**
 235  * Helper to compute the first coord and the weight for
 236  * linear wrap repeat npot textures
 237  */
 238 void
 239 lp_build_coord_repeat_npot_linear(struct lp_build_sample_context *bld,
 240                                   LLVMValueRef coord_f,
 241                                   LLVMValueRef length_i,
 242                                   LLVMValueRef length_f,
 243                                   LLVMValueRef *coord0_i,
 244                                   LLVMValueRef *weight_f)
 245 {
 246    struct lp_build_context *coord_bld = &bld->coord_bld;
 247    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 248    LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
 249    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length_i,
 250                                                 int_coord_bld->one);
 251    LLVMValueRef mask;
 252    /* wrap with normalized floats is just fract */
 253    coord_f = lp_build_fract(coord_bld, coord_f);
 254    /* mul by size and subtract 0.5 */
 255    coord_f = lp_build_mul(coord_bld, coord_f, length_f);
 256    coord_f = lp_build_sub(coord_bld, coord_f, half);
 257    /*
 258     * we avoided the 0.5/length division before the repeat wrap,
 259     * now need to fix up edge cases with selects
 260     */
 261    /* convert to int, compute lerp weight */
 262    lp_build_ifloor_fract(coord_bld, coord_f, coord0_i, weight_f);
 263    mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
 264                            PIPE_FUNC_LESS, *coord0_i, int_coord_bld->zero);
 265    *coord0_i = lp_build_select(int_coord_bld, mask, length_minus_one, *coord0_i);
 266 }
 267
 268
 269 /**
 270  * Build LLVM code for texture wrap mode for linear filtering.
 271  * \param x0_out  returns first integer texcoord
 272  * \param x1_out  returns second integer texcoord
 273  * \param weight_out  returns linear interpolation weight
 274  */
 275 static void
 276 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
 277                             LLVMValueRef coord,
 278                             LLVMValueRef length,
 279                             LLVMValueRef length_f,
 280                             LLVMValueRef offset,
 281                             boolean is_pot,
 282                             unsigned wrap_mode,
 283                             LLVMValueRef *x0_out,
 284                             LLVMValueRef *x1_out,
 285                             LLVMValueRef *weight_out)
 286 {
 287    struct lp_build_context *coord_bld = &bld->coord_bld;
 288    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 289    LLVMBuilderRef builder = bld->gallivm->builder;
 290    LLVMValueRef half = lp_build_const_vec(bld->gallivm, coord_bld->type, 0.5);
 291    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
 292    LLVMValueRef coord0, coord1, weight;
 293
 294    switch(wrap_mode) {
 295    case PIPE_TEX_WRAP_REPEAT:
 296       if (is_pot) {
 297          /* mul by size and subtract 0.5 */
 298          coord = lp_build_mul(coord_bld, coord, length_f);
 299          coord = lp_build_sub(coord_bld, coord, half);
 300          if (offset) {
 301             offset = lp_build_int_to_float(coord_bld, offset);
 302             coord = lp_build_add(coord_bld, coord, offset);
 303          }
 304          /* convert to int, compute lerp weight */
 305          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 306          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 307          /* repeat wrap */
 308          coord0 = LLVMBuildAnd(builder, coord0, length_minus_one, "");
 309          coord1 = LLVMBuildAnd(builder, coord1, length_minus_one, "");
 310       }
 311       else {
 312          LLVMValueRef mask;
 313          if (offset) {
 314             offset = lp_build_int_to_float(coord_bld, offset);
 315             offset = lp_build_div(coord_bld, offset, length_f);
 316             coord = lp_build_add(coord_bld, coord, offset);
 317          }
 318          lp_build_coord_repeat_npot_linear(bld, coord,
 319                                            length, length_f,
 320                                            &coord0, &weight);
 321          mask = lp_build_compare(int_coord_bld->gallivm, int_coord_bld->type,
 322                                  PIPE_FUNC_NOTEQUAL, coord0, length_minus_one);
 323          coord1 = LLVMBuildAnd(builder,
 324                                lp_build_add(int_coord_bld, coord0, int_coord_bld->one),
 325                                mask, "");
 326       }
 327       break;
 328
 329    case PIPE_TEX_WRAP_CLAMP:
 330       if (bld->static_sampler_state->normalized_coords) {
 331          /* scale coord to length */
 332          coord = lp_build_mul(coord_bld, coord, length_f);
 333       }
 334       if (offset) {
 335          offset = lp_build_int_to_float(coord_bld, offset);
 336          coord = lp_build_add(coord_bld, coord, offset);
 337       }
 338
 339       /* clamp to [0, length] */
 340       coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
 341
 342       coord = lp_build_sub(coord_bld, coord, half);
 343
 344       /* convert to int, compute lerp weight */
 345       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 346       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 347       break;
 348
 349    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 350       {
 351          struct lp_build_context abs_coord_bld = bld->coord_bld;
 352          abs_coord_bld.type.sign = FALSE;
 353
 354          if (bld->static_sampler_state->normalized_coords) {
 355             /* mul by tex size */
 356             coord = lp_build_mul(coord_bld, coord, length_f);
 357          }
 358          if (offset) {
 359             offset = lp_build_int_to_float(coord_bld, offset);
 360             coord = lp_build_add(coord_bld, coord, offset);
 361          }
 362
 363          /* clamp to length max */
 364          coord = lp_build_min(coord_bld, coord, length_f);
 365          /* subtract 0.5 */
 366          coord = lp_build_sub(coord_bld, coord, half);
 367          /* clamp to [0, length - 0.5] */
 368          coord = lp_build_max(coord_bld, coord, coord_bld->zero);
 369          /* convert to int, compute lerp weight */
 370          lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
 371          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 372          /* coord1 = min(coord1, length-1) */
 373          coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
 374          break;
 375       }
 376
 377    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 378       if (bld->static_sampler_state->normalized_coords) {
 379          /* scale coord to length */
 380          coord = lp_build_mul(coord_bld, coord, length_f);
 381       }
 382       if (offset) {
 383          offset = lp_build_int_to_float(coord_bld, offset);
 384          coord = lp_build_add(coord_bld, coord, offset);
 385       }
 386       /* was: clamp to [-0.5, length + 0.5], then sub 0.5 */
 387       /* can skip clamp (though might not work for very large coord values */
 388       coord = lp_build_sub(coord_bld, coord, half);
 389       /* convert to int, compute lerp weight */
 390       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 391       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 392       break;
 393
 394    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 395       /* compute mirror function */
 396       coord = lp_build_coord_mirror(bld, coord);
 397
 398       /* scale coord to length */
 399       coord = lp_build_mul(coord_bld, coord, length_f);
 400       coord = lp_build_sub(coord_bld, coord, half);
 401       if (offset) {
 402          offset = lp_build_int_to_float(coord_bld, offset);
 403          coord = lp_build_add(coord_bld, coord, offset);
 404       }
 405
 406       /* convert to int, compute lerp weight */
 407       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 408       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 409
 410       /* coord0 = max(coord0, 0) */
 411       coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
 412       /* coord1 = min(coord1, length-1) */
 413       coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
 414       break;
 415
 416    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 417       if (bld->static_sampler_state->normalized_coords) {
 418          /* scale coord to length */
 419          coord = lp_build_mul(coord_bld, coord, length_f);
 420       }
 421       if (offset) {
 422          offset = lp_build_int_to_float(coord_bld, offset);
 423          coord = lp_build_add(coord_bld, coord, offset);
 424       }
 425       coord = lp_build_abs(coord_bld, coord);
 426
 427       /* clamp to [0, length] */
 428       coord = lp_build_min(coord_bld, coord, length_f);
 429
 430       coord = lp_build_sub(coord_bld, coord, half);
 431
 432       /* convert to int, compute lerp weight */
 433       lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 434       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 435       break;
 436
 437    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 438       {
 439          LLVMValueRef min, max;
 440          struct lp_build_context abs_coord_bld = bld->coord_bld;
 441          abs_coord_bld.type.sign = FALSE;
 442
 443          if (bld->static_sampler_state->normalized_coords) {
 444             /* scale coord to length */
 445             coord = lp_build_mul(coord_bld, coord, length_f);
 446          }
 447          if (offset) {
 448             offset = lp_build_int_to_float(coord_bld, offset);
 449             coord = lp_build_add(coord_bld, coord, offset);
 450          }
 451          coord = lp_build_abs(coord_bld, coord);
 452
 453          /* clamp to [0.5, length - 0.5] */
 454          min = half;
 455          max = lp_build_sub(coord_bld, length_f, min);
 456          coord = lp_build_clamp(coord_bld, coord, min, max);
 457
 458          coord = lp_build_sub(coord_bld, coord, half);
 459
 460          /* convert to int, compute lerp weight */
 461          lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
 462          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 463       }
 464       break;
 465
 466    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 467       {
 468          if (bld->static_sampler_state->normalized_coords) {
 469             /* scale coord to length */
 470             coord = lp_build_mul(coord_bld, coord, length_f);
 471          }
 472          if (offset) {
 473             offset = lp_build_int_to_float(coord_bld, offset);
 474             coord = lp_build_add(coord_bld, coord, offset);
 475          }
 476          coord = lp_build_abs(coord_bld, coord);
 477
 478          /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
 479          /* skip clamp - always positive, and other side
 480             only potentially matters for very large coords */
 481          coord = lp_build_sub(coord_bld, coord, half);
 482
 483          /* convert to int, compute lerp weight */
 484          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 485          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 486       }
 487       break;
 488
 489    default:
 490       assert(0);
 491       coord0 = NULL;
 492       coord1 = NULL;
 493       weight = NULL;
 494    }
 495
 496    *x0_out = coord0;
 497    *x1_out = coord1;
 498    *weight_out = weight;
 499 }
 500
 501
 502 /**
 503  * Build LLVM code for texture wrap mode for nearest filtering.
 504  * \param coord  the incoming texcoord (nominally in [0,1])
 505  * \param length  the texture size along one dimension, as int vector
 506  * \param length_f  the texture size along one dimension, as float vector
 507  * \param offset  texel offset along one dimension (as int vector)
 508  * \param is_pot  if TRUE, length is a power of two
 509  * \param wrap_mode  one of PIPE_TEX_WRAP_x
 510  */
 511 static LLVMValueRef
 512 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
 513                              LLVMValueRef coord,
 514                              LLVMValueRef length,
 515                              LLVMValueRef length_f,
 516                              LLVMValueRef offset,
 517                              boolean is_pot,
 518                              unsigned wrap_mode)
 519 {
 520    struct lp_build_context *coord_bld = &bld->coord_bld;
 521    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 522    LLVMBuilderRef builder = bld->gallivm->builder;
 523    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
 524    LLVMValueRef icoord;
 525
 526    switch(wrap_mode) {
 527    case PIPE_TEX_WRAP_REPEAT:
 528       if (is_pot) {
 529          coord = lp_build_mul(coord_bld, coord, length_f);
 530          icoord = lp_build_ifloor(coord_bld, coord);
 531          if (offset) {
 532             icoord = lp_build_add(int_coord_bld, icoord, offset);
 533          }
 534          icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
 535       }
 536       else {
 537           if (offset) {
 538              offset = lp_build_int_to_float(coord_bld, offset);
 539              offset = lp_build_div(coord_bld, offset, length_f);
 540              coord = lp_build_add(coord_bld, coord, offset);
 541           }
 542           /* take fraction, unnormalize */
 543           coord = lp_build_fract_safe(coord_bld, coord);
 544           coord = lp_build_mul(coord_bld, coord, length_f);
 545           icoord = lp_build_itrunc(coord_bld, coord);
 546       }
 547       break;
 548
 549    case PIPE_TEX_WRAP_CLAMP:
 550    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 551       if (bld->static_sampler_state->normalized_coords) {
 552          /* scale coord to length */
 553          coord = lp_build_mul(coord_bld, coord, length_f);
 554       }
 555
 556       /* floor */
 557       /* use itrunc instead since we clamp to 0 anyway */
 558       icoord = lp_build_itrunc(coord_bld, coord);
 559       if (offset) {
 560          icoord = lp_build_add(int_coord_bld, icoord, offset);
 561       }
 562
 563       /* clamp to [0, length - 1]. */
 564       icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
 565                               length_minus_one);
 566       break;
 567
 568    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 569       if (bld->static_sampler_state->normalized_coords) {
 570          /* scale coord to length */
 571          coord = lp_build_mul(coord_bld, coord, length_f);
 572       }
 573       /* no clamp necessary, border masking will handle this */
 574       icoord = lp_build_ifloor(coord_bld, coord);
 575       if (offset) {
 576          icoord = lp_build_add(int_coord_bld, icoord, offset);
 577       }
 578       break;
 579
 580    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 581       if (offset) {
 582          offset = lp_build_int_to_float(coord_bld, offset);
 583          offset = lp_build_div(coord_bld, offset, length_f);
 584          coord = lp_build_add(coord_bld, coord, offset);
 585       }
 586       /* compute mirror function */
 587       coord = lp_build_coord_mirror(bld, coord);
 588
 589       /* scale coord to length */
 590       assert(bld->static_sampler_state->normalized_coords);
 591       coord = lp_build_mul(coord_bld, coord, length_f);
 592
 593       /* itrunc == ifloor here */
 594       icoord = lp_build_itrunc(coord_bld, coord);
 595
 596       /* clamp to [0, length - 1] */
 597       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 598       break;
 599
 600    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 601    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 602       if (bld->static_sampler_state->normalized_coords) {
 603          /* scale coord to length */
 604          coord = lp_build_mul(coord_bld, coord, length_f);
 605       }
 606       if (offset) {
 607          offset = lp_build_int_to_float(coord_bld, offset);
 608          coord = lp_build_add(coord_bld, coord, offset);
 609       }
 610       coord = lp_build_abs(coord_bld, coord);
 611
 612       /* itrunc == ifloor here */
 613       icoord = lp_build_itrunc(coord_bld, coord);
 614
 615       /* clamp to [0, length - 1] */
 616       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 617       break;
 618
 619    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 620       if (bld->static_sampler_state->normalized_coords) {
 621          /* scale coord to length */
 622          coord = lp_build_mul(coord_bld, coord, length_f);
 623       }
 624       if (offset) {
 625          offset = lp_build_int_to_float(coord_bld, offset);
 626          coord = lp_build_add(coord_bld, coord, offset);
 627       }
 628       coord = lp_build_abs(coord_bld, coord);
 629
 630       /* itrunc == ifloor here */
 631       icoord = lp_build_itrunc(coord_bld, coord);
 632       break;
 633
 634    default:
 635       assert(0);
 636       icoord = NULL;
 637    }
 638
 639    return icoord;
 640 }
 641
 642
 643 /**
 644  * Generate code to sample a mipmap level with nearest filtering.
 645  * If sampling a cube texture, r = cube face in [0,5].
 646  */
 647 static void
 648 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
 649                               unsigned sampler_unit,
 650                               LLVMValueRef size,
 651                               LLVMValueRef row_stride_vec,
 652                               LLVMValueRef img_stride_vec,
 653                               LLVMValueRef data_ptr,
 654                               LLVMValueRef mipoffsets,
 655                               LLVMValueRef s,
 656                               LLVMValueRef t,
 657                               LLVMValueRef r,
 658                               const LLVMValueRef *offsets,
 659                               LLVMValueRef colors_out[4])
 660 {
 661    const unsigned dims = bld->dims;
 662    LLVMValueRef width_vec;
 663    LLVMValueRef height_vec;
 664    LLVMValueRef depth_vec;
 665    LLVMValueRef flt_size;
 666    LLVMValueRef flt_width_vec;
 667    LLVMValueRef flt_height_vec;
 668    LLVMValueRef flt_depth_vec;
 669    LLVMValueRef x, y = NULL, z = NULL;
 670
 671    lp_build_extract_image_sizes(bld,
 672                                 &bld->int_size_bld,
 673                                 bld->int_coord_type,
 674                                 size,
 675                                 &width_vec, &height_vec, &depth_vec);
 676
 677    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
 678
 679    lp_build_extract_image_sizes(bld,
 680                                 &bld->float_size_bld,
 681                                 bld->coord_type,
 682                                 flt_size,
 683                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
 684
 685    /*
 686     * Compute integer texcoords.
 687     */
 688    x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, offsets[0],
 689                                     bld->static_texture_state->pot_width,
 690                                     bld->static_sampler_state->wrap_s);
 691    lp_build_name(x, "tex.x.wrapped");
 692
 693    if (dims >= 2) {
 694       y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, offsets[1],
 695                                        bld->static_texture_state->pot_height,
 696                                        bld->static_sampler_state->wrap_t);
 697       lp_build_name(y, "tex.y.wrapped");
 698
 699       if (dims == 3) {
 700          z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, offsets[2],
 701                                           bld->static_texture_state->pot_depth,
 702                                           bld->static_sampler_state->wrap_r);
 703          lp_build_name(z, "tex.z.wrapped");
 704       }
 705    }
 706    if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
 707        bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
 708        bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
 709       z = r;
 710       lp_build_name(z, "tex.z.layer");
 711    }
 712
 713    /*
 714     * Get texture colors.
 715     */
 716    lp_build_sample_texel_soa(bld, sampler_unit,
 717                              width_vec, height_vec, depth_vec,
 718                              x, y, z,
 719                              row_stride_vec, img_stride_vec,
 720                              data_ptr, mipoffsets, colors_out);
 721 }
 722
 723
 724 /**
 725  * Generate code to sample a mipmap level with linear filtering.
 726  * If sampling a cube texture, r = cube face in [0,5].
 727  */
 728 static void
 729 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
 730                              unsigned sampler_unit,
 731                              LLVMValueRef size,
 732                              LLVMValueRef row_stride_vec,
 733                              LLVMValueRef img_stride_vec,
 734                              LLVMValueRef data_ptr,
 735                              LLVMValueRef mipoffsets,
 736                              LLVMValueRef s,
 737                              LLVMValueRef t,
 738                              LLVMValueRef r,
 739                              const LLVMValueRef *offsets,
 740                              LLVMValueRef colors_out[4])
 741 {
 742    const unsigned dims = bld->dims;
 743    LLVMValueRef width_vec;
 744    LLVMValueRef height_vec;
 745    LLVMValueRef depth_vec;
 746    LLVMValueRef flt_size;
 747    LLVMValueRef flt_width_vec;
 748    LLVMValueRef flt_height_vec;
 749    LLVMValueRef flt_depth_vec;
 750    LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
 751    LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
 752    LLVMValueRef neighbors[2][2][4];
 753    int chan;
 754
 755    lp_build_extract_image_sizes(bld,
 756                                 &bld->int_size_bld,
 757                                 bld->int_coord_type,
 758                                 size,
 759                                 &width_vec, &height_vec, &depth_vec);
 760
 761    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
 762
 763    lp_build_extract_image_sizes(bld,
 764                                 &bld->float_size_bld,
 765                                 bld->coord_type,
 766                                 flt_size,
 767                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
 768
 769    /*
 770     * Compute integer texcoords.
 771     */
 772    lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec, offsets[0],
 773                                bld->static_texture_state->pot_width,
 774                                bld->static_sampler_state->wrap_s,
 775                                &x0, &x1, &s_fpart);
 776    lp_build_name(x0, "tex.x0.wrapped");
 777    lp_build_name(x1, "tex.x1.wrapped");
 778
 779    if (dims >= 2) {
 780       lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec, offsets[1],
 781                                   bld->static_texture_state->pot_height,
 782                                   bld->static_sampler_state->wrap_t,
 783                                   &y0, &y1, &t_fpart);
 784       lp_build_name(y0, "tex.y0.wrapped");
 785       lp_build_name(y1, "tex.y1.wrapped");
 786
 787       if (dims == 3) {
 788          lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec, offsets[2],
 789                                      bld->static_texture_state->pot_depth,
 790                                      bld->static_sampler_state->wrap_r,
 791                                      &z0, &z1, &r_fpart);
 792          lp_build_name(z0, "tex.z0.wrapped");
 793          lp_build_name(z1, "tex.z1.wrapped");
 794       }
 795    }
 796    if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
 797        bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
 798        bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
 799       z0 = z1 = r;  /* cube face or array layer */
 800       lp_build_name(z0, "tex.z0.layer");
 801       lp_build_name(z1, "tex.z1.layer");
 802    }
 803
 804
 805    /*
 806     * Get texture colors.
 807     */
 808    /* get x0/x1 texels */
 809    lp_build_sample_texel_soa(bld, sampler_unit,
 810                              width_vec, height_vec, depth_vec,
 811                              x0, y0, z0,
 812                              row_stride_vec, img_stride_vec,
 813                              data_ptr, mipoffsets, neighbors[0][0]);
 814    lp_build_sample_texel_soa(bld, sampler_unit,
 815                              width_vec, height_vec, depth_vec,
 816                              x1, y0, z0,
 817                              row_stride_vec, img_stride_vec,
 818                              data_ptr, mipoffsets, neighbors[0][1]);
 819
 820    if (dims == 1) {
 821       /* Interpolate two samples from 1D image to produce one color */
 822       for (chan = 0; chan < 4; chan++) {
 823          colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
 824                                           neighbors[0][0][chan],
 825                                           neighbors[0][1][chan],
 826                                           0);
 827       }
 828    }
 829    else {
 830       /* 2D/3D texture */
 831       LLVMValueRef colors0[4];
 832
 833       /* get x0/x1 texels at y1 */
 834       lp_build_sample_texel_soa(bld, sampler_unit,
 835                                 width_vec, height_vec, depth_vec,
 836                                 x0, y1, z0,
 837                                 row_stride_vec, img_stride_vec,
 838                                 data_ptr, mipoffsets, neighbors[1][0]);
 839       lp_build_sample_texel_soa(bld, sampler_unit,
 840                                 width_vec, height_vec, depth_vec,
 841                                 x1, y1, z0,
 842                                 row_stride_vec, img_stride_vec,
 843                                 data_ptr, mipoffsets, neighbors[1][1]);
 844
 845       /* Bilinear interpolate the four samples from the 2D image / 3D slice */
 846       for (chan = 0; chan < 4; chan++) {
 847          colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
 848                                           s_fpart, t_fpart,
 849                                           neighbors[0][0][chan],
 850                                           neighbors[0][1][chan],
 851                                           neighbors[1][0][chan],
 852                                           neighbors[1][1][chan],
 853                                           0);
 854       }
 855
 856       if (dims == 3) {
 857          LLVMValueRef neighbors1[2][2][4];
 858          LLVMValueRef colors1[4];
 859
 860          /* get x0/x1/y0/y1 texels at z1 */
 861          lp_build_sample_texel_soa(bld, sampler_unit,
 862                                    width_vec, height_vec, depth_vec,
 863                                    x0, y0, z1,
 864                                    row_stride_vec, img_stride_vec,
 865                                    data_ptr, mipoffsets, neighbors1[0][0]);
 866          lp_build_sample_texel_soa(bld, sampler_unit,
 867                                    width_vec, height_vec, depth_vec,
 868                                    x1, y0, z1,
 869                                    row_stride_vec, img_stride_vec,
 870                                    data_ptr, mipoffsets, neighbors1[0][1]);
 871          lp_build_sample_texel_soa(bld, sampler_unit,
 872                                    width_vec, height_vec, depth_vec,
 873                                    x0, y1, z1,
 874                                    row_stride_vec, img_stride_vec,
 875                                    data_ptr, mipoffsets, neighbors1[1][0]);
 876          lp_build_sample_texel_soa(bld, sampler_unit,
 877                                    width_vec, height_vec, depth_vec,
 878                                    x1, y1, z1,
 879                                    row_stride_vec, img_stride_vec,
 880                                    data_ptr, mipoffsets, neighbors1[1][1]);
 881
 882          /* Bilinear interpolate the four samples from the second Z slice */
 883          for (chan = 0; chan < 4; chan++) {
 884             colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
 885                                              s_fpart, t_fpart,
 886                                              neighbors1[0][0][chan],
 887                                              neighbors1[0][1][chan],
 888                                              neighbors1[1][0][chan],
 889                                              neighbors1[1][1][chan],
 890                                              0);
 891          }
 892
 893          /* Linearly interpolate the two samples from the two 3D slices */
 894          for (chan = 0; chan < 4; chan++) {
 895             colors_out[chan] = lp_build_lerp(&bld->texel_bld,
 896                                              r_fpart,
 897                                              colors0[chan], colors1[chan],
 898                                              0);
 899          }
 900       }
 901       else {
 902          /* 2D tex */
 903          for (chan = 0; chan < 4; chan++) {
 904             colors_out[chan] = colors0[chan];
 905          }
 906       }
 907    }
 908 }
 909
 910
 911 /**
 912  * Sample the texture/mipmap using given image filter and mip filter.
 913  * data0_ptr and data1_ptr point to the two mipmap levels to sample
 914  * from.  width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
 915  * If we're using nearest miplevel sampling the '1' values will be null/unused.
 916  */
 917 static void
 918 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
 919                        unsigned sampler_unit,
 920                        unsigned img_filter,
 921                        unsigned mip_filter,
 922                        LLVMValueRef s,
 923                        LLVMValueRef t,
 924                        LLVMValueRef r,
 925                        const LLVMValueRef *offsets,
 926                        LLVMValueRef ilevel0,
 927                        LLVMValueRef ilevel1,
 928                        LLVMValueRef lod_fpart,
 929                        LLVMValueRef *colors_out)
 930 {
 931    LLVMBuilderRef builder = bld->gallivm->builder;
 932    LLVMValueRef size0 = NULL;
 933    LLVMValueRef size1 = NULL;
 934    LLVMValueRef row_stride0_vec = NULL;
 935    LLVMValueRef row_stride1_vec = NULL;
 936    LLVMValueRef img_stride0_vec = NULL;
 937    LLVMValueRef img_stride1_vec = NULL;
 938    LLVMValueRef data_ptr0 = NULL;
 939    LLVMValueRef data_ptr1 = NULL;
 940    LLVMValueRef mipoff0 = NULL;
 941    LLVMValueRef mipoff1 = NULL;
 942    LLVMValueRef colors0[4], colors1[4];
 943    unsigned chan;
 944
 945    /* sample the first mipmap level */
 946    lp_build_mipmap_level_sizes(bld, ilevel0,
 947                                &size0,
 948                                &row_stride0_vec, &img_stride0_vec);
 949    if (bld->num_lods == 1) {
 950       data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
 951    }
 952    else {
 953       /* This path should work for num_lods 1 too but slightly less efficient */
 954       data_ptr0 = bld->base_ptr;
 955       mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
 956    }
 957    if (img_filter == PIPE_TEX_FILTER_NEAREST) {
 958       lp_build_sample_image_nearest(bld, sampler_unit,
 959                                     size0,
 960                                     row_stride0_vec, img_stride0_vec,
 961                                     data_ptr0, mipoff0, s, t, r, offsets,
 962                                     colors0);
 963    }
 964    else {
 965       assert(img_filter == PIPE_TEX_FILTER_LINEAR);
 966       lp_build_sample_image_linear(bld, sampler_unit,
 967                                    size0,
 968                                    row_stride0_vec, img_stride0_vec,
 969                                    data_ptr0, mipoff0, s, t, r, offsets,
 970                                    colors0);
 971    }
 972
 973    /* Store the first level's colors in the output variables */
 974    for (chan = 0; chan < 4; chan++) {
 975        LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
 976    }
 977
 978    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
 979       struct lp_build_if_state if_ctx;
 980       LLVMValueRef need_lerp;
 981       unsigned num_quads = bld->coord_bld.type.length / 4;
 982
 983       /* need_lerp = lod_fpart > 0 */
 984       if (num_quads == 1) {
 985          need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
 986                                    lod_fpart, bld->perquadf_bld.zero,
 987                                    "need_lerp");
 988       }
 989       else {
 990          /*
 991           * We'll do mip filtering if any of the quads need it.
 992           * It might be better to split the vectors here and only fetch/filter
 993           * quads which need it.
 994           */
 995          /*
 996           * We unfortunately need to clamp lod_fpart here since we can get
 997           * negative values which would screw up filtering if not all
 998           * lod_fpart values have same sign.
 999           */
1000          lod_fpart = lp_build_max(&bld->perquadf_bld, lod_fpart,
1001                                   bld->perquadf_bld.zero);
1002          need_lerp = lp_build_compare(bld->gallivm, bld->perquadf_bld.type,
1003                                       PIPE_FUNC_GREATER,
1004                                       lod_fpart, bld->perquadf_bld.zero);
1005          need_lerp = lp_build_any_true_range(&bld->perquadi_bld, num_quads, need_lerp);
1006      }
1007
1008       lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1009       {
1010          /* sample the second mipmap level */
1011          lp_build_mipmap_level_sizes(bld, ilevel1,
1012                                      &size1,
1013                                      &row_stride1_vec, &img_stride1_vec);
1014          if (bld->num_lods == 1) {
1015             data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1016          }
1017          else {
1018             data_ptr1 = bld->base_ptr;
1019             mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1020          }
1021          if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1022             lp_build_sample_image_nearest(bld, sampler_unit,
1023                                           size1,
1024                                           row_stride1_vec, img_stride1_vec,
1025                                           data_ptr1, mipoff1, s, t, r, offsets,
1026                                           colors1);
1027          }
1028          else {
1029             lp_build_sample_image_linear(bld, sampler_unit,
1030                                          size1,
1031                                          row_stride1_vec, img_stride1_vec,
1032                                          data_ptr1, mipoff1, s, t, r, offsets,
1033                                          colors1);
1034          }
1035
1036          /* interpolate samples from the two mipmap levels */
1037
1038          lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1039                                                            bld->perquadf_bld.type,
1040                                                            bld->texel_bld.type,
1041                                                            lod_fpart);
1042
1043          for (chan = 0; chan < 4; chan++) {
1044             colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1045                                           colors0[chan], colors1[chan],
1046                                           0);
1047             LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1048          }
1049       }
1050       lp_build_endif(&if_ctx);
1051    }
1052 }
1053
1054
1055 /**
1056  * Clamp layer coord to valid values.
1057  */
1058 static LLVMValueRef
1059 lp_build_layer_coord(struct lp_build_sample_context *bld,
1060                      unsigned texture_unit,
1061                      LLVMValueRef layer)
1062 {
1063    LLVMValueRef maxlayer;
1064
1065    maxlayer = bld->dynamic_state->depth(bld->dynamic_state,
1066                                         bld->gallivm, texture_unit);
1067    maxlayer = lp_build_sub(&bld->int_bld, maxlayer, bld->int_bld.one);
1068    maxlayer = lp_build_broadcast_scalar(&bld->int_coord_bld, maxlayer);
1069    return lp_build_clamp(&bld->int_coord_bld, layer,
1070                          bld->int_coord_bld.zero, maxlayer);
1071
1072 }
1073
1074
1075 /**
1076  * Calculate cube face, lod, mip levels.
1077  */
1078 static void
1079 lp_build_sample_common(struct lp_build_sample_context *bld,
1080                        unsigned texture_index,
1081                        unsigned sampler_index,
1082                        LLVMValueRef *s,
1083                        LLVMValueRef *t,
1084                        LLVMValueRef *r,
1085                        const struct lp_derivatives *derivs, /* optional */
1086                        LLVMValueRef lod_bias, /* optional */
1087                        LLVMValueRef explicit_lod, /* optional */
1088                        LLVMValueRef *lod_ipart,
1089                        LLVMValueRef *lod_fpart,
1090                        LLVMValueRef *ilevel0,
1091                        LLVMValueRef *ilevel1)
1092 {
1093    const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1094    const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1095    const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1096    const unsigned target = bld->static_texture_state->target;
1097    LLVMValueRef first_level, cube_rho = NULL;
1098
1099    /*
1100    printf("%s mip %d  min %d  mag %d\n", __FUNCTION__,
1101           mip_filter, min_filter, mag_filter);
1102    */
1103
1104    /*
1105     * Choose cube face, recompute texcoords for the chosen face and
1106     * compute rho here too (as it requires transform of derivatives).
1107     */
1108    if (target == PIPE_TEXTURE_CUBE) {
1109       LLVMValueRef face, face_s, face_t;
1110       boolean need_derivs;
1111       need_derivs = ((min_filter != mag_filter ||
1112                       mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
1113                       !bld->static_sampler_state->min_max_lod_equal &&
1114                       !explicit_lod);
1115       lp_build_cube_lookup(bld, *s, *t, *r, derivs, &face, &face_s, &face_t,
1116                            &cube_rho, need_derivs);
1117       *s = face_s; /* vec */
1118       *t = face_t; /* vec */
1119       /* use 'r' to indicate cube face */
1120       *r = face; /* vec */
1121    }
1122    else if (target == PIPE_TEXTURE_1D_ARRAY) {
1123       *r = lp_build_iround(&bld->coord_bld, *t);
1124       *r = lp_build_layer_coord(bld, texture_index, *r);
1125    }
1126    else if (target == PIPE_TEXTURE_2D_ARRAY) {
1127       *r = lp_build_iround(&bld->coord_bld, *r);
1128       *r = lp_build_layer_coord(bld, texture_index, *r);
1129    }
1130
1131    /*
1132     * Compute the level of detail (float).
1133     */
1134    if (min_filter != mag_filter ||
1135        mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1136       /* Need to compute lod either to choose mipmap levels or to
1137        * distinguish between minification/magnification with one mipmap level.
1138        */
1139       lp_build_lod_selector(bld, texture_index, sampler_index,
1140                             *s, *t, *r, cube_rho,
1141                             derivs, lod_bias, explicit_lod,
1142                             mip_filter,
1143                             lod_ipart, lod_fpart);
1144    } else {
1145       *lod_ipart = bld->perquadi_bld.zero;
1146    }
1147
1148    /*
1149     * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1150     */
1151    switch (mip_filter) {
1152    default:
1153       assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1154       /* fall-through */
1155    case PIPE_TEX_MIPFILTER_NONE:
1156       /* always use mip level 0 */
1157       if (HAVE_LLVM == 0x0207 && target == PIPE_TEXTURE_CUBE) {
1158          /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1159           * We should be able to set ilevel0 = const(0) but that causes
1160           * bad x86 code to be emitted.
1161           */
1162          assert(*lod_ipart);
1163          lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1164       }
1165       else {
1166          first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1167                                                        bld->gallivm, texture_index);
1168          first_level = lp_build_broadcast_scalar(&bld->perquadi_bld, first_level);
1169          *ilevel0 = first_level;
1170       }
1171       break;
1172    case PIPE_TEX_MIPFILTER_NEAREST:
1173       assert(*lod_ipart);
1174       lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1175       break;
1176    case PIPE_TEX_MIPFILTER_LINEAR:
1177       assert(*lod_ipart);
1178       assert(*lod_fpart);
1179       lp_build_linear_mip_levels(bld, texture_index,
1180                                  *lod_ipart, lod_fpart,
1181                                  ilevel0, ilevel1);
1182       break;
1183    }
1184 }
1185
1186 /**
1187  * General texture sampling codegen.
1188  * This function handles texture sampling for all texture targets (1D,
1189  * 2D, 3D, cube) and all filtering modes.
1190  */
1191 static void
1192 lp_build_sample_general(struct lp_build_sample_context *bld,
1193                         unsigned sampler_unit,
1194                         LLVMValueRef s,
1195                         LLVMValueRef t,
1196                         LLVMValueRef r,
1197                         const LLVMValueRef *offsets,
1198                         LLVMValueRef lod_ipart,
1199                         LLVMValueRef lod_fpart,
1200                         LLVMValueRef ilevel0,
1201                         LLVMValueRef ilevel1,
1202                         LLVMValueRef *colors_out)
1203 {
1204    struct lp_build_context *int_bld = &bld->int_bld;
1205    LLVMBuilderRef builder = bld->gallivm->builder;
1206    const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1207    const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1208    const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1209    LLVMValueRef texels[4];
1210    unsigned chan;
1211
1212    /*
1213     * Get/interpolate texture colors.
1214     */
1215
1216    for (chan = 0; chan < 4; ++chan) {
1217      texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1218      lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
1219    }
1220
1221    if (min_filter == mag_filter) {
1222       /* no need to distinguish between minification and magnification */
1223       lp_build_sample_mipmap(bld, sampler_unit,
1224                              min_filter, mip_filter,
1225                              s, t, r, offsets,
1226                              ilevel0, ilevel1, lod_fpart,
1227                              texels);
1228    }
1229    else {
1230       /* Emit conditional to choose min image filter or mag image filter
1231        * depending on the lod being > 0 or <= 0, respectively.
1232        */
1233       struct lp_build_if_state if_ctx;
1234       LLVMValueRef minify;
1235
1236       /*
1237        * XXX this should to all lods into account, if some are min
1238        * some max probably could hack up the coords/weights in the linear
1239        * path with selects to work for nearest.
1240        * If that's just two quads sitting next to each other it seems
1241        * quite ok to do the same filtering method on both though, at
1242        * least unless we have explicit lod (and who uses different
1243        * min/mag filter with that?)
1244        */
1245       if (bld->num_lods > 1)
1246          lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
1247                                              lp_build_const_int32(bld->gallivm, 0), "");
1248
1249       /* minify = lod >= 0.0 */
1250       minify = LLVMBuildICmp(builder, LLVMIntSGE,
1251                              lod_ipart, int_bld->zero, "");
1252
1253       lp_build_if(&if_ctx, bld->gallivm, minify);
1254       {
1255          /* Use the minification filter */
1256          lp_build_sample_mipmap(bld, sampler_unit,
1257                                 min_filter, mip_filter,
1258                                 s, t, r, offsets,
1259                                 ilevel0, ilevel1, lod_fpart,
1260                                 texels);
1261       }
1262       lp_build_else(&if_ctx);
1263       {
1264          /* Use the magnification filter */
1265          lp_build_sample_mipmap(bld, sampler_unit,
1266                                 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1267                                 s, t, r, offsets,
1268                                 ilevel0, NULL, NULL,
1269                                 texels);
1270       }
1271       lp_build_endif(&if_ctx);
1272    }
1273
1274    for (chan = 0; chan < 4; ++chan) {
1275      colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1276      lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
1277    }
1278 }
1279
1280
1281 /**
1282  * Texel fetch function.
1283  * In contrast to general sampling there is no filtering, no coord minification,
1284  * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
1285  * directly to be applied to the selected mip level (after adding texel offsets).
1286  * This function handles texel fetch for all targets where texel fetch is supported
1287  * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
1288  */
1289 static void
1290 lp_build_fetch_texel(struct lp_build_sample_context *bld,
1291                      unsigned texture_unit,
1292                      const LLVMValueRef *coords,
1293                      LLVMValueRef explicit_lod,
1294                      const LLVMValueRef *offsets,
1295                      LLVMValueRef *colors_out)
1296 {
1297    struct lp_build_context *perquadi_bld = &bld->perquadi_bld;
1298    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1299    unsigned dims = bld->dims, chan;
1300    unsigned target = bld->static_texture_state->target;
1301    LLVMValueRef size, ilevel;
1302    LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
1303    LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
1304    LLVMValueRef width, height, depth, i, j;
1305    LLVMValueRef offset, out_of_bounds, out1;
1306
1307    /* XXX just like ordinary sampling, we don't handle per-pixel lod (yet). */
1308    if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
1309       ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
1310                                          perquadi_bld->type, explicit_lod, 0);
1311       lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel);
1312    }
1313    else {
1314       bld->num_lods = 1;
1315       ilevel = lp_build_const_int32(bld->gallivm, 0);
1316    }
1317    lp_build_mipmap_level_sizes(bld, ilevel,
1318                                &size,
1319                                &row_stride_vec, &img_stride_vec);
1320    lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
1321                                 size, &width, &height, &depth);
1322
1323    if (target == PIPE_TEXTURE_1D_ARRAY ||
1324        target == PIPE_TEXTURE_2D_ARRAY) {
1325       if (target == PIPE_TEXTURE_1D_ARRAY) {
1326          z = lp_build_layer_coord(bld, texture_unit, y);
1327       }
1328       else {
1329          z = lp_build_layer_coord(bld, texture_unit, z);
1330       }
1331    }
1332
1333    /* This is a lot like border sampling */
1334    if (offsets[0]) {
1335       /* XXX coords are really unsigned, offsets are signed */
1336       x = lp_build_add(int_coord_bld, x, offsets[0]);
1337    }
1338    out_of_bounds = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
1339    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
1340    out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1341
1342    if (dims >= 2) {
1343       if (offsets[1]) {
1344          y = lp_build_add(int_coord_bld, y, offsets[1]);
1345       }
1346       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
1347       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1348       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
1349       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1350
1351       if (dims >= 3) {
1352          if (offsets[2]) {
1353             z = lp_build_add(int_coord_bld, z, offsets[2]);
1354          }
1355          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
1356          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1357          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
1358          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1359       }
1360    }
1361
1362    lp_build_sample_offset(int_coord_bld,
1363                           bld->format_desc,
1364                           x, y, z, row_stride_vec, img_stride_vec,
1365                           &offset, &i, &j);
1366
1367    if (bld->static_texture_state->target != PIPE_BUFFER) {
1368       offset = lp_build_add(int_coord_bld, offset,
1369                             lp_build_get_mip_offsets(bld, ilevel));
1370    }
1371
1372    offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
1373
1374    lp_build_fetch_rgba_soa(bld->gallivm,
1375                            bld->format_desc,
1376                            bld->texel_type,
1377                            bld->base_ptr, offset,
1378                            i, j,
1379                            colors_out);
1380
1381    if (0) {
1382       /*
1383        * Not needed except for ARB_robust_buffer_access_behavior.
1384        * Could use min/max above instead of out-of-bounds comparisons
1385        * (in fact cast to unsigned and min only is sufficient)
1386        * if we don't care about the result returned for out-of-bounds.
1387        */
1388       for (chan = 0; chan < 4; chan++) {
1389          colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
1390                                             bld->texel_bld.zero, colors_out[chan]);
1391       }
1392    }
1393 }
1394
1395
1396 /**
1397  * Do shadow test/comparison.
1398  * \param coords  incoming texcoords
1399  * \param texel  the texel to compare against (use the X channel)
1400  * Ideally this should really be done per-sample.
1401  */
1402 static void
1403 lp_build_sample_compare(struct lp_build_sample_context *bld,
1404                         const LLVMValueRef *coords,
1405                         LLVMValueRef texel[4])
1406 {
1407    struct lp_build_context *texel_bld = &bld->texel_bld;
1408    LLVMBuilderRef builder = bld->gallivm->builder;
1409    LLVMValueRef res, p;
1410    const unsigned chan = 0;
1411
1412    if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1413       return;
1414
1415    if (bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY ||
1416        bld->static_texture_state->target == PIPE_TEXTURE_CUBE) {
1417       p = coords[3];
1418    }
1419    else {
1420       p = coords[2];
1421    }
1422
1423    /* debug code */
1424    if (0) {
1425       LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
1426       LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
1427       LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
1428       lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
1429                       coord, tex);
1430    }
1431
1432    /* Clamp p coords to [0,1] */
1433    p = lp_build_clamp(&bld->coord_bld, p,
1434                       bld->coord_bld.zero,
1435                       bld->coord_bld.one);
1436
1437    /* result = (p FUNC texel) ? 1 : 0 */
1438    res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
1439                       p, texel[chan]);
1440    res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1441
1442    /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1443    texel[0] =
1444    texel[1] =
1445    texel[2] = res;
1446    texel[3] = texel_bld->one;
1447 }
1448
1449
1450 /**
1451  * Just set texels to white instead of actually sampling the texture.
1452  * For debugging.
1453  */
1454 void
1455 lp_build_sample_nop(struct gallivm_state *gallivm,
1456                     struct lp_type type,
1457                     const LLVMValueRef *coords,
1458                     LLVMValueRef texel_out[4])
1459 {
1460    LLVMValueRef one = lp_build_one(gallivm, type);
1461    unsigned chan;
1462
1463    for (chan = 0; chan < 4; chan++) {
1464       texel_out[chan] = one;
1465    }
1466 }
1467
1468
1469 /**
1470  * Build texture sampling code.
1471  * 'texel' will return a vector of four LLVMValueRefs corresponding to
1472  * R, G, B, A.
1473  * \param type  vector float type to use for coords, etc.
1474  * \param is_fetch  if this is a texel fetch instruction.
1475  * \param derivs  partial derivatives of (s,t,r,q) with respect to x and y
1476  */
1477 void
1478 lp_build_sample_soa(struct gallivm_state *gallivm,
1479                     const struct lp_static_texture_state *static_texture_state,
1480                     const struct lp_static_sampler_state *static_sampler_state,
1481                     struct lp_sampler_dynamic_state *dynamic_state,
1482                     struct lp_type type,
1483                     boolean is_fetch,
1484                     unsigned texture_index,
1485                     unsigned sampler_index,
1486                     const LLVMValueRef *coords,
1487                     const LLVMValueRef *offsets,
1488                     const struct lp_derivatives *derivs, /* optional */
1489                     LLVMValueRef lod_bias, /* optional */
1490                     LLVMValueRef explicit_lod, /* optional */
1491                     LLVMValueRef texel_out[4])
1492 {
1493    unsigned dims = texture_dims(static_texture_state->target);
1494    unsigned num_quads = type.length / 4;
1495    unsigned mip_filter;
1496    struct lp_build_sample_context bld;
1497    struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
1498    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1499    LLVMBuilderRef builder = gallivm->builder;
1500    LLVMValueRef tex_width;
1501    LLVMValueRef s;
1502    LLVMValueRef t;
1503    LLVMValueRef r;
1504
1505    if (0) {
1506       enum pipe_format fmt = static_texture_state->format;
1507       debug_printf("Sample from %s\n", util_format_name(fmt));
1508    }
1509
1510    assert(type.floating);
1511
1512    /* Setup our build context */
1513    memset(&bld, 0, sizeof bld);
1514    bld.gallivm = gallivm;
1515    bld.static_sampler_state = &derived_sampler_state;
1516    bld.static_texture_state = static_texture_state;
1517    bld.dynamic_state = dynamic_state;
1518    bld.format_desc = util_format_description(static_texture_state->format);
1519    bld.dims = dims;
1520
1521    bld.vector_width = lp_type_width(type);
1522
1523    bld.float_type = lp_type_float(32);
1524    bld.int_type = lp_type_int(32);
1525    bld.coord_type = type;
1526    bld.int_coord_type = lp_int_type(type);
1527    bld.float_size_in_type = lp_type_float(32);
1528    bld.float_size_in_type.length = dims > 1 ? 4 : 1;
1529    bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
1530    bld.texel_type = type;
1531    bld.perquadf_type = type;
1532    /* we want native vector size to be able to use our intrinsics */
1533    bld.perquadf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
1534    bld.perquadi_type = lp_int_type(bld.perquadf_type);
1535
1536    /* always using the first channel hopefully should be safe,
1537     * if not things WILL break in other places anyway.
1538     */
1539    if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1540        bld.format_desc->channel[0].pure_integer) {
1541       if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1542          bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1543       }
1544       else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1545          bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
1546       }
1547    }
1548    else if (util_format_has_stencil(bld.format_desc) &&
1549        !util_format_has_depth(bld.format_desc)) {
1550       /* for stencil only formats, sample stencil (uint) */
1551       bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1552    }
1553
1554    if (!static_texture_state->level_zero_only) {
1555       derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
1556    } else {
1557       derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
1558    }
1559    mip_filter = derived_sampler_state.min_mip_filter;
1560
1561    if (0) {
1562       debug_printf("  .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
1563    }
1564
1565    /*
1566     * There are other situations where at least the multiple int lods could be
1567     * avoided like min and max lod being equal.
1568     */
1569    if ((is_fetch && explicit_lod && bld.static_texture_state->target != PIPE_BUFFER) ||
1570        (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
1571       bld.num_lods = num_quads;
1572    }
1573    else {
1574       bld.num_lods = 1;
1575    }
1576
1577    bld.float_size_type = bld.float_size_in_type;
1578    bld.float_size_type.length = bld.num_lods > 1 ? type.length :
1579                                    bld.float_size_in_type.length;
1580    bld.int_size_type = lp_int_type(bld.float_size_type);
1581
1582    lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1583    lp_build_context_init(&bld.float_vec_bld, gallivm, type);
1584    lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1585    lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1586    lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1587    lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
1588    lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
1589    lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1590    lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1591    lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1592    lp_build_context_init(&bld.perquadf_bld, gallivm, bld.perquadf_type);
1593    lp_build_context_init(&bld.perquadi_bld, gallivm, bld.perquadi_type);
1594
1595    /* Get the dynamic state */
1596    tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index);
1597    bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index);
1598    bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index);
1599    bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index);
1600    bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index);
1601    /* Note that mip_offsets is an array[level] of offsets to texture images */
1602
1603    s = coords[0];
1604    t = coords[1];
1605    r = coords[2];
1606
1607    /* width, height, depth as single int vector */
1608    if (dims <= 1) {
1609       bld.int_size = tex_width;
1610    }
1611    else {
1612       bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
1613                                             tex_width, LLVMConstInt(i32t, 0, 0), "");
1614       if (dims >= 2) {
1615          LLVMValueRef tex_height =
1616             dynamic_state->height(dynamic_state, gallivm, texture_index);
1617          bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1618                                                tex_height, LLVMConstInt(i32t, 1, 0), "");
1619          if (dims >= 3) {
1620             LLVMValueRef tex_depth =
1621                dynamic_state->depth(dynamic_state, gallivm, texture_index);
1622             bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1623                                                   tex_depth, LLVMConstInt(i32t, 2, 0), "");
1624          }
1625       }
1626    }
1627
1628    if (0) {
1629       /* For debug: no-op texture sampling */
1630       lp_build_sample_nop(gallivm,
1631                           bld.texel_type,
1632                           coords,
1633                           texel_out);
1634    }
1635
1636    else if (is_fetch) {
1637       lp_build_fetch_texel(&bld, texture_index, coords,
1638                            explicit_lod, offsets,
1639                            texel_out);
1640    }
1641
1642    else {
1643       LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
1644       LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
1645       boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
1646                         lp_is_simple_wrap_mode(static_sampler_state->wrap_s) &&
1647                         lp_is_simple_wrap_mode(static_sampler_state->wrap_t);
1648
1649       if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1650           !use_aos && util_format_fits_8unorm(bld.format_desc)) {
1651          debug_printf("%s: using floating point linear filtering for %s\n",
1652                       __FUNCTION__, bld.format_desc->short_name);
1653          debug_printf("  min_img %d  mag_img %d  mip %d  wraps %d  wrapt %d\n",
1654                       static_sampler_state->min_img_filter,
1655                       static_sampler_state->mag_img_filter,
1656                       static_sampler_state->min_mip_filter,
1657                       static_sampler_state->wrap_s,
1658                       static_sampler_state->wrap_t);
1659       }
1660
1661       lp_build_sample_common(&bld, texture_index, sampler_index,
1662                              &s, &t, &r,
1663                              derivs, lod_bias, explicit_lod,
1664                              &lod_ipart, &lod_fpart,
1665                              &ilevel0, &ilevel1);
1666
1667       /*
1668        * we only try 8-wide sampling with soa as it appears to
1669        * be a loss with aos with AVX (but it should work).
1670        * (It should be faster if we'd support avx2)
1671        */
1672       if (num_quads == 1 || !use_aos) {
1673
1674          if (num_quads > 1) {
1675             if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1676                LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
1677                /*
1678                 * These parameters are the same for all quads,
1679                 * could probably simplify.
1680                 */
1681                lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
1682                ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
1683             }
1684          }
1685          if (use_aos) {
1686             /* do sampling/filtering with fixed pt arithmetic */
1687             lp_build_sample_aos(&bld, sampler_index,
1688                                 s, t, r, offsets,
1689                                 lod_ipart, lod_fpart,
1690                                 ilevel0, ilevel1,
1691                                 texel_out);
1692          }
1693
1694          else {
1695             lp_build_sample_general(&bld, sampler_index,
1696                                     s, t, r, offsets,
1697                                     lod_ipart, lod_fpart,
1698                                     ilevel0, ilevel1,
1699                                     texel_out);
1700          }
1701       }
1702       else {
1703          unsigned j;
1704          struct lp_build_sample_context bld4;
1705          struct lp_type type4 = type;
1706          unsigned i;
1707          LLVMValueRef texelout4[4];
1708          LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
1709
1710          type4.length = 4;
1711
1712          /* Setup our build context */
1713          memset(&bld4, 0, sizeof bld4);
1714          bld4.gallivm = bld.gallivm;
1715          bld4.static_texture_state = bld.static_texture_state;
1716          bld4.static_sampler_state = bld.static_sampler_state;
1717          bld4.dynamic_state = bld.dynamic_state;
1718          bld4.format_desc = bld.format_desc;
1719          bld4.dims = bld.dims;
1720          bld4.row_stride_array = bld.row_stride_array;
1721          bld4.img_stride_array = bld.img_stride_array;
1722          bld4.base_ptr = bld.base_ptr;
1723          bld4.mip_offsets = bld.mip_offsets;
1724          bld4.int_size = bld.int_size;
1725
1726          bld4.vector_width = lp_type_width(type4);
1727
1728          bld4.float_type = lp_type_float(32);
1729          bld4.int_type = lp_type_int(32);
1730          bld4.coord_type = type4;
1731          bld4.int_coord_type = lp_int_type(type4);
1732          bld4.float_size_in_type = lp_type_float(32);
1733          bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
1734          bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
1735          bld4.texel_type = bld.texel_type;
1736          bld4.texel_type.length = 4;
1737          bld4.perquadf_type = type4;
1738          /* we want native vector size to be able to use our intrinsics */
1739          bld4.perquadf_type.length = 1;
1740          bld4.perquadi_type = lp_int_type(bld4.perquadf_type);
1741
1742          bld4.num_lods = 1;
1743          bld4.int_size_type = bld4.int_size_in_type;
1744          bld4.float_size_type = bld4.float_size_in_type;
1745
1746          lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
1747          lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
1748          lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
1749          lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
1750          lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
1751          lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
1752          lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
1753          lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
1754          lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
1755          lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
1756          lp_build_context_init(&bld4.perquadf_bld, gallivm, bld4.perquadf_type);
1757          lp_build_context_init(&bld4.perquadi_bld, gallivm, bld4.perquadi_type);
1758
1759          for (i = 0; i < num_quads; i++) {
1760             LLVMValueRef s4, t4, r4;
1761             LLVMValueRef lod_iparts, lod_fparts = NULL;
1762             LLVMValueRef ilevel0s, ilevel1s = NULL;
1763             LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
1764             LLVMValueRef offsets4[4] = { NULL };
1765
1766             s4 = lp_build_extract_range(gallivm, s, 4*i, 4);
1767             t4 = lp_build_extract_range(gallivm, t, 4*i, 4);
1768             r4 = lp_build_extract_range(gallivm, r, 4*i, 4);
1769
1770             if (offsets[0]) {
1771                offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
1772                if (dims > 1) {
1773                   offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
1774                   if (dims > 2) {
1775                      offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
1776                   }
1777                }
1778             }
1779             lod_iparts = LLVMBuildExtractElement(builder, lod_ipart, indexi, "");
1780             ilevel0s = LLVMBuildExtractElement(builder, ilevel0, indexi, "");
1781             if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1782                ilevel1s = LLVMBuildExtractElement(builder, ilevel1, indexi, "");
1783                lod_fparts = LLVMBuildExtractElement(builder, lod_fpart, indexi, "");
1784             }
1785
1786             if (use_aos) {
1787                /* do sampling/filtering with fixed pt arithmetic */
1788                lp_build_sample_aos(&bld4, sampler_index,
1789                                    s4, t4, r4, offsets4,
1790                                    lod_iparts, lod_fparts,
1791                                    ilevel0s, ilevel1s,
1792                                    texelout4);
1793             }
1794
1795             else {
1796                lp_build_sample_general(&bld4, sampler_index,
1797                                        s4, t4, r4, offsets4,
1798                                        lod_iparts, lod_fparts,
1799                                        ilevel0s, ilevel1s,
1800                                        texelout4);
1801             }
1802             for (j = 0; j < 4; j++) {
1803                texelouttmp[j][i] = texelout4[j];
1804             }
1805          }
1806
1807          for (j = 0; j < 4; j++) {
1808             texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
1809          }
1810       }
1811
1812       lp_build_sample_compare(&bld, coords, texel_out);
1813    }
1814
1815    if (static_texture_state->target != PIPE_BUFFER) {
1816       apply_sampler_swizzle(&bld, texel_out);
1817    }
1818
1819    /*
1820     * texel type can be a (32bit) int/uint (for pure int formats only),
1821     * however we are expected to always return floats (storage is untyped).
1822     */
1823    if (!bld.texel_type.floating) {
1824       unsigned chan;
1825       for (chan = 0; chan < 4; chan++) {
1826          texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
1827                                             lp_build_vec_type(gallivm, type), "");
1828       }
1829    }
1830 }
1831
1832 void
1833 lp_build_size_query_soa(struct gallivm_state *gallivm,
1834                         const struct lp_static_texture_state *static_state,
1835                         struct lp_sampler_dynamic_state *dynamic_state,
1836                         struct lp_type int_type,
1837                         unsigned texture_unit,
1838                         boolean need_nr_mips,
1839                         LLVMValueRef explicit_lod,
1840                         LLVMValueRef *sizes_out)
1841 {
1842    LLVMValueRef lod;
1843    LLVMValueRef size;
1844    LLVMValueRef first_level = NULL;
1845    int dims, i;
1846    boolean has_array;
1847    struct lp_build_context bld_int_vec;
1848
1849    dims = texture_dims(static_state->target);
1850
1851    switch (static_state->target) {
1852    case PIPE_TEXTURE_1D_ARRAY:
1853    case PIPE_TEXTURE_2D_ARRAY:
1854       has_array = TRUE;
1855       break;
1856    default:
1857       has_array = FALSE;
1858       break;
1859    }
1860
1861    assert(!int_type.floating);
1862
1863    lp_build_context_init(&bld_int_vec, gallivm, lp_type_int_vec(32, 128));
1864
1865    if (explicit_lod) {
1866       lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), "");
1867       first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit);
1868       lod = lp_build_broadcast_scalar(&bld_int_vec,
1869                                       LLVMBuildAdd(gallivm->builder, lod, first_level, "lod"));
1870
1871    } else {
1872       lod = bld_int_vec.zero;
1873    }
1874
1875    if (need_nr_mips) {
1876       size = bld_int_vec.zero;
1877    }
1878    else {
1879       size = bld_int_vec.undef;
1880    }
1881
1882    size = LLVMBuildInsertElement(gallivm->builder, size,
1883                                  dynamic_state->width(dynamic_state, gallivm, texture_unit),
1884                                  lp_build_const_int32(gallivm, 0), "");
1885
1886    if (dims >= 2) {
1887       size = LLVMBuildInsertElement(gallivm->builder, size,
1888                                     dynamic_state->height(dynamic_state, gallivm, texture_unit),
1889                                     lp_build_const_int32(gallivm, 1), "");
1890    }
1891
1892    if (dims >= 3) {
1893       size = LLVMBuildInsertElement(gallivm->builder, size,
1894                                     dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1895                                     lp_build_const_int32(gallivm, 2), "");
1896    }
1897
1898    size = lp_build_minify(&bld_int_vec, size, lod);
1899
1900    if (has_array)
1901       size = LLVMBuildInsertElement(gallivm->builder, size,
1902                                     dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1903                                     lp_build_const_int32(gallivm, dims), "");
1904
1905    /*
1906     * XXX for out-of-bounds lod, should set size to zero vector here
1907     * (for dx10-style only, i.e. need_nr_mips)
1908     */
1909
1910    for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
1911       sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec.type, int_type,
1912                                                 size,
1913                                                 lp_build_const_int32(gallivm, i));
1914    }
1915
1916    /*
1917     * if there's no explicit_lod (buffers, rects) queries requiring nr of
1918     * mips would be illegal.
1919     */
1920    if (need_nr_mips && explicit_lod) {
1921       struct lp_build_context bld_int_scalar;
1922       LLVMValueRef num_levels;
1923       lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
1924
1925       if (static_state->level_zero_only) {
1926          num_levels = bld_int_scalar.one;
1927       }
1928       else {
1929          LLVMValueRef last_level;
1930
1931          last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit);
1932          num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
1933          num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one);
1934       }
1935       sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type),
1936                                         num_levels);
1937    }
1938 }