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          struct lp_build_context abs_coord_bld = bld->coord_bld;
 440          abs_coord_bld.type.sign = FALSE;
 441
 442          if (bld->static_sampler_state->normalized_coords) {
 443             /* scale coord to length */
 444             coord = lp_build_mul(coord_bld, coord, length_f);
 445          }
 446          if (offset) {
 447             offset = lp_build_int_to_float(coord_bld, offset);
 448             coord = lp_build_add(coord_bld, coord, offset);
 449          }
 450          coord = lp_build_abs(coord_bld, coord);
 451
 452          /* clamp to length max */
 453          coord = lp_build_min(coord_bld, coord, length_f);
 454          /* subtract 0.5 */
 455          coord = lp_build_sub(coord_bld, coord, half);
 456          /* clamp to [0, length - 0.5] */
 457          coord = lp_build_max(coord_bld, coord, coord_bld->zero);
 458
 459          /* convert to int, compute lerp weight */
 460          lp_build_ifloor_fract(&abs_coord_bld, coord, &coord0, &weight);
 461          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 462          /* coord1 = min(coord1, length-1) */
 463          coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
 464       }
 465       break;
 466
 467    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 468       {
 469          if (bld->static_sampler_state->normalized_coords) {
 470             /* scale coord to length */
 471             coord = lp_build_mul(coord_bld, coord, length_f);
 472          }
 473          if (offset) {
 474             offset = lp_build_int_to_float(coord_bld, offset);
 475             coord = lp_build_add(coord_bld, coord, offset);
 476          }
 477          coord = lp_build_abs(coord_bld, coord);
 478
 479          /* was: clamp to [-0.5, length + 0.5] then sub 0.5 */
 480          /* skip clamp - always positive, and other side
 481             only potentially matters for very large coords */
 482          coord = lp_build_sub(coord_bld, coord, half);
 483
 484          /* convert to int, compute lerp weight */
 485          lp_build_ifloor_fract(coord_bld, coord, &coord0, &weight);
 486          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 487       }
 488       break;
 489
 490    default:
 491       assert(0);
 492       coord0 = NULL;
 493       coord1 = NULL;
 494       weight = NULL;
 495    }
 496
 497    *x0_out = coord0;
 498    *x1_out = coord1;
 499    *weight_out = weight;
 500 }
 501
 502
 503 /**
 504  * Build LLVM code for texture wrap mode for nearest filtering.
 505  * \param coord  the incoming texcoord (nominally in [0,1])
 506  * \param length  the texture size along one dimension, as int vector
 507  * \param length_f  the texture size along one dimension, as float vector
 508  * \param offset  texel offset along one dimension (as int vector)
 509  * \param is_pot  if TRUE, length is a power of two
 510  * \param wrap_mode  one of PIPE_TEX_WRAP_x
 511  */
 512 static LLVMValueRef
 513 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
 514                              LLVMValueRef coord,
 515                              LLVMValueRef length,
 516                              LLVMValueRef length_f,
 517                              LLVMValueRef offset,
 518                              boolean is_pot,
 519                              unsigned wrap_mode)
 520 {
 521    struct lp_build_context *coord_bld = &bld->coord_bld;
 522    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 523    LLVMBuilderRef builder = bld->gallivm->builder;
 524    LLVMValueRef length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
 525    LLVMValueRef icoord;
 526
 527    switch(wrap_mode) {
 528    case PIPE_TEX_WRAP_REPEAT:
 529       if (is_pot) {
 530          coord = lp_build_mul(coord_bld, coord, length_f);
 531          icoord = lp_build_ifloor(coord_bld, coord);
 532          if (offset) {
 533             icoord = lp_build_add(int_coord_bld, icoord, offset);
 534          }
 535          icoord = LLVMBuildAnd(builder, icoord, length_minus_one, "");
 536       }
 537       else {
 538           if (offset) {
 539              offset = lp_build_int_to_float(coord_bld, offset);
 540              offset = lp_build_div(coord_bld, offset, length_f);
 541              coord = lp_build_add(coord_bld, coord, offset);
 542           }
 543           /* take fraction, unnormalize */
 544           coord = lp_build_fract_safe(coord_bld, coord);
 545           coord = lp_build_mul(coord_bld, coord, length_f);
 546           icoord = lp_build_itrunc(coord_bld, coord);
 547       }
 548       break;
 549
 550    case PIPE_TEX_WRAP_CLAMP:
 551    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 552       if (bld->static_sampler_state->normalized_coords) {
 553          /* scale coord to length */
 554          coord = lp_build_mul(coord_bld, coord, length_f);
 555       }
 556
 557       /* floor */
 558       /* use itrunc instead since we clamp to 0 anyway */
 559       icoord = lp_build_itrunc(coord_bld, coord);
 560       if (offset) {
 561          icoord = lp_build_add(int_coord_bld, icoord, offset);
 562       }
 563
 564       /* clamp to [0, length - 1]. */
 565       icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
 566                               length_minus_one);
 567       break;
 568
 569    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 570       if (bld->static_sampler_state->normalized_coords) {
 571          /* scale coord to length */
 572          coord = lp_build_mul(coord_bld, coord, length_f);
 573       }
 574       /* no clamp necessary, border masking will handle this */
 575       icoord = lp_build_ifloor(coord_bld, coord);
 576       if (offset) {
 577          icoord = lp_build_add(int_coord_bld, icoord, offset);
 578       }
 579       break;
 580
 581    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 582       if (offset) {
 583          offset = lp_build_int_to_float(coord_bld, offset);
 584          offset = lp_build_div(coord_bld, offset, length_f);
 585          coord = lp_build_add(coord_bld, coord, offset);
 586       }
 587       /* compute mirror function */
 588       coord = lp_build_coord_mirror(bld, coord);
 589
 590       /* scale coord to length */
 591       assert(bld->static_sampler_state->normalized_coords);
 592       coord = lp_build_mul(coord_bld, coord, length_f);
 593
 594       /* itrunc == ifloor here */
 595       icoord = lp_build_itrunc(coord_bld, coord);
 596
 597       /* clamp to [0, length - 1] */
 598       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 599       break;
 600
 601    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 602    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 603       if (bld->static_sampler_state->normalized_coords) {
 604          /* scale coord to length */
 605          coord = lp_build_mul(coord_bld, coord, length_f);
 606       }
 607       if (offset) {
 608          offset = lp_build_int_to_float(coord_bld, offset);
 609          coord = lp_build_add(coord_bld, coord, offset);
 610       }
 611       coord = lp_build_abs(coord_bld, coord);
 612
 613       /* itrunc == ifloor here */
 614       icoord = lp_build_itrunc(coord_bld, coord);
 615
 616       /* clamp to [0, length - 1] */
 617       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 618       break;
 619
 620    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 621       if (bld->static_sampler_state->normalized_coords) {
 622          /* scale coord to length */
 623          coord = lp_build_mul(coord_bld, coord, length_f);
 624       }
 625       if (offset) {
 626          offset = lp_build_int_to_float(coord_bld, offset);
 627          coord = lp_build_add(coord_bld, coord, offset);
 628       }
 629       coord = lp_build_abs(coord_bld, coord);
 630
 631       /* itrunc == ifloor here */
 632       icoord = lp_build_itrunc(coord_bld, coord);
 633       break;
 634
 635    default:
 636       assert(0);
 637       icoord = NULL;
 638    }
 639
 640    return icoord;
 641 }
 642
 643
 644 /**
 645  * Generate code to sample a mipmap level with nearest filtering.
 646  * If sampling a cube texture, r = cube face in [0,5].
 647  */
 648 static void
 649 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
 650                               unsigned sampler_unit,
 651                               LLVMValueRef size,
 652                               LLVMValueRef row_stride_vec,
 653                               LLVMValueRef img_stride_vec,
 654                               LLVMValueRef data_ptr,
 655                               LLVMValueRef mipoffsets,
 656                               LLVMValueRef s,
 657                               LLVMValueRef t,
 658                               LLVMValueRef r,
 659                               const LLVMValueRef *offsets,
 660                               LLVMValueRef colors_out[4])
 661 {
 662    const unsigned dims = bld->dims;
 663    LLVMValueRef width_vec;
 664    LLVMValueRef height_vec;
 665    LLVMValueRef depth_vec;
 666    LLVMValueRef flt_size;
 667    LLVMValueRef flt_width_vec;
 668    LLVMValueRef flt_height_vec;
 669    LLVMValueRef flt_depth_vec;
 670    LLVMValueRef x, y = NULL, z = NULL;
 671
 672    lp_build_extract_image_sizes(bld,
 673                                 &bld->int_size_bld,
 674                                 bld->int_coord_type,
 675                                 size,
 676                                 &width_vec, &height_vec, &depth_vec);
 677
 678    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
 679
 680    lp_build_extract_image_sizes(bld,
 681                                 &bld->float_size_bld,
 682                                 bld->coord_type,
 683                                 flt_size,
 684                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
 685
 686    /*
 687     * Compute integer texcoords.
 688     */
 689    x = lp_build_sample_wrap_nearest(bld, s, width_vec, flt_width_vec, offsets[0],
 690                                     bld->static_texture_state->pot_width,
 691                                     bld->static_sampler_state->wrap_s);
 692    lp_build_name(x, "tex.x.wrapped");
 693
 694    if (dims >= 2) {
 695       y = lp_build_sample_wrap_nearest(bld, t, height_vec, flt_height_vec, offsets[1],
 696                                        bld->static_texture_state->pot_height,
 697                                        bld->static_sampler_state->wrap_t);
 698       lp_build_name(y, "tex.y.wrapped");
 699
 700       if (dims == 3) {
 701          z = lp_build_sample_wrap_nearest(bld, r, depth_vec, flt_depth_vec, offsets[2],
 702                                           bld->static_texture_state->pot_depth,
 703                                           bld->static_sampler_state->wrap_r);
 704          lp_build_name(z, "tex.z.wrapped");
 705       }
 706    }
 707    if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
 708        bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
 709        bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
 710       z = r;
 711       lp_build_name(z, "tex.z.layer");
 712    }
 713
 714    /*
 715     * Get texture colors.
 716     */
 717    lp_build_sample_texel_soa(bld, sampler_unit,
 718                              width_vec, height_vec, depth_vec,
 719                              x, y, z,
 720                              row_stride_vec, img_stride_vec,
 721                              data_ptr, mipoffsets, colors_out);
 722 }
 723
 724
 725 /**
 726  * Generate code to sample a mipmap level with linear filtering.
 727  * If sampling a cube texture, r = cube face in [0,5].
 728  */
 729 static void
 730 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
 731                              unsigned sampler_unit,
 732                              LLVMValueRef size,
 733                              LLVMValueRef row_stride_vec,
 734                              LLVMValueRef img_stride_vec,
 735                              LLVMValueRef data_ptr,
 736                              LLVMValueRef mipoffsets,
 737                              LLVMValueRef s,
 738                              LLVMValueRef t,
 739                              LLVMValueRef r,
 740                              const LLVMValueRef *offsets,
 741                              LLVMValueRef colors_out[4])
 742 {
 743    const unsigned dims = bld->dims;
 744    LLVMValueRef width_vec;
 745    LLVMValueRef height_vec;
 746    LLVMValueRef depth_vec;
 747    LLVMValueRef flt_size;
 748    LLVMValueRef flt_width_vec;
 749    LLVMValueRef flt_height_vec;
 750    LLVMValueRef flt_depth_vec;
 751    LLVMValueRef x0, y0 = NULL, z0 = NULL, x1, y1 = NULL, z1 = NULL;
 752    LLVMValueRef s_fpart, t_fpart = NULL, r_fpart = NULL;
 753    LLVMValueRef neighbors[2][2][4];
 754    int chan;
 755
 756    lp_build_extract_image_sizes(bld,
 757                                 &bld->int_size_bld,
 758                                 bld->int_coord_type,
 759                                 size,
 760                                 &width_vec, &height_vec, &depth_vec);
 761
 762    flt_size = lp_build_int_to_float(&bld->float_size_bld, size);
 763
 764    lp_build_extract_image_sizes(bld,
 765                                 &bld->float_size_bld,
 766                                 bld->coord_type,
 767                                 flt_size,
 768                                 &flt_width_vec, &flt_height_vec, &flt_depth_vec);
 769
 770    /*
 771     * Compute integer texcoords.
 772     */
 773    lp_build_sample_wrap_linear(bld, s, width_vec, flt_width_vec, offsets[0],
 774                                bld->static_texture_state->pot_width,
 775                                bld->static_sampler_state->wrap_s,
 776                                &x0, &x1, &s_fpart);
 777    lp_build_name(x0, "tex.x0.wrapped");
 778    lp_build_name(x1, "tex.x1.wrapped");
 779
 780    if (dims >= 2) {
 781       lp_build_sample_wrap_linear(bld, t, height_vec, flt_height_vec, offsets[1],
 782                                   bld->static_texture_state->pot_height,
 783                                   bld->static_sampler_state->wrap_t,
 784                                   &y0, &y1, &t_fpart);
 785       lp_build_name(y0, "tex.y0.wrapped");
 786       lp_build_name(y1, "tex.y1.wrapped");
 787
 788       if (dims == 3) {
 789          lp_build_sample_wrap_linear(bld, r, depth_vec, flt_depth_vec, offsets[2],
 790                                      bld->static_texture_state->pot_depth,
 791                                      bld->static_sampler_state->wrap_r,
 792                                      &z0, &z1, &r_fpart);
 793          lp_build_name(z0, "tex.z0.wrapped");
 794          lp_build_name(z1, "tex.z1.wrapped");
 795       }
 796    }
 797    if (bld->static_texture_state->target == PIPE_TEXTURE_CUBE ||
 798        bld->static_texture_state->target == PIPE_TEXTURE_1D_ARRAY ||
 799        bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY) {
 800       z0 = z1 = r;  /* cube face or array layer */
 801       lp_build_name(z0, "tex.z0.layer");
 802       lp_build_name(z1, "tex.z1.layer");
 803    }
 804
 805
 806    /*
 807     * Get texture colors.
 808     */
 809    /* get x0/x1 texels */
 810    lp_build_sample_texel_soa(bld, sampler_unit,
 811                              width_vec, height_vec, depth_vec,
 812                              x0, y0, z0,
 813                              row_stride_vec, img_stride_vec,
 814                              data_ptr, mipoffsets, neighbors[0][0]);
 815    lp_build_sample_texel_soa(bld, sampler_unit,
 816                              width_vec, height_vec, depth_vec,
 817                              x1, y0, z0,
 818                              row_stride_vec, img_stride_vec,
 819                              data_ptr, mipoffsets, neighbors[0][1]);
 820
 821    if (dims == 1) {
 822       /* Interpolate two samples from 1D image to produce one color */
 823       for (chan = 0; chan < 4; chan++) {
 824          colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
 825                                           neighbors[0][0][chan],
 826                                           neighbors[0][1][chan],
 827                                           0);
 828       }
 829    }
 830    else {
 831       /* 2D/3D texture */
 832       LLVMValueRef colors0[4];
 833
 834       /* get x0/x1 texels at y1 */
 835       lp_build_sample_texel_soa(bld, sampler_unit,
 836                                 width_vec, height_vec, depth_vec,
 837                                 x0, y1, z0,
 838                                 row_stride_vec, img_stride_vec,
 839                                 data_ptr, mipoffsets, neighbors[1][0]);
 840       lp_build_sample_texel_soa(bld, sampler_unit,
 841                                 width_vec, height_vec, depth_vec,
 842                                 x1, y1, z0,
 843                                 row_stride_vec, img_stride_vec,
 844                                 data_ptr, mipoffsets, neighbors[1][1]);
 845
 846       /* Bilinear interpolate the four samples from the 2D image / 3D slice */
 847       for (chan = 0; chan < 4; chan++) {
 848          colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
 849                                           s_fpart, t_fpart,
 850                                           neighbors[0][0][chan],
 851                                           neighbors[0][1][chan],
 852                                           neighbors[1][0][chan],
 853                                           neighbors[1][1][chan],
 854                                           0);
 855       }
 856
 857       if (dims == 3) {
 858          LLVMValueRef neighbors1[2][2][4];
 859          LLVMValueRef colors1[4];
 860
 861          /* get x0/x1/y0/y1 texels at z1 */
 862          lp_build_sample_texel_soa(bld, sampler_unit,
 863                                    width_vec, height_vec, depth_vec,
 864                                    x0, y0, z1,
 865                                    row_stride_vec, img_stride_vec,
 866                                    data_ptr, mipoffsets, neighbors1[0][0]);
 867          lp_build_sample_texel_soa(bld, sampler_unit,
 868                                    width_vec, height_vec, depth_vec,
 869                                    x1, y0, z1,
 870                                    row_stride_vec, img_stride_vec,
 871                                    data_ptr, mipoffsets, neighbors1[0][1]);
 872          lp_build_sample_texel_soa(bld, sampler_unit,
 873                                    width_vec, height_vec, depth_vec,
 874                                    x0, y1, z1,
 875                                    row_stride_vec, img_stride_vec,
 876                                    data_ptr, mipoffsets, neighbors1[1][0]);
 877          lp_build_sample_texel_soa(bld, sampler_unit,
 878                                    width_vec, height_vec, depth_vec,
 879                                    x1, y1, z1,
 880                                    row_stride_vec, img_stride_vec,
 881                                    data_ptr, mipoffsets, neighbors1[1][1]);
 882
 883          /* Bilinear interpolate the four samples from the second Z slice */
 884          for (chan = 0; chan < 4; chan++) {
 885             colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
 886                                              s_fpart, t_fpart,
 887                                              neighbors1[0][0][chan],
 888                                              neighbors1[0][1][chan],
 889                                              neighbors1[1][0][chan],
 890                                              neighbors1[1][1][chan],
 891                                              0);
 892          }
 893
 894          /* Linearly interpolate the two samples from the two 3D slices */
 895          for (chan = 0; chan < 4; chan++) {
 896             colors_out[chan] = lp_build_lerp(&bld->texel_bld,
 897                                              r_fpart,
 898                                              colors0[chan], colors1[chan],
 899                                              0);
 900          }
 901       }
 902       else {
 903          /* 2D tex */
 904          for (chan = 0; chan < 4; chan++) {
 905             colors_out[chan] = colors0[chan];
 906          }
 907       }
 908    }
 909 }
 910
 911
 912 /**
 913  * Sample the texture/mipmap using given image filter and mip filter.
 914  * data0_ptr and data1_ptr point to the two mipmap levels to sample
 915  * from.  width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
 916  * If we're using nearest miplevel sampling the '1' values will be null/unused.
 917  */
 918 static void
 919 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
 920                        unsigned sampler_unit,
 921                        unsigned img_filter,
 922                        unsigned mip_filter,
 923                        LLVMValueRef s,
 924                        LLVMValueRef t,
 925                        LLVMValueRef r,
 926                        const LLVMValueRef *offsets,
 927                        LLVMValueRef ilevel0,
 928                        LLVMValueRef ilevel1,
 929                        LLVMValueRef lod_fpart,
 930                        LLVMValueRef *colors_out)
 931 {
 932    LLVMBuilderRef builder = bld->gallivm->builder;
 933    LLVMValueRef size0 = NULL;
 934    LLVMValueRef size1 = NULL;
 935    LLVMValueRef row_stride0_vec = NULL;
 936    LLVMValueRef row_stride1_vec = NULL;
 937    LLVMValueRef img_stride0_vec = NULL;
 938    LLVMValueRef img_stride1_vec = NULL;
 939    LLVMValueRef data_ptr0 = NULL;
 940    LLVMValueRef data_ptr1 = NULL;
 941    LLVMValueRef mipoff0 = NULL;
 942    LLVMValueRef mipoff1 = NULL;
 943    LLVMValueRef colors0[4], colors1[4];
 944    unsigned chan;
 945
 946    /* sample the first mipmap level */
 947    lp_build_mipmap_level_sizes(bld, ilevel0,
 948                                &size0,
 949                                &row_stride0_vec, &img_stride0_vec);
 950    if (bld->num_lods == 1) {
 951       data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
 952    }
 953    else {
 954       /* This path should work for num_lods 1 too but slightly less efficient */
 955       data_ptr0 = bld->base_ptr;
 956       mipoff0 = lp_build_get_mip_offsets(bld, ilevel0);
 957    }
 958    if (img_filter == PIPE_TEX_FILTER_NEAREST) {
 959       lp_build_sample_image_nearest(bld, sampler_unit,
 960                                     size0,
 961                                     row_stride0_vec, img_stride0_vec,
 962                                     data_ptr0, mipoff0, s, t, r, offsets,
 963                                     colors0);
 964    }
 965    else {
 966       assert(img_filter == PIPE_TEX_FILTER_LINEAR);
 967       lp_build_sample_image_linear(bld, sampler_unit,
 968                                    size0,
 969                                    row_stride0_vec, img_stride0_vec,
 970                                    data_ptr0, mipoff0, s, t, r, offsets,
 971                                    colors0);
 972    }
 973
 974    /* Store the first level's colors in the output variables */
 975    for (chan = 0; chan < 4; chan++) {
 976        LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
 977    }
 978
 979    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
 980       struct lp_build_if_state if_ctx;
 981       LLVMValueRef need_lerp;
 982
 983       /* need_lerp = lod_fpart > 0 */
 984       if (bld->num_lods == 1) {
 985          need_lerp = LLVMBuildFCmp(builder, LLVMRealUGT,
 986                                    lod_fpart, bld->levelf_bld.zero,
 987                                    "need_lerp");
 988       }
 989       else {
 990          /*
 991           * We'll do mip filtering if any of the quads (or individual
 992           * pixel in case of per-pixel lod) need it.
 993           * It might be better to split the vectors here and only fetch/filter
 994           * quads which need it.
 995           */
 996          /*
 997           * We unfortunately need to clamp lod_fpart here since we can get
 998           * negative values which would screw up filtering if not all
 999           * lod_fpart values have same sign.
1000           */
1001          lod_fpart = lp_build_max(&bld->levelf_bld, lod_fpart,
1002                                   bld->levelf_bld.zero);
1003          need_lerp = lp_build_compare(bld->gallivm, bld->levelf_bld.type,
1004                                       PIPE_FUNC_GREATER,
1005                                       lod_fpart, bld->levelf_bld.zero);
1006          need_lerp = lp_build_any_true_range(&bld->leveli_bld, bld->num_lods, need_lerp);
1007       }
1008
1009       lp_build_if(&if_ctx, bld->gallivm, need_lerp);
1010       {
1011          /* sample the second mipmap level */
1012          lp_build_mipmap_level_sizes(bld, ilevel1,
1013                                      &size1,
1014                                      &row_stride1_vec, &img_stride1_vec);
1015          if (bld->num_lods == 1) {
1016             data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
1017          }
1018          else {
1019             data_ptr1 = bld->base_ptr;
1020             mipoff1 = lp_build_get_mip_offsets(bld, ilevel1);
1021          }
1022          if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1023             lp_build_sample_image_nearest(bld, sampler_unit,
1024                                           size1,
1025                                           row_stride1_vec, img_stride1_vec,
1026                                           data_ptr1, mipoff1, s, t, r, offsets,
1027                                           colors1);
1028          }
1029          else {
1030             lp_build_sample_image_linear(bld, sampler_unit,
1031                                          size1,
1032                                          row_stride1_vec, img_stride1_vec,
1033                                          data_ptr1, mipoff1, s, t, r, offsets,
1034                                          colors1);
1035          }
1036
1037          /* interpolate samples from the two mipmap levels */
1038
1039          if (bld->num_lods != bld->coord_type.length)
1040             lod_fpart = lp_build_unpack_broadcast_aos_scalars(bld->gallivm,
1041                                                               bld->levelf_bld.type,
1042                                                               bld->texel_bld.type,
1043                                                               lod_fpart);
1044
1045          for (chan = 0; chan < 4; chan++) {
1046             colors0[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1047                                           colors0[chan], colors1[chan],
1048                                           0);
1049             LLVMBuildStore(builder, colors0[chan], colors_out[chan]);
1050          }
1051       }
1052       lp_build_endif(&if_ctx);
1053    }
1054 }
1055
1056
1057 /**
1058  * Clamp layer coord to valid values.
1059  */
1060 static LLVMValueRef
1061 lp_build_layer_coord(struct lp_build_sample_context *bld,
1062                      unsigned texture_unit,
1063                      LLVMValueRef layer)
1064 {
1065    LLVMValueRef maxlayer;
1066
1067    maxlayer = bld->dynamic_state->depth(bld->dynamic_state,
1068                                         bld->gallivm, texture_unit);
1069    maxlayer = lp_build_sub(&bld->int_bld, maxlayer, bld->int_bld.one);
1070    maxlayer = lp_build_broadcast_scalar(&bld->int_coord_bld, maxlayer);
1071    return lp_build_clamp(&bld->int_coord_bld, layer,
1072                          bld->int_coord_bld.zero, maxlayer);
1073
1074 }
1075
1076
1077 /**
1078  * Calculate cube face, lod, mip levels.
1079  */
1080 static void
1081 lp_build_sample_common(struct lp_build_sample_context *bld,
1082                        unsigned texture_index,
1083                        unsigned sampler_index,
1084                        LLVMValueRef *s,
1085                        LLVMValueRef *t,
1086                        LLVMValueRef *r,
1087                        const struct lp_derivatives *derivs, /* optional */
1088                        LLVMValueRef lod_bias, /* optional */
1089                        LLVMValueRef explicit_lod, /* optional */
1090                        LLVMValueRef *lod_ipart,
1091                        LLVMValueRef *lod_fpart,
1092                        LLVMValueRef *ilevel0,
1093                        LLVMValueRef *ilevel1)
1094 {
1095    const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1096    const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1097    const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1098    const unsigned target = bld->static_texture_state->target;
1099    LLVMValueRef first_level, cube_rho = NULL;
1100
1101    /*
1102    printf("%s mip %d  min %d  mag %d\n", __FUNCTION__,
1103           mip_filter, min_filter, mag_filter);
1104    */
1105
1106    /*
1107     * Choose cube face, recompute texcoords for the chosen face and
1108     * compute rho here too (as it requires transform of derivatives).
1109     */
1110    if (target == PIPE_TEXTURE_CUBE) {
1111       LLVMValueRef face, face_s, face_t;
1112       boolean need_derivs;
1113       need_derivs = ((min_filter != mag_filter ||
1114                       mip_filter != PIPE_TEX_MIPFILTER_NONE) &&
1115                       !bld->static_sampler_state->min_max_lod_equal &&
1116                       !explicit_lod);
1117       lp_build_cube_lookup(bld, *s, *t, *r, derivs, &face, &face_s, &face_t,
1118                            &cube_rho, need_derivs);
1119       *s = face_s; /* vec */
1120       *t = face_t; /* vec */
1121       /* use 'r' to indicate cube face */
1122       *r = face; /* vec */
1123    }
1124    else if (target == PIPE_TEXTURE_1D_ARRAY) {
1125       *r = lp_build_iround(&bld->coord_bld, *t);
1126       *r = lp_build_layer_coord(bld, texture_index, *r);
1127    }
1128    else if (target == PIPE_TEXTURE_2D_ARRAY) {
1129       *r = lp_build_iround(&bld->coord_bld, *r);
1130       *r = lp_build_layer_coord(bld, texture_index, *r);
1131    }
1132
1133    /*
1134     * Compute the level of detail (float).
1135     */
1136    if (min_filter != mag_filter ||
1137        mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1138       /* Need to compute lod either to choose mipmap levels or to
1139        * distinguish between minification/magnification with one mipmap level.
1140        */
1141       lp_build_lod_selector(bld, texture_index, sampler_index,
1142                             *s, *t, *r, cube_rho,
1143                             derivs, lod_bias, explicit_lod,
1144                             mip_filter,
1145                             lod_ipart, lod_fpart);
1146    } else {
1147       *lod_ipart = bld->leveli_bld.zero;
1148    }
1149
1150    /*
1151     * Compute integer mipmap level(s) to fetch texels from: ilevel0, ilevel1
1152     */
1153    switch (mip_filter) {
1154    default:
1155       assert(0 && "bad mip_filter value in lp_build_sample_soa()");
1156       /* fall-through */
1157    case PIPE_TEX_MIPFILTER_NONE:
1158       /* always use mip level 0 */
1159       if (HAVE_LLVM == 0x0207 && target == PIPE_TEXTURE_CUBE) {
1160          /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1161           * We should be able to set ilevel0 = const(0) but that causes
1162           * bad x86 code to be emitted.
1163           */
1164          assert(*lod_ipart);
1165          lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1166       }
1167       else {
1168          first_level = bld->dynamic_state->first_level(bld->dynamic_state,
1169                                                        bld->gallivm, texture_index);
1170          first_level = lp_build_broadcast_scalar(&bld->leveli_bld, first_level);
1171          *ilevel0 = first_level;
1172       }
1173       break;
1174    case PIPE_TEX_MIPFILTER_NEAREST:
1175       assert(*lod_ipart);
1176       lp_build_nearest_mip_level(bld, texture_index, *lod_ipart, ilevel0);
1177       break;
1178    case PIPE_TEX_MIPFILTER_LINEAR:
1179       assert(*lod_ipart);
1180       assert(*lod_fpart);
1181       lp_build_linear_mip_levels(bld, texture_index,
1182                                  *lod_ipart, lod_fpart,
1183                                  ilevel0, ilevel1);
1184       break;
1185    }
1186 }
1187
1188 /**
1189  * General texture sampling codegen.
1190  * This function handles texture sampling for all texture targets (1D,
1191  * 2D, 3D, cube) and all filtering modes.
1192  */
1193 static void
1194 lp_build_sample_general(struct lp_build_sample_context *bld,
1195                         unsigned sampler_unit,
1196                         LLVMValueRef s,
1197                         LLVMValueRef t,
1198                         LLVMValueRef r,
1199                         const LLVMValueRef *offsets,
1200                         LLVMValueRef lod_ipart,
1201                         LLVMValueRef lod_fpart,
1202                         LLVMValueRef ilevel0,
1203                         LLVMValueRef ilevel1,
1204                         LLVMValueRef *colors_out)
1205 {
1206    struct lp_build_context *int_bld = &bld->int_bld;
1207    LLVMBuilderRef builder = bld->gallivm->builder;
1208    const unsigned mip_filter = bld->static_sampler_state->min_mip_filter;
1209    const unsigned min_filter = bld->static_sampler_state->min_img_filter;
1210    const unsigned mag_filter = bld->static_sampler_state->mag_img_filter;
1211    LLVMValueRef texels[4];
1212    unsigned chan;
1213
1214    /*
1215     * Get/interpolate texture colors.
1216     */
1217
1218    for (chan = 0; chan < 4; ++chan) {
1219      texels[chan] = lp_build_alloca(bld->gallivm, bld->texel_bld.vec_type, "");
1220      lp_build_name(texels[chan], "sampler%u_texel_%c_var", sampler_unit, "xyzw"[chan]);
1221    }
1222
1223    if (min_filter == mag_filter) {
1224       /* no need to distinguish between minification and magnification */
1225       lp_build_sample_mipmap(bld, sampler_unit,
1226                              min_filter, mip_filter,
1227                              s, t, r, offsets,
1228                              ilevel0, ilevel1, lod_fpart,
1229                              texels);
1230    }
1231    else {
1232       /* Emit conditional to choose min image filter or mag image filter
1233        * depending on the lod being > 0 or <= 0, respectively.
1234        */
1235       struct lp_build_if_state if_ctx;
1236       LLVMValueRef minify;
1237
1238       /*
1239        * XXX this should to all lods into account, if some are min
1240        * some max probably could hack up the coords/weights in the linear
1241        * path with selects to work for nearest.
1242        * If that's just two quads sitting next to each other it seems
1243        * quite ok to do the same filtering method on both though, at
1244        * least unless we have explicit lod (and who uses different
1245        * min/mag filter with that?)
1246        */
1247       if (bld->num_lods > 1)
1248          lod_ipart = LLVMBuildExtractElement(builder, lod_ipart,
1249                                              lp_build_const_int32(bld->gallivm, 0), "");
1250
1251       /* minify = lod >= 0.0 */
1252       minify = LLVMBuildICmp(builder, LLVMIntSGE,
1253                              lod_ipart, int_bld->zero, "");
1254
1255       lp_build_if(&if_ctx, bld->gallivm, minify);
1256       {
1257          /* Use the minification filter */
1258          lp_build_sample_mipmap(bld, sampler_unit,
1259                                 min_filter, mip_filter,
1260                                 s, t, r, offsets,
1261                                 ilevel0, ilevel1, lod_fpart,
1262                                 texels);
1263       }
1264       lp_build_else(&if_ctx);
1265       {
1266          /* Use the magnification filter */
1267          lp_build_sample_mipmap(bld, sampler_unit,
1268                                 mag_filter, PIPE_TEX_MIPFILTER_NONE,
1269                                 s, t, r, offsets,
1270                                 ilevel0, NULL, NULL,
1271                                 texels);
1272       }
1273       lp_build_endif(&if_ctx);
1274    }
1275
1276    for (chan = 0; chan < 4; ++chan) {
1277      colors_out[chan] = LLVMBuildLoad(builder, texels[chan], "");
1278      lp_build_name(colors_out[chan], "sampler%u_texel_%c", sampler_unit, "xyzw"[chan]);
1279    }
1280 }
1281
1282
1283 /**
1284  * Texel fetch function.
1285  * In contrast to general sampling there is no filtering, no coord minification,
1286  * lod (if any) is always explicit uint, coords are uints (in terms of texel units)
1287  * directly to be applied to the selected mip level (after adding texel offsets).
1288  * This function handles texel fetch for all targets where texel fetch is supported
1289  * (no cube maps, but 1d, 2d, 3d are supported, arrays and buffers should be too).
1290  */
1291 static void
1292 lp_build_fetch_texel(struct lp_build_sample_context *bld,
1293                      unsigned texture_unit,
1294                      const LLVMValueRef *coords,
1295                      LLVMValueRef explicit_lod,
1296                      const LLVMValueRef *offsets,
1297                      LLVMValueRef *colors_out)
1298 {
1299    struct lp_build_context *perquadi_bld = &bld->leveli_bld;
1300    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
1301    unsigned dims = bld->dims, chan;
1302    unsigned target = bld->static_texture_state->target;
1303    LLVMValueRef size, ilevel;
1304    LLVMValueRef row_stride_vec = NULL, img_stride_vec = NULL;
1305    LLVMValueRef x = coords[0], y = coords[1], z = coords[2];
1306    LLVMValueRef width, height, depth, i, j;
1307    LLVMValueRef offset, out_of_bounds, out1;
1308
1309    if (explicit_lod && bld->static_texture_state->target != PIPE_BUFFER) {
1310       if (bld->num_lods != int_coord_bld->type.length) {
1311          ilevel = lp_build_pack_aos_scalars(bld->gallivm, int_coord_bld->type,
1312                                             perquadi_bld->type, explicit_lod, 0);
1313       }
1314       else {
1315          ilevel = explicit_lod;
1316       }
1317       lp_build_nearest_mip_level(bld, texture_unit, ilevel, &ilevel);
1318    }
1319    else {
1320       bld->num_lods = 1;
1321       ilevel = lp_build_const_int32(bld->gallivm, 0);
1322    }
1323    lp_build_mipmap_level_sizes(bld, ilevel,
1324                                &size,
1325                                &row_stride_vec, &img_stride_vec);
1326    lp_build_extract_image_sizes(bld, &bld->int_size_bld, int_coord_bld->type,
1327                                 size, &width, &height, &depth);
1328
1329    if (target == PIPE_TEXTURE_1D_ARRAY ||
1330        target == PIPE_TEXTURE_2D_ARRAY) {
1331       if (target == PIPE_TEXTURE_1D_ARRAY) {
1332          z = lp_build_layer_coord(bld, texture_unit, y);
1333       }
1334       else {
1335          z = lp_build_layer_coord(bld, texture_unit, z);
1336       }
1337    }
1338
1339    /* This is a lot like border sampling */
1340    if (offsets[0]) {
1341       /* XXX coords are really unsigned, offsets are signed */
1342       x = lp_build_add(int_coord_bld, x, offsets[0]);
1343    }
1344    out_of_bounds = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
1345    out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
1346    out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1347
1348    if (dims >= 2) {
1349       if (offsets[1]) {
1350          y = lp_build_add(int_coord_bld, y, offsets[1]);
1351       }
1352       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
1353       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1354       out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
1355       out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1356
1357       if (dims >= 3) {
1358          if (offsets[2]) {
1359             z = lp_build_add(int_coord_bld, z, offsets[2]);
1360          }
1361          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
1362          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1363          out1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
1364          out_of_bounds = lp_build_or(int_coord_bld, out_of_bounds, out1);
1365       }
1366    }
1367
1368    lp_build_sample_offset(int_coord_bld,
1369                           bld->format_desc,
1370                           x, y, z, row_stride_vec, img_stride_vec,
1371                           &offset, &i, &j);
1372
1373    if (bld->static_texture_state->target != PIPE_BUFFER) {
1374       offset = lp_build_add(int_coord_bld, offset,
1375                             lp_build_get_mip_offsets(bld, ilevel));
1376    }
1377
1378    offset = lp_build_andnot(int_coord_bld, offset, out_of_bounds);
1379
1380    lp_build_fetch_rgba_soa(bld->gallivm,
1381                            bld->format_desc,
1382                            bld->texel_type,
1383                            bld->base_ptr, offset,
1384                            i, j,
1385                            colors_out);
1386
1387    if (0) {
1388       /*
1389        * Not needed except for ARB_robust_buffer_access_behavior.
1390        * Could use min/max above instead of out-of-bounds comparisons
1391        * (in fact cast to unsigned and min only is sufficient)
1392        * if we don't care about the result returned for out-of-bounds.
1393        */
1394       for (chan = 0; chan < 4; chan++) {
1395          colors_out[chan] = lp_build_select(&bld->texel_bld, out_of_bounds,
1396                                             bld->texel_bld.zero, colors_out[chan]);
1397       }
1398    }
1399 }
1400
1401
1402 /**
1403  * Do shadow test/comparison.
1404  * \param coords  incoming texcoords
1405  * \param texel  the texel to compare against (use the X channel)
1406  * Ideally this should really be done per-sample.
1407  */
1408 static void
1409 lp_build_sample_compare(struct lp_build_sample_context *bld,
1410                         const LLVMValueRef *coords,
1411                         LLVMValueRef texel[4])
1412 {
1413    struct lp_build_context *texel_bld = &bld->texel_bld;
1414    LLVMBuilderRef builder = bld->gallivm->builder;
1415    LLVMValueRef res, p;
1416    const unsigned chan = 0;
1417
1418    if (bld->static_sampler_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1419       return;
1420
1421    if (bld->static_texture_state->target == PIPE_TEXTURE_2D_ARRAY ||
1422        bld->static_texture_state->target == PIPE_TEXTURE_CUBE) {
1423       p = coords[3];
1424    }
1425    else {
1426       p = coords[2];
1427    }
1428
1429    /* debug code */
1430    if (0) {
1431       LLVMValueRef indx = lp_build_const_int32(bld->gallivm, 0);
1432       LLVMValueRef coord = LLVMBuildExtractElement(builder, p, indx, "");
1433       LLVMValueRef tex = LLVMBuildExtractElement(builder, texel[chan], indx, "");
1434       lp_build_printf(bld->gallivm, "shadow compare coord %f to texture %f\n",
1435                       coord, tex);
1436    }
1437
1438    /* Clamp p coords to [0,1] */
1439    p = lp_build_clamp(&bld->coord_bld, p,
1440                       bld->coord_bld.zero,
1441                       bld->coord_bld.one);
1442
1443    /* result = (p FUNC texel) ? 1 : 0 */
1444    res = lp_build_cmp(texel_bld, bld->static_sampler_state->compare_func,
1445                       p, texel[chan]);
1446    res = lp_build_select(texel_bld, res, texel_bld->one, texel_bld->zero);
1447
1448    /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1449    texel[0] =
1450    texel[1] =
1451    texel[2] = res;
1452    texel[3] = texel_bld->one;
1453 }
1454
1455
1456 /**
1457  * Just set texels to white instead of actually sampling the texture.
1458  * For debugging.
1459  */
1460 void
1461 lp_build_sample_nop(struct gallivm_state *gallivm,
1462                     struct lp_type type,
1463                     const LLVMValueRef *coords,
1464                     LLVMValueRef texel_out[4])
1465 {
1466    LLVMValueRef one = lp_build_one(gallivm, type);
1467    unsigned chan;
1468
1469    for (chan = 0; chan < 4; chan++) {
1470       texel_out[chan] = one;
1471    }
1472 }
1473
1474
1475 /**
1476  * Build texture sampling code.
1477  * 'texel' will return a vector of four LLVMValueRefs corresponding to
1478  * R, G, B, A.
1479  * \param type  vector float type to use for coords, etc.
1480  * \param is_fetch  if this is a texel fetch instruction.
1481  * \param derivs  partial derivatives of (s,t,r,q) with respect to x and y
1482  */
1483 void
1484 lp_build_sample_soa(struct gallivm_state *gallivm,
1485                     const struct lp_static_texture_state *static_texture_state,
1486                     const struct lp_static_sampler_state *static_sampler_state,
1487                     struct lp_sampler_dynamic_state *dynamic_state,
1488                     struct lp_type type,
1489                     boolean is_fetch,
1490                     unsigned texture_index,
1491                     unsigned sampler_index,
1492                     const LLVMValueRef *coords,
1493                     const LLVMValueRef *offsets,
1494                     const struct lp_derivatives *derivs, /* optional */
1495                     LLVMValueRef lod_bias, /* optional */
1496                     LLVMValueRef explicit_lod, /* optional */
1497                     boolean scalar_lod,
1498                     LLVMValueRef texel_out[4])
1499 {
1500    unsigned dims = texture_dims(static_texture_state->target);
1501    unsigned num_quads = type.length / 4;
1502    unsigned mip_filter;
1503    struct lp_build_sample_context bld;
1504    struct lp_static_sampler_state derived_sampler_state = *static_sampler_state;
1505    LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
1506    LLVMBuilderRef builder = gallivm->builder;
1507    LLVMValueRef tex_width;
1508    LLVMValueRef s;
1509    LLVMValueRef t;
1510    LLVMValueRef r;
1511
1512    if (0) {
1513       enum pipe_format fmt = static_texture_state->format;
1514       debug_printf("Sample from %s\n", util_format_name(fmt));
1515    }
1516
1517    assert(type.floating);
1518
1519    /* Setup our build context */
1520    memset(&bld, 0, sizeof bld);
1521    bld.gallivm = gallivm;
1522    bld.static_sampler_state = &derived_sampler_state;
1523    bld.static_texture_state = static_texture_state;
1524    bld.dynamic_state = dynamic_state;
1525    bld.format_desc = util_format_description(static_texture_state->format);
1526    bld.dims = dims;
1527
1528    bld.vector_width = lp_type_width(type);
1529
1530    bld.float_type = lp_type_float(32);
1531    bld.int_type = lp_type_int(32);
1532    bld.coord_type = type;
1533    bld.int_coord_type = lp_int_type(type);
1534    bld.float_size_in_type = lp_type_float(32);
1535    bld.float_size_in_type.length = dims > 1 ? 4 : 1;
1536    bld.int_size_in_type = lp_int_type(bld.float_size_in_type);
1537    bld.texel_type = type;
1538
1539    /* always using the first channel hopefully should be safe,
1540     * if not things WILL break in other places anyway.
1541     */
1542    if (bld.format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
1543        bld.format_desc->channel[0].pure_integer) {
1544       if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
1545          bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1546       }
1547       else if (bld.format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
1548          bld.texel_type = lp_type_uint_vec(type.width, type.width * type.length);
1549       }
1550    }
1551    else if (util_format_has_stencil(bld.format_desc) &&
1552        !util_format_has_depth(bld.format_desc)) {
1553       /* for stencil only formats, sample stencil (uint) */
1554       bld.texel_type = lp_type_int_vec(type.width, type.width * type.length);
1555    }
1556
1557    if (!static_texture_state->level_zero_only) {
1558       derived_sampler_state.min_mip_filter = static_sampler_state->min_mip_filter;
1559    } else {
1560       derived_sampler_state.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
1561    }
1562    mip_filter = derived_sampler_state.min_mip_filter;
1563
1564    if (0) {
1565       debug_printf("  .min_mip_filter = %u\n", derived_sampler_state.min_mip_filter);
1566    }
1567
1568    /*
1569     * This is all a bit complicated different paths are chosen for performance
1570     * reasons.
1571     * Essentially, there can be 1 lod per element, 1 lod per quad or 1 lod for
1572     * everything (the last two options are equivalent for 4-wide case).
1573     * If there's per-quad lod but we split to 4-wide so we can use AoS, per-quad
1574     * lod is calculated then the lod value extracted afterwards so making this
1575     * case basically the same as far as lod handling is concerned for the
1576     * further sample/filter code as the 1 lod for everything case.
1577     * Different lod handling mostly shows up when building mipmap sizes
1578     * (lp_build_mipmap_level_sizes() and friends) and also in filtering
1579     * (getting the fractional part of the lod to the right texels).
1580     */
1581
1582    /*
1583     * There are other situations where at least the multiple int lods could be
1584     * avoided like min and max lod being equal.
1585     */
1586    if (explicit_lod && !scalar_lod &&
1587        ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) ||
1588         (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
1589       bld.num_lods = type.length;
1590    /* TODO: for true scalar_lod should only use 1 lod value */
1591    else if ((is_fetch && explicit_lod && bld.static_texture_state->target != PIPE_BUFFER ) ||
1592             (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)) {
1593       bld.num_lods = num_quads;
1594    }
1595    else {
1596       bld.num_lods = 1;
1597    }
1598
1599    bld.levelf_type = type;
1600    /* we want native vector size to be able to use our intrinsics */
1601    if (bld.num_lods != type.length) {
1602       bld.levelf_type.length = type.length > 4 ? ((type.length + 15) / 16) * 4 : 1;
1603    }
1604    bld.leveli_type = lp_int_type(bld.levelf_type);
1605    bld.float_size_type = bld.float_size_in_type;
1606    /* Note: size vectors may not be native. They contain minified w/h/d/_ values,
1607     * with per-element lod that is w0/h0/d0/_/w1/h1/d1_/... so up to 8x4f32 */
1608    if (bld.num_lods > 1) {
1609       bld.float_size_type.length = bld.num_lods == type.length ?
1610                                       bld.num_lods * bld.float_size_in_type.length :
1611                                       type.length;
1612    }
1613    bld.int_size_type = lp_int_type(bld.float_size_type);
1614
1615    lp_build_context_init(&bld.float_bld, gallivm, bld.float_type);
1616    lp_build_context_init(&bld.float_vec_bld, gallivm, type);
1617    lp_build_context_init(&bld.int_bld, gallivm, bld.int_type);
1618    lp_build_context_init(&bld.coord_bld, gallivm, bld.coord_type);
1619    lp_build_context_init(&bld.int_coord_bld, gallivm, bld.int_coord_type);
1620    lp_build_context_init(&bld.int_size_in_bld, gallivm, bld.int_size_in_type);
1621    lp_build_context_init(&bld.float_size_in_bld, gallivm, bld.float_size_in_type);
1622    lp_build_context_init(&bld.int_size_bld, gallivm, bld.int_size_type);
1623    lp_build_context_init(&bld.float_size_bld, gallivm, bld.float_size_type);
1624    lp_build_context_init(&bld.texel_bld, gallivm, bld.texel_type);
1625    lp_build_context_init(&bld.levelf_bld, gallivm, bld.levelf_type);
1626    lp_build_context_init(&bld.leveli_bld, gallivm, bld.leveli_type);
1627
1628    /* Get the dynamic state */
1629    tex_width = dynamic_state->width(dynamic_state, gallivm, texture_index);
1630    bld.row_stride_array = dynamic_state->row_stride(dynamic_state, gallivm, texture_index);
1631    bld.img_stride_array = dynamic_state->img_stride(dynamic_state, gallivm, texture_index);
1632    bld.base_ptr = dynamic_state->base_ptr(dynamic_state, gallivm, texture_index);
1633    bld.mip_offsets = dynamic_state->mip_offsets(dynamic_state, gallivm, texture_index);
1634    /* Note that mip_offsets is an array[level] of offsets to texture images */
1635
1636    s = coords[0];
1637    t = coords[1];
1638    r = coords[2];
1639
1640    /* width, height, depth as single int vector */
1641    if (dims <= 1) {
1642       bld.int_size = tex_width;
1643    }
1644    else {
1645       bld.int_size = LLVMBuildInsertElement(builder, bld.int_size_in_bld.undef,
1646                                             tex_width, LLVMConstInt(i32t, 0, 0), "");
1647       if (dims >= 2) {
1648          LLVMValueRef tex_height =
1649             dynamic_state->height(dynamic_state, gallivm, texture_index);
1650          bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1651                                                tex_height, LLVMConstInt(i32t, 1, 0), "");
1652          if (dims >= 3) {
1653             LLVMValueRef tex_depth =
1654                dynamic_state->depth(dynamic_state, gallivm, texture_index);
1655             bld.int_size = LLVMBuildInsertElement(builder, bld.int_size,
1656                                                   tex_depth, LLVMConstInt(i32t, 2, 0), "");
1657          }
1658       }
1659    }
1660
1661    if (0) {
1662       /* For debug: no-op texture sampling */
1663       lp_build_sample_nop(gallivm,
1664                           bld.texel_type,
1665                           coords,
1666                           texel_out);
1667    }
1668
1669    else if (is_fetch) {
1670       lp_build_fetch_texel(&bld, texture_index, coords,
1671                            explicit_lod, offsets,
1672                            texel_out);
1673    }
1674
1675    else {
1676       LLVMValueRef lod_ipart = NULL, lod_fpart = NULL;
1677       LLVMValueRef ilevel0 = NULL, ilevel1 = NULL;
1678       boolean use_aos = util_format_fits_8unorm(bld.format_desc) &&
1679                         lp_is_simple_wrap_mode(static_sampler_state->wrap_s) &&
1680                         lp_is_simple_wrap_mode(static_sampler_state->wrap_t);
1681
1682       if ((gallivm_debug & GALLIVM_DEBUG_PERF) &&
1683           !use_aos && util_format_fits_8unorm(bld.format_desc)) {
1684          debug_printf("%s: using floating point linear filtering for %s\n",
1685                       __FUNCTION__, bld.format_desc->short_name);
1686          debug_printf("  min_img %d  mag_img %d  mip %d  wraps %d  wrapt %d\n",
1687                       static_sampler_state->min_img_filter,
1688                       static_sampler_state->mag_img_filter,
1689                       static_sampler_state->min_mip_filter,
1690                       static_sampler_state->wrap_s,
1691                       static_sampler_state->wrap_t);
1692       }
1693
1694       lp_build_sample_common(&bld, texture_index, sampler_index,
1695                              &s, &t, &r,
1696                              derivs, lod_bias, explicit_lod,
1697                              &lod_ipart, &lod_fpart,
1698                              &ilevel0, &ilevel1);
1699
1700       /*
1701        * we only try 8-wide sampling with soa as it appears to
1702        * be a loss with aos with AVX (but it should work).
1703        * (It should be faster if we'd support avx2)
1704        */
1705       if (num_quads == 1 || !use_aos) {
1706
1707          if (num_quads > 1) {
1708             if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1709                LLVMValueRef index0 = lp_build_const_int32(gallivm, 0);
1710                /*
1711                 * These parameters are the same for all quads,
1712                 * could probably simplify.
1713                 */
1714                lod_ipart = LLVMBuildExtractElement(builder, lod_ipart, index0, "");
1715                ilevel0 = LLVMBuildExtractElement(builder, ilevel0, index0, "");
1716             }
1717          }
1718          if (use_aos) {
1719             /* do sampling/filtering with fixed pt arithmetic */
1720             lp_build_sample_aos(&bld, sampler_index,
1721                                 s, t, r, offsets,
1722                                 lod_ipart, lod_fpart,
1723                                 ilevel0, ilevel1,
1724                                 texel_out);
1725          }
1726
1727          else {
1728             lp_build_sample_general(&bld, sampler_index,
1729                                     s, t, r, offsets,
1730                                     lod_ipart, lod_fpart,
1731                                     ilevel0, ilevel1,
1732                                     texel_out);
1733          }
1734       }
1735       else {
1736          unsigned j;
1737          struct lp_build_sample_context bld4;
1738          struct lp_type type4 = type;
1739          unsigned i;
1740          LLVMValueRef texelout4[4];
1741          LLVMValueRef texelouttmp[4][LP_MAX_VECTOR_LENGTH/16];
1742
1743          type4.length = 4;
1744
1745          /* Setup our build context */
1746          memset(&bld4, 0, sizeof bld4);
1747          bld4.gallivm = bld.gallivm;
1748          bld4.static_texture_state = bld.static_texture_state;
1749          bld4.static_sampler_state = bld.static_sampler_state;
1750          bld4.dynamic_state = bld.dynamic_state;
1751          bld4.format_desc = bld.format_desc;
1752          bld4.dims = bld.dims;
1753          bld4.row_stride_array = bld.row_stride_array;
1754          bld4.img_stride_array = bld.img_stride_array;
1755          bld4.base_ptr = bld.base_ptr;
1756          bld4.mip_offsets = bld.mip_offsets;
1757          bld4.int_size = bld.int_size;
1758
1759          bld4.vector_width = lp_type_width(type4);
1760
1761          bld4.float_type = lp_type_float(32);
1762          bld4.int_type = lp_type_int(32);
1763          bld4.coord_type = type4;
1764          bld4.int_coord_type = lp_int_type(type4);
1765          bld4.float_size_in_type = lp_type_float(32);
1766          bld4.float_size_in_type.length = dims > 1 ? 4 : 1;
1767          bld4.int_size_in_type = lp_int_type(bld4.float_size_in_type);
1768          bld4.texel_type = bld.texel_type;
1769          bld4.texel_type.length = 4;
1770          bld4.levelf_type = type4;
1771          /* we want native vector size to be able to use our intrinsics */
1772          bld4.levelf_type.length = 1;
1773          bld4.leveli_type = lp_int_type(bld4.levelf_type);
1774
1775          if (explicit_lod && !scalar_lod &&
1776              ((is_fetch && bld.static_texture_state->target != PIPE_BUFFER) ||
1777               (!is_fetch && mip_filter != PIPE_TEX_MIPFILTER_NONE)))
1778             bld4.num_lods = type4.length;
1779          else
1780             bld4.num_lods = 1;
1781
1782          bld4.levelf_type = type4;
1783          /* we want native vector size to be able to use our intrinsics */
1784          if (bld4.num_lods != type4.length) {
1785             bld4.levelf_type.length = 1;
1786          }
1787          bld4.leveli_type = lp_int_type(bld4.levelf_type);
1788          bld4.float_size_type = bld4.float_size_in_type;
1789          if (bld4.num_lods > 1) {
1790             bld4.float_size_type.length = bld4.num_lods == type4.length ?
1791                                             bld4.num_lods * bld4.float_size_in_type.length :
1792                                             type4.length;
1793          }
1794          bld4.int_size_type = lp_int_type(bld4.float_size_type);
1795
1796          lp_build_context_init(&bld4.float_bld, gallivm, bld4.float_type);
1797          lp_build_context_init(&bld4.float_vec_bld, gallivm, type4);
1798          lp_build_context_init(&bld4.int_bld, gallivm, bld4.int_type);
1799          lp_build_context_init(&bld4.coord_bld, gallivm, bld4.coord_type);
1800          lp_build_context_init(&bld4.int_coord_bld, gallivm, bld4.int_coord_type);
1801          lp_build_context_init(&bld4.int_size_in_bld, gallivm, bld4.int_size_in_type);
1802          lp_build_context_init(&bld4.float_size_in_bld, gallivm, bld4.float_size_in_type);
1803          lp_build_context_init(&bld4.int_size_bld, gallivm, bld4.int_size_type);
1804          lp_build_context_init(&bld4.float_size_bld, gallivm, bld4.float_size_type);
1805          lp_build_context_init(&bld4.texel_bld, gallivm, bld4.texel_type);
1806          lp_build_context_init(&bld4.levelf_bld, gallivm, bld4.levelf_type);
1807          lp_build_context_init(&bld4.leveli_bld, gallivm, bld4.leveli_type);
1808
1809          for (i = 0; i < num_quads; i++) {
1810             LLVMValueRef s4, t4, r4;
1811             LLVMValueRef lod_ipart4, lod_fpart4 = NULL;
1812             LLVMValueRef ilevel04, ilevel14 = NULL;
1813             LLVMValueRef offsets4[4] = { NULL };
1814             unsigned num_lods = bld4.num_lods;
1815
1816             s4 = lp_build_extract_range(gallivm, s, 4*i, 4);
1817             t4 = lp_build_extract_range(gallivm, t, 4*i, 4);
1818             r4 = lp_build_extract_range(gallivm, r, 4*i, 4);
1819
1820             if (offsets[0]) {
1821                offsets4[0] = lp_build_extract_range(gallivm, offsets[0], 4*i, 4);
1822                if (dims > 1) {
1823                   offsets4[1] = lp_build_extract_range(gallivm, offsets[1], 4*i, 4);
1824                   if (dims > 2) {
1825                      offsets4[2] = lp_build_extract_range(gallivm, offsets[2], 4*i, 4);
1826                   }
1827                }
1828             }
1829             lod_ipart4 = lp_build_extract_range(gallivm, lod_ipart, num_lods * i, num_lods);
1830             ilevel04 = lp_build_extract_range(gallivm, ilevel0, num_lods * i, num_lods);
1831             if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1832                ilevel14 = lp_build_extract_range(gallivm, ilevel1, num_lods * i, num_lods);
1833                lod_fpart4 = lp_build_extract_range(gallivm, lod_fpart, num_lods * i, num_lods);
1834             }
1835
1836             if (use_aos) {
1837                /* do sampling/filtering with fixed pt arithmetic */
1838                lp_build_sample_aos(&bld4, sampler_index,
1839                                    s4, t4, r4, offsets4,
1840                                    lod_ipart4, lod_fpart4,
1841                                    ilevel04, ilevel14,
1842                                    texelout4);
1843             }
1844
1845             else {
1846                lp_build_sample_general(&bld4, sampler_index,
1847                                        s4, t4, r4, offsets4,
1848                                        lod_ipart4, lod_fpart4,
1849                                        ilevel04, ilevel14,
1850                                        texelout4);
1851             }
1852             for (j = 0; j < 4; j++) {
1853                texelouttmp[j][i] = texelout4[j];
1854             }
1855          }
1856
1857          for (j = 0; j < 4; j++) {
1858             texel_out[j] = lp_build_concat(gallivm, texelouttmp[j], type4, num_quads);
1859          }
1860       }
1861
1862       lp_build_sample_compare(&bld, coords, texel_out);
1863    }
1864
1865    if (static_texture_state->target != PIPE_BUFFER) {
1866       apply_sampler_swizzle(&bld, texel_out);
1867    }
1868
1869    /*
1870     * texel type can be a (32bit) int/uint (for pure int formats only),
1871     * however we are expected to always return floats (storage is untyped).
1872     */
1873    if (!bld.texel_type.floating) {
1874       unsigned chan;
1875       for (chan = 0; chan < 4; chan++) {
1876          texel_out[chan] = LLVMBuildBitCast(builder, texel_out[chan],
1877                                             lp_build_vec_type(gallivm, type), "");
1878       }
1879    }
1880 }
1881
1882 void
1883 lp_build_size_query_soa(struct gallivm_state *gallivm,
1884                         const struct lp_static_texture_state *static_state,
1885                         struct lp_sampler_dynamic_state *dynamic_state,
1886                         struct lp_type int_type,
1887                         unsigned texture_unit,
1888                         boolean need_nr_mips,
1889                         LLVMValueRef explicit_lod,
1890                         LLVMValueRef *sizes_out)
1891 {
1892    LLVMValueRef lod;
1893    LLVMValueRef size;
1894    LLVMValueRef first_level = NULL;
1895    int dims, i;
1896    boolean has_array;
1897    struct lp_build_context bld_int_vec;
1898
1899    dims = texture_dims(static_state->target);
1900
1901    switch (static_state->target) {
1902    case PIPE_TEXTURE_1D_ARRAY:
1903    case PIPE_TEXTURE_2D_ARRAY:
1904       has_array = TRUE;
1905       break;
1906    default:
1907       has_array = FALSE;
1908       break;
1909    }
1910
1911    assert(!int_type.floating);
1912
1913    lp_build_context_init(&bld_int_vec, gallivm, lp_type_int_vec(32, 128));
1914
1915    if (explicit_lod) {
1916       /* FIXME: this needs to honor per-element lod */
1917       lod = LLVMBuildExtractElement(gallivm->builder, explicit_lod, lp_build_const_int32(gallivm, 0), "");
1918       first_level = dynamic_state->first_level(dynamic_state, gallivm, texture_unit);
1919       lod = lp_build_broadcast_scalar(&bld_int_vec,
1920                                       LLVMBuildAdd(gallivm->builder, lod, first_level, "lod"));
1921
1922    } else {
1923       lod = bld_int_vec.zero;
1924    }
1925
1926    if (need_nr_mips) {
1927       size = bld_int_vec.zero;
1928    }
1929    else {
1930       size = bld_int_vec.undef;
1931    }
1932
1933    size = LLVMBuildInsertElement(gallivm->builder, size,
1934                                  dynamic_state->width(dynamic_state, gallivm, texture_unit),
1935                                  lp_build_const_int32(gallivm, 0), "");
1936
1937    if (dims >= 2) {
1938       size = LLVMBuildInsertElement(gallivm->builder, size,
1939                                     dynamic_state->height(dynamic_state, gallivm, texture_unit),
1940                                     lp_build_const_int32(gallivm, 1), "");
1941    }
1942
1943    if (dims >= 3) {
1944       size = LLVMBuildInsertElement(gallivm->builder, size,
1945                                     dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1946                                     lp_build_const_int32(gallivm, 2), "");
1947    }
1948
1949    size = lp_build_minify(&bld_int_vec, size, lod);
1950
1951    if (has_array)
1952       size = LLVMBuildInsertElement(gallivm->builder, size,
1953                                     dynamic_state->depth(dynamic_state, gallivm, texture_unit),
1954                                     lp_build_const_int32(gallivm, dims), "");
1955
1956    /*
1957     * XXX for out-of-bounds lod, should set size to zero vector here
1958     * (for dx10-style only, i.e. need_nr_mips)
1959     */
1960
1961    for (i = 0; i < dims + (has_array ? 1 : 0); i++) {
1962       sizes_out[i] = lp_build_extract_broadcast(gallivm, bld_int_vec.type, int_type,
1963                                                 size,
1964                                                 lp_build_const_int32(gallivm, i));
1965    }
1966
1967    /*
1968     * if there's no explicit_lod (buffers, rects) queries requiring nr of
1969     * mips would be illegal.
1970     */
1971    if (need_nr_mips && explicit_lod) {
1972       struct lp_build_context bld_int_scalar;
1973       LLVMValueRef num_levels;
1974       lp_build_context_init(&bld_int_scalar, gallivm, lp_type_int(32));
1975
1976       if (static_state->level_zero_only) {
1977          num_levels = bld_int_scalar.one;
1978       }
1979       else {
1980          LLVMValueRef last_level;
1981
1982          last_level = dynamic_state->last_level(dynamic_state, gallivm, texture_unit);
1983          num_levels = lp_build_sub(&bld_int_scalar, last_level, first_level);
1984          num_levels = lp_build_add(&bld_int_scalar, num_levels, bld_int_scalar.one);
1985       }
1986       sizes_out[3] = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, int_type),
1987                                         num_levels);
1988    }
1989 }