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 "util/u_debug.h"
  39 #include "util/u_dump.h"
  40 #include "util/u_memory.h"
  41 #include "util/u_math.h"
  42 #include "util/u_format.h"
  43 #include "util/u_cpu_detect.h"
  44 #include "lp_bld_debug.h"
  45 #include "lp_bld_type.h"
  46 #include "lp_bld_const.h"
  47 #include "lp_bld_conv.h"
  48 #include "lp_bld_arit.h"
  49 #include "lp_bld_logic.h"
  50 #include "lp_bld_swizzle.h"
  51 #include "lp_bld_pack.h"
  52 #include "lp_bld_flow.h"
  53 #include "lp_bld_gather.h"
  54 #include "lp_bld_format.h"
  55 #include "lp_bld_sample.h"
  56 #include "lp_bld_quad.h"
  57
  58
  59 /**
  60  * Keep all information for sampling code generation in a single place.
  61  */
  62 struct lp_build_sample_context
  63 {
  64    LLVMBuilderRef builder;
  65
  66    const struct lp_sampler_static_state *static_state;
  67
  68    struct lp_sampler_dynamic_state *dynamic_state;
  69
  70    const struct util_format_description *format_desc;
  71
  72    /** regular scalar float type */
  73    struct lp_type float_type;
  74    struct lp_build_context float_bld;
  75
  76    /** regular scalar float type */
  77    struct lp_type int_type;
  78    struct lp_build_context int_bld;
  79
  80    /** Incoming coordinates type and build context */
  81    struct lp_type coord_type;
  82    struct lp_build_context coord_bld;
  83
  84    /** Unsigned integer coordinates */
  85    struct lp_type uint_coord_type;
  86    struct lp_build_context uint_coord_bld;
  87
  88    /** Signed integer coordinates */
  89    struct lp_type int_coord_type;
  90    struct lp_build_context int_coord_bld;
  91
  92    /** Output texels type and build context */
  93    struct lp_type texel_type;
  94    struct lp_build_context texel_bld;
  95 };
  96
  97
  98 /**
  99  * Does the given texture wrap mode allow sampling the texture border color?
 100  * XXX maybe move this into gallium util code.
 101  */
 102 static boolean
 103 wrap_mode_uses_border_color(unsigned mode)
 104 {
 105    switch (mode) {
 106    case PIPE_TEX_WRAP_REPEAT:
 107    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 108    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 109    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 110       return FALSE;
 111    case PIPE_TEX_WRAP_CLAMP:
 112    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 113    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 114    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 115       return TRUE;
 116    default:
 117       assert(0 && "unexpected wrap mode");
 118       return FALSE;
 119    }
 120 }
 121
 122
 123 static LLVMValueRef
 124 lp_build_get_mipmap_level(struct lp_build_sample_context *bld,
 125                           LLVMValueRef data_array, LLVMValueRef level)
 126 {
 127    LLVMValueRef indexes[2], data_ptr;
 128    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 129    indexes[1] = level;
 130    data_ptr = LLVMBuildGEP(bld->builder, data_array, indexes, 2, "");
 131    data_ptr = LLVMBuildLoad(bld->builder, data_ptr, "");
 132    return data_ptr;
 133 }
 134
 135
 136 static LLVMValueRef
 137 lp_build_get_const_mipmap_level(struct lp_build_sample_context *bld,
 138                                 LLVMValueRef data_array, int level)
 139 {
 140    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
 141    return lp_build_get_mipmap_level(bld, data_array, lvl);
 142 }
 143
 144
 145 /**
 146  * Dereference stride_array[mipmap_level] array to get a stride.
 147  * Return stride as a vector.
 148  */
 149 static LLVMValueRef
 150 lp_build_get_level_stride_vec(struct lp_build_sample_context *bld,
 151                               LLVMValueRef stride_array, LLVMValueRef level)
 152 {
 153    LLVMValueRef indexes[2], stride;
 154    indexes[0] = LLVMConstInt(LLVMInt32Type(), 0, 0);
 155    indexes[1] = level;
 156    stride = LLVMBuildGEP(bld->builder, stride_array, indexes, 2, "");
 157    stride = LLVMBuildLoad(bld->builder, stride, "");
 158    stride = lp_build_broadcast_scalar(&bld->int_coord_bld, stride);
 159    return stride;
 160 }
 161
 162
 163 /** Dereference stride_array[0] array to get a stride (as vector). */
 164 static LLVMValueRef
 165 lp_build_get_const_level_stride_vec(struct lp_build_sample_context *bld,
 166                                     LLVMValueRef stride_array, int level)
 167 {
 168    LLVMValueRef lvl = LLVMConstInt(LLVMInt32Type(), level, 0);
 169    return lp_build_get_level_stride_vec(bld, stride_array, lvl);
 170 }
 171
 172
 173 static int
 174 texture_dims(enum pipe_texture_target tex)
 175 {
 176    switch (tex) {
 177    case PIPE_TEXTURE_1D:
 178       return 1;
 179    case PIPE_TEXTURE_2D:
 180    case PIPE_TEXTURE_CUBE:
 181       return 2;
 182    case PIPE_TEXTURE_3D:
 183       return 3;
 184    default:
 185       assert(0 && "bad texture target in texture_dims()");
 186       return 2;
 187    }
 188 }
 189
 190
 191 static void
 192 apply_sampler_swizzle(struct lp_build_sample_context *bld,
 193                       LLVMValueRef *texel)
 194 {
 195    unsigned char swizzles[4];
 196
 197    swizzles[0] = bld->static_state->swizzle_r;
 198    swizzles[1] = bld->static_state->swizzle_g;
 199    swizzles[2] = bld->static_state->swizzle_b;
 200    swizzles[3] = bld->static_state->swizzle_a;
 201
 202    lp_build_swizzle_soa_inplace(&bld->texel_bld, texel, swizzles);
 203 }
 204
 205
 206
 207 /**
 208  * Generate code to fetch a texel from a texture at int coords (x, y, z).
 209  * The computation depends on whether the texture is 1D, 2D or 3D.
 210  * The result, texel, will be:
 211  *   texel[0] = red values
 212  *   texel[1] = green values
 213  *   texel[2] = blue values
 214  *   texel[3] = alpha values
 215  */
 216 static void
 217 lp_build_sample_texel_soa(struct lp_build_sample_context *bld,
 218                           LLVMValueRef width,
 219                           LLVMValueRef height,
 220                           LLVMValueRef depth,
 221                           LLVMValueRef x,
 222                           LLVMValueRef y,
 223                           LLVMValueRef z,
 224                           LLVMValueRef y_stride,
 225                           LLVMValueRef z_stride,
 226                           LLVMValueRef data_ptr,
 227                           LLVMValueRef texel_out[4])
 228 {
 229    const int dims = texture_dims(bld->static_state->target);
 230    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 231    LLVMValueRef offset;
 232    LLVMValueRef i, j;
 233    LLVMValueRef use_border = NULL;
 234
 235    /* use_border = x < 0 || x >= width || y < 0 || y >= height */
 236    if (wrap_mode_uses_border_color(bld->static_state->wrap_s)) {
 237       LLVMValueRef b1, b2;
 238       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, x, int_coord_bld->zero);
 239       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, x, width);
 240       use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
 241    }
 242
 243    if (dims >= 2 && wrap_mode_uses_border_color(bld->static_state->wrap_t)) {
 244       LLVMValueRef b1, b2;
 245       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, y, int_coord_bld->zero);
 246       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, y, height);
 247       if (use_border) {
 248          use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
 249          use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
 250       }
 251       else {
 252          use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
 253       }
 254    }
 255
 256    if (dims == 3 && wrap_mode_uses_border_color(bld->static_state->wrap_r)) {
 257       LLVMValueRef b1, b2;
 258       b1 = lp_build_cmp(int_coord_bld, PIPE_FUNC_LESS, z, int_coord_bld->zero);
 259       b2 = lp_build_cmp(int_coord_bld, PIPE_FUNC_GEQUAL, z, depth);
 260       if (use_border) {
 261          use_border = LLVMBuildOr(bld->builder, use_border, b1, "ub_or_b1");
 262          use_border = LLVMBuildOr(bld->builder, use_border, b2, "ub_or_b2");
 263       }
 264       else {
 265          use_border = LLVMBuildOr(bld->builder, b1, b2, "b1_or_b2");
 266       }
 267    }
 268
 269    /* convert x,y,z coords to linear offset from start of texture, in bytes */
 270    lp_build_sample_offset(&bld->uint_coord_bld,
 271                           bld->format_desc,
 272                           x, y, z, y_stride, z_stride,
 273                           &offset, &i, &j);
 274
 275    if (use_border) {
 276       /* If we can sample the border color, it means that texcoords may
 277        * lie outside the bounds of the texture image.  We need to do
 278        * something to prevent reading out of bounds and causing a segfault.
 279        *
 280        * Simply AND the texture coords with !use_border.  This will cause
 281        * coords which are out of bounds to become zero.  Zero's guaranteed
 282        * to be inside the texture image.
 283        */
 284       offset = lp_build_andc(&bld->uint_coord_bld, offset, use_border);
 285    }
 286
 287    lp_build_fetch_rgba_soa(bld->builder,
 288                            bld->format_desc,
 289                            bld->texel_type,
 290                            data_ptr, offset,
 291                            i, j,
 292                            texel_out);
 293
 294    apply_sampler_swizzle(bld, texel_out);
 295
 296    /*
 297     * Note: if we find an app which frequently samples the texture border
 298     * we might want to implement a true conditional here to avoid sampling
 299     * the texture whenever possible (since that's quite a bit of code).
 300     * Ex:
 301     *   if (use_border) {
 302     *      texel = border_color;
 303     *   }
 304     *   else {
 305     *      texel = sample_texture(coord);
 306     *   }
 307     * As it is now, we always sample the texture, then selectively replace
 308     * the texel color results with the border color.
 309     */
 310
 311    if (use_border) {
 312       /* select texel color or border color depending on use_border */
 313       int chan;
 314       for (chan = 0; chan < 4; chan++) {
 315          LLVMValueRef border_chan =
 316             lp_build_const_vec(bld->texel_type,
 317                                   bld->static_state->border_color[chan]);
 318          texel_out[chan] = lp_build_select(&bld->texel_bld, use_border,
 319                                            border_chan, texel_out[chan]);
 320       }
 321    }
 322 }
 323
 324
 325 /**
 326  * Fetch the texels as <4n x i8> in AoS form.
 327  */
 328 static LLVMValueRef
 329 lp_build_sample_packed(struct lp_build_sample_context *bld,
 330                        LLVMValueRef x,
 331                        LLVMValueRef y,
 332                        LLVMValueRef y_stride,
 333                        LLVMValueRef data_array)
 334 {
 335    LLVMValueRef offset, i, j;
 336    LLVMValueRef data_ptr;
 337    LLVMValueRef res;
 338
 339    /* convert x,y,z coords to linear offset from start of texture, in bytes */
 340    lp_build_sample_offset(&bld->uint_coord_bld,
 341                           bld->format_desc,
 342                           x, y, NULL, y_stride, NULL,
 343                           &offset, &i, &j);
 344
 345    /* get pointer to mipmap level 0 data */
 346    data_ptr = lp_build_get_const_mipmap_level(bld, data_array, 0);
 347
 348    if (util_format_is_rgba8_variant(bld->format_desc)) {
 349       /* Just fetch the data directly without swizzling */
 350       assert(bld->format_desc->block.width == 1);
 351       assert(bld->format_desc->block.height == 1);
 352       assert(bld->format_desc->block.bits <= bld->texel_type.width);
 353
 354       res = lp_build_gather(bld->builder,
 355                             bld->texel_type.length,
 356                             bld->format_desc->block.bits,
 357                             bld->texel_type.width,
 358                             data_ptr, offset);
 359    }
 360    else {
 361       struct lp_type type;
 362
 363       assert(bld->texel_type.width == 32);
 364
 365       memset(&type, 0, sizeof type);
 366       type.width = 8;
 367       type.length = bld->texel_type.length*4;
 368       type.norm = TRUE;
 369
 370       res = lp_build_fetch_rgba_aos(bld->builder, bld->format_desc, type,
 371                                     data_ptr, offset, i, j);
 372    }
 373
 374    return res;
 375 }
 376
 377
 378 /**
 379  * Helper to compute the mirror function for the PIPE_WRAP_MIRROR modes.
 380  */
 381 static LLVMValueRef
 382 lp_build_coord_mirror(struct lp_build_sample_context *bld,
 383                       LLVMValueRef coord)
 384 {
 385    struct lp_build_context *coord_bld = &bld->coord_bld;
 386    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 387    LLVMValueRef fract, flr, isOdd;
 388
 389    /* fract = coord - floor(coord) */
 390    fract = lp_build_sub(coord_bld, coord, lp_build_floor(coord_bld, coord));
 391
 392    /* flr = ifloor(coord); */
 393    flr = lp_build_ifloor(coord_bld, coord);
 394
 395    /* isOdd = flr & 1 */
 396    isOdd = LLVMBuildAnd(bld->builder, flr, int_coord_bld->one, "");
 397
 398    /* make coord positive or negative depending on isOdd */
 399    coord = lp_build_set_sign(coord_bld, fract, isOdd);
 400
 401    /* convert isOdd to float */
 402    isOdd = lp_build_int_to_float(coord_bld, isOdd);
 403
 404    /* add isOdd to coord */
 405    coord = lp_build_add(coord_bld, coord, isOdd);
 406
 407    return coord;
 408 }
 409
 410
 411 /**
 412  * We only support a few wrap modes in lp_build_sample_wrap_int() at this time.
 413  * Return whether the given mode is supported by that function.
 414  */
 415 static boolean
 416 is_simple_wrap_mode(unsigned mode)
 417 {
 418    switch (mode) {
 419    case PIPE_TEX_WRAP_REPEAT:
 420    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 421       return TRUE;
 422    default:
 423       return FALSE;
 424    }
 425 }
 426
 427
 428 /**
 429  * Build LLVM code for texture wrap mode, for scaled integer texcoords.
 430  * \param coord  the incoming texcoord (s,t,r or q) scaled to the texture size
 431  * \param length  the texture size along one dimension
 432  * \param is_pot  if TRUE, length is a power of two
 433  * \param wrap_mode  one of PIPE_TEX_WRAP_x
 434  */
 435 static LLVMValueRef
 436 lp_build_sample_wrap_int(struct lp_build_sample_context *bld,
 437                          LLVMValueRef coord,
 438                          LLVMValueRef length,
 439                          boolean is_pot,
 440                          unsigned wrap_mode)
 441 {
 442    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
 443    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 444    LLVMValueRef length_minus_one;
 445
 446    length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
 447
 448    switch(wrap_mode) {
 449    case PIPE_TEX_WRAP_REPEAT:
 450       if(is_pot)
 451          coord = LLVMBuildAnd(bld->builder, coord, length_minus_one, "");
 452       else
 453          /* Signed remainder won't give the right results for negative
 454           * dividends but unsigned remainder does.*/
 455          coord = LLVMBuildURem(bld->builder, coord, length, "");
 456       break;
 457
 458    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 459       coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
 460       coord = lp_build_min(int_coord_bld, coord, length_minus_one);
 461       break;
 462
 463    case PIPE_TEX_WRAP_CLAMP:
 464    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 465    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 466    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 467    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 468    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 469    default:
 470       assert(0);
 471    }
 472
 473    return coord;
 474 }
 475
 476
 477 /**
 478  * Build LLVM code for texture wrap mode for linear filtering.
 479  * \param x0_out  returns first integer texcoord
 480  * \param x1_out  returns second integer texcoord
 481  * \param weight_out  returns linear interpolation weight
 482  */
 483 static void
 484 lp_build_sample_wrap_linear(struct lp_build_sample_context *bld,
 485                             LLVMValueRef coord,
 486                             LLVMValueRef length,
 487                             boolean is_pot,
 488                             unsigned wrap_mode,
 489                             LLVMValueRef *x0_out,
 490                             LLVMValueRef *x1_out,
 491                             LLVMValueRef *weight_out)
 492 {
 493    struct lp_build_context *coord_bld = &bld->coord_bld;
 494    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 495    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
 496    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
 497    LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
 498    LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
 499    LLVMValueRef coord0, coord1, weight;
 500
 501    switch(wrap_mode) {
 502    case PIPE_TEX_WRAP_REPEAT:
 503       /* mul by size and subtract 0.5 */
 504       coord = lp_build_mul(coord_bld, coord, length_f);
 505       coord = lp_build_sub(coord_bld, coord, half);
 506       /* convert to int */
 507       coord0 = lp_build_ifloor(coord_bld, coord);
 508       coord1 = lp_build_add(uint_coord_bld, coord0, uint_coord_bld->one);
 509       /* compute lerp weight */
 510       weight = lp_build_fract(coord_bld, coord);
 511       /* repeat wrap */
 512       if (is_pot) {
 513          coord0 = LLVMBuildAnd(bld->builder, coord0, length_minus_one, "");
 514          coord1 = LLVMBuildAnd(bld->builder, coord1, length_minus_one, "");
 515       }
 516       else {
 517          /* Signed remainder won't give the right results for negative
 518           * dividends but unsigned remainder does.*/
 519          coord0 = LLVMBuildURem(bld->builder, coord0, length, "");
 520          coord1 = LLVMBuildURem(bld->builder, coord1, length, "");
 521       }
 522       break;
 523
 524    case PIPE_TEX_WRAP_CLAMP:
 525       if (bld->static_state->normalized_coords) {
 526          /* scale coord to length */
 527          coord = lp_build_mul(coord_bld, coord, length_f);
 528       }
 529
 530       /* clamp to [0, length] */
 531       coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, length_f);
 532
 533       coord = lp_build_sub(coord_bld, coord, half);
 534
 535       weight = lp_build_fract(coord_bld, coord);
 536       coord0 = lp_build_ifloor(coord_bld, coord);
 537       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 538       break;
 539
 540    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 541       if (bld->static_state->normalized_coords) {
 542          /* clamp to [0,1] */
 543          coord = lp_build_clamp(coord_bld, coord, coord_bld->zero, coord_bld->one);
 544          /* mul by tex size and subtract 0.5 */
 545          coord = lp_build_mul(coord_bld, coord, length_f);
 546          coord = lp_build_sub(coord_bld, coord, half);
 547       }
 548       else {
 549          LLVMValueRef min, max;
 550          /* clamp to [0.5, length - 0.5] */
 551          min = half;
 552          max = lp_build_sub(coord_bld, length_f, min);
 553          coord = lp_build_clamp(coord_bld, coord, min, max);
 554       }
 555       /* compute lerp weight */
 556       weight = lp_build_fract(coord_bld, coord);
 557       /* coord0 = floor(coord); */
 558       coord0 = lp_build_ifloor(coord_bld, coord);
 559       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 560       /* coord0 = max(coord0, 0) */
 561       coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
 562       /* coord1 = min(coord1, length-1) */
 563       coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
 564       break;
 565
 566    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 567       {
 568          LLVMValueRef min, max;
 569          if (bld->static_state->normalized_coords) {
 570             /* scale coord to length */
 571             coord = lp_build_mul(coord_bld, coord, length_f);
 572          }
 573          /* clamp to [-0.5, length + 0.5] */
 574          min = lp_build_const_vec(coord_bld->type, -0.5F);
 575          max = lp_build_sub(coord_bld, length_f, min);
 576          coord = lp_build_clamp(coord_bld, coord, min, max);
 577          coord = lp_build_sub(coord_bld, coord, half);
 578          /* compute lerp weight */
 579          weight = lp_build_fract(coord_bld, coord);
 580          /* convert to int */
 581          coord0 = lp_build_ifloor(coord_bld, coord);
 582          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 583       }
 584       break;
 585
 586    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 587       /* compute mirror function */
 588       coord = lp_build_coord_mirror(bld, coord);
 589
 590       /* scale coord to length */
 591       coord = lp_build_mul(coord_bld, coord, length_f);
 592       coord = lp_build_sub(coord_bld, coord, half);
 593
 594       /* compute lerp weight */
 595       weight = lp_build_fract(coord_bld, coord);
 596
 597       /* convert to int coords */
 598       coord0 = lp_build_ifloor(coord_bld, coord);
 599       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 600
 601       /* coord0 = max(coord0, 0) */
 602       coord0 = lp_build_max(int_coord_bld, coord0, int_coord_bld->zero);
 603       /* coord1 = min(coord1, length-1) */
 604       coord1 = lp_build_min(int_coord_bld, coord1, length_minus_one);
 605       break;
 606
 607    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 608       coord = lp_build_abs(coord_bld, coord);
 609
 610       if (bld->static_state->normalized_coords) {
 611          /* scale coord to length */
 612          coord = lp_build_mul(coord_bld, coord, length_f);
 613       }
 614
 615       /* clamp to [0, length] */
 616       coord = lp_build_min(coord_bld, coord, length_f);
 617
 618       coord = lp_build_sub(coord_bld, coord, half);
 619
 620       weight = lp_build_fract(coord_bld, coord);
 621       coord0 = lp_build_ifloor(coord_bld, coord);
 622       coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 623       break;
 624
 625    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 626       {
 627          LLVMValueRef min, max;
 628
 629          coord = lp_build_abs(coord_bld, coord);
 630
 631          if (bld->static_state->normalized_coords) {
 632             /* scale coord to length */
 633             coord = lp_build_mul(coord_bld, coord, length_f);
 634          }
 635
 636          /* clamp to [0.5, length - 0.5] */
 637          min = half;
 638          max = lp_build_sub(coord_bld, length_f, min);
 639          coord = lp_build_clamp(coord_bld, coord, min, max);
 640
 641          coord = lp_build_sub(coord_bld, coord, half);
 642
 643          weight = lp_build_fract(coord_bld, coord);
 644          coord0 = lp_build_ifloor(coord_bld, coord);
 645          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 646       }
 647       break;
 648
 649    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 650       {
 651          LLVMValueRef min, max;
 652
 653          coord = lp_build_abs(coord_bld, coord);
 654
 655          if (bld->static_state->normalized_coords) {
 656             /* scale coord to length */
 657             coord = lp_build_mul(coord_bld, coord, length_f);
 658          }
 659
 660          /* clamp to [-0.5, length + 0.5] */
 661          min = lp_build_negate(coord_bld, half);
 662          max = lp_build_sub(coord_bld, length_f, min);
 663          coord = lp_build_clamp(coord_bld, coord, min, max);
 664
 665          coord = lp_build_sub(coord_bld, coord, half);
 666
 667          weight = lp_build_fract(coord_bld, coord);
 668          coord0 = lp_build_ifloor(coord_bld, coord);
 669          coord1 = lp_build_add(int_coord_bld, coord0, int_coord_bld->one);
 670       }
 671       break;
 672
 673    default:
 674       assert(0);
 675       coord0 = NULL;
 676       coord1 = NULL;
 677       weight = NULL;
 678    }
 679
 680    *x0_out = coord0;
 681    *x1_out = coord1;
 682    *weight_out = weight;
 683 }
 684
 685
 686 /**
 687  * Build LLVM code for texture wrap mode for nearest filtering.
 688  * \param coord  the incoming texcoord (nominally in [0,1])
 689  * \param length  the texture size along one dimension, as int
 690  * \param is_pot  if TRUE, length is a power of two
 691  * \param wrap_mode  one of PIPE_TEX_WRAP_x
 692  */
 693 static LLVMValueRef
 694 lp_build_sample_wrap_nearest(struct lp_build_sample_context *bld,
 695                              LLVMValueRef coord,
 696                              LLVMValueRef length,
 697                              boolean is_pot,
 698                              unsigned wrap_mode)
 699 {
 700    struct lp_build_context *coord_bld = &bld->coord_bld;
 701    struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
 702    struct lp_build_context *uint_coord_bld = &bld->uint_coord_bld;
 703    LLVMValueRef length_f = lp_build_int_to_float(coord_bld, length);
 704    LLVMValueRef length_minus_one = lp_build_sub(uint_coord_bld, length, uint_coord_bld->one);
 705    LLVMValueRef icoord;
 706
 707    switch(wrap_mode) {
 708    case PIPE_TEX_WRAP_REPEAT:
 709       coord = lp_build_mul(coord_bld, coord, length_f);
 710       icoord = lp_build_ifloor(coord_bld, coord);
 711       if (is_pot)
 712          icoord = LLVMBuildAnd(bld->builder, icoord, length_minus_one, "");
 713       else
 714          /* Signed remainder won't give the right results for negative
 715           * dividends but unsigned remainder does.*/
 716          icoord = LLVMBuildURem(bld->builder, icoord, length, "");
 717       break;
 718
 719    case PIPE_TEX_WRAP_CLAMP:
 720    case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 721       if (bld->static_state->normalized_coords) {
 722          /* scale coord to length */
 723          coord = lp_build_mul(coord_bld, coord, length_f);
 724       }
 725
 726       /* floor */
 727       icoord = lp_build_ifloor(coord_bld, coord);
 728
 729       /* clamp to [0, length - 1]. */
 730       icoord = lp_build_clamp(int_coord_bld, icoord, int_coord_bld->zero,
 731                               length_minus_one);
 732       break;
 733
 734    case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 735       /* Note: this is the same as CLAMP_TO_EDGE, except min = -min */
 736       {
 737          LLVMValueRef min, max;
 738
 739          if (bld->static_state->normalized_coords) {
 740             /* scale coord to length */
 741             coord = lp_build_mul(coord_bld, coord, length_f);
 742          }
 743
 744          icoord = lp_build_ifloor(coord_bld, coord);
 745
 746          /* clamp to [-1, length] */
 747          min = lp_build_negate(int_coord_bld, int_coord_bld->one);
 748          max = length;
 749          icoord = lp_build_clamp(int_coord_bld, icoord, min, max);
 750       }
 751       break;
 752
 753    case PIPE_TEX_WRAP_MIRROR_REPEAT:
 754       /* compute mirror function */
 755       coord = lp_build_coord_mirror(bld, coord);
 756
 757       /* scale coord to length */
 758       assert(bld->static_state->normalized_coords);
 759       coord = lp_build_mul(coord_bld, coord, length_f);
 760
 761       icoord = lp_build_ifloor(coord_bld, coord);
 762
 763       /* clamp to [0, length - 1] */
 764       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 765       break;
 766
 767    case PIPE_TEX_WRAP_MIRROR_CLAMP:
 768    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 769       coord = lp_build_abs(coord_bld, coord);
 770
 771       if (bld->static_state->normalized_coords) {
 772          /* scale coord to length */
 773          coord = lp_build_mul(coord_bld, coord, length_f);
 774       }
 775
 776       icoord = lp_build_ifloor(coord_bld, coord);
 777
 778       /* clamp to [0, length - 1] */
 779       icoord = lp_build_min(int_coord_bld, icoord, length_minus_one);
 780       break;
 781
 782    case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 783       coord = lp_build_abs(coord_bld, coord);
 784
 785       if (bld->static_state->normalized_coords) {
 786          /* scale coord to length */
 787          coord = lp_build_mul(coord_bld, coord, length_f);
 788       }
 789
 790       icoord = lp_build_ifloor(coord_bld, coord);
 791
 792       /* clamp to [0, length] */
 793       icoord = lp_build_min(int_coord_bld, icoord, length);
 794       break;
 795
 796    default:
 797       assert(0);
 798       icoord = NULL;
 799    }
 800
 801    return icoord;
 802 }
 803
 804
 805 /**
 806  * Codegen equivalent for u_minify().
 807  * Return max(1, base_size >> level);
 808  */
 809 static LLVMValueRef
 810 lp_build_minify(struct lp_build_sample_context *bld,
 811                 LLVMValueRef base_size,
 812                 LLVMValueRef level)
 813 {
 814    LLVMValueRef size = LLVMBuildAShr(bld->builder, base_size, level, "minify");
 815    size = lp_build_max(&bld->int_coord_bld, size, bld->int_coord_bld.one);
 816    return size;
 817 }
 818
 819
 820 /**
 821  * Generate code to compute texture level of detail (lambda).
 822  * \param ddx  partial derivatives of (s, t, r, q) with respect to X
 823  * \param ddy  partial derivatives of (s, t, r, q) with respect to Y
 824  * \param lod_bias  optional float vector with the shader lod bias
 825  * \param explicit_lod  optional float vector with the explicit lod
 826  * \param width  scalar int texture width
 827  * \param height  scalar int texture height
 828  * \param depth  scalar int texture depth
 829  *
 830  * XXX: The resulting lod is scalar, so ignore all but the first element of
 831  * derivatives, lod_bias, etc that are passed by the shader.
 832  */
 833 static LLVMValueRef
 834 lp_build_lod_selector(struct lp_build_sample_context *bld,
 835                       const LLVMValueRef ddx[4],
 836                       const LLVMValueRef ddy[4],
 837                       LLVMValueRef lod_bias, /* optional */
 838                       LLVMValueRef explicit_lod, /* optional */
 839                       LLVMValueRef width,
 840                       LLVMValueRef height,
 841                       LLVMValueRef depth)
 842
 843 {
 844    if (bld->static_state->min_lod == bld->static_state->max_lod) {
 845       /* User is forcing sampling from a particular mipmap level.
 846        * This is hit during mipmap generation.
 847        */
 848       return LLVMConstReal(LLVMFloatType(), bld->static_state->min_lod);
 849    }
 850    else {
 851       struct lp_build_context *float_bld = &bld->float_bld;
 852       LLVMValueRef sampler_lod_bias = LLVMConstReal(LLVMFloatType(),
 853                                                     bld->static_state->lod_bias);
 854       LLVMValueRef min_lod = LLVMConstReal(LLVMFloatType(),
 855                                            bld->static_state->min_lod);
 856       LLVMValueRef max_lod = LLVMConstReal(LLVMFloatType(),
 857                                            bld->static_state->max_lod);
 858       LLVMValueRef index0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 859       LLVMValueRef lod;
 860
 861       if (explicit_lod) {
 862          lod = LLVMBuildExtractElement(bld->builder, explicit_lod,
 863                                        index0, "");
 864       }
 865       else {
 866          const int dims = texture_dims(bld->static_state->target);
 867          LLVMValueRef dsdx, dsdy;
 868          LLVMValueRef dtdx = NULL, dtdy = NULL, drdx = NULL, drdy = NULL;
 869          LLVMValueRef rho;
 870
 871          dsdx = LLVMBuildExtractElement(bld->builder, ddx[0], index0, "dsdx");
 872          dsdx = lp_build_abs(float_bld, dsdx);
 873          dsdy = LLVMBuildExtractElement(bld->builder, ddy[0], index0, "dsdy");
 874          dsdy = lp_build_abs(float_bld, dsdy);
 875          if (dims > 1) {
 876             dtdx = LLVMBuildExtractElement(bld->builder, ddx[1], index0, "dtdx");
 877             dtdx = lp_build_abs(float_bld, dtdx);
 878             dtdy = LLVMBuildExtractElement(bld->builder, ddy[1], index0, "dtdy");
 879             dtdy = lp_build_abs(float_bld, dtdy);
 880             if (dims > 2) {
 881                drdx = LLVMBuildExtractElement(bld->builder, ddx[2], index0, "drdx");
 882                drdx = lp_build_abs(float_bld, drdx);
 883                drdy = LLVMBuildExtractElement(bld->builder, ddy[2], index0, "drdy");
 884                drdy = lp_build_abs(float_bld, drdy);
 885             }
 886          }
 887
 888          /* Compute rho = max of all partial derivatives scaled by texture size.
 889           * XXX this could be vectorized somewhat
 890           */
 891          rho = LLVMBuildMul(bld->builder,
 892                             lp_build_max(float_bld, dsdx, dsdy),
 893                             lp_build_int_to_float(float_bld, width), "");
 894          if (dims > 1) {
 895             LLVMValueRef max;
 896             max = LLVMBuildMul(bld->builder,
 897                                lp_build_max(float_bld, dtdx, dtdy),
 898                                lp_build_int_to_float(float_bld, height), "");
 899             rho = lp_build_max(float_bld, rho, max);
 900             if (dims > 2) {
 901                max = LLVMBuildMul(bld->builder,
 902                                   lp_build_max(float_bld, drdx, drdy),
 903                                   lp_build_int_to_float(float_bld, depth), "");
 904                rho = lp_build_max(float_bld, rho, max);
 905             }
 906          }
 907
 908          /* compute lod = log2(rho) */
 909          lod = lp_build_log2(float_bld, rho);
 910
 911          /* add shader lod bias */
 912          if (lod_bias) {
 913             lod_bias = LLVMBuildExtractElement(bld->builder, lod_bias,
 914                                                index0, "");
 915             lod = LLVMBuildAdd(bld->builder, lod, lod_bias, "shader_lod_bias");
 916          }
 917       }
 918
 919       /* add sampler lod bias */
 920       lod = LLVMBuildAdd(bld->builder, lod, sampler_lod_bias, "sampler_lod_bias");
 921
 922       /* clamp lod */
 923       lod = lp_build_clamp(float_bld, lod, min_lod, max_lod);
 924
 925       return lod;
 926    }
 927 }
 928
 929
 930 /**
 931  * For PIPE_TEX_MIPFILTER_NEAREST, convert float LOD to integer
 932  * mipmap level index.
 933  * Note: this is all scalar code.
 934  * \param lod  scalar float texture level of detail
 935  * \param level_out  returns integer
 936  */
 937 static void
 938 lp_build_nearest_mip_level(struct lp_build_sample_context *bld,
 939                            unsigned unit,
 940                            LLVMValueRef lod,
 941                            LLVMValueRef *level_out)
 942 {
 943    struct lp_build_context *float_bld = &bld->float_bld;
 944    struct lp_build_context *int_bld = &bld->int_bld;
 945    LLVMValueRef last_level, level;
 946
 947    LLVMValueRef zero = LLVMConstInt(LLVMInt32Type(), 0, 0);
 948
 949    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 950                                                bld->builder, unit);
 951
 952    /* convert float lod to integer */
 953    level = lp_build_iround(float_bld, lod);
 954
 955    /* clamp level to legal range of levels */
 956    *level_out = lp_build_clamp(int_bld, level, zero, last_level);
 957 }
 958
 959
 960 /**
 961  * For PIPE_TEX_MIPFILTER_LINEAR, convert float LOD to integer to
 962  * two (adjacent) mipmap level indexes.  Later, we'll sample from those
 963  * two mipmap levels and interpolate between them.
 964  */
 965 static void
 966 lp_build_linear_mip_levels(struct lp_build_sample_context *bld,
 967                            unsigned unit,
 968                            LLVMValueRef lod,
 969                            LLVMValueRef *level0_out,
 970                            LLVMValueRef *level1_out,
 971                            LLVMValueRef *weight_out)
 972 {
 973    struct lp_build_context *float_bld = &bld->float_bld;
 974    struct lp_build_context *int_bld = &bld->int_bld;
 975    LLVMValueRef last_level, level;
 976
 977    last_level = bld->dynamic_state->last_level(bld->dynamic_state,
 978                                                bld->builder, unit);
 979
 980    /* convert float lod to integer */
 981    level = lp_build_ifloor(float_bld, lod);
 982
 983    /* compute level 0 and clamp to legal range of levels */
 984    *level0_out = lp_build_clamp(int_bld, level,
 985                                 int_bld->zero,
 986                                 last_level);
 987    /* compute level 1 and clamp to legal range of levels */
 988    level = lp_build_add(int_bld, level, int_bld->one);
 989    *level1_out = lp_build_clamp(int_bld, level,
 990                                 int_bld->zero,
 991                                 last_level);
 992
 993    *weight_out = lp_build_fract(float_bld, lod);
 994 }
 995
 996
 997 /**
 998  * Generate code to sample a mipmap level with nearest filtering.
 999  * If sampling a cube texture, r = cube face in [0,5].
1000  */
1001 static void
1002 lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
1003                               LLVMValueRef width_vec,
1004                               LLVMValueRef height_vec,
1005                               LLVMValueRef depth_vec,
1006                               LLVMValueRef row_stride_vec,
1007                               LLVMValueRef img_stride_vec,
1008                               LLVMValueRef data_ptr,
1009                               LLVMValueRef s,
1010                               LLVMValueRef t,
1011                               LLVMValueRef r,
1012                               LLVMValueRef colors_out[4])
1013 {
1014    const int dims = texture_dims(bld->static_state->target);
1015    LLVMValueRef x, y, z;
1016
1017    /*
1018     * Compute integer texcoords.
1019     */
1020    x = lp_build_sample_wrap_nearest(bld, s, width_vec,
1021                                     bld->static_state->pot_width,
1022                                     bld->static_state->wrap_s);
1023    lp_build_name(x, "tex.x.wrapped");
1024
1025    if (dims >= 2) {
1026       y = lp_build_sample_wrap_nearest(bld, t, height_vec,
1027                                        bld->static_state->pot_height,
1028                                        bld->static_state->wrap_t);
1029       lp_build_name(y, "tex.y.wrapped");
1030
1031       if (dims == 3) {
1032          z = lp_build_sample_wrap_nearest(bld, r, depth_vec,
1033                                           bld->static_state->pot_height,
1034                                           bld->static_state->wrap_r);
1035          lp_build_name(z, "tex.z.wrapped");
1036       }
1037       else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1038          z = r;
1039       }
1040       else {
1041          z = NULL;
1042       }
1043    }
1044    else {
1045       y = z = NULL;
1046    }
1047
1048    /*
1049     * Get texture colors.
1050     */
1051    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1052                              x, y, z,
1053                              row_stride_vec, img_stride_vec,
1054                              data_ptr, colors_out);
1055 }
1056
1057
1058 /**
1059  * Generate code to sample a mipmap level with linear filtering.
1060  * If sampling a cube texture, r = cube face in [0,5].
1061  */
1062 static void
1063 lp_build_sample_image_linear(struct lp_build_sample_context *bld,
1064                              LLVMValueRef width_vec,
1065                              LLVMValueRef height_vec,
1066                              LLVMValueRef depth_vec,
1067                              LLVMValueRef row_stride_vec,
1068                              LLVMValueRef img_stride_vec,
1069                              LLVMValueRef data_ptr,
1070                              LLVMValueRef s,
1071                              LLVMValueRef t,
1072                              LLVMValueRef r,
1073                              LLVMValueRef colors_out[4])
1074 {
1075    const int dims = texture_dims(bld->static_state->target);
1076    LLVMValueRef x0, y0, z0, x1, y1, z1;
1077    LLVMValueRef s_fpart, t_fpart, r_fpart;
1078    LLVMValueRef neighbors[2][2][4];
1079    int chan;
1080
1081    /*
1082     * Compute integer texcoords.
1083     */
1084    lp_build_sample_wrap_linear(bld, s, width_vec,
1085                                bld->static_state->pot_width,
1086                                bld->static_state->wrap_s,
1087                                &x0, &x1, &s_fpart);
1088    lp_build_name(x0, "tex.x0.wrapped");
1089    lp_build_name(x1, "tex.x1.wrapped");
1090
1091    if (dims >= 2) {
1092       lp_build_sample_wrap_linear(bld, t, height_vec,
1093                                   bld->static_state->pot_height,
1094                                   bld->static_state->wrap_t,
1095                                   &y0, &y1, &t_fpart);
1096       lp_build_name(y0, "tex.y0.wrapped");
1097       lp_build_name(y1, "tex.y1.wrapped");
1098
1099       if (dims == 3) {
1100          lp_build_sample_wrap_linear(bld, r, depth_vec,
1101                                      bld->static_state->pot_depth,
1102                                      bld->static_state->wrap_r,
1103                                      &z0, &z1, &r_fpart);
1104          lp_build_name(z0, "tex.z0.wrapped");
1105          lp_build_name(z1, "tex.z1.wrapped");
1106       }
1107       else if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1108          z0 = z1 = r;  /* cube face */
1109          r_fpart = NULL;
1110       }
1111       else {
1112          z0 = z1 = NULL;
1113          r_fpart = NULL;
1114       }
1115    }
1116    else {
1117       y0 = y1 = t_fpart = NULL;
1118       z0 = z1 = r_fpart = NULL;
1119    }
1120
1121    /*
1122     * Get texture colors.
1123     */
1124    /* get x0/x1 texels */
1125    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1126                              x0, y0, z0,
1127                              row_stride_vec, img_stride_vec,
1128                              data_ptr, neighbors[0][0]);
1129    lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1130                              x1, y0, z0,
1131                              row_stride_vec, img_stride_vec,
1132                              data_ptr, neighbors[0][1]);
1133
1134    if (dims == 1) {
1135       /* Interpolate two samples from 1D image to produce one color */
1136       for (chan = 0; chan < 4; chan++) {
1137          colors_out[chan] = lp_build_lerp(&bld->texel_bld, s_fpart,
1138                                           neighbors[0][0][chan],
1139                                           neighbors[0][1][chan]);
1140       }
1141    }
1142    else {
1143       /* 2D/3D texture */
1144       LLVMValueRef colors0[4];
1145
1146       /* get x0/x1 texels at y1 */
1147       lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1148                                 x0, y1, z0,
1149                                 row_stride_vec, img_stride_vec,
1150                                 data_ptr, neighbors[1][0]);
1151       lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1152                                 x1, y1, z0,
1153                                 row_stride_vec, img_stride_vec,
1154                                 data_ptr, neighbors[1][1]);
1155
1156       /* Bilinear interpolate the four samples from the 2D image / 3D slice */
1157       for (chan = 0; chan < 4; chan++) {
1158          colors0[chan] = lp_build_lerp_2d(&bld->texel_bld,
1159                                           s_fpart, t_fpart,
1160                                           neighbors[0][0][chan],
1161                                           neighbors[0][1][chan],
1162                                           neighbors[1][0][chan],
1163                                           neighbors[1][1][chan]);
1164       }
1165
1166       if (dims == 3) {
1167          LLVMValueRef neighbors1[2][2][4];
1168          LLVMValueRef colors1[4];
1169
1170          /* get x0/x1/y0/y1 texels at z1 */
1171          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1172                                    x0, y0, z1,
1173                                    row_stride_vec, img_stride_vec,
1174                                    data_ptr, neighbors1[0][0]);
1175          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1176                                    x1, y0, z1,
1177                                    row_stride_vec, img_stride_vec,
1178                                    data_ptr, neighbors1[0][1]);
1179          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1180                                    x0, y1, z1,
1181                                    row_stride_vec, img_stride_vec,
1182                                    data_ptr, neighbors1[1][0]);
1183          lp_build_sample_texel_soa(bld, width_vec, height_vec, depth_vec,
1184                                    x1, y1, z1,
1185                                    row_stride_vec, img_stride_vec,
1186                                    data_ptr, neighbors1[1][1]);
1187
1188          /* Bilinear interpolate the four samples from the second Z slice */
1189          for (chan = 0; chan < 4; chan++) {
1190             colors1[chan] = lp_build_lerp_2d(&bld->texel_bld,
1191                                              s_fpart, t_fpart,
1192                                              neighbors1[0][0][chan],
1193                                              neighbors1[0][1][chan],
1194                                              neighbors1[1][0][chan],
1195                                              neighbors1[1][1][chan]);
1196          }
1197
1198          /* Linearly interpolate the two samples from the two 3D slices */
1199          for (chan = 0; chan < 4; chan++) {
1200             colors_out[chan] = lp_build_lerp(&bld->texel_bld,
1201                                              r_fpart,
1202                                              colors0[chan], colors1[chan]);
1203          }
1204       }
1205       else {
1206          /* 2D tex */
1207          for (chan = 0; chan < 4; chan++) {
1208             colors_out[chan] = colors0[chan];
1209          }
1210       }
1211    }
1212 }
1213
1214
1215 /** Helper used by lp_build_cube_lookup() */
1216 static LLVMValueRef
1217 lp_build_cube_ima(struct lp_build_context *coord_bld, LLVMValueRef coord)
1218 {
1219    /* ima = -0.5 / abs(coord); */
1220    LLVMValueRef negHalf = lp_build_const_vec(coord_bld->type, -0.5);
1221    LLVMValueRef absCoord = lp_build_abs(coord_bld, coord);
1222    LLVMValueRef ima = lp_build_mul(coord_bld, negHalf,
1223                                    lp_build_rcp(coord_bld, absCoord));
1224    return ima;
1225 }
1226
1227
1228 /**
1229  * Helper used by lp_build_cube_lookup()
1230  * \param sign  scalar +1 or -1
1231  * \param coord  float vector
1232  * \param ima  float vector
1233  */
1234 static LLVMValueRef
1235 lp_build_cube_coord(struct lp_build_context *coord_bld,
1236                     LLVMValueRef sign, int negate_coord,
1237                     LLVMValueRef coord, LLVMValueRef ima)
1238 {
1239    /* return negate(coord) * ima * sign + 0.5; */
1240    LLVMValueRef half = lp_build_const_vec(coord_bld->type, 0.5);
1241    LLVMValueRef res;
1242
1243    assert(negate_coord == +1 || negate_coord == -1);
1244
1245    if (negate_coord == -1) {
1246       coord = lp_build_negate(coord_bld, coord);
1247    }
1248
1249    res = lp_build_mul(coord_bld, coord, ima);
1250    if (sign) {
1251       sign = lp_build_broadcast_scalar(coord_bld, sign);
1252       res = lp_build_mul(coord_bld, res, sign);
1253    }
1254    res = lp_build_add(coord_bld, res, half);
1255
1256    return res;
1257 }
1258
1259
1260 /** Helper used by lp_build_cube_lookup()
1261  * Return (major_coord >= 0) ? pos_face : neg_face;
1262  */
1263 static LLVMValueRef
1264 lp_build_cube_face(struct lp_build_sample_context *bld,
1265                    LLVMValueRef major_coord,
1266                    unsigned pos_face, unsigned neg_face)
1267 {
1268    LLVMValueRef cmp = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1269                                     major_coord,
1270                                     bld->float_bld.zero, "");
1271    LLVMValueRef pos = LLVMConstInt(LLVMInt32Type(), pos_face, 0);
1272    LLVMValueRef neg = LLVMConstInt(LLVMInt32Type(), neg_face, 0);
1273    LLVMValueRef res = LLVMBuildSelect(bld->builder, cmp, pos, neg, "");
1274    return res;
1275 }
1276
1277
1278
1279 /**
1280  * Generate code to do cube face selection and compute per-face texcoords.
1281  */
1282 static void
1283 lp_build_cube_lookup(struct lp_build_sample_context *bld,
1284                      LLVMValueRef s,
1285                      LLVMValueRef t,
1286                      LLVMValueRef r,
1287                      LLVMValueRef *face,
1288                      LLVMValueRef *face_s,
1289                      LLVMValueRef *face_t)
1290 {
1291    struct lp_build_context *float_bld = &bld->float_bld;
1292    struct lp_build_context *coord_bld = &bld->coord_bld;
1293    LLVMValueRef rx, ry, rz;
1294    LLVMValueRef arx, ary, arz;
1295    LLVMValueRef c25 = LLVMConstReal(LLVMFloatType(), 0.25);
1296    LLVMValueRef arx_ge_ary, arx_ge_arz;
1297    LLVMValueRef ary_ge_arx, ary_ge_arz;
1298    LLVMValueRef arx_ge_ary_arz, ary_ge_arx_arz;
1299    LLVMValueRef rx_pos, ry_pos, rz_pos;
1300
1301    assert(bld->coord_bld.type.length == 4);
1302
1303    /*
1304     * Use the average of the four pixel's texcoords to choose the face.
1305     */
1306    rx = lp_build_mul(float_bld, c25,
1307                      lp_build_sum_vector(&bld->coord_bld, s));
1308    ry = lp_build_mul(float_bld, c25,
1309                      lp_build_sum_vector(&bld->coord_bld, t));
1310    rz = lp_build_mul(float_bld, c25,
1311                      lp_build_sum_vector(&bld->coord_bld, r));
1312
1313    arx = lp_build_abs(float_bld, rx);
1314    ary = lp_build_abs(float_bld, ry);
1315    arz = lp_build_abs(float_bld, rz);
1316
1317    /*
1318     * Compare sign/magnitude of rx,ry,rz to determine face
1319     */
1320    arx_ge_ary = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, ary, "");
1321    arx_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, arx, arz, "");
1322    ary_ge_arx = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arx, "");
1323    ary_ge_arz = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ary, arz, "");
1324
1325    arx_ge_ary_arz = LLVMBuildAnd(bld->builder, arx_ge_ary, arx_ge_arz, "");
1326    ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1327
1328    rx_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rx, float_bld->zero, "");
1329    ry_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, ry, float_bld->zero, "");
1330    rz_pos = LLVMBuildFCmp(bld->builder, LLVMRealUGE, rz, float_bld->zero, "");
1331
1332    {
1333       struct lp_build_flow_context *flow_ctx;
1334       struct lp_build_if_state if_ctx;
1335
1336       flow_ctx = lp_build_flow_create(bld->builder);
1337       lp_build_flow_scope_begin(flow_ctx);
1338
1339       *face_s = bld->coord_bld.undef;
1340       *face_t = bld->coord_bld.undef;
1341       *face = bld->int_bld.undef;
1342
1343       lp_build_name(*face_s, "face_s");
1344       lp_build_name(*face_t, "face_t");
1345       lp_build_name(*face, "face");
1346
1347       lp_build_flow_scope_declare(flow_ctx, face_s);
1348       lp_build_flow_scope_declare(flow_ctx, face_t);
1349       lp_build_flow_scope_declare(flow_ctx, face);
1350
1351       lp_build_if(&if_ctx, flow_ctx, bld->builder, arx_ge_ary_arz);
1352       {
1353          /* +/- X face */
1354          LLVMValueRef sign = lp_build_sgn(float_bld, rx);
1355          LLVMValueRef ima = lp_build_cube_ima(coord_bld, s);
1356          *face_s = lp_build_cube_coord(coord_bld, sign, +1, r, ima);
1357          *face_t = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1358          *face = lp_build_cube_face(bld, rx,
1359                                     PIPE_TEX_FACE_POS_X,
1360                                     PIPE_TEX_FACE_NEG_X);
1361       }
1362       lp_build_else(&if_ctx);
1363       {
1364          struct lp_build_flow_context *flow_ctx2;
1365          struct lp_build_if_state if_ctx2;
1366
1367          LLVMValueRef face_s2 = bld->coord_bld.undef;
1368          LLVMValueRef face_t2 = bld->coord_bld.undef;
1369          LLVMValueRef face2 = bld->int_bld.undef;
1370
1371          flow_ctx2 = lp_build_flow_create(bld->builder);
1372          lp_build_flow_scope_begin(flow_ctx2);
1373          lp_build_flow_scope_declare(flow_ctx2, &face_s2);
1374          lp_build_flow_scope_declare(flow_ctx2, &face_t2);
1375          lp_build_flow_scope_declare(flow_ctx2, &face2);
1376
1377          ary_ge_arx_arz = LLVMBuildAnd(bld->builder, ary_ge_arx, ary_ge_arz, "");
1378
1379          lp_build_if(&if_ctx2, flow_ctx2, bld->builder, ary_ge_arx_arz);
1380          {
1381             /* +/- Y face */
1382             LLVMValueRef sign = lp_build_sgn(float_bld, ry);
1383             LLVMValueRef ima = lp_build_cube_ima(coord_bld, t);
1384             face_s2 = lp_build_cube_coord(coord_bld, NULL, -1, s, ima);
1385             face_t2 = lp_build_cube_coord(coord_bld, sign, -1, r, ima);
1386             face2 = lp_build_cube_face(bld, ry,
1387                                        PIPE_TEX_FACE_POS_Y,
1388                                        PIPE_TEX_FACE_NEG_Y);
1389          }
1390          lp_build_else(&if_ctx2);
1391          {
1392             /* +/- Z face */
1393             LLVMValueRef sign = lp_build_sgn(float_bld, rz);
1394             LLVMValueRef ima = lp_build_cube_ima(coord_bld, r);
1395             face_s2 = lp_build_cube_coord(coord_bld, sign, -1, s, ima);
1396             face_t2 = lp_build_cube_coord(coord_bld, NULL, +1, t, ima);
1397             face2 = lp_build_cube_face(bld, rz,
1398                                        PIPE_TEX_FACE_POS_Z,
1399                                        PIPE_TEX_FACE_NEG_Z);
1400          }
1401          lp_build_endif(&if_ctx2);
1402          lp_build_flow_scope_end(flow_ctx2);
1403          lp_build_flow_destroy(flow_ctx2);
1404          *face_s = face_s2;
1405          *face_t = face_t2;
1406          *face = face2;
1407       }
1408
1409       lp_build_endif(&if_ctx);
1410       lp_build_flow_scope_end(flow_ctx);
1411       lp_build_flow_destroy(flow_ctx);
1412    }
1413 }
1414
1415
1416
1417 /**
1418  * Sample the texture/mipmap using given image filter and mip filter.
1419  * data0_ptr and data1_ptr point to the two mipmap levels to sample
1420  * from.  width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
1421  * If we're using nearest miplevel sampling the '1' values will be null/unused.
1422  */
1423 static void
1424 lp_build_sample_mipmap(struct lp_build_sample_context *bld,
1425                        unsigned img_filter,
1426                        unsigned mip_filter,
1427                        LLVMValueRef s,
1428                        LLVMValueRef t,
1429                        LLVMValueRef r,
1430                        LLVMValueRef lod_fpart,
1431                        LLVMValueRef width0_vec,
1432                        LLVMValueRef width1_vec,
1433                        LLVMValueRef height0_vec,
1434                        LLVMValueRef height1_vec,
1435                        LLVMValueRef depth0_vec,
1436                        LLVMValueRef depth1_vec,
1437                        LLVMValueRef row_stride0_vec,
1438                        LLVMValueRef row_stride1_vec,
1439                        LLVMValueRef img_stride0_vec,
1440                        LLVMValueRef img_stride1_vec,
1441                        LLVMValueRef data_ptr0,
1442                        LLVMValueRef data_ptr1,
1443                        LLVMValueRef *colors_out)
1444 {
1445    LLVMValueRef colors0[4], colors1[4];
1446    int chan;
1447
1448    if (img_filter == PIPE_TEX_FILTER_NEAREST) {
1449       /* sample the first mipmap level */
1450       lp_build_sample_image_nearest(bld,
1451                                     width0_vec, height0_vec, depth0_vec,
1452                                     row_stride0_vec, img_stride0_vec,
1453                                     data_ptr0, s, t, r, colors0);
1454
1455       if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1456          /* sample the second mipmap level */
1457          lp_build_sample_image_nearest(bld,
1458                                        width1_vec, height1_vec, depth1_vec,
1459                                        row_stride1_vec, img_stride1_vec,
1460                                        data_ptr1, s, t, r, colors1);
1461       }
1462    }
1463    else {
1464       assert(img_filter == PIPE_TEX_FILTER_LINEAR);
1465
1466       /* sample the first mipmap level */
1467       lp_build_sample_image_linear(bld,
1468                                    width0_vec, height0_vec, depth0_vec,
1469                                    row_stride0_vec, img_stride0_vec,
1470                                    data_ptr0, s, t, r, colors0);
1471
1472       if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1473          /* sample the second mipmap level */
1474          lp_build_sample_image_linear(bld,
1475                                       width1_vec, height1_vec, depth1_vec,
1476                                       row_stride1_vec, img_stride1_vec,
1477                                       data_ptr1, s, t, r, colors1);
1478       }
1479    }
1480
1481    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1482       /* interpolate samples from the two mipmap levels */
1483       for (chan = 0; chan < 4; chan++) {
1484          colors_out[chan] = lp_build_lerp(&bld->texel_bld, lod_fpart,
1485                                           colors0[chan], colors1[chan]);
1486       }
1487    }
1488    else {
1489       /* use first/only level's colors */
1490       for (chan = 0; chan < 4; chan++) {
1491          colors_out[chan] = colors0[chan];
1492       }
1493    }
1494 }
1495
1496
1497
1498 /**
1499  * General texture sampling codegen.
1500  * This function handles texture sampling for all texture targets (1D,
1501  * 2D, 3D, cube) and all filtering modes.
1502  */
1503 static void
1504 lp_build_sample_general(struct lp_build_sample_context *bld,
1505                         unsigned unit,
1506                         LLVMValueRef s,
1507                         LLVMValueRef t,
1508                         LLVMValueRef r,
1509                         const LLVMValueRef *ddx,
1510                         const LLVMValueRef *ddy,
1511                         LLVMValueRef lod_bias, /* optional */
1512                         LLVMValueRef explicit_lod, /* optional */
1513                         LLVMValueRef width,
1514                         LLVMValueRef height,
1515                         LLVMValueRef depth,
1516                         LLVMValueRef width_vec,
1517                         LLVMValueRef height_vec,
1518                         LLVMValueRef depth_vec,
1519                         LLVMValueRef row_stride_array,
1520                         LLVMValueRef img_stride_array,
1521                         LLVMValueRef data_array,
1522                         LLVMValueRef *colors_out)
1523 {
1524    struct lp_build_context *float_bld = &bld->float_bld;
1525    const unsigned mip_filter = bld->static_state->min_mip_filter;
1526    const unsigned min_filter = bld->static_state->min_img_filter;
1527    const unsigned mag_filter = bld->static_state->mag_img_filter;
1528    const int dims = texture_dims(bld->static_state->target);
1529    LLVMValueRef lod = NULL, lod_fpart = NULL;
1530    LLVMValueRef ilevel0, ilevel1 = NULL, ilevel0_vec, ilevel1_vec = NULL;
1531    LLVMValueRef width0_vec = NULL, height0_vec = NULL, depth0_vec = NULL;
1532    LLVMValueRef width1_vec = NULL, height1_vec = NULL, depth1_vec = NULL;
1533    LLVMValueRef row_stride0_vec = NULL, row_stride1_vec = NULL;
1534    LLVMValueRef img_stride0_vec = NULL, img_stride1_vec = NULL;
1535    LLVMValueRef data_ptr0, data_ptr1 = NULL;
1536    LLVMValueRef face_ddx[4], face_ddy[4];
1537
1538    /*
1539    printf("%s mip %d  min %d  mag %d\n", __FUNCTION__,
1540           mip_filter, min_filter, mag_filter);
1541    */
1542
1543    /*
1544     * Choose cube face, recompute texcoords and derivatives for the chosen face.
1545     */
1546    if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1547       LLVMValueRef face, face_s, face_t;
1548       lp_build_cube_lookup(bld, s, t, r, &face, &face_s, &face_t);
1549       s = face_s; /* vec */
1550       t = face_t; /* vec */
1551       /* use 'r' to indicate cube face */
1552       r = lp_build_broadcast_scalar(&bld->int_coord_bld, face); /* vec */
1553
1554       /* recompute ddx, ddy using the new (s,t) face texcoords */
1555       face_ddx[0] = lp_build_ddx(&bld->coord_bld, s);
1556       face_ddx[1] = lp_build_ddx(&bld->coord_bld, t);
1557       face_ddx[2] = NULL;
1558       face_ddx[3] = NULL;
1559       face_ddy[0] = lp_build_ddy(&bld->coord_bld, s);
1560       face_ddy[1] = lp_build_ddy(&bld->coord_bld, t);
1561       face_ddy[2] = NULL;
1562       face_ddy[3] = NULL;
1563       ddx = face_ddx;
1564       ddy = face_ddy;
1565    }
1566
1567    /*
1568     * Compute the level of detail (float).
1569     */
1570    if (min_filter != mag_filter ||
1571        mip_filter != PIPE_TEX_MIPFILTER_NONE) {
1572       /* Need to compute lod either to choose mipmap levels or to
1573        * distinguish between minification/magnification with one mipmap level.
1574        */
1575       lod = lp_build_lod_selector(bld, ddx, ddy,
1576                                   lod_bias, explicit_lod,
1577                                   width, height, depth);
1578    }
1579
1580    /*
1581     * Compute integer mipmap level(s) to fetch texels from.
1582     */
1583    if (mip_filter == PIPE_TEX_MIPFILTER_NONE) {
1584       /* always use mip level 0 */
1585       if (bld->static_state->target == PIPE_TEXTURE_CUBE) {
1586          /* XXX this is a work-around for an apparent bug in LLVM 2.7.
1587           * We should be able to set ilevel0 = const(0) but that causes
1588           * bad x86 code to be emitted.
1589           */
1590          lod = lp_build_const_elem(bld->coord_bld.type, 0.0);
1591          lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1592       }
1593       else {
1594          ilevel0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
1595       }
1596    }
1597    else {
1598       assert(lod);
1599       if (mip_filter == PIPE_TEX_MIPFILTER_NEAREST) {
1600          lp_build_nearest_mip_level(bld, unit, lod, &ilevel0);
1601       }
1602       else {
1603          assert(mip_filter == PIPE_TEX_MIPFILTER_LINEAR);
1604          lp_build_linear_mip_levels(bld, unit, lod, &ilevel0, &ilevel1,
1605                                     &lod_fpart);
1606          lod_fpart = lp_build_broadcast_scalar(&bld->coord_bld, lod_fpart);
1607       }
1608    }
1609
1610    /*
1611     * Convert scalar integer mipmap levels into vectors.
1612     */
1613    ilevel0_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel0);
1614    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR)
1615       ilevel1_vec = lp_build_broadcast_scalar(&bld->int_coord_bld, ilevel1);
1616
1617    /*
1618     * Compute width, height at mipmap level 'ilevel0'
1619     */
1620    width0_vec = lp_build_minify(bld, width_vec, ilevel0_vec);
1621    if (dims >= 2) {
1622       height0_vec = lp_build_minify(bld, height_vec, ilevel0_vec);
1623       row_stride0_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1624                                                       ilevel0);
1625       if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1626          img_stride0_vec = lp_build_get_level_stride_vec(bld,
1627                                                          img_stride_array,
1628                                                          ilevel0);
1629          if (dims == 3) {
1630             depth0_vec = lp_build_minify(bld, depth_vec, ilevel0_vec);
1631          }
1632       }
1633    }
1634    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1635       /* compute width, height, depth for second mipmap level at 'ilevel1' */
1636       width1_vec = lp_build_minify(bld, width_vec, ilevel1_vec);
1637       if (dims >= 2) {
1638          height1_vec = lp_build_minify(bld, height_vec, ilevel1_vec);
1639          row_stride1_vec = lp_build_get_level_stride_vec(bld, row_stride_array,
1640                                                          ilevel1);
1641          if (dims == 3 || bld->static_state->target == PIPE_TEXTURE_CUBE) {
1642             img_stride1_vec = lp_build_get_level_stride_vec(bld,
1643                                                             img_stride_array,
1644                                                             ilevel1);
1645             if (dims ==3) {
1646                depth1_vec = lp_build_minify(bld, depth_vec, ilevel1_vec);
1647             }
1648          }
1649       }
1650    }
1651
1652    /*
1653     * Get pointer(s) to image data for mipmap level(s).
1654     */
1655    data_ptr0 = lp_build_get_mipmap_level(bld, data_array, ilevel0);
1656    if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
1657       data_ptr1 = lp_build_get_mipmap_level(bld, data_array, ilevel1);
1658    }
1659
1660    /*
1661     * Get/interpolate texture colors.
1662     */
1663    if (min_filter == mag_filter) {
1664       /* no need to distinquish between minification and magnification */
1665       lp_build_sample_mipmap(bld, min_filter, mip_filter, s, t, r, lod_fpart,
1666                              width0_vec, width1_vec,
1667                              height0_vec, height1_vec,
1668                              depth0_vec, depth1_vec,
1669                              row_stride0_vec, row_stride1_vec,
1670                              img_stride0_vec, img_stride1_vec,
1671                              data_ptr0, data_ptr1,
1672                              colors_out);
1673    }
1674    else {
1675       /* Emit conditional to choose min image filter or mag image filter
1676        * depending on the lod being >0 or <= 0, respectively.
1677        */
1678       struct lp_build_flow_context *flow_ctx;
1679       struct lp_build_if_state if_ctx;
1680       LLVMValueRef minify;
1681
1682       flow_ctx = lp_build_flow_create(bld->builder);
1683       lp_build_flow_scope_begin(flow_ctx);
1684
1685       lp_build_flow_scope_declare(flow_ctx, &colors_out[0]);
1686       lp_build_flow_scope_declare(flow_ctx, &colors_out[1]);
1687       lp_build_flow_scope_declare(flow_ctx, &colors_out[2]);
1688       lp_build_flow_scope_declare(flow_ctx, &colors_out[3]);
1689
1690       /* minify = lod > 0.0 */
1691       minify = LLVMBuildFCmp(bld->builder, LLVMRealUGE,
1692                              lod, float_bld->zero, "");
1693
1694       lp_build_if(&if_ctx, flow_ctx, bld->builder, minify);
1695       {
1696          /* Use the minification filter */
1697          lp_build_sample_mipmap(bld, min_filter, mip_filter,
1698                                 s, t, r, lod_fpart,
1699                                 width0_vec, width1_vec,
1700                                 height0_vec, height1_vec,
1701                                 depth0_vec, depth1_vec,
1702                                 row_stride0_vec, row_stride1_vec,
1703                                 img_stride0_vec, img_stride1_vec,
1704                                 data_ptr0, data_ptr1,
1705                                 colors_out);
1706       }
1707       lp_build_else(&if_ctx);
1708       {
1709          /* Use the magnification filter */
1710          lp_build_sample_mipmap(bld, mag_filter, mip_filter,
1711                                 s, t, r, lod_fpart,
1712                                 width0_vec, width1_vec,
1713                                 height0_vec, height1_vec,
1714                                 depth0_vec, depth1_vec,
1715                                 row_stride0_vec, row_stride1_vec,
1716                                 img_stride0_vec, img_stride1_vec,
1717                                 data_ptr0, data_ptr1,
1718                                 colors_out);
1719       }
1720       lp_build_endif(&if_ctx);
1721
1722       lp_build_flow_scope_end(flow_ctx);
1723       lp_build_flow_destroy(flow_ctx);
1724    }
1725 }
1726
1727
1728
1729 static void
1730 lp_build_sample_2d_linear_aos(struct lp_build_sample_context *bld,
1731                               LLVMValueRef s,
1732                               LLVMValueRef t,
1733                               LLVMValueRef width,
1734                               LLVMValueRef height,
1735                               LLVMValueRef stride_array,
1736                               LLVMValueRef data_array,
1737                               LLVMValueRef texel_out[4])
1738 {
1739    LLVMBuilderRef builder = bld->builder;
1740    struct lp_build_context i32, h16, u8n;
1741    LLVMTypeRef i32_vec_type, h16_vec_type, u8n_vec_type;
1742    LLVMValueRef i32_c8, i32_c128, i32_c255;
1743    LLVMValueRef s_ipart, s_fpart, s_fpart_lo, s_fpart_hi;
1744    LLVMValueRef t_ipart, t_fpart, t_fpart_lo, t_fpart_hi;
1745    LLVMValueRef x0, x1;
1746    LLVMValueRef y0, y1;
1747    LLVMValueRef neighbors[2][2];
1748    LLVMValueRef neighbors_lo[2][2];
1749    LLVMValueRef neighbors_hi[2][2];
1750    LLVMValueRef packed, packed_lo, packed_hi;
1751    LLVMValueRef unswizzled[4];
1752    LLVMValueRef stride;
1753
1754    assert(bld->static_state->target == PIPE_TEXTURE_2D);
1755    assert(bld->static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR);
1756    assert(bld->static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR);
1757    assert(bld->static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE);
1758
1759    lp_build_context_init(&i32, builder, lp_type_int_vec(32));
1760    lp_build_context_init(&h16, builder, lp_type_ufixed(16));
1761    lp_build_context_init(&u8n, builder, lp_type_unorm(8));
1762
1763    i32_vec_type = lp_build_vec_type(i32.type);
1764    h16_vec_type = lp_build_vec_type(h16.type);
1765    u8n_vec_type = lp_build_vec_type(u8n.type);
1766
1767    if (bld->static_state->normalized_coords) {
1768       LLVMTypeRef coord_vec_type = lp_build_vec_type(bld->coord_type);
1769       LLVMValueRef fp_width = LLVMBuildSIToFP(bld->builder, width, coord_vec_type, "");
1770       LLVMValueRef fp_height = LLVMBuildSIToFP(bld->builder, height, coord_vec_type, "");
1771       s = lp_build_mul(&bld->coord_bld, s, fp_width);
1772       t = lp_build_mul(&bld->coord_bld, t, fp_height);
1773    }
1774
1775    /* scale coords by 256 (8 fractional bits) */
1776    s = lp_build_mul_imm(&bld->coord_bld, s, 256);
1777    t = lp_build_mul_imm(&bld->coord_bld, t, 256);
1778
1779    /* convert float to int */
1780    s = LLVMBuildFPToSI(builder, s, i32_vec_type, "");
1781    t = LLVMBuildFPToSI(builder, t, i32_vec_type, "");
1782
1783    /* subtract 0.5 (add -128) */
1784    i32_c128 = lp_build_const_int_vec(i32.type, -128);
1785    s = LLVMBuildAdd(builder, s, i32_c128, "");
1786    t = LLVMBuildAdd(builder, t, i32_c128, "");
1787
1788    /* compute floor (shift right 8) */
1789    i32_c8 = lp_build_const_int_vec(i32.type, 8);
1790    s_ipart = LLVMBuildAShr(builder, s, i32_c8, "");
1791    t_ipart = LLVMBuildAShr(builder, t, i32_c8, "");
1792
1793    /* compute fractional part (AND with 0xff) */
1794    i32_c255 = lp_build_const_int_vec(i32.type, 255);
1795    s_fpart = LLVMBuildAnd(builder, s, i32_c255, "");
1796    t_fpart = LLVMBuildAnd(builder, t, i32_c255, "");
1797
1798    x0 = s_ipart;
1799    y0 = t_ipart;
1800
1801    x1 = lp_build_add(&bld->int_coord_bld, x0, bld->int_coord_bld.one);
1802    y1 = lp_build_add(&bld->int_coord_bld, y0, bld->int_coord_bld.one);
1803
1804    x0 = lp_build_sample_wrap_int(bld, x0, width,  bld->static_state->pot_width,
1805                                  bld->static_state->wrap_s);
1806    y0 = lp_build_sample_wrap_int(bld, y0, height, bld->static_state->pot_height,
1807                                  bld->static_state->wrap_t);
1808
1809    x1 = lp_build_sample_wrap_int(bld, x1, width,  bld->static_state->pot_width,
1810                                  bld->static_state->wrap_s);
1811    y1 = lp_build_sample_wrap_int(bld, y1, height, bld->static_state->pot_height,
1812                                  bld->static_state->wrap_t);
1813
1814    /*
1815     * Transform 4 x i32 in
1816     *
1817     *   s_fpart = {s0, s1, s2, s3}
1818     *
1819     * into 8 x i16
1820     *
1821     *   s_fpart = {00, s0, 00, s1, 00, s2, 00, s3}
1822     *
1823     * into two 8 x i16
1824     *
1825     *   s_fpart_lo = {s0, s0, s0, s0, s1, s1, s1, s1}
1826     *   s_fpart_hi = {s2, s2, s2, s2, s3, s3, s3, s3}
1827     *
1828     * and likewise for t_fpart. There is no risk of loosing precision here
1829     * since the fractional parts only use the lower 8bits.
1830     */
1831
1832    s_fpart = LLVMBuildBitCast(builder, s_fpart, h16_vec_type, "");
1833    t_fpart = LLVMBuildBitCast(builder, t_fpart, h16_vec_type, "");
1834
1835    {
1836       LLVMTypeRef elem_type = LLVMInt32Type();
1837       LLVMValueRef shuffles_lo[LP_MAX_VECTOR_LENGTH];
1838       LLVMValueRef shuffles_hi[LP_MAX_VECTOR_LENGTH];
1839       LLVMValueRef shuffle_lo;
1840       LLVMValueRef shuffle_hi;
1841       unsigned i, j;
1842
1843       for(j = 0; j < h16.type.length; j += 4) {
1844          unsigned subindex = util_cpu_caps.little_endian ? 0 : 1;
1845          LLVMValueRef index;
1846
1847          index = LLVMConstInt(elem_type, j/2 + subindex, 0);
1848          for(i = 0; i < 4; ++i)
1849             shuffles_lo[j + i] = index;
1850
1851          index = LLVMConstInt(elem_type, h16.type.length/2 + j/2 + subindex, 0);
1852          for(i = 0; i < 4; ++i)
1853             shuffles_hi[j + i] = index;
1854       }
1855
1856       shuffle_lo = LLVMConstVector(shuffles_lo, h16.type.length);
1857       shuffle_hi = LLVMConstVector(shuffles_hi, h16.type.length);
1858
1859       s_fpart_lo = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_lo, "");
1860       t_fpart_lo = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_lo, "");
1861       s_fpart_hi = LLVMBuildShuffleVector(builder, s_fpart, h16.undef, shuffle_hi, "");
1862       t_fpart_hi = LLVMBuildShuffleVector(builder, t_fpart, h16.undef, shuffle_hi, "");
1863    }
1864
1865    stride = lp_build_get_const_level_stride_vec(bld, stride_array, 0);
1866
1867    /*
1868     * Fetch the pixels as 4 x 32bit (rgba order might differ):
1869     *
1870     *   rgba0 rgba1 rgba2 rgba3
1871     *
1872     * bit cast them into 16 x u8
1873     *
1874     *   r0 g0 b0 a0 r1 g1 b1 a1 r2 g2 b2 a2 r3 g3 b3 a3
1875     *
1876     * unpack them into two 8 x i16:
1877     *
1878     *   r0 g0 b0 a0 r1 g1 b1 a1
1879     *   r2 g2 b2 a2 r3 g3 b3 a3
1880     *
1881     * The higher 8 bits of the resulting elements will be zero.
1882     */
1883
1884    neighbors[0][0] = lp_build_sample_packed(bld, x0, y0, stride, data_array);
1885    neighbors[0][1] = lp_build_sample_packed(bld, x1, y0, stride, data_array);
1886    neighbors[1][0] = lp_build_sample_packed(bld, x0, y1, stride, data_array);
1887    neighbors[1][1] = lp_build_sample_packed(bld, x1, y1, stride, data_array);
1888
1889    neighbors[0][0] = LLVMBuildBitCast(builder, neighbors[0][0], u8n_vec_type, "");
1890    neighbors[0][1] = LLVMBuildBitCast(builder, neighbors[0][1], u8n_vec_type, "");
1891    neighbors[1][0] = LLVMBuildBitCast(builder, neighbors[1][0], u8n_vec_type, "");
1892    neighbors[1][1] = LLVMBuildBitCast(builder, neighbors[1][1], u8n_vec_type, "");
1893
1894    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][0], &neighbors_lo[0][0], &neighbors_hi[0][0]);
1895    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[0][1], &neighbors_lo[0][1], &neighbors_hi[0][1]);
1896    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][0], &neighbors_lo[1][0], &neighbors_hi[1][0]);
1897    lp_build_unpack2(builder, u8n.type, h16.type, neighbors[1][1], &neighbors_lo[1][1], &neighbors_hi[1][1]);
1898
1899    /*
1900     * Linear interpolate with 8.8 fixed point.
1901     */
1902
1903    packed_lo = lp_build_lerp_2d(&h16,
1904                                 s_fpart_lo, t_fpart_lo,
1905                                 neighbors_lo[0][0],
1906                                 neighbors_lo[0][1],
1907                                 neighbors_lo[1][0],
1908                                 neighbors_lo[1][1]);
1909
1910    packed_hi = lp_build_lerp_2d(&h16,
1911                                 s_fpart_hi, t_fpart_hi,
1912                                 neighbors_hi[0][0],
1913                                 neighbors_hi[0][1],
1914                                 neighbors_hi[1][0],
1915                                 neighbors_hi[1][1]);
1916
1917    packed = lp_build_pack2(builder, h16.type, u8n.type, packed_lo, packed_hi);
1918
1919    /*
1920     * Convert to SoA and swizzle.
1921     */
1922
1923    lp_build_rgba8_to_f32_soa(bld->builder,
1924                              bld->texel_type,
1925                              packed, unswizzled);
1926
1927    if (util_format_is_rgba8_variant(bld->format_desc)) {
1928       lp_build_format_swizzle_soa(bld->format_desc,
1929                                   &bld->texel_bld,
1930                                   unswizzled, texel_out);
1931    } else {
1932       texel_out[0] = unswizzled[0];
1933       texel_out[1] = unswizzled[1];
1934       texel_out[2] = unswizzled[2];
1935       texel_out[3] = unswizzled[3];
1936    }
1937
1938    apply_sampler_swizzle(bld, texel_out);
1939 }
1940
1941
1942 static void
1943 lp_build_sample_compare(struct lp_build_sample_context *bld,
1944                         LLVMValueRef p,
1945                         LLVMValueRef texel[4])
1946 {
1947    struct lp_build_context *texel_bld = &bld->texel_bld;
1948    LLVMValueRef res;
1949    unsigned chan;
1950
1951    if(bld->static_state->compare_mode == PIPE_TEX_COMPARE_NONE)
1952       return;
1953
1954    /* TODO: Compare before swizzling, to avoid redundant computations */
1955    res = NULL;
1956    for(chan = 0; chan < 4; ++chan) {
1957       LLVMValueRef cmp;
1958       cmp = lp_build_cmp(texel_bld, bld->static_state->compare_func, p, texel[chan]);
1959       cmp = lp_build_select(texel_bld, cmp, texel_bld->one, texel_bld->zero);
1960
1961       if(res)
1962          res = lp_build_add(texel_bld, res, cmp);
1963       else
1964          res = cmp;
1965    }
1966
1967    assert(res);
1968    res = lp_build_mul(texel_bld, res, lp_build_const_vec(texel_bld->type, 0.25));
1969
1970    /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
1971    for(chan = 0; chan < 3; ++chan)
1972       texel[chan] = res;
1973    texel[3] = texel_bld->one;
1974 }
1975
1976
1977 /**
1978  * Just set texels to white instead of actually sampling the texture.
1979  * For debugging.
1980  */
1981 static void
1982 lp_build_sample_nop(struct lp_build_sample_context *bld,
1983                     LLVMValueRef texel_out[4])
1984 {
1985    struct lp_build_context *texel_bld = &bld->texel_bld;
1986    unsigned chan;
1987
1988    for (chan = 0; chan < 4; chan++) {
1989       /*lp_bld_mov(texel_bld, texel, texel_bld->one);*/
1990       texel_out[chan] = texel_bld->one;
1991    }
1992 }
1993
1994
1995 /**
1996  * Build texture sampling code.
1997  * 'texel' will return a vector of four LLVMValueRefs corresponding to
1998  * R, G, B, A.
1999  * \param type  vector float type to use for coords, etc.
2000  * \param ddx  partial derivatives of (s,t,r,q) with respect to x
2001  * \param ddy  partial derivatives of (s,t,r,q) with respect to y
2002  */
2003 void
2004 lp_build_sample_soa(LLVMBuilderRef builder,
2005                     const struct lp_sampler_static_state *static_state,
2006                     struct lp_sampler_dynamic_state *dynamic_state,
2007                     struct lp_type type,
2008                     unsigned unit,
2009                     unsigned num_coords,
2010                     const LLVMValueRef *coords,
2011                     const LLVMValueRef ddx[4],
2012                     const LLVMValueRef ddy[4],
2013                     LLVMValueRef lod_bias, /* optional */
2014                     LLVMValueRef explicit_lod, /* optional */
2015                     LLVMValueRef texel_out[4])
2016 {
2017    struct lp_build_sample_context bld;
2018    LLVMValueRef width, width_vec;
2019    LLVMValueRef height, height_vec;
2020    LLVMValueRef depth, depth_vec;
2021    LLVMValueRef row_stride_array, img_stride_array;
2022    LLVMValueRef data_array;
2023    LLVMValueRef s;
2024    LLVMValueRef t;
2025    LLVMValueRef r;
2026
2027    if (0) {
2028       enum pipe_format fmt = static_state->format;
2029       debug_printf("Sample from %s\n", util_format_name(fmt));
2030    }
2031
2032    /* Setup our build context */
2033    memset(&bld, 0, sizeof bld);
2034    bld.builder = builder;
2035    bld.static_state = static_state;
2036    bld.dynamic_state = dynamic_state;
2037    bld.format_desc = util_format_description(static_state->format);
2038
2039    bld.float_type = lp_type_float(32);
2040    bld.int_type = lp_type_int(32);
2041    bld.coord_type = type;
2042    bld.uint_coord_type = lp_uint_type(type);
2043    bld.int_coord_type = lp_int_type(type);
2044    bld.texel_type = type;
2045
2046    lp_build_context_init(&bld.float_bld, builder, bld.float_type);
2047    lp_build_context_init(&bld.int_bld, builder, bld.int_type);
2048    lp_build_context_init(&bld.coord_bld, builder, bld.coord_type);
2049    lp_build_context_init(&bld.uint_coord_bld, builder, bld.uint_coord_type);
2050    lp_build_context_init(&bld.int_coord_bld, builder, bld.int_coord_type);
2051    lp_build_context_init(&bld.texel_bld, builder, bld.texel_type);
2052
2053    /* Get the dynamic state */
2054    width = dynamic_state->width(dynamic_state, builder, unit);
2055    height = dynamic_state->height(dynamic_state, builder, unit);
2056    depth = dynamic_state->depth(dynamic_state, builder, unit);
2057    row_stride_array = dynamic_state->row_stride(dynamic_state, builder, unit);
2058    img_stride_array = dynamic_state->img_stride(dynamic_state, builder, unit);
2059    data_array = dynamic_state->data_ptr(dynamic_state, builder, unit);
2060    /* Note that data_array is an array[level] of pointers to texture images */
2061
2062    s = coords[0];
2063    t = coords[1];
2064    r = coords[2];
2065
2066    width_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, width);
2067    height_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, height);
2068    depth_vec = lp_build_broadcast_scalar(&bld.uint_coord_bld, depth);
2069
2070    if (0) {
2071       /* For debug: no-op texture sampling */
2072       lp_build_sample_nop(&bld, texel_out);
2073    }
2074    else if (util_format_fits_8unorm(bld.format_desc) &&
2075             bld.format_desc->nr_channels > 1 &&
2076             static_state->target == PIPE_TEXTURE_2D &&
2077             static_state->min_img_filter == PIPE_TEX_FILTER_LINEAR &&
2078             static_state->mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
2079             static_state->min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
2080             is_simple_wrap_mode(static_state->wrap_s) &&
2081             is_simple_wrap_mode(static_state->wrap_t)) {
2082       /* special case */
2083       lp_build_sample_2d_linear_aos(&bld, s, t, width_vec, height_vec,
2084                                     row_stride_array, data_array, texel_out);
2085    }
2086    else {
2087       lp_build_sample_general(&bld, unit, s, t, r, ddx, ddy,
2088                               lod_bias, explicit_lod,
2089                               width, height, depth,
2090                               width_vec, height_vec, depth_vec,
2091                               row_stride_array, img_stride_array,
2092                               data_array,
2093                               texel_out);
2094    }
2095
2096    lp_build_sample_compare(&bld, r, texel_out);
2097 }