src/gallium/drivers/llvmpipe/lp_bld_interp.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 VMware, Inc.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Position and shader input interpolation.
  32  *
  33  * @author Jose Fonseca <jfonseca@vmware.com>
  34  */
  35
  36 #include "pipe/p_shader_tokens.h"
  37 #include "util/u_debug.h"
  38 #include "util/u_memory.h"
  39 #include "util/u_math.h"
  40 #include "tgsi/tgsi_scan.h"
  41 #include "gallivm/lp_bld_debug.h"
  42 #include "gallivm/lp_bld_const.h"
  43 #include "gallivm/lp_bld_arit.h"
  44 #include "gallivm/lp_bld_swizzle.h"
  45 #include "gallivm/lp_bld_flow.h"
  46 #include "lp_bld_interp.h"
  47
  48
  49 /*
  50  * The shader JIT function operates on blocks of quads.
  51  * Each block has 2x2 quads and each quad has 2x2 pixels.
  52  *
  53  * We iterate over the quads in order 0, 1, 2, 3:
  54  *
  55  * #################
  56  * #   |   #   |   #
  57  * #---0---#---1---#
  58  * #   |   #   |   #
  59  * #################
  60  * #   |   #   |   #
  61  * #---2---#---3---#
  62  * #   |   #   |   #
  63  * #################
  64  *
  65  * If we iterate over multiple quads at once, quads 01 and 23 are processed
  66  * together.
  67  *
  68  * Within each quad, we have four pixels which are represented in SOA
  69  * order:
  70  *
  71  * #########
  72  * # 0 | 1 #
  73  * #---+---#
  74  * # 2 | 3 #
  75  * #########
  76  *
  77  * So the green channel (for example) of the four pixels is stored in
  78  * a single vector register: {g0, g1, g2, g3}.
  79  * The order stays the same even with multiple quads:
  80  * 0 1 4 5
  81  * 2 3 6 7
  82  * is stored as g0..g7
  83  */
  84
  85
  86 /**
  87  * Do one perspective divide per quad.
  88  *
  89  * For perspective interpolation, the final attribute value is given
  90  *
  91  *  a' = a/w = a * oow
  92  *
  93  * where
  94  *
  95  *  a = a0 + dadx*x + dady*y
  96  *  w = w0 + dwdx*x + dwdy*y
  97  *  oow = 1/w = 1/(w0 + dwdx*x + dwdy*y)
  98  *
  99  * Instead of computing the division per pixel, with this macro we compute the
 100  * division on the upper left pixel of each quad, and use a linear
 101  * approximation in the remaining pixels, given by:
 102  *
 103  *  da'dx = (dadx - dwdx*a)*oow
 104  *  da'dy = (dady - dwdy*a)*oow
 105  *
 106  * Ironically, this actually makes things slower -- probably because the
 107  * divide hardware unit is rarely used, whereas the multiply unit is typically
 108  * already saturated.
 109  */
 110 #define PERSPECTIVE_DIVIDE_PER_QUAD 0
 111
 112
 113 static const unsigned char quad_offset_x[16] = {0, 1, 0, 1, 2, 3, 2, 3, 0, 1, 0, 1, 2, 3, 2, 3};
 114 static const unsigned char quad_offset_y[16] = {0, 0, 1, 1, 0, 0, 1, 1, 2, 2, 3, 3, 2, 2, 3, 3};
 115
 116
 117 static void
 118 attrib_name(LLVMValueRef val, unsigned attrib, unsigned chan, const char *suffix)
 119 {
 120    if(attrib == 0)
 121       lp_build_name(val, "pos.%c%s", "xyzw"[chan], suffix);
 122    else
 123       lp_build_name(val, "input%u.%c%s", attrib - 1, "xyzw"[chan], suffix);
 124 }
 125
 126 static void
 127 calc_offsets(struct lp_build_context *coeff_bld,
 128              unsigned quad_start_index,
 129              LLVMValueRef *pixoffx,
 130              LLVMValueRef *pixoffy)
 131 {
 132    unsigned i;
 133    unsigned num_pix = coeff_bld->type.length;
 134    struct gallivm_state *gallivm = coeff_bld->gallivm;
 135    LLVMBuilderRef builder = coeff_bld->gallivm->builder;
 136    LLVMValueRef nr, pixxf, pixyf;
 137
 138    *pixoffx = coeff_bld->undef;
 139    *pixoffy = coeff_bld->undef;
 140
 141    for (i = 0; i < num_pix; i++) {
 142       nr = lp_build_const_int32(gallivm, i);
 143       pixxf = lp_build_const_float(gallivm, quad_offset_x[i % num_pix] +
 144                                    (quad_start_index & 1) * 2);
 145       pixyf = lp_build_const_float(gallivm, quad_offset_y[i % num_pix] +
 146                                    (quad_start_index & 2));
 147       *pixoffx = LLVMBuildInsertElement(builder, *pixoffx, pixxf, nr, "");
 148       *pixoffy = LLVMBuildInsertElement(builder, *pixoffy, pixyf, nr, "");
 149    }
 150 }
 151
 152
 153 /* Much easier, and significantly less instructions in the per-stamp
 154  * part (less than half) but overall more instructions so a loss if
 155  * most quads are active. Might be a win though with larger vectors.
 156  * No ability to do per-quad divide (doable but not implemented)
 157  * Could be made to work with passed in pixel offsets (i.e. active quad merging).
 158  */
 159 static void
 160 coeffs_init_simple(struct lp_build_interp_soa_context *bld,
 161                    LLVMValueRef a0_ptr,
 162                    LLVMValueRef dadx_ptr,
 163                    LLVMValueRef dady_ptr)
 164 {
 165    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 166    struct lp_build_context *setup_bld = &bld->setup_bld;
 167    struct gallivm_state *gallivm = coeff_bld->gallivm;
 168    LLVMBuilderRef builder = gallivm->builder;
 169    unsigned attrib;
 170
 171    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 172       /*
 173        * always fetch all 4 values for performance/simplicity
 174        * Note: we do that here because it seems to generate better
 175        * code. It generates a lot of moves initially but less
 176        * moves later. As far as I can tell this looks like a
 177        * llvm issue, instead of simply reloading the values from
 178        * the passed in pointers it if it runs out of registers
 179        * it spills/reloads them. Maybe some optimization passes
 180        * would help.
 181        * Might want to investigate this again later.
 182        */
 183       const unsigned interp = bld->interp[attrib];
 184       LLVMValueRef index = lp_build_const_int32(gallivm,
 185                                 attrib * TGSI_NUM_CHANNELS);
 186       LLVMValueRef ptr;
 187       LLVMValueRef dadxaos = setup_bld->zero;
 188       LLVMValueRef dadyaos = setup_bld->zero;
 189       LLVMValueRef a0aos = setup_bld->zero;
 190
 191       switch (interp) {
 192       case LP_INTERP_PERSPECTIVE:
 193          /* fall-through */
 194
 195       case LP_INTERP_LINEAR:
 196          ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
 197          ptr = LLVMBuildBitCast(builder, ptr,
 198                LLVMPointerType(setup_bld->vec_type, 0), "");
 199          dadxaos = LLVMBuildLoad(builder, ptr, "");
 200
 201          ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
 202          ptr = LLVMBuildBitCast(builder, ptr,
 203                LLVMPointerType(setup_bld->vec_type, 0), "");
 204          dadyaos = LLVMBuildLoad(builder, ptr, "");
 205
 206          attrib_name(dadxaos, attrib, 0, ".dadxaos");
 207          attrib_name(dadyaos, attrib, 0, ".dadyaos");
 208          /* fall-through */
 209
 210       case LP_INTERP_CONSTANT:
 211       case LP_INTERP_FACING:
 212          ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
 213          ptr = LLVMBuildBitCast(builder, ptr,
 214                LLVMPointerType(setup_bld->vec_type, 0), "");
 215          a0aos = LLVMBuildLoad(builder, ptr, "");
 216          attrib_name(a0aos, attrib, 0, ".a0aos");
 217          break;
 218
 219       case LP_INTERP_POSITION:
 220          /* Nothing to do as the position coeffs are already setup in slot 0 */
 221          continue;
 222
 223       default:
 224          assert(0);
 225          break;
 226       }
 227       bld->a0aos[attrib] = a0aos;
 228       bld->dadxaos[attrib] = dadxaos;
 229       bld->dadyaos[attrib] = dadyaos;
 230    }
 231 }
 232
 233 /**
 234  * Interpolate the shader input attribute values.
 235  * This is called for each (group of) quad(s).
 236  */
 237 static void
 238 attribs_update_simple(struct lp_build_interp_soa_context *bld,
 239                       struct gallivm_state *gallivm,
 240                       LLVMValueRef loop_iter,
 241                       int start,
 242                       int end)
 243 {
 244    LLVMBuilderRef builder = gallivm->builder;
 245    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 246    struct lp_build_context *setup_bld = &bld->setup_bld;
 247    LLVMValueRef oow = NULL;
 248    unsigned attrib;
 249    LLVMValueRef pixoffx;
 250    LLVMValueRef pixoffy;
 251    LLVMValueRef ptr;
 252
 253    /* could do this with code-generated passed in pixel offsets too */
 254
 255    assert(loop_iter);
 256    ptr = LLVMBuildGEP(builder, bld->xoffset_store, &loop_iter, 1, "");
 257    pixoffx = LLVMBuildLoad(builder, ptr, "");
 258    ptr = LLVMBuildGEP(builder, bld->yoffset_store, &loop_iter, 1, "");
 259    pixoffy = LLVMBuildLoad(builder, ptr, "");
 260
 261    pixoffx = LLVMBuildFAdd(builder, pixoffx,
 262                            lp_build_broadcast_scalar(coeff_bld, bld->x), "");
 263    pixoffy = LLVMBuildFAdd(builder, pixoffy,
 264                            lp_build_broadcast_scalar(coeff_bld, bld->y), "");
 265
 266    for (attrib = start; attrib < end; attrib++) {
 267       const unsigned mask = bld->mask[attrib];
 268       const unsigned interp = bld->interp[attrib];
 269       unsigned chan;
 270
 271       for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
 272          if (mask & (1 << chan)) {
 273             LLVMValueRef index;
 274             LLVMValueRef dadx = coeff_bld->zero;
 275             LLVMValueRef dady = coeff_bld->zero;
 276             LLVMValueRef a = coeff_bld->zero;
 277
 278             index = lp_build_const_int32(gallivm, chan);
 279             switch (interp) {
 280             case LP_INTERP_PERSPECTIVE:
 281                /* fall-through */
 282
 283             case LP_INTERP_LINEAR:
 284                if (attrib == 0 && chan == 0) {
 285                   dadx = coeff_bld->one;
 286                   if (bld->pos_offset) {
 287                      a = lp_build_const_vec(gallivm, coeff_bld->type, bld->pos_offset);
 288                   }
 289                }
 290                else if (attrib == 0 && chan == 1) {
 291                   dady = coeff_bld->one;
 292                   if (bld->pos_offset) {
 293                      a = lp_build_const_vec(gallivm, coeff_bld->type, bld->pos_offset);
 294                   }
 295                }
 296                else {
 297                   dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
 298                                                     coeff_bld->type, bld->dadxaos[attrib],
 299                                                     index);
 300                   dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
 301                                                     coeff_bld->type, bld->dadyaos[attrib],
 302                                                     index);
 303                   a = lp_build_extract_broadcast(gallivm, setup_bld->type,
 304                                                  coeff_bld->type, bld->a0aos[attrib],
 305                                                  index);
 306                }
 307                /*
 308                 * a = a0 + (x * dadx + y * dady)
 309                 */
 310                dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
 311                dady = LLVMBuildFMul(builder, dady, pixoffy, "");
 312                a = LLVMBuildFAdd(builder, a, dadx, "");
 313                a = LLVMBuildFAdd(builder, a, dady, "");
 314
 315                if (interp == LP_INTERP_PERSPECTIVE) {
 316                   if (oow == NULL) {
 317                      LLVMValueRef w = bld->attribs[0][3];
 318                      assert(attrib != 0);
 319                      assert(bld->mask[0] & TGSI_WRITEMASK_W);
 320                      oow = lp_build_rcp(coeff_bld, w);
 321                   }
 322                   a = lp_build_mul(coeff_bld, a, oow);
 323                }
 324                break;
 325
 326             case LP_INTERP_CONSTANT:
 327             case LP_INTERP_FACING:
 328                a = lp_build_extract_broadcast(gallivm, setup_bld->type,
 329                                               coeff_bld->type, bld->a0aos[attrib],
 330                                               index);
 331                break;
 332
 333             case LP_INTERP_POSITION:
 334                assert(attrib > 0);
 335                a = bld->attribs[0][chan];
 336                break;
 337
 338             default:
 339                assert(0);
 340                break;
 341             }
 342
 343             if ((attrib == 0) && (chan == 2)){
 344                /* FIXME: Depth values can exceed 1.0, due to the fact that
 345                 * setup interpolation coefficients refer to (0,0) which causes
 346                 * precision loss. So we must clamp to 1.0 here to avoid artifacts
 347                 */
 348                a = lp_build_min(coeff_bld, a, coeff_bld->one);
 349             }
 350             bld->attribs[attrib][chan] = a;
 351          }
 352       }
 353    }
 354 }
 355
 356 /**
 357  * Initialize the bld->a, dadq fields.  This involves fetching
 358  * those values from the arrays which are passed into the JIT function.
 359  */
 360 static void
 361 coeffs_init(struct lp_build_interp_soa_context *bld,
 362             LLVMValueRef a0_ptr,
 363             LLVMValueRef dadx_ptr,
 364             LLVMValueRef dady_ptr)
 365 {
 366    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 367    struct lp_build_context *setup_bld = &bld->setup_bld;
 368    struct gallivm_state *gallivm = coeff_bld->gallivm;
 369    LLVMBuilderRef builder = gallivm->builder;
 370    LLVMValueRef pixoffx, pixoffy;
 371    unsigned attrib;
 372    unsigned chan;
 373    unsigned i;
 374
 375    pixoffx = coeff_bld->undef;
 376    pixoffy = coeff_bld->undef;
 377    for (i = 0; i < coeff_bld->type.length; i++) {
 378       LLVMValueRef nr = lp_build_const_int32(gallivm, i);
 379       LLVMValueRef pixxf = lp_build_const_float(gallivm, quad_offset_x[i]);
 380       LLVMValueRef pixyf = lp_build_const_float(gallivm, quad_offset_y[i]);
 381       pixoffx = LLVMBuildInsertElement(builder, pixoffx, pixxf, nr, "");
 382       pixoffy = LLVMBuildInsertElement(builder, pixoffy, pixyf, nr, "");
 383    }
 384
 385
 386    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 387       const unsigned mask = bld->mask[attrib];
 388       const unsigned interp = bld->interp[attrib];
 389       LLVMValueRef index = lp_build_const_int32(gallivm,
 390                                 attrib * TGSI_NUM_CHANNELS);
 391       LLVMValueRef ptr;
 392       LLVMValueRef dadxaos = setup_bld->zero;
 393       LLVMValueRef dadyaos = setup_bld->zero;
 394       LLVMValueRef a0aos = setup_bld->zero;
 395
 396       /* always fetch all 4 values for performance/simplicity */
 397       switch (interp) {
 398       case LP_INTERP_PERSPECTIVE:
 399          /* fall-through */
 400
 401       case LP_INTERP_LINEAR:
 402          ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
 403          ptr = LLVMBuildBitCast(builder, ptr,
 404                LLVMPointerType(setup_bld->vec_type, 0), "");
 405          dadxaos = LLVMBuildLoad(builder, ptr, "");
 406
 407          ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
 408          ptr = LLVMBuildBitCast(builder, ptr,
 409                LLVMPointerType(setup_bld->vec_type, 0), "");
 410          dadyaos = LLVMBuildLoad(builder, ptr, "");
 411
 412          attrib_name(dadxaos, attrib, 0, ".dadxaos");
 413          attrib_name(dadyaos, attrib, 0, ".dadyaos");
 414          /* fall-through */
 415
 416       case LP_INTERP_CONSTANT:
 417       case LP_INTERP_FACING:
 418          ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
 419          ptr = LLVMBuildBitCast(builder, ptr,
 420                LLVMPointerType(setup_bld->vec_type, 0), "");
 421          a0aos = LLVMBuildLoad(builder, ptr, "");
 422          attrib_name(a0aos, attrib, 0, ".a0aos");
 423          break;
 424
 425       case LP_INTERP_POSITION:
 426          /* Nothing to do as the position coeffs are already setup in slot 0 */
 427          continue;
 428
 429       default:
 430          assert(0);
 431          break;
 432       }
 433
 434       /*
 435        * a = a0 + (x * dadx + y * dady)
 436        * a0aos is the attrib value at top left corner of stamp
 437        */
 438       if (interp != LP_INTERP_CONSTANT &&
 439           interp != LP_INTERP_FACING) {
 440          LLVMValueRef axaos, ayaos;
 441          axaos = LLVMBuildFMul(builder, lp_build_broadcast_scalar(setup_bld, bld->x),
 442                                dadxaos, "");
 443          ayaos = LLVMBuildFMul(builder, lp_build_broadcast_scalar(setup_bld, bld->y),
 444                                dadyaos, "");
 445          a0aos = LLVMBuildFAdd(builder, a0aos, ayaos, "");
 446          a0aos = LLVMBuildFAdd(builder, a0aos, axaos, "");
 447       }
 448
 449       /*
 450        * dadq = {0, dadx, dady, dadx + dady}
 451        * for two quads (side by side) this is:
 452        * {0, dadx, dady, dadx+dady, 2*dadx, 2*dadx+dady, 3*dadx+dady}
 453        */
 454       for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 455          /* this generates a CRAPLOAD of shuffles... */
 456          if (mask & (1 << chan)) {
 457             LLVMValueRef dadx, dady;
 458             LLVMValueRef dadq, dadq2;
 459             LLVMValueRef a;
 460             LLVMValueRef chan_index = lp_build_const_int32(gallivm, chan);
 461
 462             if (attrib == 0 && chan == 0) {
 463                a = bld->x;
 464                if (bld->pos_offset) {
 465                   a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
 466                }
 467                a = lp_build_broadcast_scalar(coeff_bld, a);
 468                dadx = coeff_bld->one;
 469                dady = coeff_bld->zero;
 470             }
 471             else if (attrib == 0 && chan == 1) {
 472                a = bld->y;
 473                if (bld->pos_offset) {
 474                   a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
 475                }
 476                a = lp_build_broadcast_scalar(coeff_bld, a);
 477                dady = coeff_bld->one;
 478                dadx = coeff_bld->zero;
 479             }
 480             else {
 481                dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
 482                                               coeff_bld->type, dadxaos, chan_index);
 483                dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
 484                                               coeff_bld->type, dadyaos, chan_index);
 485
 486                /*
 487                 * a = {a, a, a, a}
 488                 */
 489                a = lp_build_extract_broadcast(gallivm, setup_bld->type,
 490                                               coeff_bld->type, a0aos, chan_index);
 491             }
 492
 493             dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
 494             dady = LLVMBuildFMul(builder, dady, pixoffy, "");
 495             dadq = LLVMBuildFAdd(builder, dadx, dady, "");
 496
 497             /*
 498              * Compute the attrib values on the upper-left corner of each
 499              * group of quads.
 500              * Note that if we process 2 quads at once this doesn't
 501              * really exactly to what we want.
 502              * We need to access elem 0 and 2 respectively later if we process
 503              * 2 quads at once.
 504              */
 505
 506             if (interp != LP_INTERP_CONSTANT &&
 507                 interp != LP_INTERP_FACING) {
 508                dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
 509                a = LLVMBuildFAdd(builder, a, dadq2, "");
 510             }
 511
 512 #if PERSPECTIVE_DIVIDE_PER_QUAD
 513             /*
 514              * a *= 1 / w
 515              */
 516
 517             /*
 518              * XXX since we're only going to access elements 0,2 out of 8
 519              * if we have 8-wide vectors we should do the division only 4-wide.
 520              * a is really a 2-elements in a 4-wide vector disguised as 8-wide
 521              * in this case.
 522              */
 523             if (interp == LP_INTERP_PERSPECTIVE) {
 524                LLVMValueRef w = bld->a[0][3];
 525                assert(attrib != 0);
 526                assert(bld->mask[0] & TGSI_WRITEMASK_W);
 527                if (!bld->oow) {
 528                   bld->oow = lp_build_rcp(coeff_bld, w);
 529                   lp_build_name(bld->oow, "oow");
 530                }
 531                a = lp_build_mul(coeff_bld, a, bld->oow);
 532             }
 533 #endif
 534
 535             attrib_name(a, attrib, chan, ".a");
 536             attrib_name(dadq, attrib, chan, ".dadq");
 537
 538             bld->a[attrib][chan] = lp_build_alloca(gallivm,
 539                                                    LLVMTypeOf(a), "");
 540             LLVMBuildStore(builder, a, bld->a[attrib][chan]);
 541             bld->dadq[attrib][chan] = dadq;
 542          }
 543       }
 544    }
 545 }
 546
 547
 548 /**
 549  * Increment the shader input attribute values.
 550  * This is called when we move from one quad to the next.
 551  */
 552 static void
 553 attribs_update(struct lp_build_interp_soa_context *bld,
 554                struct gallivm_state *gallivm,
 555                LLVMValueRef loop_iter,
 556                int start,
 557                int end)
 558 {
 559    LLVMBuilderRef builder = gallivm->builder;
 560    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 561    LLVMValueRef oow = NULL;
 562    unsigned attrib;
 563    unsigned chan;
 564
 565    for(attrib = start; attrib < end; ++attrib) {
 566       const unsigned mask = bld->mask[attrib];
 567       const unsigned interp = bld->interp[attrib];
 568       for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 569          if(mask & (1 << chan)) {
 570             LLVMValueRef a;
 571             if (interp == LP_INTERP_CONSTANT ||
 572                 interp == LP_INTERP_FACING) {
 573                a = LLVMBuildLoad(builder, bld->a[attrib][chan], "");
 574             }
 575             else if (interp == LP_INTERP_POSITION) {
 576                assert(attrib > 0);
 577                a = bld->attribs[0][chan];
 578             }
 579             else {
 580                LLVMValueRef dadq;
 581
 582                a = bld->a[attrib][chan];
 583
 584                /*
 585                 * Broadcast the attribute value for this quad into all elements
 586                 */
 587
 588                {
 589                   /* stored as vector load as float */
 590                   LLVMTypeRef ptr_type = LLVMPointerType(LLVMFloatTypeInContext(
 591                                                             gallivm->context), 0);
 592                   LLVMValueRef ptr;
 593                   a = LLVMBuildBitCast(builder, a, ptr_type, "");
 594                   ptr = LLVMBuildGEP(builder, a, &loop_iter, 1, "");
 595                   a = LLVMBuildLoad(builder, ptr, "");
 596                   a = lp_build_broadcast_scalar(&bld->coeff_bld, a);
 597                }
 598
 599                /*
 600                 * Get the derivatives.
 601                 */
 602
 603                dadq = bld->dadq[attrib][chan];
 604
 605 #if PERSPECTIVE_DIVIDE_PER_QUAD
 606                if (interp == LP_INTERP_PERSPECTIVE) {
 607                   LLVMValueRef dwdq = bld->dadq[0][3];
 608
 609                   if (oow == NULL) {
 610                      assert(bld->oow);
 611                      oow = LLVMBuildShuffleVector(coeff_bld->builder,
 612                                                   bld->oow, coeff_bld->undef,
 613                                                   shuffle, "");
 614                   }
 615
 616                   dadq = lp_build_sub(coeff_bld,
 617                                       dadq,
 618                                       lp_build_mul(coeff_bld, a, dwdq));
 619                   dadq = lp_build_mul(coeff_bld, dadq, oow);
 620                }
 621 #endif
 622
 623                /*
 624                 * Add the derivatives
 625                 */
 626
 627                a = lp_build_add(coeff_bld, a, dadq);
 628
 629 #if !PERSPECTIVE_DIVIDE_PER_QUAD
 630                if (interp == LP_INTERP_PERSPECTIVE) {
 631                   if (oow == NULL) {
 632                      LLVMValueRef w = bld->attribs[0][3];
 633                      assert(attrib != 0);
 634                      assert(bld->mask[0] & TGSI_WRITEMASK_W);
 635                      oow = lp_build_rcp(coeff_bld, w);
 636                   }
 637                   a = lp_build_mul(coeff_bld, a, oow);
 638                }
 639 #endif
 640
 641                if (attrib == 0 && chan == 2) {
 642                   /* FIXME: Depth values can exceed 1.0, due to the fact that
 643                    * setup interpolation coefficients refer to (0,0) which causes
 644                    * precision loss. So we must clamp to 1.0 here to avoid artifacts
 645                    */
 646                   a = lp_build_min(coeff_bld, a, coeff_bld->one);
 647                }
 648
 649                attrib_name(a, attrib, chan, "");
 650             }
 651             bld->attribs[attrib][chan] = a;
 652          }
 653       }
 654    }
 655 }
 656
 657
 658 /**
 659  * Generate the position vectors.
 660  *
 661  * Parameter x0, y0 are the integer values with upper left coordinates.
 662  */
 663 static void
 664 pos_init(struct lp_build_interp_soa_context *bld,
 665          LLVMValueRef x0,
 666          LLVMValueRef y0)
 667 {
 668    LLVMBuilderRef builder = bld->coeff_bld.gallivm->builder;
 669    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 670
 671    bld->x = LLVMBuildSIToFP(builder, x0, coeff_bld->elem_type, "");
 672    bld->y = LLVMBuildSIToFP(builder, y0, coeff_bld->elem_type, "");
 673 }
 674
 675
 676 /**
 677  * Initialize fragment shader input attribute info.
 678  */
 679 void
 680 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
 681                          struct gallivm_state *gallivm,
 682                          unsigned num_inputs,
 683                          const struct lp_shader_input *inputs,
 684                          boolean pixel_center_integer,
 685                          LLVMBuilderRef builder,
 686                          struct lp_type type,
 687                          LLVMValueRef a0_ptr,
 688                          LLVMValueRef dadx_ptr,
 689                          LLVMValueRef dady_ptr,
 690                          LLVMValueRef x0,
 691                          LLVMValueRef y0)
 692 {
 693    struct lp_type coeff_type;
 694    struct lp_type setup_type;
 695    unsigned attrib;
 696    unsigned chan;
 697
 698    memset(bld, 0, sizeof *bld);
 699
 700    memset(&coeff_type, 0, sizeof coeff_type);
 701    coeff_type.floating = TRUE;
 702    coeff_type.sign = TRUE;
 703    coeff_type.width = 32;
 704    coeff_type.length = type.length;
 705
 706    memset(&setup_type, 0, sizeof setup_type);
 707    setup_type.floating = TRUE;
 708    setup_type.sign = TRUE;
 709    setup_type.width = 32;
 710    setup_type.length = TGSI_NUM_CHANNELS;
 711
 712
 713    /* XXX: we don't support interpolating into any other types */
 714    assert(memcmp(&coeff_type, &type, sizeof coeff_type) == 0);
 715
 716    lp_build_context_init(&bld->coeff_bld, gallivm, coeff_type);
 717    lp_build_context_init(&bld->setup_bld, gallivm, setup_type);
 718
 719    /* For convenience */
 720    bld->pos = bld->attribs[0];
 721    bld->inputs = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) bld->attribs[1];
 722
 723    /* Position */
 724    bld->mask[0] = TGSI_WRITEMASK_XYZW;
 725    bld->interp[0] = LP_INTERP_LINEAR;
 726
 727    /* Inputs */
 728    for (attrib = 0; attrib < num_inputs; ++attrib) {
 729       bld->mask[1 + attrib] = inputs[attrib].usage_mask;
 730       bld->interp[1 + attrib] = inputs[attrib].interp;
 731    }
 732    bld->num_attribs = 1 + num_inputs;
 733
 734    /* Ensure all masked out input channels have a valid value */
 735    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 736       for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 737          bld->attribs[attrib][chan] = bld->coeff_bld.undef;
 738       }
 739    }
 740
 741    if (pixel_center_integer) {
 742       bld->pos_offset = 0.0;
 743    } else {
 744       bld->pos_offset = 0.5;
 745    }
 746
 747    pos_init(bld, x0, y0);
 748
 749    if (coeff_type.length > 4) {
 750       bld->simple_interp = TRUE;
 751       {
 752          /* XXX this should use a global static table */
 753          unsigned i;
 754          unsigned num_loops = 16 / type.length;
 755          LLVMValueRef pixoffx, pixoffy, index;
 756          LLVMValueRef ptr;
 757
 758          bld->xoffset_store = lp_build_array_alloca(gallivm,
 759                                                     lp_build_vec_type(gallivm, type),
 760                                                     lp_build_const_int32(gallivm, num_loops),
 761                                                     "");
 762          bld->yoffset_store = lp_build_array_alloca(gallivm,
 763                                                     lp_build_vec_type(gallivm, type),
 764                                                     lp_build_const_int32(gallivm, num_loops),
 765                                                     "");
 766          for (i = 0; i < num_loops; i++) {
 767             index = lp_build_const_int32(gallivm, i);
 768             calc_offsets(&bld->coeff_bld, i*type.length/4, &pixoffx, &pixoffy);
 769             ptr = LLVMBuildGEP(builder, bld->xoffset_store, &index, 1, "");
 770             LLVMBuildStore(builder, pixoffx, ptr);
 771             ptr = LLVMBuildGEP(builder, bld->yoffset_store, &index, 1, "");
 772             LLVMBuildStore(builder, pixoffy, ptr);
 773          }
 774       }
 775       coeffs_init_simple(bld, a0_ptr, dadx_ptr, dady_ptr);
 776    }
 777    else {
 778       bld->simple_interp = FALSE;
 779       coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
 780    }
 781
 782 }
 783
 784
 785 /*
 786  * Advance the position and inputs to the given quad within the block.
 787  */
 788
 789 void
 790 lp_build_interp_soa_update_inputs_dyn(struct lp_build_interp_soa_context *bld,
 791                                       struct gallivm_state *gallivm,
 792                                       LLVMValueRef quad_start_index)
 793 {
 794    if (bld->simple_interp) {
 795       attribs_update_simple(bld, gallivm, quad_start_index, 1, bld->num_attribs);
 796    }
 797    else {
 798       attribs_update(bld, gallivm, quad_start_index, 1, bld->num_attribs);
 799    }
 800 }
 801
 802 void
 803 lp_build_interp_soa_update_pos_dyn(struct lp_build_interp_soa_context *bld,
 804                                    struct gallivm_state *gallivm,
 805                                    LLVMValueRef quad_start_index)
 806 {
 807    if (bld->simple_interp) {
 808       attribs_update_simple(bld, gallivm, quad_start_index, 0, 1);
 809    }
 810    else {
 811       attribs_update(bld, gallivm, quad_start_index, 0, 1);
 812    }
 813 }
 814