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                a = lp_build_fmuladd(builder, dadx, pixoffx, a);
 311                a = lp_build_fmuladd(builder, dady, pixoffy, a);
 312
 313                if (interp == LP_INTERP_PERSPECTIVE) {
 314                   if (oow == NULL) {
 315                      LLVMValueRef w = bld->attribs[0][3];
 316                      assert(attrib != 0);
 317                      assert(bld->mask[0] & TGSI_WRITEMASK_W);
 318                      oow = lp_build_rcp(coeff_bld, w);
 319                   }
 320                   a = lp_build_mul(coeff_bld, a, oow);
 321                }
 322                break;
 323
 324             case LP_INTERP_CONSTANT:
 325             case LP_INTERP_FACING:
 326                a = lp_build_extract_broadcast(gallivm, setup_bld->type,
 327                                               coeff_bld->type, bld->a0aos[attrib],
 328                                               index);
 329                break;
 330
 331             case LP_INTERP_POSITION:
 332                assert(attrib > 0);
 333                a = bld->attribs[0][chan];
 334                break;
 335
 336             default:
 337                assert(0);
 338                break;
 339             }
 340
 341             if ((attrib == 0) && (chan == 2) && !bld->depth_clamp){
 342                /* FIXME: Depth values can exceed 1.0, due to the fact that
 343                 * setup interpolation coefficients refer to (0,0) which causes
 344                 * precision loss. So we must clamp to 1.0 here to avoid artifacts.
 345                 * Note though values outside [0,1] are perfectly valid with
 346                 * depth clip disabled.
 347                 * XXX: If depth clip is disabled but we force depth clamp
 348                 * we may get values larger than 1.0 in the fs (but not in
 349                 * depth test). Not sure if that's an issue...
 350                 * Also, on a similar note, it is not obvious if the depth values
 351                 * appearing in fs (with depth clip disabled) should be clamped
 352                 * to [0,1], clamped to near/far or not be clamped at all...
 353                 */
 354                a = lp_build_min(coeff_bld, a, coeff_bld->one);
 355             }
 356             bld->attribs[attrib][chan] = a;
 357          }
 358       }
 359    }
 360 }
 361
 362 /**
 363  * Initialize the bld->a, dadq fields.  This involves fetching
 364  * those values from the arrays which are passed into the JIT function.
 365  */
 366 static void
 367 coeffs_init(struct lp_build_interp_soa_context *bld,
 368             LLVMValueRef a0_ptr,
 369             LLVMValueRef dadx_ptr,
 370             LLVMValueRef dady_ptr)
 371 {
 372    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 373    struct lp_build_context *setup_bld = &bld->setup_bld;
 374    struct gallivm_state *gallivm = coeff_bld->gallivm;
 375    LLVMBuilderRef builder = gallivm->builder;
 376    LLVMValueRef pixoffx, pixoffy;
 377    unsigned attrib;
 378    unsigned chan;
 379    unsigned i;
 380
 381    pixoffx = coeff_bld->undef;
 382    pixoffy = coeff_bld->undef;
 383    for (i = 0; i < coeff_bld->type.length; i++) {
 384       LLVMValueRef nr = lp_build_const_int32(gallivm, i);
 385       LLVMValueRef pixxf = lp_build_const_float(gallivm, quad_offset_x[i]);
 386       LLVMValueRef pixyf = lp_build_const_float(gallivm, quad_offset_y[i]);
 387       pixoffx = LLVMBuildInsertElement(builder, pixoffx, pixxf, nr, "");
 388       pixoffy = LLVMBuildInsertElement(builder, pixoffy, pixyf, nr, "");
 389    }
 390
 391
 392    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 393       const unsigned mask = bld->mask[attrib];
 394       const unsigned interp = bld->interp[attrib];
 395       LLVMValueRef index = lp_build_const_int32(gallivm,
 396                                 attrib * TGSI_NUM_CHANNELS);
 397       LLVMValueRef ptr;
 398       LLVMValueRef dadxaos = setup_bld->zero;
 399       LLVMValueRef dadyaos = setup_bld->zero;
 400       LLVMValueRef a0aos = setup_bld->zero;
 401
 402       /* always fetch all 4 values for performance/simplicity */
 403       switch (interp) {
 404       case LP_INTERP_PERSPECTIVE:
 405          /* fall-through */
 406
 407       case LP_INTERP_LINEAR:
 408          ptr = LLVMBuildGEP(builder, dadx_ptr, &index, 1, "");
 409          ptr = LLVMBuildBitCast(builder, ptr,
 410                LLVMPointerType(setup_bld->vec_type, 0), "");
 411          dadxaos = LLVMBuildLoad(builder, ptr, "");
 412
 413          ptr = LLVMBuildGEP(builder, dady_ptr, &index, 1, "");
 414          ptr = LLVMBuildBitCast(builder, ptr,
 415                LLVMPointerType(setup_bld->vec_type, 0), "");
 416          dadyaos = LLVMBuildLoad(builder, ptr, "");
 417
 418          attrib_name(dadxaos, attrib, 0, ".dadxaos");
 419          attrib_name(dadyaos, attrib, 0, ".dadyaos");
 420          /* fall-through */
 421
 422       case LP_INTERP_CONSTANT:
 423       case LP_INTERP_FACING:
 424          ptr = LLVMBuildGEP(builder, a0_ptr, &index, 1, "");
 425          ptr = LLVMBuildBitCast(builder, ptr,
 426                LLVMPointerType(setup_bld->vec_type, 0), "");
 427          a0aos = LLVMBuildLoad(builder, ptr, "");
 428          attrib_name(a0aos, attrib, 0, ".a0aos");
 429          break;
 430
 431       case LP_INTERP_POSITION:
 432          /* Nothing to do as the position coeffs are already setup in slot 0 */
 433          continue;
 434
 435       default:
 436          assert(0);
 437          break;
 438       }
 439
 440       /*
 441        * a = a0 + (x * dadx + y * dady)
 442        * a0aos is the attrib value at top left corner of stamp
 443        */
 444       if (interp != LP_INTERP_CONSTANT &&
 445           interp != LP_INTERP_FACING) {
 446          LLVMValueRef x = lp_build_broadcast_scalar(setup_bld, bld->x);
 447          LLVMValueRef y = lp_build_broadcast_scalar(setup_bld, bld->y);
 448          a0aos = lp_build_fmuladd(builder, x, dadxaos, a0aos);
 449          a0aos = lp_build_fmuladd(builder, y, dadyaos, a0aos);
 450       }
 451
 452       /*
 453        * dadq = {0, dadx, dady, dadx + dady}
 454        * for two quads (side by side) this is:
 455        * {0, dadx, dady, dadx+dady, 2*dadx, 2*dadx+dady, 3*dadx+dady}
 456        */
 457       for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 458          /* this generates a CRAPLOAD of shuffles... */
 459          if (mask & (1 << chan)) {
 460             LLVMValueRef dadx, dady;
 461             LLVMValueRef dadq, dadq2;
 462             LLVMValueRef a;
 463             LLVMValueRef chan_index = lp_build_const_int32(gallivm, chan);
 464
 465             if (attrib == 0 && chan == 0) {
 466                a = bld->x;
 467                if (bld->pos_offset) {
 468                   a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
 469                }
 470                a = lp_build_broadcast_scalar(coeff_bld, a);
 471                dadx = coeff_bld->one;
 472                dady = coeff_bld->zero;
 473             }
 474             else if (attrib == 0 && chan == 1) {
 475                a = bld->y;
 476                if (bld->pos_offset) {
 477                   a = LLVMBuildFAdd(builder, a, lp_build_const_float(gallivm, bld->pos_offset), "");
 478                }
 479                a = lp_build_broadcast_scalar(coeff_bld, a);
 480                dady = coeff_bld->one;
 481                dadx = coeff_bld->zero;
 482             }
 483             else {
 484                dadx = lp_build_extract_broadcast(gallivm, setup_bld->type,
 485                                               coeff_bld->type, dadxaos, chan_index);
 486                dady = lp_build_extract_broadcast(gallivm, setup_bld->type,
 487                                               coeff_bld->type, dadyaos, chan_index);
 488
 489                /*
 490                 * a = {a, a, a, a}
 491                 */
 492                a = lp_build_extract_broadcast(gallivm, setup_bld->type,
 493                                               coeff_bld->type, a0aos, chan_index);
 494             }
 495
 496             dadx = LLVMBuildFMul(builder, dadx, pixoffx, "");
 497             dady = LLVMBuildFMul(builder, dady, pixoffy, "");
 498             dadq = LLVMBuildFAdd(builder, dadx, dady, "");
 499
 500             /*
 501              * Compute the attrib values on the upper-left corner of each
 502              * group of quads.
 503              * Note that if we process 2 quads at once this doesn't
 504              * really exactly to what we want.
 505              * We need to access elem 0 and 2 respectively later if we process
 506              * 2 quads at once.
 507              */
 508
 509             if (interp != LP_INTERP_CONSTANT &&
 510                 interp != LP_INTERP_FACING) {
 511                dadq2 = LLVMBuildFAdd(builder, dadq, dadq, "");
 512                a = LLVMBuildFAdd(builder, a, dadq2, "");
 513             }
 514
 515 #if PERSPECTIVE_DIVIDE_PER_QUAD
 516             /*
 517              * a *= 1 / w
 518              */
 519
 520             /*
 521              * XXX since we're only going to access elements 0,2 out of 8
 522              * if we have 8-wide vectors we should do the division only 4-wide.
 523              * a is really a 2-elements in a 4-wide vector disguised as 8-wide
 524              * in this case.
 525              */
 526             if (interp == LP_INTERP_PERSPECTIVE) {
 527                LLVMValueRef w = bld->a[0][3];
 528                assert(attrib != 0);
 529                assert(bld->mask[0] & TGSI_WRITEMASK_W);
 530                if (!bld->oow) {
 531                   bld->oow = lp_build_rcp(coeff_bld, w);
 532                   lp_build_name(bld->oow, "oow");
 533                }
 534                a = lp_build_mul(coeff_bld, a, bld->oow);
 535             }
 536 #endif
 537
 538             attrib_name(a, attrib, chan, ".a");
 539             attrib_name(dadq, attrib, chan, ".dadq");
 540
 541             bld->a[attrib][chan] = lp_build_alloca(gallivm,
 542                                                    LLVMTypeOf(a), "");
 543             LLVMBuildStore(builder, a, bld->a[attrib][chan]);
 544             bld->dadq[attrib][chan] = dadq;
 545          }
 546       }
 547    }
 548 }
 549
 550
 551 /**
 552  * Increment the shader input attribute values.
 553  * This is called when we move from one quad to the next.
 554  */
 555 static void
 556 attribs_update(struct lp_build_interp_soa_context *bld,
 557                struct gallivm_state *gallivm,
 558                LLVMValueRef loop_iter,
 559                int start,
 560                int end)
 561 {
 562    LLVMBuilderRef builder = gallivm->builder;
 563    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 564    LLVMValueRef oow = NULL;
 565    unsigned attrib;
 566    unsigned chan;
 567
 568    for(attrib = start; attrib < end; ++attrib) {
 569       const unsigned mask = bld->mask[attrib];
 570       const unsigned interp = bld->interp[attrib];
 571       for(chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 572          if(mask & (1 << chan)) {
 573             LLVMValueRef a;
 574             if (interp == LP_INTERP_CONSTANT ||
 575                 interp == LP_INTERP_FACING) {
 576                a = LLVMBuildLoad(builder, bld->a[attrib][chan], "");
 577             }
 578             else if (interp == LP_INTERP_POSITION) {
 579                assert(attrib > 0);
 580                a = bld->attribs[0][chan];
 581             }
 582             else {
 583                LLVMValueRef dadq;
 584
 585                a = bld->a[attrib][chan];
 586
 587                /*
 588                 * Broadcast the attribute value for this quad into all elements
 589                 */
 590
 591                {
 592                   /* stored as vector load as float */
 593                   LLVMTypeRef ptr_type = LLVMPointerType(LLVMFloatTypeInContext(
 594                                                             gallivm->context), 0);
 595                   LLVMValueRef ptr;
 596                   a = LLVMBuildBitCast(builder, a, ptr_type, "");
 597                   ptr = LLVMBuildGEP(builder, a, &loop_iter, 1, "");
 598                   a = LLVMBuildLoad(builder, ptr, "");
 599                   a = lp_build_broadcast_scalar(&bld->coeff_bld, a);
 600                }
 601
 602                /*
 603                 * Get the derivatives.
 604                 */
 605
 606                dadq = bld->dadq[attrib][chan];
 607
 608 #if PERSPECTIVE_DIVIDE_PER_QUAD
 609                if (interp == LP_INTERP_PERSPECTIVE) {
 610                   LLVMValueRef dwdq = bld->dadq[0][3];
 611
 612                   if (oow == NULL) {
 613                      assert(bld->oow);
 614                      oow = LLVMBuildShuffleVector(coeff_bld->builder,
 615                                                   bld->oow, coeff_bld->undef,
 616                                                   shuffle, "");
 617                   }
 618
 619                   dadq = lp_build_sub(coeff_bld,
 620                                       dadq,
 621                                       lp_build_mul(coeff_bld, a, dwdq));
 622                   dadq = lp_build_mul(coeff_bld, dadq, oow);
 623                }
 624 #endif
 625
 626                /*
 627                 * Add the derivatives
 628                 */
 629
 630                a = lp_build_add(coeff_bld, a, dadq);
 631
 632 #if !PERSPECTIVE_DIVIDE_PER_QUAD
 633                if (interp == LP_INTERP_PERSPECTIVE) {
 634                   if (oow == NULL) {
 635                      LLVMValueRef w = bld->attribs[0][3];
 636                      assert(attrib != 0);
 637                      assert(bld->mask[0] & TGSI_WRITEMASK_W);
 638                      oow = lp_build_rcp(coeff_bld, w);
 639                   }
 640                   a = lp_build_mul(coeff_bld, a, oow);
 641                }
 642 #endif
 643
 644                if (attrib == 0 && chan == 2 && !bld->depth_clamp) {
 645                   /* FIXME: Depth values can exceed 1.0, due to the fact that
 646                    * setup interpolation coefficients refer to (0,0) which causes
 647                    * precision loss. So we must clamp to 1.0 here to avoid artifacts.
 648                    * Note though values outside [0,1] are perfectly valid with
 649                    * depth clip disabled..
 650                    * XXX: If depth clip is disabled but we force depth clamp
 651                    * we may get values larger than 1.0 in the fs (but not in
 652                    * depth test). Not sure if that's an issue...
 653                    * Also, on a similar note, it is not obvious if the depth values
 654                    * appearing in fs (with depth clip disabled) should be clamped
 655                    * to [0,1], clamped to near/far or not be clamped at all...
 656                    */
 657                   a = lp_build_min(coeff_bld, a, coeff_bld->one);
 658                }
 659
 660                attrib_name(a, attrib, chan, "");
 661             }
 662             bld->attribs[attrib][chan] = a;
 663          }
 664       }
 665    }
 666 }
 667
 668
 669 /**
 670  * Generate the position vectors.
 671  *
 672  * Parameter x0, y0 are the integer values with upper left coordinates.
 673  */
 674 static void
 675 pos_init(struct lp_build_interp_soa_context *bld,
 676          LLVMValueRef x0,
 677          LLVMValueRef y0)
 678 {
 679    LLVMBuilderRef builder = bld->coeff_bld.gallivm->builder;
 680    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 681
 682    bld->x = LLVMBuildSIToFP(builder, x0, coeff_bld->elem_type, "");
 683    bld->y = LLVMBuildSIToFP(builder, y0, coeff_bld->elem_type, "");
 684 }
 685
 686
 687 /**
 688  * Initialize fragment shader input attribute info.
 689  */
 690 void
 691 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
 692                          struct gallivm_state *gallivm,
 693                          unsigned num_inputs,
 694                          const struct lp_shader_input *inputs,
 695                          boolean pixel_center_integer,
 696                          boolean depth_clamp,
 697                          LLVMBuilderRef builder,
 698                          struct lp_type type,
 699                          LLVMValueRef a0_ptr,
 700                          LLVMValueRef dadx_ptr,
 701                          LLVMValueRef dady_ptr,
 702                          LLVMValueRef x0,
 703                          LLVMValueRef y0)
 704 {
 705    struct lp_type coeff_type;
 706    struct lp_type setup_type;
 707    unsigned attrib;
 708    unsigned chan;
 709
 710    memset(bld, 0, sizeof *bld);
 711
 712    memset(&coeff_type, 0, sizeof coeff_type);
 713    coeff_type.floating = TRUE;
 714    coeff_type.sign = TRUE;
 715    coeff_type.width = 32;
 716    coeff_type.length = type.length;
 717
 718    memset(&setup_type, 0, sizeof setup_type);
 719    setup_type.floating = TRUE;
 720    setup_type.sign = TRUE;
 721    setup_type.width = 32;
 722    setup_type.length = TGSI_NUM_CHANNELS;
 723
 724
 725    /* XXX: we don't support interpolating into any other types */
 726    assert(memcmp(&coeff_type, &type, sizeof coeff_type) == 0);
 727
 728    lp_build_context_init(&bld->coeff_bld, gallivm, coeff_type);
 729    lp_build_context_init(&bld->setup_bld, gallivm, setup_type);
 730
 731    /* For convenience */
 732    bld->pos = bld->attribs[0];
 733    bld->inputs = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) bld->attribs[1];
 734
 735    /* Position */
 736    bld->mask[0] = TGSI_WRITEMASK_XYZW;
 737    bld->interp[0] = LP_INTERP_LINEAR;
 738
 739    /* Inputs */
 740    for (attrib = 0; attrib < num_inputs; ++attrib) {
 741       bld->mask[1 + attrib] = inputs[attrib].usage_mask;
 742       bld->interp[1 + attrib] = inputs[attrib].interp;
 743    }
 744    bld->num_attribs = 1 + num_inputs;
 745
 746    /* Ensure all masked out input channels have a valid value */
 747    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 748       for (chan = 0; chan < TGSI_NUM_CHANNELS; ++chan) {
 749          bld->attribs[attrib][chan] = bld->coeff_bld.undef;
 750       }
 751    }
 752
 753    if (pixel_center_integer) {
 754       bld->pos_offset = 0.0;
 755    } else {
 756       bld->pos_offset = 0.5;
 757    }
 758    bld->depth_clamp = depth_clamp;
 759
 760    pos_init(bld, x0, y0);
 761
 762    /*
 763     * Simple method (single step interpolation) may be slower if vector length
 764     * is just 4, but the results are different (generally less accurate) with
 765     * the other method, so always use more accurate version.
 766     */
 767    if (1) {
 768       bld->simple_interp = TRUE;
 769       {
 770          /* XXX this should use a global static table */
 771          unsigned i;
 772          unsigned num_loops = 16 / type.length;
 773          LLVMValueRef pixoffx, pixoffy, index;
 774          LLVMValueRef ptr;
 775
 776          bld->xoffset_store = lp_build_array_alloca(gallivm,
 777                                                     lp_build_vec_type(gallivm, type),
 778                                                     lp_build_const_int32(gallivm, num_loops),
 779                                                     "");
 780          bld->yoffset_store = lp_build_array_alloca(gallivm,
 781                                                     lp_build_vec_type(gallivm, type),
 782                                                     lp_build_const_int32(gallivm, num_loops),
 783                                                     "");
 784          for (i = 0; i < num_loops; i++) {
 785             index = lp_build_const_int32(gallivm, i);
 786             calc_offsets(&bld->coeff_bld, i*type.length/4, &pixoffx, &pixoffy);
 787             ptr = LLVMBuildGEP(builder, bld->xoffset_store, &index, 1, "");
 788             LLVMBuildStore(builder, pixoffx, ptr);
 789             ptr = LLVMBuildGEP(builder, bld->yoffset_store, &index, 1, "");
 790             LLVMBuildStore(builder, pixoffy, ptr);
 791          }
 792       }
 793       coeffs_init_simple(bld, a0_ptr, dadx_ptr, dady_ptr);
 794    }
 795    else {
 796       bld->simple_interp = FALSE;
 797       coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
 798    }
 799
 800 }
 801
 802
 803 /*
 804  * Advance the position and inputs to the given quad within the block.
 805  */
 806
 807 void
 808 lp_build_interp_soa_update_inputs_dyn(struct lp_build_interp_soa_context *bld,
 809                                       struct gallivm_state *gallivm,
 810                                       LLVMValueRef quad_start_index)
 811 {
 812    if (bld->simple_interp) {
 813       attribs_update_simple(bld, gallivm, quad_start_index, 1, bld->num_attribs);
 814    }
 815    else {
 816       attribs_update(bld, gallivm, quad_start_index, 1, bld->num_attribs);
 817    }
 818 }
 819
 820 void
 821 lp_build_interp_soa_update_pos_dyn(struct lp_build_interp_soa_context *bld,
 822                                    struct gallivm_state *gallivm,
 823                                    LLVMValueRef quad_start_index)
 824 {
 825    if (bld->simple_interp) {
 826       attribs_update_simple(bld, gallivm, quad_start_index, 0, 1);
 827    }
 828    else {
 829       attribs_update(bld, gallivm, quad_start_index, 0, 1);
 830    }
 831 }
 832