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