src/gallium/drivers/llvmpipe/lp_bld_interp.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007-2008 Tungsten Graphics, Inc., Cedar Park, Texas.
   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 TUNGSTEN GRAPHICS 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 "lp_bld_interp.h"
  46
  47
  48 /*
  49  * The shader JIT function operates on blocks of quads.
  50  * Each block has 2x2 quads and each quad has 2x2 pixels.
  51  *
  52  * We iterate over the quads in order 0, 1, 2, 3:
  53  *
  54  * #################
  55  * #   |   #   |   #
  56  * #---0---#---1---#
  57  * #   |   #   |   #
  58  * #################
  59  * #   |   #   |   #
  60  * #---2---#---3---#
  61  * #   |   #   |   #
  62  * #################
  63  *
  64  * Within each quad, we have four pixels which are represented in SOA
  65  * order:
  66  *
  67  * #########
  68  * # 0 | 1 #
  69  * #---+---#
  70  * # 2 | 3 #
  71  * #########
  72  *
  73  * So the green channel (for example) of the four pixels is stored in
  74  * a single vector register: {g0, g1, g2, g3}.
  75  */
  76
  77
  78 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  79 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  80
  81
  82 static void
  83 attrib_name(LLVMValueRef val, unsigned attrib, unsigned chan, const char *suffix)
  84 {
  85    if(attrib == 0)
  86       lp_build_name(val, "pos.%c%s", "xyzw"[chan], suffix);
  87    else
  88       lp_build_name(val, "input%u.%c%s", attrib - 1, "xyzw"[chan], suffix);
  89 }
  90
  91
  92 /**
  93  * Initialize the bld->a0, dadx, dady fields.  This involves fetching
  94  * those values from the arrays which are passed into the JIT function.
  95  */
  96 static void
  97 coeffs_init(struct lp_build_interp_soa_context *bld,
  98             LLVMValueRef a0_ptr,
  99             LLVMValueRef dadx_ptr,
 100             LLVMValueRef dady_ptr)
 101 {
 102    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 103    LLVMBuilderRef builder = coeff_bld->builder;
 104    LLVMValueRef zero = LLVMConstNull(coeff_bld->elem_type);
 105    LLVMValueRef one = LLVMConstReal(coeff_bld->elem_type, 1.0);
 106    LLVMValueRef i0 = LLVMConstInt(LLVMInt32Type(), 0, 0);
 107    LLVMValueRef i1 = LLVMConstInt(LLVMInt32Type(), 1, 0);
 108    LLVMValueRef i2 = LLVMConstInt(LLVMInt32Type(), 2, 0);
 109    LLVMValueRef i3 = LLVMConstInt(LLVMInt32Type(), 3, 0);
 110    LLVMValueRef oow = NULL;
 111    unsigned attrib;
 112    unsigned chan;
 113
 114    /* TODO: Use more vector operations */
 115
 116    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 117       const unsigned mask = bld->mask[attrib];
 118       const unsigned interp = bld->interp[attrib];
 119       for (chan = 0; chan < NUM_CHANNELS; ++chan) {
 120          if (mask & (1 << chan)) {
 121             LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), attrib*NUM_CHANNELS + chan, 0);
 122             LLVMValueRef a0 = zero;
 123             LLVMValueRef dadx = zero;
 124             LLVMValueRef dady = zero;
 125             LLVMValueRef dadxy = zero;
 126             LLVMValueRef dadq;
 127             LLVMValueRef dadq2;
 128             LLVMValueRef a;
 129
 130             switch (interp) {
 131             case LP_INTERP_PERSPECTIVE:
 132                /* fall-through */
 133
 134             case LP_INTERP_LINEAR:
 135                if (attrib == 0 && chan == 0) {
 136                   dadxy = dadx = one;
 137                }
 138                else if (attrib == 0 && chan == 1) {
 139                   dadxy = dady = one;
 140                }
 141                else {
 142                   dadx = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dadx_ptr, &index, 1, ""), "");
 143                   dady = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dady_ptr, &index, 1, ""), "");
 144                   dadxy = LLVMBuildAdd(builder, dadx, dady, "");
 145                   attrib_name(dadx, attrib, chan, ".dadx");
 146                   attrib_name(dady, attrib, chan, ".dady");
 147                   attrib_name(dadxy, attrib, chan, ".dadxy");
 148                }
 149                /* fall-through */
 150
 151             case LP_INTERP_CONSTANT:
 152             case LP_INTERP_FACING:
 153                a0 = LLVMBuildLoad(builder, LLVMBuildGEP(builder, a0_ptr, &index, 1, ""), "");
 154                attrib_name(a0, attrib, chan, ".a0");
 155                break;
 156
 157             case LP_INTERP_POSITION:
 158                /* Nothing to do as the position coeffs are already setup in slot 0 */
 159                continue;
 160
 161             default:
 162                assert(0);
 163                break;
 164             }
 165
 166             /*
 167              * dadq = {0, dadx, dady, dadx + dady}
 168              */
 169
 170             dadq = coeff_bld->undef;
 171             dadq = LLVMBuildInsertElement(builder, dadq, zero,  i0, "");
 172             dadq = LLVMBuildInsertElement(builder, dadq, dadx,  i1, "");
 173             dadq = LLVMBuildInsertElement(builder, dadq, dady,  i2, "");
 174             dadq = LLVMBuildInsertElement(builder, dadq, dadxy, i3, "");
 175
 176             /*
 177              * dadq2 = 2 * dq
 178              */
 179
 180             dadq2 = LLVMBuildAdd(builder, dadq, dadq, "");
 181
 182             /*
 183              * a = a0 + x * dadx + y * dady
 184              */
 185
 186             if (attrib == 0 && chan == 0) {
 187                a = bld->x;
 188             }
 189             else if (attrib == 0 && chan == 1) {
 190                a = bld->y;
 191             }
 192             else {
 193                a = a0;
 194                if (interp != LP_INTERP_CONSTANT &&
 195                    interp != LP_INTERP_FACING) {
 196                   a = LLVMBuildAdd(builder, a,
 197                                    LLVMBuildMul(builder, bld->x, dadx, ""),
 198                                    "");
 199                   a = LLVMBuildAdd(builder, a,
 200                                    LLVMBuildMul(builder, bld->y, dady, ""),
 201                                    "");
 202                }
 203             }
 204
 205             /*
 206              * a = {a, a, a, a}
 207              */
 208
 209             a = lp_build_broadcast(builder, coeff_bld->vec_type, a);
 210
 211             /*
 212              * Compute the attrib values on the upper-left corner of each quad.
 213              */
 214
 215             a = LLVMBuildAdd(builder, a, dadq2, "");
 216
 217             /*
 218              * a    *= 1 / w
 219              * dadq *= 1 / w
 220              */
 221
 222             if (interp == LP_INTERP_PERSPECTIVE) {
 223                LLVMValueRef w = bld->a[0][3];
 224                assert(attrib != 0);
 225                assert(bld->mask[0] & TGSI_WRITEMASK_W);
 226                if (!oow) {
 227                   oow = lp_build_rcp(coeff_bld, w);
 228                   lp_build_name(oow, "oow");
 229                }
 230                a = lp_build_mul(coeff_bld, a, oow);
 231                dadq = lp_build_mul(coeff_bld, dadq, oow);
 232             }
 233
 234             attrib_name(a, attrib, chan, ".a");
 235             attrib_name(dadq, attrib, chan, ".dadq");
 236
 237             bld->a   [attrib][chan] = a;
 238             bld->dadq[attrib][chan] = dadq;
 239          }
 240       }
 241    }
 242 }
 243
 244
 245 /**
 246  * Increment the shader input attribute values.
 247  * This is called when we move from one quad to the next.
 248  */
 249 static void
 250 attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
 251 {
 252    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 253    LLVMValueRef shuffle = lp_build_const_int_vec(coeff_bld->type, quad_index);
 254    unsigned attrib;
 255    unsigned chan;
 256
 257    assert(quad_index < 4);
 258
 259    for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
 260       const unsigned mask = bld->mask[attrib];
 261       const unsigned interp = bld->interp[attrib];
 262       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 263          if(mask & (1 << chan)) {
 264             LLVMValueRef a = coeff_bld->undef;
 265             if (interp == LP_INTERP_CONSTANT ||
 266                 interp == LP_INTERP_FACING) {
 267                a = bld->a[attrib][chan];
 268             }
 269             else if (interp == LP_INTERP_POSITION) {
 270                assert(attrib > 0);
 271                a = bld->attribs[0][chan];
 272             }
 273             else {
 274                a = bld->a[attrib][chan];
 275
 276                /*
 277                 * Broadcast the attribute value for this quad into all elements
 278                 */
 279
 280                a = LLVMBuildShuffleVector(coeff_bld->builder,
 281                                           a, coeff_bld->undef, shuffle, "");
 282
 283                /*
 284                 * Add the derivatives
 285                 */
 286
 287                a = lp_build_add(coeff_bld, a, bld->dadq[attrib][chan]);
 288
 289                attrib_name(a, attrib, chan, "");
 290             }
 291             bld->attribs[attrib][chan] = a;
 292          }
 293       }
 294    }
 295 }
 296
 297
 298 /**
 299  * Generate the position vectors.
 300  *
 301  * Parameter x0, y0 are the integer values with upper left coordinates.
 302  */
 303 static void
 304 pos_init(struct lp_build_interp_soa_context *bld,
 305          LLVMValueRef x0,
 306          LLVMValueRef y0)
 307 {
 308    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 309
 310    bld->x = LLVMBuildSIToFP(coeff_bld->builder, x0, coeff_bld->elem_type, "");
 311    bld->y = LLVMBuildSIToFP(coeff_bld->builder, y0, coeff_bld->elem_type, "");
 312 }
 313
 314
 315 /**
 316  * Initialize fragment shader input attribute info.
 317  */
 318 void
 319 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
 320                          unsigned num_inputs,
 321                          const struct lp_shader_input *inputs,
 322                          LLVMBuilderRef builder,
 323                          struct lp_type type,
 324                          LLVMValueRef a0_ptr,
 325                          LLVMValueRef dadx_ptr,
 326                          LLVMValueRef dady_ptr,
 327                          LLVMValueRef x0,
 328                          LLVMValueRef y0)
 329 {
 330    struct lp_type coeff_type;
 331    unsigned attrib;
 332    unsigned chan;
 333
 334    memset(bld, 0, sizeof *bld);
 335
 336    memset(&coeff_type, 0, sizeof coeff_type);
 337    coeff_type.floating = TRUE;
 338    coeff_type.sign = TRUE;
 339    coeff_type.width = 32;
 340    coeff_type.length = QUAD_SIZE;
 341
 342    /* XXX: we don't support interpolating into any other types */
 343    assert(memcmp(&coeff_type, &type, sizeof coeff_type) == 0);
 344
 345    lp_build_context_init(&bld->coeff_bld, builder, coeff_type);
 346
 347    /* For convenience */
 348    bld->pos = bld->attribs[0];
 349    bld->inputs = (const LLVMValueRef (*)[NUM_CHANNELS]) bld->attribs[1];
 350
 351    /* Position */
 352    bld->num_attribs = 1;
 353    bld->mask[0] = TGSI_WRITEMASK_XYZW;
 354    bld->interp[0] = LP_INTERP_LINEAR;
 355
 356    /* Inputs */
 357    for (attrib = 0; attrib < num_inputs; ++attrib) {
 358       bld->mask[1 + attrib] = inputs[attrib].usage_mask;
 359       bld->interp[1 + attrib] = inputs[attrib].interp;
 360    }
 361    bld->num_attribs = 1 + num_inputs;
 362
 363    /* Ensure all masked out input channels have a valid value */
 364    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 365       for (chan = 0; chan < NUM_CHANNELS; ++chan) {
 366          bld->attribs[attrib][chan] = bld->coeff_bld.undef;
 367       }
 368    }
 369
 370    pos_init(bld, x0, y0);
 371
 372    coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
 373
 374    attribs_update(bld, 0);
 375 }
 376
 377
 378 /**
 379  * Advance the position and inputs to the given quad within the block.
 380  */
 381 void
 382 lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld,
 383                            int quad_index)
 384 {
 385    assert(quad_index < 4);
 386
 387    attribs_update(bld, quad_index);
 388 }