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 = LLVMBuildFAdd(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 = LLVMBuildFAdd(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                   LLVMValueRef tmp;
 197                   tmp = LLVMBuildFMul(builder, bld->x, dadx, "");
 198                   a = LLVMBuildFAdd(builder, a, tmp, "");
 199                   tmp = LLVMBuildFMul(builder, bld->y, dady, "");
 200                   a = LLVMBuildFAdd(builder, a, tmp, "");
 201                }
 202             }
 203
 204             /*
 205              * a = {a, a, a, a}
 206              */
 207
 208             a = lp_build_broadcast(builder, coeff_bld->vec_type, a);
 209
 210             /*
 211              * Compute the attrib values on the upper-left corner of each quad.
 212              */
 213
 214             a = LLVMBuildFAdd(builder, a, dadq2, "");
 215
 216             /*
 217              * a    *= 1 / w
 218              * dadq *= 1 / w
 219              */
 220
 221             if (interp == LP_INTERP_PERSPECTIVE) {
 222                LLVMValueRef w = bld->a[0][3];
 223                assert(attrib != 0);
 224                assert(bld->mask[0] & TGSI_WRITEMASK_W);
 225                if (!oow) {
 226                   oow = lp_build_rcp(coeff_bld, w);
 227                   lp_build_name(oow, "oow");
 228                }
 229                a = lp_build_mul(coeff_bld, a, oow);
 230                dadq = lp_build_mul(coeff_bld, dadq, oow);
 231             }
 232
 233             attrib_name(a, attrib, chan, ".a");
 234             attrib_name(dadq, attrib, chan, ".dadq");
 235
 236             bld->a   [attrib][chan] = a;
 237             bld->dadq[attrib][chan] = dadq;
 238          }
 239       }
 240    }
 241 }
 242
 243
 244 /**
 245  * Increment the shader input attribute values.
 246  * This is called when we move from one quad to the next.
 247  */
 248 static void
 249 attribs_update(struct lp_build_interp_soa_context *bld, int quad_index)
 250 {
 251    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 252    LLVMValueRef shuffle = lp_build_const_int_vec(coeff_bld->type, quad_index);
 253    unsigned attrib;
 254    unsigned chan;
 255
 256    assert(quad_index < 4);
 257
 258    for(attrib = 0; attrib < bld->num_attribs; ++attrib) {
 259       const unsigned mask = bld->mask[attrib];
 260       const unsigned interp = bld->interp[attrib];
 261       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 262          if(mask & (1 << chan)) {
 263             LLVMValueRef a;
 264             if (interp == LP_INTERP_CONSTANT ||
 265                 interp == LP_INTERP_FACING) {
 266                a = bld->a[attrib][chan];
 267             }
 268             else if (interp == LP_INTERP_POSITION) {
 269                assert(attrib > 0);
 270                a = bld->attribs[0][chan];
 271             }
 272             else {
 273                a = bld->a[attrib][chan];
 274
 275                /*
 276                 * Broadcast the attribute value for this quad into all elements
 277                 */
 278
 279                a = LLVMBuildShuffleVector(coeff_bld->builder,
 280                                           a, coeff_bld->undef, shuffle, "");
 281
 282                /*
 283                 * Add the derivatives
 284                 */
 285
 286                a = lp_build_add(coeff_bld, a, bld->dadq[attrib][chan]);
 287
 288                attrib_name(a, attrib, chan, "");
 289             }
 290             bld->attribs[attrib][chan] = a;
 291          }
 292       }
 293    }
 294 }
 295
 296
 297 /**
 298  * Generate the position vectors.
 299  *
 300  * Parameter x0, y0 are the integer values with upper left coordinates.
 301  */
 302 static void
 303 pos_init(struct lp_build_interp_soa_context *bld,
 304          LLVMValueRef x0,
 305          LLVMValueRef y0)
 306 {
 307    struct lp_build_context *coeff_bld = &bld->coeff_bld;
 308
 309    bld->x = LLVMBuildSIToFP(coeff_bld->builder, x0, coeff_bld->elem_type, "");
 310    bld->y = LLVMBuildSIToFP(coeff_bld->builder, y0, coeff_bld->elem_type, "");
 311 }
 312
 313
 314 /**
 315  * Initialize fragment shader input attribute info.
 316  */
 317 void
 318 lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
 319                          unsigned num_inputs,
 320                          const struct lp_shader_input *inputs,
 321                          LLVMBuilderRef builder,
 322                          struct lp_type type,
 323                          LLVMValueRef a0_ptr,
 324                          LLVMValueRef dadx_ptr,
 325                          LLVMValueRef dady_ptr,
 326                          LLVMValueRef x0,
 327                          LLVMValueRef y0)
 328 {
 329    struct lp_type coeff_type;
 330    unsigned attrib;
 331    unsigned chan;
 332
 333    memset(bld, 0, sizeof *bld);
 334
 335    memset(&coeff_type, 0, sizeof coeff_type);
 336    coeff_type.floating = TRUE;
 337    coeff_type.sign = TRUE;
 338    coeff_type.width = 32;
 339    coeff_type.length = QUAD_SIZE;
 340
 341    /* XXX: we don't support interpolating into any other types */
 342    assert(memcmp(&coeff_type, &type, sizeof coeff_type) == 0);
 343
 344    lp_build_context_init(&bld->coeff_bld, builder, coeff_type);
 345
 346    /* For convenience */
 347    bld->pos = bld->attribs[0];
 348    bld->inputs = (const LLVMValueRef (*)[NUM_CHANNELS]) bld->attribs[1];
 349
 350    /* Position */
 351    bld->num_attribs = 1;
 352    bld->mask[0] = TGSI_WRITEMASK_XYZW;
 353    bld->interp[0] = LP_INTERP_LINEAR;
 354
 355    /* Inputs */
 356    for (attrib = 0; attrib < num_inputs; ++attrib) {
 357       bld->mask[1 + attrib] = inputs[attrib].usage_mask;
 358       bld->interp[1 + attrib] = inputs[attrib].interp;
 359    }
 360    bld->num_attribs = 1 + num_inputs;
 361
 362    /* Ensure all masked out input channels have a valid value */
 363    for (attrib = 0; attrib < bld->num_attribs; ++attrib) {
 364       for (chan = 0; chan < NUM_CHANNELS; ++chan) {
 365          bld->attribs[attrib][chan] = bld->coeff_bld.undef;
 366       }
 367    }
 368
 369    pos_init(bld, x0, y0);
 370
 371    coeffs_init(bld, a0_ptr, dadx_ptr, dady_ptr);
 372
 373    attribs_update(bld, 0);
 374 }
 375
 376
 377 /**
 378  * Advance the position and inputs to the given quad within the block.
 379  */
 380 void
 381 lp_build_interp_soa_update(struct lp_build_interp_soa_context *bld,
 382                            int quad_index)
 383 {
 384    assert(quad_index < 4);
 385
 386    attribs_update(bld, quad_index);
 387 }