src/gallium/auxiliary/gallivm/lp_bld_conv.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 /**
  30  * @file
  31  * Helper functions for type conversions.
  32  *
  33  * We want to use the fastest type for a given computation whenever feasible.
  34  * The other side of this is that we need to be able convert between several
  35  * types accurately and efficiently.
  36  *
  37  * Conversion between types of different bit width is quite complex since a
  38  *
  39  * To remember there are a few invariants in type conversions:
  40  *
  41  * - register width must remain constant:
  42  *
  43  *     src_type.width * src_type.length == dst_type.width * dst_type.length
  44  *
  45  * - total number of elements must remain constant:
  46  *
  47  *     src_type.length * num_srcs == dst_type.length * num_dsts
  48  *
  49  * It is not always possible to do the conversion both accurately and
  50  * efficiently, usually due to lack of adequate machine instructions. In these
  51  * cases it is important not to cut shortcuts here and sacrifice accuracy, as
  52  * there this functions can be used anywhere. In the future we might have a
  53  * precision parameter which can gauge the accuracy vs efficiency compromise,
  54  * but for now if the data conversion between two stages happens to be the
  55  * bottleneck, then most likely should just avoid converting at all and run
  56  * both stages with the same type.
  57  *
  58  * Make sure to run lp_test_conv unit test after any change to this file.
  59  *
  60  * @author Jose Fonseca <jfonseca@vmware.com>
  61  */
  62
  63
  64 #include "util/u_debug.h"
  65 #include "util/u_math.h"
  66
  67 #include "lp_bld_type.h"
  68 #include "lp_bld_const.h"
  69 #include "lp_bld_arit.h"
  70 #include "lp_bld_pack.h"
  71 #include "lp_bld_conv.h"
  72
  73
  74 /**
  75  * Special case for converting clamped IEEE-754 floats to unsigned norms.
  76  *
  77  * The mathematical voodoo below may seem excessive but it is actually
  78  * paramount we do it this way for several reasons. First, there is no single
  79  * precision FP to unsigned integer conversion Intel SSE instruction. Second,
  80  * secondly, even if there was, since the FP's mantissa takes only a fraction
  81  * of register bits the typically scale and cast approach would require double
  82  * precision for accurate results, and therefore half the throughput
  83  *
  84  * Although the result values can be scaled to an arbitrary bit width specified
  85  * by dst_width, the actual result type will have the same width.
  86  */
  87 LLVMValueRef
  88 lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
  89                                         struct lp_type src_type,
  90                                         unsigned dst_width,
  91                                         LLVMValueRef src)
  92 {
  93    LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
  94    LLVMValueRef res;
  95    unsigned mantissa;
  96    unsigned n;
  97    unsigned long long ubound;
  98    unsigned long long mask;
  99    double scale;
 100    double bias;
 101
 102    assert(src_type.floating);
 103
 104    mantissa = lp_mantissa(src_type);
 105
 106    /* We cannot carry more bits than the mantissa */
 107    n = MIN2(mantissa, dst_width);
 108
 109    /* This magic coefficients will make the desired result to appear in the
 110     * lowest significant bits of the mantissa.
 111     */
 112    ubound = ((unsigned long long)1 << n);
 113    mask = ubound - 1;
 114    scale = (double)mask/ubound;
 115    bias = (double)((unsigned long long)1 << (mantissa - n));
 116
 117    res = LLVMBuildMul(builder, src, lp_build_const_vec(src_type, scale), "");
 118    res = LLVMBuildAdd(builder, res, lp_build_const_vec(src_type, bias), "");
 119    res = LLVMBuildBitCast(builder, res, int_vec_type, "");
 120
 121    if(dst_width > n) {
 122       int shift = dst_width - n;
 123       res = LLVMBuildShl(builder, res, lp_build_const_int_vec(src_type, shift), "");
 124
 125       /* TODO: Fill in the empty lower bits for additional precision? */
 126       /* YES: this fixes progs/trivial/tri-z-eq.c.
 127        * Otherwise vertex Z=1.0 values get converted to something like
 128        * 0xfffffb00 and the test for equality with 0xffffffff fails.
 129        */
 130 #if 0
 131       {
 132          LLVMValueRef msb;
 133          msb = LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, dst_width - 1), "");
 134          msb = LLVMBuildShl(builder, msb, lp_build_const_int_vec(src_type, shift), "");
 135          msb = LLVMBuildSub(builder, msb, lp_build_const_int_vec(src_type, 1), "");
 136          res = LLVMBuildOr(builder, res, msb, "");
 137       }
 138 #elif 0
 139       while(shift > 0) {
 140          res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_const_int_vec(src_type, n), ""), "");
 141          shift -= n;
 142          n *= 2;
 143       }
 144 #endif
 145    }
 146    else
 147       res = LLVMBuildAnd(builder, res, lp_build_const_int_vec(src_type, mask), "");
 148
 149    return res;
 150 }
 151
 152
 153 /**
 154  * Inverse of lp_build_clamped_float_to_unsigned_norm above.
 155  */
 156 LLVMValueRef
 157 lp_build_unsigned_norm_to_float(LLVMBuilderRef builder,
 158                                 unsigned src_width,
 159                                 struct lp_type dst_type,
 160                                 LLVMValueRef src)
 161 {
 162    LLVMTypeRef vec_type = lp_build_vec_type(dst_type);
 163    LLVMTypeRef int_vec_type = lp_build_int_vec_type(dst_type);
 164    LLVMValueRef bias_;
 165    LLVMValueRef res;
 166    unsigned mantissa;
 167    unsigned n;
 168    unsigned long long ubound;
 169    unsigned long long mask;
 170    double scale;
 171    double bias;
 172
 173    mantissa = lp_mantissa(dst_type);
 174
 175    n = MIN2(mantissa, src_width);
 176
 177    ubound = ((unsigned long long)1 << n);
 178    mask = ubound - 1;
 179    scale = (double)ubound/mask;
 180    bias = (double)((unsigned long long)1 << (mantissa - n));
 181
 182    res = src;
 183
 184    if(src_width > mantissa) {
 185       int shift = src_width - mantissa;
 186       res = LLVMBuildLShr(builder, res, lp_build_const_int_vec(dst_type, shift), "");
 187    }
 188
 189    bias_ = lp_build_const_vec(dst_type, bias);
 190
 191    res = LLVMBuildOr(builder,
 192                      res,
 193                      LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");
 194
 195    res = LLVMBuildBitCast(builder, res, vec_type, "");
 196
 197    res = LLVMBuildSub(builder, res, bias_, "");
 198    res = LLVMBuildMul(builder, res, lp_build_const_vec(dst_type, scale), "");
 199
 200    return res;
 201 }
 202
 203
 204 /**
 205  * Generic type conversion.
 206  *
 207  * TODO: Take a precision argument, or even better, add a new precision member
 208  * to the lp_type union.
 209  */
 210 void
 211 lp_build_conv(LLVMBuilderRef builder,
 212               struct lp_type src_type,
 213               struct lp_type dst_type,
 214               const LLVMValueRef *src, unsigned num_srcs,
 215               LLVMValueRef *dst, unsigned num_dsts)
 216 {
 217    struct lp_type tmp_type;
 218    LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
 219    unsigned num_tmps;
 220    unsigned i;
 221
 222    /* Register width must remain constant */
 223    assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
 224
 225    /* We must not loose or gain channels. Only precision */
 226    assert(src_type.length * num_srcs == dst_type.length * num_dsts);
 227
 228    assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
 229    assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
 230
 231    tmp_type = src_type;
 232    for(i = 0; i < num_srcs; ++i)
 233       tmp[i] = src[i];
 234    num_tmps = num_srcs;
 235
 236    /*
 237     * Clamp if necessary
 238     */
 239
 240    if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
 241       struct lp_build_context bld;
 242       double src_min = lp_const_min(src_type);
 243       double dst_min = lp_const_min(dst_type);
 244       double src_max = lp_const_max(src_type);
 245       double dst_max = lp_const_max(dst_type);
 246       LLVMValueRef thres;
 247
 248       lp_build_context_init(&bld, builder, tmp_type);
 249
 250       if(src_min < dst_min) {
 251          if(dst_min == 0.0)
 252             thres = bld.zero;
 253          else
 254             thres = lp_build_const_vec(src_type, dst_min);
 255          for(i = 0; i < num_tmps; ++i)
 256             tmp[i] = lp_build_max(&bld, tmp[i], thres);
 257       }
 258
 259       if(src_max > dst_max) {
 260          if(dst_max == 1.0)
 261             thres = bld.one;
 262          else
 263             thres = lp_build_const_vec(src_type, dst_max);
 264          for(i = 0; i < num_tmps; ++i)
 265             tmp[i] = lp_build_min(&bld, tmp[i], thres);
 266       }
 267    }
 268
 269    /*
 270     * Scale to the narrowest range
 271     */
 272
 273    if(dst_type.floating) {
 274       /* Nothing to do */
 275    }
 276    else if(tmp_type.floating) {
 277       if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
 278          for(i = 0; i < num_tmps; ++i) {
 279             tmp[i] = lp_build_clamped_float_to_unsigned_norm(builder,
 280                                                              tmp_type,
 281                                                              dst_type.width,
 282                                                              tmp[i]);
 283          }
 284          tmp_type.floating = FALSE;
 285       }
 286       else {
 287          double dst_scale = lp_const_scale(dst_type);
 288          LLVMTypeRef tmp_vec_type;
 289
 290          if (dst_scale != 1.0) {
 291             LLVMValueRef scale = lp_build_const_vec(tmp_type, dst_scale);
 292             for(i = 0; i < num_tmps; ++i)
 293                tmp[i] = LLVMBuildMul(builder, tmp[i], scale, "");
 294          }
 295
 296          /* Use an equally sized integer for intermediate computations */
 297          tmp_type.floating = FALSE;
 298          tmp_vec_type = lp_build_vec_type(tmp_type);
 299          for(i = 0; i < num_tmps; ++i) {
 300 #if 0
 301             if(dst_type.sign)
 302                tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
 303             else
 304                tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
 305 #else
 306            /* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
 307             tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
 308 #endif
 309          }
 310       }
 311    }
 312    else {
 313       unsigned src_shift = lp_const_shift(src_type);
 314       unsigned dst_shift = lp_const_shift(dst_type);
 315
 316       /* FIXME: compensate different offsets too */
 317       if(src_shift > dst_shift) {
 318          LLVMValueRef shift = lp_build_const_int_vec(tmp_type, src_shift - dst_shift);
 319          for(i = 0; i < num_tmps; ++i)
 320             if(src_type.sign)
 321                tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
 322             else
 323                tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
 324       }
 325    }
 326
 327    /*
 328     * Truncate or expand bit width
 329     */
 330
 331    assert(!tmp_type.floating || tmp_type.width == dst_type.width);
 332
 333    if(tmp_type.width > dst_type.width) {
 334       assert(num_dsts == 1);
 335       tmp[0] = lp_build_pack(builder, tmp_type, dst_type, TRUE, tmp, num_tmps);
 336       tmp_type.width = dst_type.width;
 337       tmp_type.length = dst_type.length;
 338       num_tmps = 1;
 339    }
 340
 341    if(tmp_type.width < dst_type.width) {
 342       assert(num_tmps == 1);
 343       lp_build_unpack(builder, tmp_type, dst_type, tmp[0], tmp, num_dsts);
 344       tmp_type.width = dst_type.width;
 345       tmp_type.length = dst_type.length;
 346       num_tmps = num_dsts;
 347    }
 348
 349    assert(tmp_type.width == dst_type.width);
 350    assert(tmp_type.length == dst_type.length);
 351    assert(num_tmps == num_dsts);
 352
 353    /*
 354     * Scale to the widest range
 355     */
 356
 357    if(src_type.floating) {
 358       /* Nothing to do */
 359    }
 360    else if(!src_type.floating && dst_type.floating) {
 361       if(!src_type.fixed && !src_type.sign && src_type.norm) {
 362          for(i = 0; i < num_tmps; ++i) {
 363             tmp[i] = lp_build_unsigned_norm_to_float(builder,
 364                                                      src_type.width,
 365                                                      dst_type,
 366                                                      tmp[i]);
 367          }
 368          tmp_type.floating = TRUE;
 369       }
 370       else {
 371          double src_scale = lp_const_scale(src_type);
 372          LLVMTypeRef tmp_vec_type;
 373
 374          /* Use an equally sized integer for intermediate computations */
 375          tmp_type.floating = TRUE;
 376          tmp_type.sign = TRUE;
 377          tmp_vec_type = lp_build_vec_type(tmp_type);
 378          for(i = 0; i < num_tmps; ++i) {
 379 #if 0
 380             if(dst_type.sign)
 381                tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
 382             else
 383                tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
 384 #else
 385             /* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
 386             tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
 387 #endif
 388           }
 389
 390           if (src_scale != 1.0) {
 391              LLVMValueRef scale = lp_build_const_vec(tmp_type, 1.0/src_scale);
 392              for(i = 0; i < num_tmps; ++i)
 393                 tmp[i] = LLVMBuildMul(builder, tmp[i], scale, "");
 394           }
 395       }
 396     }
 397     else {
 398        unsigned src_shift = lp_const_shift(src_type);
 399        unsigned dst_shift = lp_const_shift(dst_type);
 400
 401        /* FIXME: compensate different offsets too */
 402        if(src_shift < dst_shift) {
 403           LLVMValueRef shift = lp_build_const_int_vec(tmp_type, dst_shift - src_shift);
 404           for(i = 0; i < num_tmps; ++i)
 405              tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
 406        }
 407     }
 408
 409    for(i = 0; i < num_dsts; ++i)
 410       dst[i] = tmp[i];
 411 }
 412
 413
 414 /**
 415  * Bit mask conversion.
 416  *
 417  * This will convert the integer masks that match the given types.
 418  *
 419  * The mask values should 0 or -1, i.e., all bits either set to zero or one.
 420  * Any other value will likely cause in unpredictable results.
 421  *
 422  * This is basically a very trimmed down version of lp_build_conv.
 423  */
 424 void
 425 lp_build_conv_mask(LLVMBuilderRef builder,
 426                    struct lp_type src_type,
 427                    struct lp_type dst_type,
 428                    const LLVMValueRef *src, unsigned num_srcs,
 429                    LLVMValueRef *dst, unsigned num_dsts)
 430 {
 431    /* Register width must remain constant */
 432    assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
 433
 434    /* We must not loose or gain channels. Only precision */
 435    assert(src_type.length * num_srcs == dst_type.length * num_dsts);
 436
 437    /*
 438     * Drop
 439     *
 440     * We assume all values are 0 or -1
 441     */
 442
 443    src_type.floating = FALSE;
 444    src_type.fixed = FALSE;
 445    src_type.sign = TRUE;
 446    src_type.norm = FALSE;
 447
 448    dst_type.floating = FALSE;
 449    dst_type.fixed = FALSE;
 450    dst_type.sign = TRUE;
 451    dst_type.norm = FALSE;
 452
 453    /*
 454     * Truncate or expand bit width
 455     */
 456
 457    if(src_type.width > dst_type.width) {
 458       assert(num_dsts == 1);
 459       dst[0] = lp_build_pack(builder, src_type, dst_type, TRUE, src, num_srcs);
 460    }
 461    else if(src_type.width < dst_type.width) {
 462       assert(num_srcs == 1);
 463       lp_build_unpack(builder, src_type, dst_type, src[0], dst, num_dsts);
 464    }
 465    else {
 466       assert(num_srcs == num_dsts);
 467       memcpy(dst, src, num_dsts * sizeof *dst);
 468    }
 469 }