src/gallium/drivers/llvmpipe/lp_test_arit.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2011 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 #include <limits.h>
  30 #include <stdio.h>
  31 #include <stdlib.h>
  32
  33 #include "util/u_pointer.h"
  34 #include "util/u_memory.h"
  35 #include "util/u_math.h"
  36
  37 #include "gallivm/lp_bld.h"
  38 #include "gallivm/lp_bld_debug.h"
  39 #include "gallivm/lp_bld_init.h"
  40 #include "gallivm/lp_bld_arit.h"
  41
  42 #include "lp_test.h"
  43
  44
  45 void
  46 write_tsv_header(FILE *fp)
  47 {
  48    fprintf(fp,
  49            "result\t"
  50            "format\n");
  51
  52    fflush(fp);
  53 }
  54
  55
  56 typedef void (*unary_func_t)(float *out, const float *in);
  57
  58
  59 /**
  60  * Describe a test case of one unary function.
  61  */
  62 struct unary_test_t
  63 {
  64    /*
  65     * Test name -- name of the mathematical function under test.
  66     */
  67
  68    const char *name;
  69
  70    LLVMValueRef
  71    (*builder)(struct lp_build_context *bld, LLVMValueRef a);
  72
  73    /*
  74     * Reference (pure-C) function.
  75     */
  76    float
  77    (*ref)(float a);
  78
  79    /*
  80     * Test values.
  81     */
  82    const float *values;
  83    unsigned num_values;
  84
  85    /*
  86     * Required precision in bits.
  87     */
  88    double precision;
  89 };
  90
  91
  92 static float negf(float x)
  93 {
  94    return -x;
  95 }
  96
  97
  98 static float sgnf(float x)
  99 {
 100    if (x > 0.0f) {
 101       return 1.0f;
 102    }
 103    if (x < 0.0f) {
 104       return -1.0f;
 105    }
 106    return 0.0f;
 107 }
 108
 109
 110 const float exp2_values[] = {
 111    -INFINITY,
 112    -60,
 113    -4,
 114    -2,
 115    -1,
 116    -1e-007,
 117    0,
 118    1e-007,
 119    0.01,
 120    0.1,
 121    0.9,
 122    0.99,
 123    1,
 124    2,
 125    4,
 126    60,
 127    INFINITY,
 128    NAN
 129 };
 130
 131
 132 const float log2_values[] = {
 133 #if 0
 134    /*
 135     * Smallest denormalized number; meant just for experimentation, but not
 136     * validation.
 137     */
 138    1.4012984643248171e-45,
 139 #endif
 140    -INFINITY,
 141    0,
 142    1e-007,
 143    0.1,
 144    0.5,
 145    0.99,
 146    1,
 147    1.01,
 148    1.1,
 149    1.9,
 150    1.99,
 151    2,
 152    4,
 153    100000,
 154    1e+018,
 155    INFINITY,
 156    NAN
 157 };
 158
 159
 160 static float rcpf(float x)
 161 {
 162    return 1.0/x;
 163 }
 164
 165
 166 const float rcp_values[] = {
 167    -0.0, 0.0,
 168    -1.0, 1.0,
 169    -1e-007, 1e-007,
 170    -4.0, 4.0,
 171    -1e+035, -100000,
 172    100000, 1e+035,
 173    5.88e-39f, // denormal
 174 #if (__STDC_VERSION__ >= 199901L)
 175    INFINITY, -INFINITY,
 176 #endif
 177 };
 178
 179
 180 static float rsqrtf(float x)
 181 {
 182    return 1.0/(float)sqrt(x);
 183 }
 184
 185
 186 const float rsqrt_values[] = {
 187    // http://msdn.microsoft.com/en-us/library/windows/desktop/bb147346.aspx
 188    0.0, // must yield infinity
 189    1.0, // must yield 1.0
 190    1e-007, 4.0,
 191    100000, 1e+035,
 192    5.88e-39f, // denormal
 193 #if (__STDC_VERSION__ >= 199901L)
 194    INFINITY,
 195 #endif
 196 };
 197
 198
 199 const float sincos_values[] = {
 200    -INFINITY,
 201    -5*M_PI/4,
 202    -4*M_PI/4,
 203    -4*M_PI/4,
 204    -3*M_PI/4,
 205    -2*M_PI/4,
 206    -1*M_PI/4,
 207    1*M_PI/4,
 208    2*M_PI/4,
 209    3*M_PI/4,
 210    4*M_PI/4,
 211    5*M_PI/4,
 212    INFINITY,
 213    NAN
 214 };
 215
 216 const float round_values[] = {
 217       -10.0, -1, 0.0, 12.0,
 218       -1.49, -0.25, 1.25, 2.51,
 219       -0.99, -0.01, 0.01, 0.99,
 220       1.401298464324817e-45f, // smallest denormal
 221       -1.401298464324817e-45f,
 222       1.62981451e-08f,
 223       -1.62981451e-08f,
 224       1.62981451e15f, // large number not representable as 32bit int
 225       -1.62981451e15f,
 226       FLT_EPSILON,
 227       -FLT_EPSILON,
 228       1.0f - 0.5f*FLT_EPSILON,
 229       -1.0f + FLT_EPSILON,
 230       FLT_MAX,
 231       -FLT_MAX
 232 };
 233
 234 static float fractf(float x)
 235 {
 236    x -= floorf(x);
 237    if (x >= 1.0f) {
 238       // clamp to the largest number smaller than one
 239       x = 1.0f - 0.5f*FLT_EPSILON;
 240    }
 241    return x;
 242 }
 243
 244
 245 const float fract_values[] = {
 246    // http://en.wikipedia.org/wiki/IEEE_754-1985#Examples
 247    0.0f,
 248    -0.0f,
 249    1.0f,
 250    -1.0f,
 251    0.5f,
 252    -0.5f,
 253    1.401298464324817e-45f, // smallest denormal
 254    -1.401298464324817e-45f,
 255    5.88e-39f, // middle denormal
 256    1.18e-38f, // largest denormal
 257    -1.18e-38f,
 258    -1.62981451e-08f,
 259    FLT_EPSILON,
 260    -FLT_EPSILON,
 261    1.0f - 0.5f*FLT_EPSILON,
 262    -1.0f + FLT_EPSILON,
 263    FLT_MAX,
 264    -FLT_MAX
 265 };
 266
 267
 268 /*
 269  * Unary test cases.
 270  */
 271
 272 static const struct unary_test_t
 273 unary_tests[] = {
 274    {"neg", &lp_build_negate, &negf, exp2_values, Elements(exp2_values), 20.0 },
 275    {"exp2", &lp_build_exp2, &exp2f, exp2_values, Elements(exp2_values), 20.0 },
 276    {"log2", &lp_build_log2_safe, &log2f, log2_values, Elements(log2_values), 20.0 },
 277    {"exp", &lp_build_exp, &expf, exp2_values, Elements(exp2_values), 18.0 },
 278    {"log", &lp_build_log_safe, &logf, log2_values, Elements(log2_values), 20.0 },
 279    {"rcp", &lp_build_rcp, &rcpf, rcp_values, Elements(rcp_values), 20.0 },
 280    {"rsqrt", &lp_build_rsqrt, &rsqrtf, rsqrt_values, Elements(rsqrt_values), 20.0 },
 281    {"sin", &lp_build_sin, &sinf, sincos_values, Elements(sincos_values), 20.0 },
 282    {"cos", &lp_build_cos, &cosf, sincos_values, Elements(sincos_values), 20.0 },
 283    {"sgn", &lp_build_sgn, &sgnf, exp2_values, Elements(exp2_values), 20.0 },
 284    {"round", &lp_build_round, &roundf, round_values, Elements(round_values), 24.0 },
 285    {"trunc", &lp_build_trunc, &truncf, round_values, Elements(round_values), 24.0 },
 286    {"floor", &lp_build_floor, &floorf, round_values, Elements(round_values), 24.0 },
 287    {"ceil", &lp_build_ceil, &ceilf, round_values, Elements(round_values), 24.0 },
 288    {"fract", &lp_build_fract_safe, &fractf, fract_values, Elements(fract_values), 24.0 },
 289 };
 290
 291
 292 /*
 293  * Build LLVM function that exercises the unary operator builder.
 294  */
 295 static LLVMValueRef
 296 build_unary_test_func(struct gallivm_state *gallivm,
 297                       const struct unary_test_t *test)
 298 {
 299    struct lp_type type = lp_type_float_vec(32, lp_native_vector_width);
 300    LLVMContextRef context = gallivm->context;
 301    LLVMModuleRef module = gallivm->module;
 302    LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
 303    LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
 304    LLVMValueRef func = LLVMAddFunction(module, test->name,
 305                                        LLVMFunctionType(LLVMVoidTypeInContext(context),
 306                                                         args, Elements(args), 0));
 307    LLVMValueRef arg0 = LLVMGetParam(func, 0);
 308    LLVMValueRef arg1 = LLVMGetParam(func, 1);
 309    LLVMBuilderRef builder = gallivm->builder;
 310    LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
 311    LLVMValueRef ret;
 312
 313    struct lp_build_context bld;
 314
 315    lp_build_context_init(&bld, gallivm, type);
 316
 317    LLVMSetFunctionCallConv(func, LLVMCCallConv);
 318
 319    LLVMPositionBuilderAtEnd(builder, block);
 320
 321    arg1 = LLVMBuildLoad(builder, arg1, "");
 322
 323    ret = test->builder(&bld, arg1);
 324
 325    LLVMBuildStore(builder, ret, arg0);
 326
 327    LLVMBuildRetVoid(builder);
 328
 329    gallivm_verify_function(gallivm, func);
 330
 331    return func;
 332 }
 333
 334
 335 /*
 336  * Test one LLVM unary arithmetic builder function.
 337  */
 338 static boolean
 339 test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test)
 340 {
 341    struct gallivm_state *gallivm;
 342    LLVMValueRef test_func;
 343    unary_func_t test_func_jit;
 344    boolean success = TRUE;
 345    int i, j;
 346    int length = lp_native_vector_width / 32;
 347    float *in, *out;
 348
 349    in = align_malloc(length * 4, length * 4);
 350    out = align_malloc(length * 4, length * 4);
 351
 352    /* random NaNs or 0s could wreak havoc */
 353    for (i = 0; i < length; i++) {
 354       in[i] = 1.0;
 355    }
 356
 357    gallivm = gallivm_create("test_module", LLVMGetGlobalContext());
 358
 359    test_func = build_unary_test_func(gallivm, test);
 360
 361    gallivm_compile_module(gallivm);
 362
 363    test_func_jit = (unary_func_t) gallivm_jit_function(gallivm, test_func);
 364
 365    gallivm_free_ir(gallivm);
 366
 367    for (j = 0; j < (test->num_values + length - 1) / length; j++) {
 368       int num_vals = ((j + 1) * length <= test->num_values) ? length :
 369                                                               test->num_values % length;
 370
 371       for (i = 0; i < num_vals; ++i) {
 372          in[i] = test->values[i+j*length];
 373       }
 374
 375       test_func_jit(out, in);
 376       for (i = 0; i < num_vals; ++i) {
 377          float ref = test->ref(in[i]);
 378          double error, precision;
 379          bool pass;
 380
 381          if (util_inf_sign(ref) && util_inf_sign(out[i]) == util_inf_sign(ref)) {
 382             error = 0;
 383          } else {
 384             error = fabs(out[i] - ref);
 385          }
 386          precision = error ? -log2(error/fabs(ref)) : FLT_MANT_DIG;
 387
 388          pass = precision >= test->precision;
 389
 390          if (isnan(ref)) {
 391             continue;
 392          }
 393
 394          if (!pass || verbose) {
 395             printf("%s(%.9g): ref = %.9g, out = %.9g, precision = %f bits, %s\n",
 396                   test->name, in[i], ref, out[i], precision,
 397                   pass ? "PASS" : "FAIL");
 398             fflush(stdout);
 399          }
 400
 401          if (!pass) {
 402             success = FALSE;
 403          }
 404       }
 405    }
 406
 407    gallivm_destroy(gallivm);
 408
 409    align_free(in);
 410    align_free(out);
 411
 412    return success;
 413 }
 414
 415
 416 boolean
 417 test_all(unsigned verbose, FILE *fp)
 418 {
 419    boolean success = TRUE;
 420    int i;
 421
 422    for (i = 0; i < Elements(unary_tests); ++i) {
 423       if (!test_unary(verbose, fp, &unary_tests[i])) {
 424          success = FALSE;
 425       }
 426    }
 427
 428    return success;
 429 }
 430
 431
 432 boolean
 433 test_some(unsigned verbose, FILE *fp,
 434           unsigned long n)
 435 {
 436    /*
 437     * Not randomly generated test cases, so test all.
 438     */
 439
 440    return test_all(verbose, fp);
 441 }
 442
 443
 444 boolean
 445 test_single(unsigned verbose, FILE *fp)
 446 {
 447    return TRUE;
 448 }