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 #include "util/u_cpu_detect.h"
  37
  38 #include "gallivm/lp_bld.h"
  39 #include "gallivm/lp_bld_debug.h"
  40 #include "gallivm/lp_bld_init.h"
  41 #include "gallivm/lp_bld_arit.h"
  42
  43 #include "lp_test.h"
  44
  45
  46 void
  47 write_tsv_header(FILE *fp)
  48 {
  49    fprintf(fp,
  50            "result\t"
  51            "format\n");
  52
  53    fflush(fp);
  54 }
  55
  56
  57 typedef void (*unary_func_t)(float *out, const float *in);
  58
  59
  60 /**
  61  * Describe a test case of one unary function.
  62  */
  63 struct unary_test_t
  64 {
  65    /*
  66     * Test name -- name of the mathematical function under test.
  67     */
  68
  69    const char *name;
  70
  71    LLVMValueRef
  72    (*builder)(struct lp_build_context *bld, LLVMValueRef a);
  73
  74    /*
  75     * Reference (pure-C) function.
  76     */
  77    float
  78    (*ref)(float a);
  79
  80    /*
  81     * Test values.
  82     */
  83    const float *values;
  84    unsigned num_values;
  85
  86    /*
  87     * Required precision in bits.
  88     */
  89    double precision;
  90 };
  91
  92
  93 static float negf(float x)
  94 {
  95    return -x;
  96 }
  97
  98
  99 static float sgnf(float x)
 100 {
 101    if (x > 0.0f) {
 102       return 1.0f;
 103    }
 104    if (x < 0.0f) {
 105       return -1.0f;
 106    }
 107    return 0.0f;
 108 }
 109
 110
 111 const float exp2_values[] = {
 112    -INFINITY,
 113    -60,
 114    -4,
 115    -2,
 116    -1,
 117    -1e-007,
 118    0,
 119    1e-007,
 120    0.01,
 121    0.1,
 122    0.9,
 123    0.99,
 124    1,
 125    2,
 126    4,
 127    60,
 128    INFINITY,
 129    NAN
 130 };
 131
 132
 133 const float log2_values[] = {
 134 #if 0
 135    /*
 136     * Smallest denormalized number; meant just for experimentation, but not
 137     * validation.
 138     */
 139    1.4012984643248171e-45,
 140 #endif
 141    -INFINITY,
 142    0,
 143    1e-007,
 144    0.1,
 145    0.5,
 146    0.99,
 147    1,
 148    1.01,
 149    1.1,
 150    1.9,
 151    1.99,
 152    2,
 153    4,
 154    100000,
 155    1e+018,
 156    INFINITY,
 157    NAN
 158 };
 159
 160
 161 static float rcpf(float x)
 162 {
 163    return 1.0/x;
 164 }
 165
 166
 167 const float rcp_values[] = {
 168    -0.0, 0.0,
 169    -1.0, 1.0,
 170    -1e-007, 1e-007,
 171    -4.0, 4.0,
 172    -1e+035, -100000,
 173    100000, 1e+035,
 174    5.88e-39f, // denormal
 175 #if (__STDC_VERSION__ >= 199901L)
 176    INFINITY, -INFINITY,
 177 #endif
 178 };
 179
 180
 181 static float rsqrtf(float x)
 182 {
 183    return 1.0/(float)sqrt(x);
 184 }
 185
 186
 187 const float rsqrt_values[] = {
 188    // http://msdn.microsoft.com/en-us/library/windows/desktop/bb147346.aspx
 189    0.0, // must yield infinity
 190    1.0, // must yield 1.0
 191    1e-007, 4.0,
 192    100000, 1e+035,
 193    5.88e-39f, // denormal
 194 #if (__STDC_VERSION__ >= 199901L)
 195    INFINITY,
 196 #endif
 197 };
 198
 199
 200 const float sincos_values[] = {
 201    -INFINITY,
 202    -5*M_PI/4,
 203    -4*M_PI/4,
 204    -4*M_PI/4,
 205    -3*M_PI/4,
 206    -2*M_PI/4,
 207    -1*M_PI/4,
 208    1*M_PI/4,
 209    2*M_PI/4,
 210    3*M_PI/4,
 211    4*M_PI/4,
 212    5*M_PI/4,
 213    INFINITY,
 214    NAN
 215 };
 216
 217 const float round_values[] = {
 218       -10.0, -1, 0.0, 12.0,
 219       -1.49, -0.25, 1.25, 2.51,
 220       -0.99, -0.01, 0.01, 0.99,
 221       1.401298464324817e-45f, // smallest denormal
 222       -1.401298464324817e-45f,
 223       1.62981451e-08f,
 224       -1.62981451e-08f,
 225       1.62981451e15f, // large number not representable as 32bit int
 226       -1.62981451e15f,
 227       FLT_EPSILON,
 228       -FLT_EPSILON,
 229       1.0f - 0.5f*FLT_EPSILON,
 230       -1.0f + FLT_EPSILON,
 231       FLT_MAX,
 232       -FLT_MAX
 233 };
 234
 235 static float fractf(float x)
 236 {
 237    x -= floorf(x);
 238    if (x >= 1.0f) {
 239       // clamp to the largest number smaller than one
 240       x = 1.0f - 0.5f*FLT_EPSILON;
 241    }
 242    return x;
 243 }
 244
 245
 246 const float fract_values[] = {
 247    // http://en.wikipedia.org/wiki/IEEE_754-1985#Examples
 248    0.0f,
 249    -0.0f,
 250    1.0f,
 251    -1.0f,
 252    0.5f,
 253    -0.5f,
 254    1.401298464324817e-45f, // smallest denormal
 255    -1.401298464324817e-45f,
 256    5.88e-39f, // middle denormal
 257    1.18e-38f, // largest denormal
 258    -1.18e-38f,
 259    -1.62981451e-08f,
 260    FLT_EPSILON,
 261    -FLT_EPSILON,
 262    1.0f - 0.5f*FLT_EPSILON,
 263    -1.0f + FLT_EPSILON,
 264    FLT_MAX,
 265    -FLT_MAX
 266 };
 267
 268
 269 /*
 270  * Unary test cases.
 271  */
 272
 273 static const struct unary_test_t
 274 unary_tests[] = {
 275    {"abs", &lp_build_abs, &fabsf, exp2_values, Elements(exp2_values), 20.0 },
 276    {"neg", &lp_build_negate, &negf, exp2_values, Elements(exp2_values), 20.0 },
 277    {"exp2", &lp_build_exp2, &exp2f, exp2_values, Elements(exp2_values), 20.0 },
 278    {"log2", &lp_build_log2_safe, &log2f, log2_values, Elements(log2_values), 20.0 },
 279    {"exp", &lp_build_exp, &expf, exp2_values, Elements(exp2_values), 18.0 },
 280    {"log", &lp_build_log_safe, &logf, log2_values, Elements(log2_values), 20.0 },
 281    {"rcp", &lp_build_rcp, &rcpf, rcp_values, Elements(rcp_values), 20.0 },
 282    {"rsqrt", &lp_build_rsqrt, &rsqrtf, rsqrt_values, Elements(rsqrt_values), 20.0 },
 283    {"sin", &lp_build_sin, &sinf, sincos_values, Elements(sincos_values), 20.0 },
 284    {"cos", &lp_build_cos, &cosf, sincos_values, Elements(sincos_values), 20.0 },
 285    {"sgn", &lp_build_sgn, &sgnf, exp2_values, Elements(exp2_values), 20.0 },
 286    {"round", &lp_build_round, &roundf, round_values, Elements(round_values), 24.0 },
 287    {"trunc", &lp_build_trunc, &truncf, round_values, Elements(round_values), 24.0 },
 288    {"floor", &lp_build_floor, &floorf, round_values, Elements(round_values), 24.0 },
 289    {"ceil", &lp_build_ceil, &ceilf, round_values, Elements(round_values), 24.0 },
 290    {"fract", &lp_build_fract_safe, &fractf, fract_values, Elements(fract_values), 24.0 },
 291 };
 292
 293
 294 /*
 295  * Build LLVM function that exercises the unary operator builder.
 296  */
 297 static LLVMValueRef
 298 build_unary_test_func(struct gallivm_state *gallivm,
 299                       const struct unary_test_t *test)
 300 {
 301    struct lp_type type = lp_type_float_vec(32, lp_native_vector_width);
 302    LLVMContextRef context = gallivm->context;
 303    LLVMModuleRef module = gallivm->module;
 304    LLVMTypeRef vf32t = lp_build_vec_type(gallivm, type);
 305    LLVMTypeRef args[2] = { LLVMPointerType(vf32t, 0), LLVMPointerType(vf32t, 0) };
 306    LLVMValueRef func = LLVMAddFunction(module, test->name,
 307                                        LLVMFunctionType(LLVMVoidTypeInContext(context),
 308                                                         args, Elements(args), 0));
 309    LLVMValueRef arg0 = LLVMGetParam(func, 0);
 310    LLVMValueRef arg1 = LLVMGetParam(func, 1);
 311    LLVMBuilderRef builder = gallivm->builder;
 312    LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(context, func, "entry");
 313    LLVMValueRef ret;
 314
 315    struct lp_build_context bld;
 316
 317    lp_build_context_init(&bld, gallivm, type);
 318
 319    LLVMSetFunctionCallConv(func, LLVMCCallConv);
 320
 321    LLVMPositionBuilderAtEnd(builder, block);
 322
 323    arg1 = LLVMBuildLoad(builder, arg1, "");
 324
 325    ret = test->builder(&bld, arg1);
 326
 327    LLVMBuildStore(builder, ret, arg0);
 328
 329    LLVMBuildRetVoid(builder);
 330
 331    gallivm_verify_function(gallivm, func);
 332
 333    return func;
 334 }
 335
 336
 337 /*
 338  * Flush denorms to zero.
 339  */
 340 static float
 341 flush_denorm_to_zero(float val)
 342 {
 343    /*
 344     * If we have a denorm manually set it to (+-)0.
 345     * This is because the reference may or may not do the right thing
 346     * otherwise because we want the result according to treating all
 347     * denormals as zero (FTZ/DAZ). Not using fpclassify because
 348     * a) some compilers are stuck at c89 (msvc)
 349     * b) not sure it reliably works with non-standard ftz/daz mode
 350     * And, right now we only disable denorms with jited code on x86/sse
 351     * (albeit this should be classified as a bug) so to get results which
 352     * match we must only flush them to zero here in that case too.
 353     */
 354    union fi fi_val;
 355
 356    fi_val.f = val;
 357
 358 #if defined(PIPE_ARCH_SSE)
 359    if (util_cpu_caps.has_sse) {
 360       if ((fi_val.ui & 0x7f800000) == 0) {
 361          fi_val.ui &= 0xff800000;
 362       }
 363    }
 364 #endif
 365
 366    return fi_val.f;
 367 }
 368
 369 /*
 370  * Test one LLVM unary arithmetic builder function.
 371  */
 372 static boolean
 373 test_unary(unsigned verbose, FILE *fp, const struct unary_test_t *test)
 374 {
 375    struct gallivm_state *gallivm;
 376    LLVMValueRef test_func;
 377    unary_func_t test_func_jit;
 378    boolean success = TRUE;
 379    int i, j;
 380    int length = lp_native_vector_width / 32;
 381    float *in, *out;
 382
 383    in = align_malloc(length * 4, length * 4);
 384    out = align_malloc(length * 4, length * 4);
 385
 386    /* random NaNs or 0s could wreak havoc */
 387    for (i = 0; i < length; i++) {
 388       in[i] = 1.0;
 389    }
 390
 391    gallivm = gallivm_create("test_module", LLVMGetGlobalContext());
 392
 393    test_func = build_unary_test_func(gallivm, test);
 394
 395    gallivm_compile_module(gallivm);
 396
 397    test_func_jit = (unary_func_t) gallivm_jit_function(gallivm, test_func);
 398
 399    gallivm_free_ir(gallivm);
 400
 401    for (j = 0; j < (test->num_values + length - 1) / length; j++) {
 402       int num_vals = ((j + 1) * length <= test->num_values) ? length :
 403                                                               test->num_values % length;
 404
 405       for (i = 0; i < num_vals; ++i) {
 406          in[i] = test->values[i+j*length];
 407       }
 408
 409       test_func_jit(out, in);
 410       for (i = 0; i < num_vals; ++i) {
 411          float testval, ref;
 412          double error, precision;
 413          bool pass;
 414
 415          testval = flush_denorm_to_zero(in[i]);
 416          ref = flush_denorm_to_zero(test->ref(testval));
 417
 418          if (util_inf_sign(ref) && util_inf_sign(out[i]) == util_inf_sign(ref)) {
 419             error = 0;
 420          } else {
 421             error = fabs(out[i] - ref);
 422          }
 423          precision = error ? -log2(error/fabs(ref)) : FLT_MANT_DIG;
 424
 425          pass = precision >= test->precision;
 426
 427          if (isnan(ref)) {
 428             continue;
 429          }
 430
 431          if (!pass || verbose) {
 432             printf("%s(%.9g): ref = %.9g, out = %.9g, precision = %f bits, %s\n",
 433                   test->name, in[i], ref, out[i], precision,
 434                   pass ? "PASS" : "FAIL");
 435             fflush(stdout);
 436          }
 437
 438          if (!pass) {
 439             success = FALSE;
 440          }
 441       }
 442    }
 443
 444    gallivm_destroy(gallivm);
 445
 446    align_free(in);
 447    align_free(out);
 448
 449    return success;
 450 }
 451
 452
 453 boolean
 454 test_all(unsigned verbose, FILE *fp)
 455 {
 456    boolean success = TRUE;
 457    int i;
 458
 459    for (i = 0; i < Elements(unary_tests); ++i) {
 460       if (!test_unary(verbose, fp, &unary_tests[i])) {
 461          success = FALSE;
 462       }
 463    }
 464
 465    return success;
 466 }
 467
 468
 469 boolean
 470 test_some(unsigned verbose, FILE *fp,
 471           unsigned long n)
 472 {
 473    /*
 474     * Not randomly generated test cases, so test all.
 475     */
 476
 477    return test_all(verbose, fp);
 478 }
 479
 480
 481 boolean
 482 test_single(unsigned verbose, FILE *fp)
 483 {
 484    return TRUE;
 485 }