src/mesa/math/m_debug_util.h

   1 /*
   2  * Mesa 3-D graphics library
   3  * Version:  6.1
   4  *
   5  * Copyright (C) 1999-2004  Brian Paul   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 "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  11  * and/or sell copies of the Software, and to permit persons to whom the
  12  * Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included
  15  * in all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  18  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  21  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  22  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Gareth Hughes
  26  */
  27
  28 #ifndef __M_DEBUG_UTIL_H__
  29 #define __M_DEBUG_UTIL_H__
  30
  31
  32 #ifdef DEBUG  /* This code only used for debugging */
  33
  34
  35 /* Comment this out to deactivate the cycle counter.
  36  * NOTE: it works only on CPUs which know the 'rdtsc' command (586 or higher)
  37  * (hope, you don't try to debug Mesa on a 386 ;)
  38  */
  39 #if defined(__GNUC__) && \
  40     ((defined(__i386__) && defined(USE_X86_ASM)) || \
  41      (defined(__sparc__) && defined(USE_SPARC_ASM)))
  42 #define  RUN_DEBUG_BENCHMARK
  43 #endif
  44
  45 #define TEST_COUNT              128     /* size of the tested vector array   */
  46
  47 #define REQUIRED_PRECISION      10      /* allow 4 bits to miss              */
  48 #define MAX_PRECISION           24      /* max. precision possible           */
  49
  50
  51 #ifdef  RUN_DEBUG_BENCHMARK
  52 /* Overhead of profiling counter in cycles.  Automatically adjusted to
  53  * your machine at run time - counter initialization should give very
  54  * consistent results.
  55  */
  56 extern long counter_overhead;
  57
  58 /* This is the value of the environment variable MESA_PROFILE, and is
  59  * used to determine if we should benchmark the functions as well as
  60  * verify their correctness.
  61  */
  62 extern char *mesa_profile;
  63
  64 /* Modify the the number of tests if you like.
  65  * We take the minimum of all results, because every error should be
  66  * positive (time used by other processes, task switches etc).
  67  * It is assumed that all calculations are done in the cache.
  68  */
  69
  70 #if defined(__i386__)
  71
  72 #if 1 /* PPro, PII, PIII version */
  73
  74 /* Profiling on the P6 architecture requires a little more work, due to
  75  * the internal out-of-order execution.  We must perform a serializing
  76  * 'cpuid' instruction before and after the 'rdtsc' instructions to make
  77  * sure no other uops are executed when we sample the timestamp counter.
  78  */
  79 #define  INIT_COUNTER()                                                 \
  80    do {                                                                 \
  81       int cycle_i;                                                      \
  82       counter_overhead = LONG_MAX;                                      \
  83       for ( cycle_i = 0 ; cycle_i < 8 ; cycle_i++ ) {                   \
  84          long cycle_tmp1 = 0, cycle_tmp2 = 0;                           \
  85          __asm__ __volatile__ ( "push %%ebx       \n"                   \
  86                                 "xor %%eax, %%eax \n"                   \
  87                                 "cpuid            \n"                   \
  88                                 "rdtsc            \n"                   \
  89                                 "mov %%eax, %0    \n"                   \
  90                                 "xor %%eax, %%eax \n"                   \
  91                                 "cpuid            \n"                   \
  92                                 "pop %%ebx        \n"                   \
  93                                 "push %%ebx       \n"                   \
  94                                 "xor %%eax, %%eax \n"                   \
  95                                 "cpuid            \n"                   \
  96                                 "rdtsc            \n"                   \
  97                                 "mov %%eax, %1    \n"                   \
  98                                 "xor %%eax, %%eax \n"                   \
  99                                 "cpuid            \n"                   \
 100                                 "pop %%ebx        \n"                   \
 101                                 : "=m" (cycle_tmp1), "=m" (cycle_tmp2)  \
 102                                 : : "eax", "ecx", "edx" );              \
 103          if ( counter_overhead > (cycle_tmp2 - cycle_tmp1) ) {          \
 104             counter_overhead = cycle_tmp2 - cycle_tmp1;                 \
 105          }                                                              \
 106       }                                                                 \
 107    } while (0)
 108
 109 #define  BEGIN_RACE(x)                                                  \
 110    x = LONG_MAX;                                                        \
 111    for ( cycle_i = 0 ; cycle_i < 10 ; cycle_i++ ) {                     \
 112       long cycle_tmp1 = 0, cycle_tmp2 = 0;                              \
 113       __asm__ __volatile__ ( "push %%ebx       \n"                      \
 114                              "xor %%eax, %%eax \n"                      \
 115                              "cpuid            \n"                      \
 116                              "rdtsc            \n"                      \
 117                              "mov %%eax, %0    \n"                      \
 118                              "xor %%eax, %%eax \n"                      \
 119                              "cpuid            \n"                      \
 120                              "pop %%ebx        \n"                      \
 121                              : "=m" (cycle_tmp1)                        \
 122                              : : "eax", "ecx", "edx" );
 123
 124 #define END_RACE(x)                                                     \
 125       __asm__ __volatile__ ( "push %%ebx       \n"                      \
 126                              "xor %%eax, %%eax \n"                      \
 127                              "cpuid            \n"                      \
 128                              "rdtsc            \n"                      \
 129                              "mov %%eax, %0    \n"                      \
 130                              "xor %%eax, %%eax \n"                      \
 131                              "cpuid            \n"                      \
 132                              "pop %%ebx        \n"                      \
 133                              : "=m" (cycle_tmp2)                        \
 134                              : : "eax", "ecx", "edx" );                 \
 135       if ( x > (cycle_tmp2 - cycle_tmp1) ) {                            \
 136          x = cycle_tmp2 - cycle_tmp1;                                   \
 137       }                                                                 \
 138    }                                                                    \
 139    x -= counter_overhead;
 140
 141 #else /* PPlain, PMMX version */
 142
 143 /* To ensure accurate results, we stall the pipelines with the
 144  * non-pairable 'cdq' instruction.  This ensures all the code being
 145  * profiled is complete when the 'rdtsc' instruction executes.
 146  */
 147 #define  INIT_COUNTER(x)                                                \
 148    do {                                                                 \
 149       int cycle_i;                                                      \
 150       x = LONG_MAX;                                                     \
 151       for ( cycle_i = 0 ; cycle_i < 32 ; cycle_i++ ) {                  \
 152          long cycle_tmp1, cycle_tmp2, dummy;                            \
 153          __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp1) );               \
 154          __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp2) );               \
 155          __asm__ ( "cdq" );                                             \
 156          __asm__ ( "cdq" );                                             \
 157          __asm__ ( "rdtsc" : "=a" (cycle_tmp1), "=d" (dummy) );         \
 158          __asm__ ( "cdq" );                                             \
 159          __asm__ ( "cdq" );                                             \
 160          __asm__ ( "rdtsc" : "=a" (cycle_tmp2), "=d" (dummy) );         \
 161          if ( x > (cycle_tmp2 - cycle_tmp1) )                           \
 162             x = cycle_tmp2 - cycle_tmp1;                                \
 163       }                                                                 \
 164    } while (0)
 165
 166 #define  BEGIN_RACE(x)                                                  \
 167    x = LONG_MAX;                                                        \
 168    for ( cycle_i = 0 ; cycle_i < 16 ; cycle_i++ ) {                     \
 169       long cycle_tmp1, cycle_tmp2, dummy;                               \
 170       __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp1) );                  \
 171       __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp2) );                  \
 172       __asm__ ( "cdq" );                                                \
 173       __asm__ ( "cdq" );                                                \
 174       __asm__ ( "rdtsc" : "=a" (cycle_tmp1), "=d" (dummy) );
 175
 176
 177 #define END_RACE(x)                                                     \
 178       __asm__ ( "cdq" );                                                \
 179       __asm__ ( "cdq" );                                                \
 180       __asm__ ( "rdtsc" : "=a" (cycle_tmp2), "=d" (dummy) );            \
 181       if ( x > (cycle_tmp2 - cycle_tmp1) )                              \
 182          x = cycle_tmp2 - cycle_tmp1;                                   \
 183    }                                                                    \
 184    x -= counter_overhead;
 185
 186 #endif
 187
 188 #elif defined(__sparc__)
 189
 190 #define  INIT_COUNTER() \
 191          do { counter_overhead = 5; } while(0)
 192
 193 #define  BEGIN_RACE(x)                                                        \
 194 x = LONG_MAX;                                                                 \
 195 for (cycle_i = 0; cycle_i <10; cycle_i++) {                                   \
 196    register long cycle_tmp1 asm("l0");                                        \
 197    register long cycle_tmp2 asm("l1");                                        \
 198    /* rd %tick, %l0 */                                                        \
 199    __asm__ __volatile__ (".word 0xa1410000" : "=r" (cycle_tmp1));  /*  save timestamp   */
 200
 201 #define END_RACE(x)                                                           \
 202    /* rd %tick, %l1 */                                                        \
 203    __asm__ __volatile__ (".word 0xa3410000" : "=r" (cycle_tmp2));             \
 204    if (x > (cycle_tmp2-cycle_tmp1)) x = cycle_tmp2 - cycle_tmp1;              \
 205 }                                                                             \
 206 x -= counter_overhead;
 207
 208 #else
 209 #error Your processor is not supported for RUN_XFORM_BENCHMARK
 210 #endif
 211
 212 #else
 213
 214 #define BEGIN_RACE(x)
 215 #define END_RACE(x)
 216
 217 #endif
 218
 219
 220 /* =============================================================
 221  * Helper functions
 222  */
 223
 224 static GLfloat rnd( void )
 225 {
 226    GLfloat f = (GLfloat)rand() / (GLfloat)RAND_MAX;
 227    GLfloat gran = (GLfloat)(1 << 13);
 228
 229    f = (GLfloat)(GLint)(f * gran) / gran;
 230
 231    return f * 2.0 - 1.0;
 232 }
 233
 234 static int significand_match( GLfloat a, GLfloat b )
 235 {
 236    GLfloat d = a - b;
 237    int a_ex, b_ex, d_ex;
 238
 239    if ( d == 0.0F ) {
 240       return MAX_PRECISION;   /* Exact match */
 241    }
 242
 243    if ( a == 0.0F || b == 0.0F ) {
 244       /* It would probably be better to check if the
 245        * non-zero number is denormalized and return
 246        * the index of the highest set bit here.
 247        */
 248       return 0;
 249    }
 250
 251    frexp( a, &a_ex );
 252    frexp( b, &b_ex );
 253    frexp( d, &d_ex );
 254
 255    if ( a_ex < b_ex ) {
 256       return a_ex - d_ex;
 257    } else {
 258       return b_ex - d_ex;
 259    }
 260 }
 261
 262 enum { NIL = 0, ONE = 1, NEG = -1, VAR = 2 };
 263
 264 /* Ensure our arrays are correctly aligned.
 265  */
 266 #if defined(__GNUC__)
 267 #  define ALIGN16(type, array)  type array __attribute__ ((aligned (16)))
 268 #elif defined(__MSC__)
 269 #  define ALIGN16(type, array)  type array __declspec(align(16)) /* GH: Does this work? */
 270 #elif defined(__WATCOMC__)
 271 #  define ALIGN16(type, array)                      /* Watcom does not support this */
 272 #elif defined(__xlC__)
 273 #  define ALIGN16(type, array)       type __align (16) array
 274 #else
 275 #  warning "ALIGN16 will not 16-byte align!\n"
 276 #  define ALIGN16
 277 #endif
 278
 279
 280 #endif /* DEBUG */
 281
 282 #endif /* __M_DEBUG_UTIL_H__ */