src/mesa/math/m_debug_util.h

   1 /* $Id: m_debug_util.h,v 1.3 2001/03/30 14:44:43 gareth Exp $ */
   2
   3 /*
   4  * Mesa 3-D graphics library
   5  * Version:  3.5
   6  *
   7  * Copyright (C) 1999-2001  Brian Paul   All Rights Reserved.
   8  *
   9  * Permission is hereby granted, free of charge, to any person obtaining a
  10  * copy of this software and associated documentation files (the "Software"),
  11  * to deal in the Software without restriction, including without limitation
  12  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  13  * and/or sell copies of the Software, and to permit persons to whom the
  14  * Software is furnished to do so, subject to the following conditions:
  15  *
  16  * The above copyright notice and this permission notice shall be included
  17  * in all copies or substantial portions 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 MERCHANTABILITY,
  21  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  22  * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
  23  * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  24  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  * Authors:
  27  *    Gareth Hughes <gareth@valinux.com>
  28  */
  29
  30 #ifndef __M_DEBUG_UTIL_H__
  31 #define __M_DEBUG_UTIL_H__
  32
  33
  34 #ifdef DEBUG  /* This code only used for debugging */
  35
  36
  37 /* Comment this out to deactivate the cycle counter.
  38  * NOTE: it works only on CPUs which know the 'rdtsc' command (586 or higher)
  39  * (hope, you don't try to debug Mesa on a 386 ;)
  40  */
  41 #if defined(__GNUC__) && defined(__i386__) && defined(USE_X86_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 1 /* PPro, PII, PIII version */
  71
  72 /* Profiling on the P6 architecture requires a little more work, due to
  73  * the internal out-of-order execution.  We must perform a serializing
  74  * 'cpuid' instruction before and after the 'rdtsc' instructions to make
  75  * sure no other uops are executed when we sample the timestamp counter.
  76  */
  77 #define  INIT_COUNTER()                                                 \
  78    do {                                                                 \
  79       int cycle_i;                                                      \
  80       counter_overhead = LONG_MAX;                                      \
  81       for ( cycle_i = 0 ; cycle_i < 8 ; cycle_i++ ) {                   \
  82          long cycle_tmp1 = 0, cycle_tmp2 = 0;                           \
  83          __asm__ __volatile__ ( "push %%ebx       \n"                   \
  84                                 "xor %%eax, %%eax \n"                   \
  85                                 "cpuid            \n"                   \
  86                                 "rdtsc            \n"                   \
  87                                 "mov %%eax, %0    \n"                   \
  88                                 "xor %%eax, %%eax \n"                   \
  89                                 "cpuid            \n"                   \
  90                                 "pop %%ebx        \n"                   \
  91                                 "push %%ebx       \n"                   \
  92                                 "xor %%eax, %%eax \n"                   \
  93                                 "cpuid            \n"                   \
  94                                 "rdtsc            \n"                   \
  95                                 "mov %%eax, %1    \n"                   \
  96                                 "xor %%eax, %%eax \n"                   \
  97                                 "cpuid            \n"                   \
  98                                 "pop %%ebx        \n"                   \
  99                                 : "=m" (cycle_tmp1), "=m" (cycle_tmp2)  \
 100                                 : : "eax", "ecx", "edx" );              \
 101          if ( counter_overhead > (cycle_tmp2 - cycle_tmp1) ) {          \
 102             counter_overhead = cycle_tmp2 - cycle_tmp1;                 \
 103          }                                                              \
 104       }                                                                 \
 105    } while (0)
 106
 107 #define  BEGIN_RACE(x)                                                  \
 108    x = LONG_MAX;                                                        \
 109    for ( cycle_i = 0 ; cycle_i < 10 ; cycle_i++ ) {                     \
 110       long cycle_tmp1 = 0, cycle_tmp2 = 0;                              \
 111       __asm__ __volatile__ ( "push %%ebx       \n"                      \
 112                              "xor %%eax, %%eax \n"                      \
 113                              "cpuid            \n"                      \
 114                              "rdtsc            \n"                      \
 115                              "mov %%eax, %0    \n"                      \
 116                              "xor %%eax, %%eax \n"                      \
 117                              "cpuid            \n"                      \
 118                              "pop %%ebx        \n"                      \
 119                              : "=m" (cycle_tmp1)                        \
 120                              : : "eax", "ecx", "edx" );
 121
 122 #define END_RACE(x)                                                     \
 123       __asm__ __volatile__ ( "push %%ebx       \n"                      \
 124                              "xor %%eax, %%eax \n"                      \
 125                              "cpuid            \n"                      \
 126                              "rdtsc            \n"                      \
 127                              "mov %%eax, %0    \n"                      \
 128                              "xor %%eax, %%eax \n"                      \
 129                              "cpuid            \n"                      \
 130                              "pop %%ebx        \n"                      \
 131                              : "=m" (cycle_tmp2)                        \
 132                              : : "eax", "ecx", "edx" );                 \
 133       if ( x > (cycle_tmp2 - cycle_tmp1) ) {                            \
 134          x = cycle_tmp2 - cycle_tmp1;                                   \
 135       }                                                                 \
 136    }                                                                    \
 137    x -= counter_overhead;
 138
 139 #else /* PPlain, PMMX version */
 140
 141 /* To ensure accurate results, we stall the pipelines with the
 142  * non-pairable 'cdq' instruction.  This ensures all the code being
 143  * profiled is complete when the 'rdtsc' instruction executes.
 144  */
 145 #define  INIT_COUNTER(x)                                                \
 146    do {                                                                 \
 147       int cycle_i;                                                      \
 148       x = LONG_MAX;                                                     \
 149       for ( cycle_i = 0 ; cycle_i < 32 ; cycle_i++ ) {                  \
 150          long cycle_tmp1, cycle_tmp2, dummy;                            \
 151          __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp1) );               \
 152          __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp2) );               \
 153          __asm__ ( "cdq" );                                             \
 154          __asm__ ( "cdq" );                                             \
 155          __asm__ ( "rdtsc" : "=a" (cycle_tmp1), "=d" (dummy) );         \
 156          __asm__ ( "cdq" );                                             \
 157          __asm__ ( "cdq" );                                             \
 158          __asm__ ( "rdtsc" : "=a" (cycle_tmp2), "=d" (dummy) );         \
 159          if ( x > (cycle_tmp2 - cycle_tmp1) )                           \
 160             x = cycle_tmp2 - cycle_tmp1;                                \
 161       }                                                                 \
 162    } while (0)
 163
 164 #define  BEGIN_RACE(x)                                                  \
 165    x = LONG_MAX;                                                        \
 166    for ( cycle_i = 0 ; cycle_i < 16 ; cycle_i++ ) {                     \
 167       long cycle_tmp1, cycle_tmp2, dummy;                               \
 168       __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp1) );                  \
 169       __asm__ ( "mov %%eax, %0" : "=a" (cycle_tmp2) );                  \
 170       __asm__ ( "cdq" );                                                \
 171       __asm__ ( "cdq" );                                                \
 172       __asm__ ( "rdtsc" : "=a" (cycle_tmp1), "=d" (dummy) );
 173
 174
 175 #define END_RACE(x)                                                     \
 176       __asm__ ( "cdq" );                                                \
 177       __asm__ ( "cdq" );                                                \
 178       __asm__ ( "rdtsc" : "=a" (cycle_tmp2), "=d" (dummy) );            \
 179       if ( x > (cycle_tmp2 - cycle_tmp1) )                              \
 180          x = cycle_tmp2 - cycle_tmp1;                                   \
 181    }                                                                    \
 182    x -= counter_overhead;
 183
 184 #endif
 185
 186 #else
 187
 188 #define BEGIN_RACE(x)
 189 #define END_RACE(x)
 190
 191 #endif
 192
 193
 194 /* =============================================================
 195  * Helper functions
 196  */
 197
 198 static GLfloat rnd( void )
 199 {
 200    GLfloat f = (GLfloat)rand() / (GLfloat)RAND_MAX;
 201    GLfloat gran = (GLfloat)(1 << 13);
 202
 203    f = (GLfloat)(GLint)(f * gran) / gran;
 204
 205    return f * 2.0 - 1.0;
 206 }
 207
 208 static int significand_match( GLfloat a, GLfloat b )
 209 {
 210    GLfloat d = a - b;
 211    int a_ex, b_ex, d_ex;
 212
 213    if ( d == 0.0F ) {
 214       return MAX_PRECISION;   /* Exact match */
 215    }
 216
 217    if ( a == 0.0F || b == 0.0F ) {
 218       /* It would probably be better to check if the
 219        * non-zero number is denormalized and return
 220        * the index of the highest set bit here.
 221        */
 222       return 0;
 223    }
 224
 225    frexp( a, &a_ex );
 226    frexp( b, &b_ex );
 227    frexp( d, &d_ex );
 228
 229    if ( a_ex < b_ex ) {
 230       return a_ex - d_ex;
 231    } else {
 232       return b_ex - d_ex;
 233    }
 234 }
 235
 236 enum { NIL = 0, ONE = 1, NEG = -1, VAR = 2 };
 237
 238 static void init_matrix( GLfloat *m )
 239 {
 240    m[0] = 63.0; m[4] = 43.0; m[ 8] = 29.0; m[12] = 43.0;
 241    m[1] = 55.0; m[5] = 17.0; m[ 9] = 31.0; m[13] =  7.0;
 242    m[2] = 44.0; m[6] =  9.0; m[10] =  7.0; m[14] =  3.0;
 243    m[3] = 11.0; m[7] = 23.0; m[11] = 91.0; m[15] =  9.0;
 244 }
 245
 246
 247 /* Ensure our arrays are correctly aligned.
 248  */
 249 #if defined(__GNUC__)
 250 #  define ALIGN16       __attribute__ ((aligned (16)))
 251 #elif defined(__MSC__)
 252 #  define ALIGN16       __declspec(align(16)) /* GH: Does this work? */
 253 #else
 254 #  warning "ALIGN16 will not 16-byte align!\n"
 255 #  define ALIGN16
 256 #endif
 257
 258
 259 #endif /* DEBUG */
 260
 261 #endif /* __M_DEBUG_UTIL_H__ */