| 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(__amd64__) |
| 189 |
|
| 190 |
#define rdtscll(val) do { \ |
| 191 |
unsigned int a,d; \ |
| 192 |
__asm__ volatile("rdtsc" : "=a" (a), "=d" (d)); \ |
| 193 |
(val) = ((unsigned long)a) | (((unsigned long)d)<<32); \ |
| 194 |
} while(0) |
| 195 |
|
| 196 |
/* Copied from i386 PIII version */ |
| 197 |
#define INIT_COUNTER() \ |
| 198 |
do { \ |
| 199 |
int cycle_i; \ |
| 200 |
counter_overhead = LONG_MAX; \ |
| 201 |
for ( cycle_i = 0 ; cycle_i < 16 ; cycle_i++ ) { \ |
| 202 |
unsigned long cycle_tmp1, cycle_tmp2; \ |
| 203 |
rdtscll(cycle_tmp1); \ |
| 204 |
rdtscll(cycle_tmp2); \ |
| 205 |
if ( counter_overhead > (cycle_tmp2 - cycle_tmp1) ) { \ |
| 206 |
counter_overhead = cycle_tmp2 - cycle_tmp1; \ |
| 207 |
} \ |
| 208 |
} \ |
| 209 |
} while (0) |
| 210 |
|
| 211 |
|
| 212 |
#define BEGIN_RACE(x) \ |
| 213 |
x = LONG_MAX; \ |
| 214 |
for ( cycle_i = 0 ; cycle_i < 10 ; cycle_i++ ) { \ |
| 215 |
unsigned long cycle_tmp1, cycle_tmp2; \ |
| 216 |
rdtscll(cycle_tmp1); \ |
| 217 |
|
| 218 |
#define END_RACE(x) \ |
| 219 |
rdtscll(cycle_tmp2); \ |
| 220 |
if ( x > (cycle_tmp2 - cycle_tmp1) ) { \ |
| 221 |
x = cycle_tmp2 - cycle_tmp1; \ |
| 222 |
} \ |
| 223 |
} \ |
| 224 |
x -= counter_overhead; |
| 225 |
|
| 226 |
#elif defined(__sparc__) |
| 227 |
|
| 228 |
#define INIT_COUNTER() \ |
| 229 |
do { counter_overhead = 5; } while(0) |
| 230 |
|
| 231 |
#define BEGIN_RACE(x) \ |
| 232 |
x = LONG_MAX; \ |
| 233 |
for (cycle_i = 0; cycle_i <10; cycle_i++) { \ |
| 234 |
register long cycle_tmp1 asm("l0"); \ |
| 235 |
register long cycle_tmp2 asm("l1"); \ |
| 236 |
/* rd %tick, %l0 */ \ |
| 237 |
__asm__ __volatile__ (".word 0xa1410000" : "=r" (cycle_tmp1)); /* save timestamp */ |
| 238 |
|
| 239 |
#define END_RACE(x) \ |
| 240 |
/* rd %tick, %l1 */ \ |
| 241 |
__asm__ __volatile__ (".word 0xa3410000" : "=r" (cycle_tmp2)); \ |
| 242 |
if (x > (cycle_tmp2-cycle_tmp1)) x = cycle_tmp2 - cycle_tmp1; \ |
| 243 |
} \ |
| 244 |
x -= counter_overhead; |
| 245 |
|
| 246 |
#else |
| 247 |
#error Your processor is not supported for RUN_XFORM_BENCHMARK |
| 248 |
#endif |
| 249 |
|
| 250 |
#else |
| 251 |
|
| 252 |
#define BEGIN_RACE(x) |
| 253 |
#define END_RACE(x) |
| 254 |
|
| 255 |
#endif |
| 256 |
|
| 257 |
|
| 258 |
/* ============================================================= |
| 259 |
* Helper functions |
| 260 |
*/ |
| 261 |
|
| 262 |
static GLfloat rnd( void ) |
| 263 |
{ |
| 264 |
GLfloat f = (GLfloat)rand() / (GLfloat)RAND_MAX; |
| 265 |
GLfloat gran = (GLfloat)(1 << 13); |
| 266 |
|
| 267 |
f = (GLfloat)(GLint)(f * gran) / gran; |
| 268 |
|
| 269 |
return f * 2.0 - 1.0; |
| 270 |
} |
| 271 |
|
| 272 |
static int significand_match( GLfloat a, GLfloat b ) |
| 273 |
{ |
| 274 |
GLfloat d = a - b; |
| 275 |
int a_ex, b_ex, d_ex; |
| 276 |
|
| 277 |
if ( d == 0.0F ) { |
| 278 |
return MAX_PRECISION; /* Exact match */ |
| 279 |
} |
| 280 |
|
| 281 |
if ( a == 0.0F || b == 0.0F ) { |
| 282 |
/* It would probably be better to check if the |
| 283 |
* non-zero number is denormalized and return |
| 284 |
* the index of the highest set bit here. |
| 285 |
*/ |
| 286 |
return 0; |
| 287 |
} |
| 288 |
|
| 289 |
FREXPF( a, &a_ex ); |
| 290 |
FREXPF( b, &b_ex ); |
| 291 |
FREXPF( d, &d_ex ); |
| 292 |
|
| 293 |
if ( a_ex < b_ex ) { |
| 294 |
return a_ex - d_ex; |
| 295 |
} else { |
| 296 |
return b_ex - d_ex; |
| 297 |
} |
| 298 |
} |
| 299 |
|
| 300 |
enum { NIL = 0, ONE = 1, NEG = -1, VAR = 2 }; |
| 301 |
|
| 302 |
/* Ensure our arrays are correctly aligned. |
| 303 |
*/ |
| 304 |
#if defined(__GNUC__) |
| 305 |
# define ALIGN16(type, array) type array __attribute__ ((aligned (16))) |
| 306 |
#elif defined(__MSC__) |
| 307 |
# define ALIGN16(type, array) type array __declspec(align(16)) /* GH: Does this work? */ |
| 308 |
#elif defined(__WATCOMC__) |
| 309 |
# define ALIGN16(type, array) /* Watcom does not support this */ |
| 310 |
#elif defined(__xlC__) |
| 311 |
# define ALIGN16(type, array) type __align (16) array |
| 312 |
#else |
| 313 |
# warning "ALIGN16 will not 16-byte align!\n" |
| 314 |
# define ALIGN16 |
| 315 |
#endif |
| 316 |
|
| 317 |
|
| 318 |
#endif /* DEBUG */ |
| 319 |
|
| 320 |
#endif /* __M_DEBUG_UTIL_H__ */ |