@samitouri / QOSamiQemu
Code Commits Issues Pulls Sessions CI/CD
master
master
c 467 lines 11.8 KB
Raw
1 /*
2 * Test Floating Point Conversion
3 */
4
5 /* we want additional float type definitions */
6 #define __STDC_WANT_IEC_60559_BFP_EXT__
7 #define __STDC_WANT_IEC_60559_TYPES_EXT__
8
9 #include <stdio.h>
10 #include <inttypes.h>
11 #include <math.h>
12 #include <float.h>
13 #include <fenv.h>
14
15 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
16
17 static char flag_str[256];
18
19 static char *get_flag_state(int flags)
20 {
21 if (flags) {
22 snprintf(flag_str, sizeof(flag_str), "%s %s %s %s %s",
23 flags & FE_OVERFLOW ? "OVERFLOW" : "",
24 flags & FE_UNDERFLOW ? "UNDERFLOW" : "",
25 flags & FE_DIVBYZERO ? "DIV0" : "",
26 flags & FE_INEXACT ? "INEXACT" : "",
27 flags & FE_INVALID ? "INVALID" : "");
28 } else {
29 snprintf(flag_str, sizeof(flag_str), "OK");
30 }
31
32 return flag_str;
33 }
34
35 static void print_double_number(int i, double num)
36 {
37 uint64_t double_as_hex = *(uint64_t *) &num;
38 int flags = fetestexcept(FE_ALL_EXCEPT);
39 char *fstr = get_flag_state(flags);
40
41 printf("%02d DOUBLE: %02.20e / %#020" PRIx64 " (%#x => %s)\n",
42 i, num, double_as_hex, flags, fstr);
43 }
44
45 static void print_single_number(int i, float num)
46 {
47 uint32_t single_as_hex = *(uint32_t *) &num;
48 int flags = fetestexcept(FE_ALL_EXCEPT);
49 char *fstr = get_flag_state(flags);
50
51 printf("%02d SINGLE: %02.20e / %#010x (%#x => %s)\n",
52 i, num, single_as_hex, flags, fstr);
53 }
54
55 static void print_half_number(int i, uint16_t num)
56 {
57 int flags = fetestexcept(FE_ALL_EXCEPT);
58 char *fstr = get_flag_state(flags);
59
60 printf("%02d HALF: %#04x (%#x => %s)\n",
61 i, num, flags, fstr);
62 }
63
64 static void print_int64(int i, int64_t num)
65 {
66 uint64_t int64_as_hex = *(uint64_t *) &num;
67 int flags = fetestexcept(FE_ALL_EXCEPT);
68 char *fstr = get_flag_state(flags);
69
70 printf("%02d INT64: %20" PRId64 "/%#020" PRIx64 " (%#x => %s)\n",
71 i, num, int64_as_hex, flags, fstr);
72 }
73
74 #ifndef SNANF
75 /* Signaling NaN macros, if supported. */
76 # define SNANF (__builtin_nansf (""))
77 # define SNAN (__builtin_nans (""))
78 # define SNANL (__builtin_nansl (""))
79 #endif
80
81 float single_numbers[] = { -SNANF,
82 -NAN,
83 -INFINITY,
84 -FLT_MAX,
85 -1.111E+31,
86 -1.111E+30,
87 -1.08700982e-12,
88 -1.78051176e-20,
89 -FLT_MIN,
90 0.0,
91 FLT_MIN,
92 2.98023224e-08,
93 5.96046E-8, /* min positive FP16 subnormal */
94 6.09756E-5, /* max subnormal FP16 */
95 6.10352E-5, /* min positive normal FP16 */
96 1.0,
97 1.0009765625, /* smallest float after 1.0 FP16 */
98 2.0,
99 M_E, M_PI,
100 65503.0,
101 65504.0, /* max FP16 */
102 65505.0,
103 131007.0,
104 131008.0, /* max AFP */
105 131009.0,
106 1.111E+30,
107 FLT_MAX,
108 INFINITY,
109 NAN,
110 SNANF };
111
112 static void convert_single_to_half(void)
113 {
114 int i;
115
116 printf("Converting single-precision to half-precision\n");
117
118 for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
119 float input = single_numbers[i];
120
121 feclearexcept(FE_ALL_EXCEPT);
122
123 print_single_number(i, input);
124 #if defined(__arm__)
125 uint32_t output;
126 asm("vcvtb.f16.f32 %0, %1" : "=t" (output) : "x" (input));
127 #else
128 uint16_t output;
129 asm("fcvt %h0, %s1" : "=w" (output) : "w" (input));
130 #endif
131 print_half_number(i, output);
132 }
133 }
134
135 static void convert_single_to_double(void)
136 {
137 int i;
138
139 printf("Converting single-precision to double-precision\n");
140
141 for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
142 float input = single_numbers[i];
143 /* uint64_t output; */
144 double output;
145
146 feclearexcept(FE_ALL_EXCEPT);
147
148 print_single_number(i, input);
149 #if defined(__arm__)
150 asm("vcvt.f64.f32 %P0, %1" : "=w" (output) : "t" (input));
151 #else
152 asm("fcvt %d0, %s1" : "=w" (output) : "w" (input));
153 #endif
154 print_double_number(i, output);
155 }
156 }
157
158 static void convert_single_to_integer(void)
159 {
160 int i;
161
162 printf("Converting single-precision to integer\n");
163
164 for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
165 float input = single_numbers[i];
166 int64_t output;
167
168 feclearexcept(FE_ALL_EXCEPT);
169
170 print_single_number(i, input);
171 #if defined(__arm__)
172 /* asm("vcvt.s32.f32 %s0, %s1" : "=t" (output) : "t" (input)); */
173 output = input;
174 #else
175 #ifdef FPRCVT
176 asm("fcvtzs d0, %s1\r\n"
177 "fmov %0, d0" :
178 "=r" (output) : "w" (input));
179 #else
180 asm("fcvtzs %0, %s1" : "=r" (output) : "w" (input));
181 #endif
182 #endif
183 print_int64(i, output);
184 }
185 }
186
187 /* This allows us to initialise some doubles as pure hex */
188 typedef union {
189 double d;
190 uint64_t h;
191 } test_doubles;
192
193 test_doubles double_numbers[] = {
194 {SNAN},
195 {-NAN},
196 {-INFINITY},
197 {-DBL_MAX},
198 {-FLT_MAX-1.0},
199 {-FLT_MAX},
200 {-1.111E+31},
201 {-1.111E+30}, /* half prec */
202 {-2.0}, {-1.0},
203 {-DBL_MIN},
204 {-FLT_MIN},
205 {0.0},
206 {FLT_MIN},
207 {2.98023224e-08},
208 {5.96046E-8}, /* min positive FP16 subnormal */
209 {6.09756E-5}, /* max subnormal FP16 */
210 {6.10352E-5}, /* min positive normal FP16 */
211 {1.0},
212 {1.0009765625}, /* smallest float after 1.0 FP16 */
213 {DBL_MIN},
214 {1.3789972848607228e-308},
215 {1.4914738736681624e-308},
216 {1.0}, {2.0},
217 {M_E}, {M_PI},
218 {65503.0},
219 {65504.0}, /* max FP16 */
220 {65505.0},
221 {131007.0},
222 {131008.0}, /* max AFP */
223 {131009.0},
224 {.h = 0x41dfffffffc00000 }, /* to int = 0x7fffffff */
225 {FLT_MAX},
226 {FLT_MAX + 1.0},
227 {DBL_MAX},
228 {INFINITY},
229 {NAN},
230 {.h = 0x7ff0000000000001}, /* SNAN */
231 {SNAN},
232 };
233
234 static void convert_double_to_half(void)
235 {
236 int i;
237
238 printf("Converting double-precision to half-precision\n");
239
240 for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
241 double input = double_numbers[i].d;
242 uint16_t output;
243
244 feclearexcept(FE_ALL_EXCEPT);
245
246 print_double_number(i, input);
247
248 /* as we don't have _Float16 support */
249 #if defined(__arm__)
250 /* asm("vcvtb.f16.f64 %0, %P1" : "=t" (output) : "x" (input)); */
251 output = input;
252 #else
253 asm("fcvt %h0, %d1" : "=w" (output) : "w" (input));
254 #endif
255 print_half_number(i, output);
256 }
257 }
258
259 static void convert_double_to_single(void)
260 {
261 int i;
262
263 printf("Converting double-precision to single-precision\n");
264
265 for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
266 double input = double_numbers[i].d;
267 float output;
268
269 feclearexcept(FE_ALL_EXCEPT);
270
271 print_double_number(i, input);
272
273 #if defined(__arm__)
274 asm("vcvt.f32.f64 %0, %P1" : "=w" (output) : "x" (input));
275 #else
276 asm("fcvt %s0, %d1" : "=w" (output) : "w" (input));
277 #endif
278
279 print_single_number(i, output);
280 }
281 }
282
283 static void convert_double_to_integer(void)
284 {
285 int i;
286
287 printf("Converting double-precision to integer\n");
288
289 for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
290 double input = double_numbers[i].d;
291 int64_t output;
292
293 feclearexcept(FE_ALL_EXCEPT);
294
295 print_double_number(i, input);
296 #if defined(__arm__)
297 /* asm("vcvt.s32.f32 %s0, %s1" : "=t" (output) : "t" (input)); */
298 output = input;
299 #else
300 asm("fcvtzs %0, %d1" : "=r" (output) : "w" (input));
301 #endif
302 print_int64(i, output);
303 }
304 }
305
306 /* no handy defines for these numbers */
307 uint16_t half_numbers[] = {
308 0xffff, /* -NaN / AHP -Max */
309 0xfcff, /* -NaN / AHP */
310 0xfc01, /* -NaN / AHP */
311 0xfc00, /* -Inf */
312 0xfbff, /* -Max */
313 0xc000, /* -2 */
314 0xbc00, /* -1 */
315 0x8001, /* -MIN subnormal */
316 0x8000, /* -0 */
317 0x0000, /* +0 */
318 0x0001, /* MIN subnormal */
319 0x3c00, /* 1 */
320 0x7bff, /* Max */
321 0x7c00, /* Inf */
322 0x7c01, /* NaN / AHP */
323 0x7cff, /* NaN / AHP */
324 0x7fff, /* NaN / AHP +Max*/
325 };
326
327 static void convert_half_to_double(void)
328 {
329 int i;
330
331 printf("Converting half-precision to double-precision\n");
332
333 for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
334 uint16_t input = half_numbers[i];
335 double output;
336
337 feclearexcept(FE_ALL_EXCEPT);
338
339 print_half_number(i, input);
340 #if defined(__arm__)
341 /* asm("vcvtb.f64.f16 %P0, %1" : "=w" (output) : "t" (input)); */
342 output = input;
343 #else
344 asm("fcvt %d0, %h1" : "=w" (output) : "w" (input));
345 #endif
346 print_double_number(i, output);
347 }
348 }
349
350 static void convert_half_to_single(void)
351 {
352 int i;
353
354 printf("Converting half-precision to single-precision\n");
355
356 for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
357 uint16_t input = half_numbers[i];
358 float output;
359
360 feclearexcept(FE_ALL_EXCEPT);
361
362 print_half_number(i, input);
363 #if defined(__arm__)
364 /*
365 * Clang refuses to allocate an integer to a fp register.
366 * Perform the move from a general register by hand.
367 */
368 asm("vmov %0, %1\n\t"
369 "vcvtb.f32.f16 %0, %0" : "=w" (output) : "r" (input));
370 #else
371 asm("fcvt %s0, %h1" : "=w" (output) : "w" (input));
372 #endif
373 print_single_number(i, output);
374 }
375 }
376
377 static void convert_half_to_integer(void)
378 {
379 int i;
380
381 printf("Converting half-precision to integer\n");
382
383 for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
384 uint16_t input = half_numbers[i];
385 int64_t output;
386
387 feclearexcept(FE_ALL_EXCEPT);
388
389 print_half_number(i, input);
390 #if defined(__arm__)
391 /* asm("vcvt.s32.f16 %0, %1" : "=t" (output) : "t" (input)); v8.2*/
392 output = input;
393 #else
394 asm("fcvt %s0, %h1" : "=w" (output) : "w" (input));
395 #endif
396 print_int64(i, output);
397 }
398 }
399
400 typedef struct {
401 int flag;
402 char *desc;
403 } float_mapping;
404
405 float_mapping round_flags[] = {
406 { FE_TONEAREST, "to nearest" },
407 { FE_UPWARD, "upwards" },
408 { FE_DOWNWARD, "downwards" },
409 { FE_TOWARDZERO, "to zero" }
410 };
411
412 int main(int argc, char *argv[argc])
413 {
414 int i;
415
416 printf("#### Enabling IEEE Half Precision\n");
417
418 for (i = 0; i < ARRAY_SIZE(round_flags); ++i) {
419 fesetround(round_flags[i].flag);
420 printf("### Rounding %s\n", round_flags[i].desc);
421 convert_single_to_half();
422 convert_single_to_double();
423 convert_double_to_half();
424 convert_double_to_single();
425 convert_half_to_single();
426 convert_half_to_double();
427 }
428
429 /* convert to integer */
430 convert_single_to_integer();
431 convert_double_to_integer();
432 convert_half_to_integer();
433
434
435 /* And now with ARM alternative FP16 */
436 #if defined(__arm__)
437 asm("vmrs r1, fpscr\n\t"
438 "orr r1, r1, %[flags]\n\t"
439 "vmsr fpscr, r1"
440 : /* no output */ : [flags] "n" (1 << 26) : "r1" );
441 #else
442 asm("mrs x1, fpcr\n\t"
443 "orr x1, x1, %[flags]\n\t"
444 "msr fpcr, x1\n\t"
445 : /* no output */ : [flags] "n" (1 << 26) : "x1" );
446 #endif
447
448 printf("#### Enabling ARM Alternative Half Precision\n");
449
450 for (i = 0; i < ARRAY_SIZE(round_flags); ++i) {
451 fesetround(round_flags[i].flag);
452 printf("### Rounding %s\n", round_flags[i].desc);
453 convert_single_to_half();
454 convert_single_to_double();
455 convert_double_to_half();
456 convert_double_to_single();
457 convert_half_to_single();
458 convert_half_to_double();
459 }
460
461 /* convert to integer */
462 convert_single_to_integer();
463 convert_double_to_integer();
464 convert_half_to_integer();
465
466 return 0;
467 }