8ec8a55e3f
This runs through the usual float to float conversions and crucially also runs with ARM Alternative Half Precision Format. Signed-off-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Tested-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
459 lines
12 KiB
C
459 lines
12 KiB
C
/*
|
|
* Test Floating Point Conversion
|
|
*/
|
|
|
|
/* we want additional float type definitions */
|
|
#define __STDC_WANT_IEC_60559_BFP_EXT__
|
|
#define __STDC_WANT_IEC_60559_TYPES_EXT__
|
|
|
|
#include <stdio.h>
|
|
#include <inttypes.h>
|
|
#include <math.h>
|
|
#include <float.h>
|
|
#include <fenv.h>
|
|
|
|
#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
|
|
|
|
static char flag_str[256];
|
|
|
|
static char *get_flag_state(int flags)
|
|
{
|
|
if (flags) {
|
|
snprintf(flag_str, sizeof(flag_str), "%s %s %s %s %s",
|
|
flags & FE_OVERFLOW ? "OVERFLOW" : "",
|
|
flags & FE_UNDERFLOW ? "UNDERFLOW" : "",
|
|
flags & FE_DIVBYZERO ? "DIV0" : "",
|
|
flags & FE_INEXACT ? "INEXACT" : "",
|
|
flags & FE_INVALID ? "INVALID" : "");
|
|
} else {
|
|
snprintf(flag_str, sizeof(flag_str), "OK");
|
|
}
|
|
|
|
return flag_str;
|
|
}
|
|
|
|
static void print_double_number(int i, double num)
|
|
{
|
|
uint64_t double_as_hex = *(uint64_t *) #
|
|
int flags = fetestexcept(FE_ALL_EXCEPT);
|
|
char *fstr = get_flag_state(flags);
|
|
|
|
printf("%02d DOUBLE: %02.20e / %#020" PRIx64 " (%#x => %s)\n",
|
|
i, num, double_as_hex, flags, fstr);
|
|
}
|
|
|
|
static void print_single_number(int i, float num)
|
|
{
|
|
uint32_t single_as_hex = *(uint32_t *) #
|
|
int flags = fetestexcept(FE_ALL_EXCEPT);
|
|
char *fstr = get_flag_state(flags);
|
|
|
|
printf("%02d SINGLE: %02.20e / %#010x (%#x => %s)\n",
|
|
i, num, single_as_hex, flags, fstr);
|
|
}
|
|
|
|
static void print_half_number(int i, uint16_t num)
|
|
{
|
|
int flags = fetestexcept(FE_ALL_EXCEPT);
|
|
char *fstr = get_flag_state(flags);
|
|
|
|
printf("%02d HALF: %#04x (%#x => %s)\n",
|
|
i, num, flags, fstr);
|
|
}
|
|
|
|
static void print_int64(int i, int64_t num)
|
|
{
|
|
uint64_t int64_as_hex = *(uint64_t *) #
|
|
int flags = fetestexcept(FE_ALL_EXCEPT);
|
|
char *fstr = get_flag_state(flags);
|
|
|
|
printf("%02d INT64: %20" PRId64 "/%#020" PRIx64 " (%#x => %s)\n",
|
|
i, num, int64_as_hex, flags, fstr);
|
|
}
|
|
|
|
#ifndef SNANF
|
|
/* Signaling NaN macros, if supported. */
|
|
# if __GNUC_PREREQ(3, 3)
|
|
# define SNANF (__builtin_nansf (""))
|
|
# define SNAN (__builtin_nans (""))
|
|
# define SNANL (__builtin_nansl (""))
|
|
# endif
|
|
#endif
|
|
|
|
float single_numbers[] = { -SNANF,
|
|
-NAN,
|
|
-INFINITY,
|
|
-FLT_MAX,
|
|
-1.111E+31,
|
|
-1.111E+30,
|
|
-1.08700982e-12,
|
|
-1.78051176e-20,
|
|
-FLT_MIN,
|
|
0.0,
|
|
FLT_MIN,
|
|
2.98023224e-08,
|
|
5.96046E-8, /* min positive FP16 subnormal */
|
|
6.09756E-5, /* max subnormal FP16 */
|
|
6.10352E-5, /* min positive normal FP16 */
|
|
1.0,
|
|
1.0009765625, /* smallest float after 1.0 FP16 */
|
|
2.0,
|
|
M_E, M_PI,
|
|
65503.0,
|
|
65504.0, /* max FP16 */
|
|
65505.0,
|
|
131007.0,
|
|
131008.0, /* max AFP */
|
|
131009.0,
|
|
1.111E+30,
|
|
FLT_MAX,
|
|
INFINITY,
|
|
NAN,
|
|
SNANF };
|
|
|
|
static void convert_single_to_half(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting single-precision to half-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
|
|
float input = single_numbers[i];
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_single_number(i, input);
|
|
#if defined(__arm__)
|
|
uint32_t output;
|
|
asm("vcvtb.f16.f32 %0, %1" : "=t" (output) : "x" (input));
|
|
#else
|
|
uint16_t output;
|
|
asm("fcvt %h0, %s1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_half_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_single_to_double(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting single-precision to double-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
|
|
float input = single_numbers[i];
|
|
/* uint64_t output; */
|
|
double output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_single_number(i, input);
|
|
#if defined(__arm__)
|
|
asm("vcvt.f64.f32 %P0, %1" : "=w" (output) : "t" (input));
|
|
#else
|
|
asm("fcvt %d0, %s1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_double_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_single_to_integer(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting single-precision to integer\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(single_numbers); ++i) {
|
|
float input = single_numbers[i];
|
|
int64_t output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_single_number(i, input);
|
|
#if defined(__arm__)
|
|
/* asm("vcvt.s32.f32 %s0, %s1" : "=t" (output) : "t" (input)); */
|
|
output = input;
|
|
#else
|
|
asm("fcvtzs %0, %s1" : "=r" (output) : "w" (input));
|
|
#endif
|
|
print_int64(i, output);
|
|
}
|
|
}
|
|
|
|
/* This allows us to initialise some doubles as pure hex */
|
|
typedef union {
|
|
double d;
|
|
uint64_t h;
|
|
} test_doubles;
|
|
|
|
test_doubles double_numbers[] = {
|
|
{SNAN},
|
|
{-NAN},
|
|
{-INFINITY},
|
|
{-DBL_MAX},
|
|
{-FLT_MAX-1.0},
|
|
{-FLT_MAX},
|
|
{-1.111E+31},
|
|
{-1.111E+30}, /* half prec */
|
|
{-2.0}, {-1.0},
|
|
{-DBL_MIN},
|
|
{-FLT_MIN},
|
|
{0.0},
|
|
{FLT_MIN},
|
|
{2.98023224e-08},
|
|
{5.96046E-8}, /* min positive FP16 subnormal */
|
|
{6.09756E-5}, /* max subnormal FP16 */
|
|
{6.10352E-5}, /* min positive normal FP16 */
|
|
{1.0},
|
|
{1.0009765625}, /* smallest float after 1.0 FP16 */
|
|
{DBL_MIN},
|
|
{1.3789972848607228e-308},
|
|
{1.4914738736681624e-308},
|
|
{1.0}, {2.0},
|
|
{M_E}, {M_PI},
|
|
{65503.0},
|
|
{65504.0}, /* max FP16 */
|
|
{65505.0},
|
|
{131007.0},
|
|
{131008.0}, /* max AFP */
|
|
{131009.0},
|
|
{.h = 0x41dfffffffc00000 }, /* to int = 0x7fffffff */
|
|
{FLT_MAX},
|
|
{FLT_MAX + 1.0},
|
|
{DBL_MAX},
|
|
{INFINITY},
|
|
{NAN},
|
|
{.h = 0x7ff0000000000001}, /* SNAN */
|
|
{SNAN},
|
|
};
|
|
|
|
static void convert_double_to_half(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting double-precision to half-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
|
|
double input = double_numbers[i].d;
|
|
uint16_t output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_double_number(i, input);
|
|
|
|
/* as we don't have _Float16 support */
|
|
#if defined(__arm__)
|
|
/* asm("vcvtb.f16.f64 %0, %P1" : "=t" (output) : "x" (input)); */
|
|
output = input;
|
|
#else
|
|
asm("fcvt %h0, %d1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_half_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_double_to_single(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting double-precision to single-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
|
|
double input = double_numbers[i].d;
|
|
uint32_t output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_double_number(i, input);
|
|
|
|
#if defined(__arm__)
|
|
asm("vcvt.f32.f64 %0, %P1" : "=w" (output) : "x" (input));
|
|
#else
|
|
asm("fcvt %s0, %d1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
|
|
print_single_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_double_to_integer(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting double-precision to integer\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(double_numbers); ++i) {
|
|
double input = double_numbers[i].d;
|
|
int64_t output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_double_number(i, input);
|
|
#if defined(__arm__)
|
|
/* asm("vcvt.s32.f32 %s0, %s1" : "=t" (output) : "t" (input)); */
|
|
output = input;
|
|
#else
|
|
asm("fcvtzs %0, %d1" : "=r" (output) : "w" (input));
|
|
#endif
|
|
print_int64(i, output);
|
|
}
|
|
}
|
|
|
|
/* no handy defines for these numbers */
|
|
uint16_t half_numbers[] = {
|
|
0xffff, /* -NaN / AHP -Max */
|
|
0xfcff, /* -NaN / AHP */
|
|
0xfc01, /* -NaN / AHP */
|
|
0xfc00, /* -Inf */
|
|
0xfbff, /* -Max */
|
|
0xc000, /* -2 */
|
|
0xbc00, /* -1 */
|
|
0x8001, /* -MIN subnormal */
|
|
0x8000, /* -0 */
|
|
0x0000, /* +0 */
|
|
0x0001, /* MIN subnormal */
|
|
0x3c00, /* 1 */
|
|
0x7bff, /* Max */
|
|
0x7c00, /* Inf */
|
|
0x7c01, /* NaN / AHP */
|
|
0x7cff, /* NaN / AHP */
|
|
0x7fff, /* NaN / AHP +Max*/
|
|
};
|
|
|
|
static void convert_half_to_double(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting half-precision to double-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
|
|
uint16_t input = half_numbers[i];
|
|
double output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_half_number(i, input);
|
|
#if defined(__arm__)
|
|
/* asm("vcvtb.f64.f16 %P0, %1" : "=w" (output) : "t" (input)); */
|
|
output = input;
|
|
#else
|
|
asm("fcvt %d0, %h1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_double_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_half_to_single(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting half-precision to single-precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
|
|
uint16_t input = half_numbers[i];
|
|
float output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_half_number(i, input);
|
|
#if defined(__arm__)
|
|
asm("vcvtb.f32.f16 %0, %1" : "=w" (output) : "x" ((uint32_t)input));
|
|
#else
|
|
asm("fcvt %s0, %h1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_single_number(i, output);
|
|
}
|
|
}
|
|
|
|
static void convert_half_to_integer(void)
|
|
{
|
|
int i;
|
|
|
|
printf("Converting half-precision to integer\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(half_numbers); ++i) {
|
|
uint16_t input = half_numbers[i];
|
|
int64_t output;
|
|
|
|
feclearexcept(FE_ALL_EXCEPT);
|
|
|
|
print_half_number(i, input);
|
|
#if defined(__arm__)
|
|
/* asm("vcvt.s32.f16 %0, %1" : "=t" (output) : "t" (input)); v8.2*/
|
|
output = input;
|
|
#else
|
|
asm("fcvt %s0, %h1" : "=w" (output) : "x" (input));
|
|
#endif
|
|
print_int64(i, output);
|
|
}
|
|
}
|
|
|
|
typedef struct {
|
|
int flag;
|
|
char *desc;
|
|
} float_mapping;
|
|
|
|
float_mapping round_flags[] = {
|
|
{ FE_TONEAREST, "to nearest" },
|
|
{ FE_UPWARD, "upwards" },
|
|
{ FE_DOWNWARD, "downwards" },
|
|
{ FE_TOWARDZERO, "to zero" }
|
|
};
|
|
|
|
int main(int argc, char *argv[argc])
|
|
{
|
|
int i;
|
|
|
|
printf("#### Enabling IEEE Half Precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(round_flags); ++i) {
|
|
fesetround(round_flags[i].flag);
|
|
printf("### Rounding %s\n", round_flags[i].desc);
|
|
convert_single_to_half();
|
|
convert_single_to_double();
|
|
convert_double_to_half();
|
|
convert_double_to_single();
|
|
convert_half_to_single();
|
|
convert_half_to_double();
|
|
}
|
|
|
|
/* convert to integer */
|
|
convert_single_to_integer();
|
|
convert_double_to_integer();
|
|
convert_half_to_integer();
|
|
|
|
/* And now with ARM alternative FP16 */
|
|
#if defined(__arm__)
|
|
/* See glibc sysdeps/arm/fpu_control.h */
|
|
asm("mrc p10, 7, r1, cr1, cr0, 0\n\t"
|
|
"orr r1, r1, %[flags]\n\t"
|
|
"mcr p10, 7, r1, cr1, cr0, 0\n\t"
|
|
: /* no output */ : [flags] "n" (1 << 26) : "r1" );
|
|
#else
|
|
asm("mrs x1, fpcr\n\t"
|
|
"orr x1, x1, %[flags]\n\t"
|
|
"msr fpcr, x1\n\t"
|
|
: /* no output */ : [flags] "n" (1 << 26) : "x1" );
|
|
#endif
|
|
|
|
printf("#### Enabling ARM Alternative Half Precision\n");
|
|
|
|
for (i = 0; i < ARRAY_SIZE(round_flags); ++i) {
|
|
fesetround(round_flags[i].flag);
|
|
printf("### Rounding %s\n", round_flags[i].desc);
|
|
convert_single_to_half();
|
|
convert_single_to_double();
|
|
convert_double_to_half();
|
|
convert_double_to_single();
|
|
convert_half_to_single();
|
|
convert_half_to_double();
|
|
}
|
|
|
|
/* convert to integer */
|
|
convert_single_to_integer();
|
|
convert_double_to_integer();
|
|
convert_half_to_integer();
|
|
|
|
return 0;
|
|
}
|