mirror of
https://gitlab.freedesktop.org/mesa/mesa.git
synced 2026-06-21 08:08:24 +02:00
util: add float8 conversion functions
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35434>
This commit is contained in:
Georg Lehmann
Marge Bot
parent
9da23499ff
commit
2237c022a2
4 changed files with 830 additions and 0 deletions
192
src/util/float8.c
Normal file
192
src/util/float8.c
Normal file
|
|
@ -0,0 +1,192 @@
|
|||
/*
|
||||
* Copyright 2025 Valve Corporation
|
||||
* SPDX-License-Identifier: MIT
|
||||
*/
|
||||
|
||||
#include "float8.h"
|
||||
|
||||
#include <assert.h>
|
||||
#include <math.h>
|
||||
#include "bitscan.h"
|
||||
#include "u_math.h"
|
||||
|
||||
#define E4M3FN_NAN 0xff
|
||||
#define E4M3FN_MAX 0x7e
|
||||
|
||||
#define E5M2_NAN 0xfe
|
||||
#define E5M2_MAX 0x7b
|
||||
#define E5M2_INF 0x7c
|
||||
|
||||
|
||||
uint8_t
|
||||
_mesa_float_to_e4m3fn(float val)
|
||||
{
|
||||
/* This is a finite only format, out of range values (after rounding)
|
||||
* are converted to NaN.
|
||||
*/
|
||||
if (fabs(val) > 464.0f || isnan(val))
|
||||
return E4M3FN_NAN;
|
||||
|
||||
bool s = fui(val) & 0x80000000;
|
||||
int e = ((fui(val) >> 23) & 0xff) - 127 + 7;
|
||||
uint32_t m = fui(val) & 0x7fffff;
|
||||
|
||||
uint8_t res = s ? 0x80 : 0;
|
||||
|
||||
/* Zero, underflow. */
|
||||
if (e < -3)
|
||||
return res;
|
||||
|
||||
bool is_denorm = e <= 0;
|
||||
|
||||
bool round_up = false;
|
||||
|
||||
if (is_denorm) {
|
||||
unsigned offset = 1 - e;
|
||||
round_up |= m & ((1 << offset) - 1);
|
||||
m = (m | 0x800000) >> offset;
|
||||
}
|
||||
|
||||
round_up |= m & 0x17ffff;
|
||||
|
||||
if ((m & 0x080000) && round_up) {
|
||||
m += 0x100000;
|
||||
|
||||
if (m & 0x800000) {
|
||||
m = 0;
|
||||
e += 1;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
if (!is_denorm)
|
||||
res |= (e << 3);
|
||||
res |= (m >> 20);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
uint8_t
|
||||
_mesa_float_to_e4m3fn_sat(float val)
|
||||
{
|
||||
if (val > 448.0f)
|
||||
return E4M3FN_MAX;
|
||||
else if (val < -448.0f)
|
||||
return 0x80 | E4M3FN_MAX;
|
||||
else
|
||||
return _mesa_float_to_e4m3fn(val);
|
||||
}
|
||||
|
||||
float
|
||||
_mesa_e4m3fn_to_float(uint8_t val)
|
||||
{
|
||||
bool s = val & 0x80;
|
||||
uint32_t e = (val >> 3) & 0xf;
|
||||
uint32_t m = val & 0x7;
|
||||
|
||||
if (e == 0xf && m == 0x7)
|
||||
return uif(0xffc00000);
|
||||
|
||||
uint32_t res = s ? 0x80000000 : 0;
|
||||
|
||||
if (e == 0 && m == 0) {
|
||||
/* Zero. */
|
||||
} else if (e == 0) {
|
||||
/* Denorm. */
|
||||
unsigned shift = (4 - util_last_bit(m));
|
||||
res |= (127 - 6 - shift) << 23;
|
||||
res |= ((m << shift) & 0x7) << (23 - 3);
|
||||
} else {
|
||||
res |= (e + (127 - 7)) << 23;
|
||||
res |= m << (23 - 3);
|
||||
}
|
||||
|
||||
return uif(res);
|
||||
}
|
||||
|
||||
uint8_t
|
||||
_mesa_float_to_e5m2(float val)
|
||||
{
|
||||
bool s = fui(val) & 0x80000000;
|
||||
uint8_t res = s ? 0x80 : 0;
|
||||
if (isnan(val))
|
||||
return E5M2_NAN;
|
||||
else if (fabs(val) >= 61440.0f)
|
||||
return res | E5M2_INF;
|
||||
|
||||
int e = ((fui(val) >> 23) & 0xff) - 127 + 15;
|
||||
uint32_t m = fui(val) & 0x7fffff;
|
||||
|
||||
/* Zero, underflow. */
|
||||
if (e < -2)
|
||||
return res;
|
||||
|
||||
bool is_denorm = e <= 0;
|
||||
|
||||
bool round_up = false;
|
||||
|
||||
if (is_denorm) {
|
||||
unsigned offset = 1 - e;
|
||||
round_up |= m & ((1 << offset) - 1);
|
||||
m = (m | 0x800000) >> offset;
|
||||
}
|
||||
|
||||
round_up |= m & 0x2fffff;
|
||||
|
||||
if ((m & 0x100000) && round_up) {
|
||||
m += 0x200000;
|
||||
|
||||
if (m & 0x800000) {
|
||||
m = 0;
|
||||
e += 1;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
if (!is_denorm)
|
||||
res |= (e << 2);
|
||||
res |= (m >> 21);
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
uint8_t
|
||||
_mesa_float_to_e5m2_sat(float val)
|
||||
{
|
||||
if (val > 57344.0f)
|
||||
return E5M2_MAX;
|
||||
else if (val < -57344.0f)
|
||||
return 0x80 | E5M2_MAX;
|
||||
else
|
||||
return _mesa_float_to_e5m2(val);
|
||||
}
|
||||
|
||||
float
|
||||
_mesa_e5m2_to_float(uint8_t val)
|
||||
{
|
||||
bool s = val & 0x80;
|
||||
uint32_t e = (val >> 2) & 0x1f;
|
||||
uint32_t m = val & 0x3;
|
||||
|
||||
if (e == 0x1f && m != 0)
|
||||
return uif(0xffc00000);
|
||||
|
||||
uint32_t res = s ? 0x80000000 : 0;
|
||||
|
||||
if (e == 0x1f) {
|
||||
/* Infinity. */
|
||||
res |= 0x7f800000;
|
||||
} else if (e == 0 && m == 0) {
|
||||
/* Zero. */
|
||||
} else if (e == 0) {
|
||||
/* Denorm. */
|
||||
unsigned shift = (3 - util_last_bit(m));
|
||||
res |= (127 - 14 - shift) << 23;
|
||||
res |= ((m << shift) & 0x3) << (23 - 2);
|
||||
} else {
|
||||
res |= (e + (127 - 15)) << 23;
|
||||
res |= m << (23 - 2);
|
||||
}
|
||||
|
||||
return uif(res);
|
||||
}
|
||||
24
src/util/float8.h
Normal file
24
src/util/float8.h
Normal file
|
|
@ -0,0 +1,24 @@
|
|||
/*
|
||||
* Copyright 2025 Valve Corporation
|
||||
* SPDX-License-Identifier: MIT
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <stdint.h>
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
uint8_t _mesa_float_to_e4m3fn(float val);
|
||||
uint8_t _mesa_float_to_e4m3fn_sat(float val);
|
||||
float _mesa_e4m3fn_to_float(uint8_t val);
|
||||
|
||||
uint8_t _mesa_float_to_e5m2(float val);
|
||||
uint8_t _mesa_float_to_e5m2_sat(float val);
|
||||
float _mesa_e5m2_to_float(uint8_t val);
|
||||
|
||||
#ifdef __cplusplus
|
||||
} /* extern C */
|
||||
#endif
|
||||
|
|
@ -44,6 +44,7 @@ files_mesa_util = files(
|
|||
'enum_operators.h',
|
||||
'fast_idiv_by_const.c',
|
||||
'fast_idiv_by_const.h',
|
||||
'float8.c',
|
||||
'format_r11g11b10f.h',
|
||||
'format_rgb9e5.h',
|
||||
'format_srgb.h',
|
||||
|
|
@ -391,6 +392,7 @@ if with_tests
|
|||
'tests/dag_test.cpp',
|
||||
'tests/fast_idiv_by_const_test.cpp',
|
||||
'tests/fast_urem_by_const_test.cpp',
|
||||
'tests/float8_test.cpp',
|
||||
'tests/gc_alloc_tests.cpp',
|
||||
'tests/half_float_test.cpp',
|
||||
'tests/int_min_max.cpp',
|
||||
|
|
|
|||
612
src/util/tests/float8_test.cpp
Normal file
612
src/util/tests/float8_test.cpp
Normal file
|
|
@ -0,0 +1,612 @@
|
|||
/*
|
||||
* Copyright 2025 Valve Corporation
|
||||
* SPDX-License-Identifier: MIT
|
||||
*/
|
||||
|
||||
|
||||
#include <math.h>
|
||||
#include <gtest/gtest.h>
|
||||
|
||||
#include "util/float8.h"
|
||||
#include "u_math.h"
|
||||
|
||||
/* Navi48 output. */
|
||||
static const uint32_t e4m3fn_lut[256] = {
|
||||
0x0, /* 0x0 */
|
||||
0x3b000000, /* 0x1 */
|
||||
0x3b800000, /* 0x2 */
|
||||
0x3bc00000, /* 0x3 */
|
||||
0x3c000000, /* 0x4 */
|
||||
0x3c200000, /* 0x5 */
|
||||
0x3c400000, /* 0x6 */
|
||||
0x3c600000, /* 0x7 */
|
||||
0x3c800000, /* 0x8 */
|
||||
0x3c900000, /* 0x9 */
|
||||
0x3ca00000, /* 0xa */
|
||||
0x3cb00000, /* 0xb */
|
||||
0x3cc00000, /* 0xc */
|
||||
0x3cd00000, /* 0xd */
|
||||
0x3ce00000, /* 0xe */
|
||||
0x3cf00000, /* 0xf */
|
||||
0x3d000000, /* 0x10 */
|
||||
0x3d100000, /* 0x11 */
|
||||
0x3d200000, /* 0x12 */
|
||||
0x3d300000, /* 0x13 */
|
||||
0x3d400000, /* 0x14 */
|
||||
0x3d500000, /* 0x15 */
|
||||
0x3d600000, /* 0x16 */
|
||||
0x3d700000, /* 0x17 */
|
||||
0x3d800000, /* 0x18 */
|
||||
0x3d900000, /* 0x19 */
|
||||
0x3da00000, /* 0x1a */
|
||||
0x3db00000, /* 0x1b */
|
||||
0x3dc00000, /* 0x1c */
|
||||
0x3dd00000, /* 0x1d */
|
||||
0x3de00000, /* 0x1e */
|
||||
0x3df00000, /* 0x1f */
|
||||
0x3e000000, /* 0x20 */
|
||||
0x3e100000, /* 0x21 */
|
||||
0x3e200000, /* 0x22 */
|
||||
0x3e300000, /* 0x23 */
|
||||
0x3e400000, /* 0x24 */
|
||||
0x3e500000, /* 0x25 */
|
||||
0x3e600000, /* 0x26 */
|
||||
0x3e700000, /* 0x27 */
|
||||
0x3e800000, /* 0x28 */
|
||||
0x3e900000, /* 0x29 */
|
||||
0x3ea00000, /* 0x2a */
|
||||
0x3eb00000, /* 0x2b */
|
||||
0x3ec00000, /* 0x2c */
|
||||
0x3ed00000, /* 0x2d */
|
||||
0x3ee00000, /* 0x2e */
|
||||
0x3ef00000, /* 0x2f */
|
||||
0x3f000000, /* 0x30 */
|
||||
0x3f100000, /* 0x31 */
|
||||
0x3f200000, /* 0x32 */
|
||||
0x3f300000, /* 0x33 */
|
||||
0x3f400000, /* 0x34 */
|
||||
0x3f500000, /* 0x35 */
|
||||
0x3f600000, /* 0x36 */
|
||||
0x3f700000, /* 0x37 */
|
||||
0x3f800000, /* 0x38 */
|
||||
0x3f900000, /* 0x39 */
|
||||
0x3fa00000, /* 0x3a */
|
||||
0x3fb00000, /* 0x3b */
|
||||
0x3fc00000, /* 0x3c */
|
||||
0x3fd00000, /* 0x3d */
|
||||
0x3fe00000, /* 0x3e */
|
||||
0x3ff00000, /* 0x3f */
|
||||
0x40000000, /* 0x40 */
|
||||
0x40100000, /* 0x41 */
|
||||
0x40200000, /* 0x42 */
|
||||
0x40300000, /* 0x43 */
|
||||
0x40400000, /* 0x44 */
|
||||
0x40500000, /* 0x45 */
|
||||
0x40600000, /* 0x46 */
|
||||
0x40700000, /* 0x47 */
|
||||
0x40800000, /* 0x48 */
|
||||
0x40900000, /* 0x49 */
|
||||
0x40a00000, /* 0x4a */
|
||||
0x40b00000, /* 0x4b */
|
||||
0x40c00000, /* 0x4c */
|
||||
0x40d00000, /* 0x4d */
|
||||
0x40e00000, /* 0x4e */
|
||||
0x40f00000, /* 0x4f */
|
||||
0x41000000, /* 0x50 */
|
||||
0x41100000, /* 0x51 */
|
||||
0x41200000, /* 0x52 */
|
||||
0x41300000, /* 0x53 */
|
||||
0x41400000, /* 0x54 */
|
||||
0x41500000, /* 0x55 */
|
||||
0x41600000, /* 0x56 */
|
||||
0x41700000, /* 0x57 */
|
||||
0x41800000, /* 0x58 */
|
||||
0x41900000, /* 0x59 */
|
||||
0x41a00000, /* 0x5a */
|
||||
0x41b00000, /* 0x5b */
|
||||
0x41c00000, /* 0x5c */
|
||||
0x41d00000, /* 0x5d */
|
||||
0x41e00000, /* 0x5e */
|
||||
0x41f00000, /* 0x5f */
|
||||
0x42000000, /* 0x60 */
|
||||
0x42100000, /* 0x61 */
|
||||
0x42200000, /* 0x62 */
|
||||
0x42300000, /* 0x63 */
|
||||
0x42400000, /* 0x64 */
|
||||
0x42500000, /* 0x65 */
|
||||
0x42600000, /* 0x66 */
|
||||
0x42700000, /* 0x67 */
|
||||
0x42800000, /* 0x68 */
|
||||
0x42900000, /* 0x69 */
|
||||
0x42a00000, /* 0x6a */
|
||||
0x42b00000, /* 0x6b */
|
||||
0x42c00000, /* 0x6c */
|
||||
0x42d00000, /* 0x6d */
|
||||
0x42e00000, /* 0x6e */
|
||||
0x42f00000, /* 0x6f */
|
||||
0x43000000, /* 0x70 */
|
||||
0x43100000, /* 0x71 */
|
||||
0x43200000, /* 0x72 */
|
||||
0x43300000, /* 0x73 */
|
||||
0x43400000, /* 0x74 */
|
||||
0x43500000, /* 0x75 */
|
||||
0x43600000, /* 0x76 */
|
||||
0x43700000, /* 0x77 */
|
||||
0x43800000, /* 0x78 */
|
||||
0x43900000, /* 0x79 */
|
||||
0x43a00000, /* 0x7a */
|
||||
0x43b00000, /* 0x7b */
|
||||
0x43c00000, /* 0x7c */
|
||||
0x43d00000, /* 0x7d */
|
||||
0x43e00000, /* 0x7e */
|
||||
0xffc00000, /* 0x7f */
|
||||
0x80000000, /* 0x80 */
|
||||
0xbb000000, /* 0x81 */
|
||||
0xbb800000, /* 0x82 */
|
||||
0xbbc00000, /* 0x83 */
|
||||
0xbc000000, /* 0x84 */
|
||||
0xbc200000, /* 0x85 */
|
||||
0xbc400000, /* 0x86 */
|
||||
0xbc600000, /* 0x87 */
|
||||
0xbc800000, /* 0x88 */
|
||||
0xbc900000, /* 0x89 */
|
||||
0xbca00000, /* 0x8a */
|
||||
0xbcb00000, /* 0x8b */
|
||||
0xbcc00000, /* 0x8c */
|
||||
0xbcd00000, /* 0x8d */
|
||||
0xbce00000, /* 0x8e */
|
||||
0xbcf00000, /* 0x8f */
|
||||
0xbd000000, /* 0x90 */
|
||||
0xbd100000, /* 0x91 */
|
||||
0xbd200000, /* 0x92 */
|
||||
0xbd300000, /* 0x93 */
|
||||
0xbd400000, /* 0x94 */
|
||||
0xbd500000, /* 0x95 */
|
||||
0xbd600000, /* 0x96 */
|
||||
0xbd700000, /* 0x97 */
|
||||
0xbd800000, /* 0x98 */
|
||||
0xbd900000, /* 0x99 */
|
||||
0xbda00000, /* 0x9a */
|
||||
0xbdb00000, /* 0x9b */
|
||||
0xbdc00000, /* 0x9c */
|
||||
0xbdd00000, /* 0x9d */
|
||||
0xbde00000, /* 0x9e */
|
||||
0xbdf00000, /* 0x9f */
|
||||
0xbe000000, /* 0xa0 */
|
||||
0xbe100000, /* 0xa1 */
|
||||
0xbe200000, /* 0xa2 */
|
||||
0xbe300000, /* 0xa3 */
|
||||
0xbe400000, /* 0xa4 */
|
||||
0xbe500000, /* 0xa5 */
|
||||
0xbe600000, /* 0xa6 */
|
||||
0xbe700000, /* 0xa7 */
|
||||
0xbe800000, /* 0xa8 */
|
||||
0xbe900000, /* 0xa9 */
|
||||
0xbea00000, /* 0xaa */
|
||||
0xbeb00000, /* 0xab */
|
||||
0xbec00000, /* 0xac */
|
||||
0xbed00000, /* 0xad */
|
||||
0xbee00000, /* 0xae */
|
||||
0xbef00000, /* 0xaf */
|
||||
0xbf000000, /* 0xb0 */
|
||||
0xbf100000, /* 0xb1 */
|
||||
0xbf200000, /* 0xb2 */
|
||||
0xbf300000, /* 0xb3 */
|
||||
0xbf400000, /* 0xb4 */
|
||||
0xbf500000, /* 0xb5 */
|
||||
0xbf600000, /* 0xb6 */
|
||||
0xbf700000, /* 0xb7 */
|
||||
0xbf800000, /* 0xb8 */
|
||||
0xbf900000, /* 0xb9 */
|
||||
0xbfa00000, /* 0xba */
|
||||
0xbfb00000, /* 0xbb */
|
||||
0xbfc00000, /* 0xbc */
|
||||
0xbfd00000, /* 0xbd */
|
||||
0xbfe00000, /* 0xbe */
|
||||
0xbff00000, /* 0xbf */
|
||||
0xc0000000, /* 0xc0 */
|
||||
0xc0100000, /* 0xc1 */
|
||||
0xc0200000, /* 0xc2 */
|
||||
0xc0300000, /* 0xc3 */
|
||||
0xc0400000, /* 0xc4 */
|
||||
0xc0500000, /* 0xc5 */
|
||||
0xc0600000, /* 0xc6 */
|
||||
0xc0700000, /* 0xc7 */
|
||||
0xc0800000, /* 0xc8 */
|
||||
0xc0900000, /* 0xc9 */
|
||||
0xc0a00000, /* 0xca */
|
||||
0xc0b00000, /* 0xcb */
|
||||
0xc0c00000, /* 0xcc */
|
||||
0xc0d00000, /* 0xcd */
|
||||
0xc0e00000, /* 0xce */
|
||||
0xc0f00000, /* 0xcf */
|
||||
0xc1000000, /* 0xd0 */
|
||||
0xc1100000, /* 0xd1 */
|
||||
0xc1200000, /* 0xd2 */
|
||||
0xc1300000, /* 0xd3 */
|
||||
0xc1400000, /* 0xd4 */
|
||||
0xc1500000, /* 0xd5 */
|
||||
0xc1600000, /* 0xd6 */
|
||||
0xc1700000, /* 0xd7 */
|
||||
0xc1800000, /* 0xd8 */
|
||||
0xc1900000, /* 0xd9 */
|
||||
0xc1a00000, /* 0xda */
|
||||
0xc1b00000, /* 0xdb */
|
||||
0xc1c00000, /* 0xdc */
|
||||
0xc1d00000, /* 0xdd */
|
||||
0xc1e00000, /* 0xde */
|
||||
0xc1f00000, /* 0xdf */
|
||||
0xc2000000, /* 0xe0 */
|
||||
0xc2100000, /* 0xe1 */
|
||||
0xc2200000, /* 0xe2 */
|
||||
0xc2300000, /* 0xe3 */
|
||||
0xc2400000, /* 0xe4 */
|
||||
0xc2500000, /* 0xe5 */
|
||||
0xc2600000, /* 0xe6 */
|
||||
0xc2700000, /* 0xe7 */
|
||||
0xc2800000, /* 0xe8 */
|
||||
0xc2900000, /* 0xe9 */
|
||||
0xc2a00000, /* 0xea */
|
||||
0xc2b00000, /* 0xeb */
|
||||
0xc2c00000, /* 0xec */
|
||||
0xc2d00000, /* 0xed */
|
||||
0xc2e00000, /* 0xee */
|
||||
0xc2f00000, /* 0xef */
|
||||
0xc3000000, /* 0xf0 */
|
||||
0xc3100000, /* 0xf1 */
|
||||
0xc3200000, /* 0xf2 */
|
||||
0xc3300000, /* 0xf3 */
|
||||
0xc3400000, /* 0xf4 */
|
||||
0xc3500000, /* 0xf5 */
|
||||
0xc3600000, /* 0xf6 */
|
||||
0xc3700000, /* 0xf7 */
|
||||
0xc3800000, /* 0xf8 */
|
||||
0xc3900000, /* 0xf9 */
|
||||
0xc3a00000, /* 0xfa */
|
||||
0xc3b00000, /* 0xfb */
|
||||
0xc3c00000, /* 0xfc */
|
||||
0xc3d00000, /* 0xfd */
|
||||
0xc3e00000, /* 0xfe */
|
||||
0xffc00000, /* 0xff */
|
||||
};
|
||||
|
||||
static const uint32_t e5m2_lut[256] = {
|
||||
0x0, /* 0x0 */
|
||||
0x37800000, /* 0x1 */
|
||||
0x38000000, /* 0x2 */
|
||||
0x38400000, /* 0x3 */
|
||||
0x38800000, /* 0x4 */
|
||||
0x38a00000, /* 0x5 */
|
||||
0x38c00000, /* 0x6 */
|
||||
0x38e00000, /* 0x7 */
|
||||
0x39000000, /* 0x8 */
|
||||
0x39200000, /* 0x9 */
|
||||
0x39400000, /* 0xa */
|
||||
0x39600000, /* 0xb */
|
||||
0x39800000, /* 0xc */
|
||||
0x39a00000, /* 0xd */
|
||||
0x39c00000, /* 0xe */
|
||||
0x39e00000, /* 0xf */
|
||||
0x3a000000, /* 0x10 */
|
||||
0x3a200000, /* 0x11 */
|
||||
0x3a400000, /* 0x12 */
|
||||
0x3a600000, /* 0x13 */
|
||||
0x3a800000, /* 0x14 */
|
||||
0x3aa00000, /* 0x15 */
|
||||
0x3ac00000, /* 0x16 */
|
||||
0x3ae00000, /* 0x17 */
|
||||
0x3b000000, /* 0x18 */
|
||||
0x3b200000, /* 0x19 */
|
||||
0x3b400000, /* 0x1a */
|
||||
0x3b600000, /* 0x1b */
|
||||
0x3b800000, /* 0x1c */
|
||||
0x3ba00000, /* 0x1d */
|
||||
0x3bc00000, /* 0x1e */
|
||||
0x3be00000, /* 0x1f */
|
||||
0x3c000000, /* 0x20 */
|
||||
0x3c200000, /* 0x21 */
|
||||
0x3c400000, /* 0x22 */
|
||||
0x3c600000, /* 0x23 */
|
||||
0x3c800000, /* 0x24 */
|
||||
0x3ca00000, /* 0x25 */
|
||||
0x3cc00000, /* 0x26 */
|
||||
0x3ce00000, /* 0x27 */
|
||||
0x3d000000, /* 0x28 */
|
||||
0x3d200000, /* 0x29 */
|
||||
0x3d400000, /* 0x2a */
|
||||
0x3d600000, /* 0x2b */
|
||||
0x3d800000, /* 0x2c */
|
||||
0x3da00000, /* 0x2d */
|
||||
0x3dc00000, /* 0x2e */
|
||||
0x3de00000, /* 0x2f */
|
||||
0x3e000000, /* 0x30 */
|
||||
0x3e200000, /* 0x31 */
|
||||
0x3e400000, /* 0x32 */
|
||||
0x3e600000, /* 0x33 */
|
||||
0x3e800000, /* 0x34 */
|
||||
0x3ea00000, /* 0x35 */
|
||||
0x3ec00000, /* 0x36 */
|
||||
0x3ee00000, /* 0x37 */
|
||||
0x3f000000, /* 0x38 */
|
||||
0x3f200000, /* 0x39 */
|
||||
0x3f400000, /* 0x3a */
|
||||
0x3f600000, /* 0x3b */
|
||||
0x3f800000, /* 0x3c */
|
||||
0x3fa00000, /* 0x3d */
|
||||
0x3fc00000, /* 0x3e */
|
||||
0x3fe00000, /* 0x3f */
|
||||
0x40000000, /* 0x40 */
|
||||
0x40200000, /* 0x41 */
|
||||
0x40400000, /* 0x42 */
|
||||
0x40600000, /* 0x43 */
|
||||
0x40800000, /* 0x44 */
|
||||
0x40a00000, /* 0x45 */
|
||||
0x40c00000, /* 0x46 */
|
||||
0x40e00000, /* 0x47 */
|
||||
0x41000000, /* 0x48 */
|
||||
0x41200000, /* 0x49 */
|
||||
0x41400000, /* 0x4a */
|
||||
0x41600000, /* 0x4b */
|
||||
0x41800000, /* 0x4c */
|
||||
0x41a00000, /* 0x4d */
|
||||
0x41c00000, /* 0x4e */
|
||||
0x41e00000, /* 0x4f */
|
||||
0x42000000, /* 0x50 */
|
||||
0x42200000, /* 0x51 */
|
||||
0x42400000, /* 0x52 */
|
||||
0x42600000, /* 0x53 */
|
||||
0x42800000, /* 0x54 */
|
||||
0x42a00000, /* 0x55 */
|
||||
0x42c00000, /* 0x56 */
|
||||
0x42e00000, /* 0x57 */
|
||||
0x43000000, /* 0x58 */
|
||||
0x43200000, /* 0x59 */
|
||||
0x43400000, /* 0x5a */
|
||||
0x43600000, /* 0x5b */
|
||||
0x43800000, /* 0x5c */
|
||||
0x43a00000, /* 0x5d */
|
||||
0x43c00000, /* 0x5e */
|
||||
0x43e00000, /* 0x5f */
|
||||
0x44000000, /* 0x60 */
|
||||
0x44200000, /* 0x61 */
|
||||
0x44400000, /* 0x62 */
|
||||
0x44600000, /* 0x63 */
|
||||
0x44800000, /* 0x64 */
|
||||
0x44a00000, /* 0x65 */
|
||||
0x44c00000, /* 0x66 */
|
||||
0x44e00000, /* 0x67 */
|
||||
0x45000000, /* 0x68 */
|
||||
0x45200000, /* 0x69 */
|
||||
0x45400000, /* 0x6a */
|
||||
0x45600000, /* 0x6b */
|
||||
0x45800000, /* 0x6c */
|
||||
0x45a00000, /* 0x6d */
|
||||
0x45c00000, /* 0x6e */
|
||||
0x45e00000, /* 0x6f */
|
||||
0x46000000, /* 0x70 */
|
||||
0x46200000, /* 0x71 */
|
||||
0x46400000, /* 0x72 */
|
||||
0x46600000, /* 0x73 */
|
||||
0x46800000, /* 0x74 */
|
||||
0x46a00000, /* 0x75 */
|
||||
0x46c00000, /* 0x76 */
|
||||
0x46e00000, /* 0x77 */
|
||||
0x47000000, /* 0x78 */
|
||||
0x47200000, /* 0x79 */
|
||||
0x47400000, /* 0x7a */
|
||||
0x47600000, /* 0x7b */
|
||||
0x7f800000, /* 0x7c */
|
||||
0xffc00000, /* 0x7d */
|
||||
0xffc00000, /* 0x7e */
|
||||
0xffc00000, /* 0x7f */
|
||||
0x80000000, /* 0x80 */
|
||||
0xb7800000, /* 0x81 */
|
||||
0xb8000000, /* 0x82 */
|
||||
0xb8400000, /* 0x83 */
|
||||
0xb8800000, /* 0x84 */
|
||||
0xb8a00000, /* 0x85 */
|
||||
0xb8c00000, /* 0x86 */
|
||||
0xb8e00000, /* 0x87 */
|
||||
0xb9000000, /* 0x88 */
|
||||
0xb9200000, /* 0x89 */
|
||||
0xb9400000, /* 0x8a */
|
||||
0xb9600000, /* 0x8b */
|
||||
0xb9800000, /* 0x8c */
|
||||
0xb9a00000, /* 0x8d */
|
||||
0xb9c00000, /* 0x8e */
|
||||
0xb9e00000, /* 0x8f */
|
||||
0xba000000, /* 0x90 */
|
||||
0xba200000, /* 0x91 */
|
||||
0xba400000, /* 0x92 */
|
||||
0xba600000, /* 0x93 */
|
||||
0xba800000, /* 0x94 */
|
||||
0xbaa00000, /* 0x95 */
|
||||
0xbac00000, /* 0x96 */
|
||||
0xbae00000, /* 0x97 */
|
||||
0xbb000000, /* 0x98 */
|
||||
0xbb200000, /* 0x99 */
|
||||
0xbb400000, /* 0x9a */
|
||||
0xbb600000, /* 0x9b */
|
||||
0xbb800000, /* 0x9c */
|
||||
0xbba00000, /* 0x9d */
|
||||
0xbbc00000, /* 0x9e */
|
||||
0xbbe00000, /* 0x9f */
|
||||
0xbc000000, /* 0xa0 */
|
||||
0xbc200000, /* 0xa1 */
|
||||
0xbc400000, /* 0xa2 */
|
||||
0xbc600000, /* 0xa3 */
|
||||
0xbc800000, /* 0xa4 */
|
||||
0xbca00000, /* 0xa5 */
|
||||
0xbcc00000, /* 0xa6 */
|
||||
0xbce00000, /* 0xa7 */
|
||||
0xbd000000, /* 0xa8 */
|
||||
0xbd200000, /* 0xa9 */
|
||||
0xbd400000, /* 0xaa */
|
||||
0xbd600000, /* 0xab */
|
||||
0xbd800000, /* 0xac */
|
||||
0xbda00000, /* 0xad */
|
||||
0xbdc00000, /* 0xae */
|
||||
0xbde00000, /* 0xaf */
|
||||
0xbe000000, /* 0xb0 */
|
||||
0xbe200000, /* 0xb1 */
|
||||
0xbe400000, /* 0xb2 */
|
||||
0xbe600000, /* 0xb3 */
|
||||
0xbe800000, /* 0xb4 */
|
||||
0xbea00000, /* 0xb5 */
|
||||
0xbec00000, /* 0xb6 */
|
||||
0xbee00000, /* 0xb7 */
|
||||
0xbf000000, /* 0xb8 */
|
||||
0xbf200000, /* 0xb9 */
|
||||
0xbf400000, /* 0xba */
|
||||
0xbf600000, /* 0xbb */
|
||||
0xbf800000, /* 0xbc */
|
||||
0xbfa00000, /* 0xbd */
|
||||
0xbfc00000, /* 0xbe */
|
||||
0xbfe00000, /* 0xbf */
|
||||
0xc0000000, /* 0xc0 */
|
||||
0xc0200000, /* 0xc1 */
|
||||
0xc0400000, /* 0xc2 */
|
||||
0xc0600000, /* 0xc3 */
|
||||
0xc0800000, /* 0xc4 */
|
||||
0xc0a00000, /* 0xc5 */
|
||||
0xc0c00000, /* 0xc6 */
|
||||
0xc0e00000, /* 0xc7 */
|
||||
0xc1000000, /* 0xc8 */
|
||||
0xc1200000, /* 0xc9 */
|
||||
0xc1400000, /* 0xca */
|
||||
0xc1600000, /* 0xcb */
|
||||
0xc1800000, /* 0xcc */
|
||||
0xc1a00000, /* 0xcd */
|
||||
0xc1c00000, /* 0xce */
|
||||
0xc1e00000, /* 0xcf */
|
||||
0xc2000000, /* 0xd0 */
|
||||
0xc2200000, /* 0xd1 */
|
||||
0xc2400000, /* 0xd2 */
|
||||
0xc2600000, /* 0xd3 */
|
||||
0xc2800000, /* 0xd4 */
|
||||
0xc2a00000, /* 0xd5 */
|
||||
0xc2c00000, /* 0xd6 */
|
||||
0xc2e00000, /* 0xd7 */
|
||||
0xc3000000, /* 0xd8 */
|
||||
0xc3200000, /* 0xd9 */
|
||||
0xc3400000, /* 0xda */
|
||||
0xc3600000, /* 0xdb */
|
||||
0xc3800000, /* 0xdc */
|
||||
0xc3a00000, /* 0xdd */
|
||||
0xc3c00000, /* 0xde */
|
||||
0xc3e00000, /* 0xdf */
|
||||
0xc4000000, /* 0xe0 */
|
||||
0xc4200000, /* 0xe1 */
|
||||
0xc4400000, /* 0xe2 */
|
||||
0xc4600000, /* 0xe3 */
|
||||
0xc4800000, /* 0xe4 */
|
||||
0xc4a00000, /* 0xe5 */
|
||||
0xc4c00000, /* 0xe6 */
|
||||
0xc4e00000, /* 0xe7 */
|
||||
0xc5000000, /* 0xe8 */
|
||||
0xc5200000, /* 0xe9 */
|
||||
0xc5400000, /* 0xea */
|
||||
0xc5600000, /* 0xeb */
|
||||
0xc5800000, /* 0xec */
|
||||
0xc5a00000, /* 0xed */
|
||||
0xc5c00000, /* 0xee */
|
||||
0xc5e00000, /* 0xef */
|
||||
0xc6000000, /* 0xf0 */
|
||||
0xc6200000, /* 0xf1 */
|
||||
0xc6400000, /* 0xf2 */
|
||||
0xc6600000, /* 0xf3 */
|
||||
0xc6800000, /* 0xf4 */
|
||||
0xc6a00000, /* 0xf5 */
|
||||
0xc6c00000, /* 0xf6 */
|
||||
0xc6e00000, /* 0xf7 */
|
||||
0xc7000000, /* 0xf8 */
|
||||
0xc7200000, /* 0xf9 */
|
||||
0xc7400000, /* 0xfa */
|
||||
0xc7600000, /* 0xfb */
|
||||
0xff800000, /* 0xfc */
|
||||
0xffc00000, /* 0xfd */
|
||||
0xffc00000, /* 0xfe */
|
||||
0xffc00000, /* 0xff */
|
||||
};
|
||||
|
||||
TEST(float8_test, e4m3fn_to_float)
|
||||
{
|
||||
for (unsigned i = 0; i < 256; i++) {
|
||||
EXPECT_EQ(fui(_mesa_e4m3fn_to_float(i)), e4m3fn_lut[i]);
|
||||
}
|
||||
}
|
||||
|
||||
TEST(float8_test, e5m2_to_float)
|
||||
{
|
||||
for (unsigned i = 0; i < 256; i++) {
|
||||
EXPECT_EQ(fui(_mesa_e5m2_to_float(i)), e5m2_lut[i]);
|
||||
}
|
||||
}
|
||||
|
||||
TEST(float8_test, float_to_e4m3fn_exact)
|
||||
{
|
||||
for (unsigned i = 0; i < 256; i++) {
|
||||
EXPECT_EQ(fui(_mesa_e4m3fn_to_float(_mesa_float_to_e4m3fn(uif(e4m3fn_lut[i])))), e4m3fn_lut[i]);
|
||||
}
|
||||
}
|
||||
|
||||
TEST(float8_test, float_to_e5m2_exact)
|
||||
{
|
||||
for (unsigned i = 0; i < 256; i++) {
|
||||
EXPECT_EQ(fui(_mesa_e5m2_to_float(_mesa_float_to_e5m2(uif(e5m2_lut[i])))), e5m2_lut[i]);
|
||||
}
|
||||
}
|
||||
|
||||
TEST(float8_test, float_to_e4m3fn_rtne)
|
||||
{
|
||||
/* underflow border */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3A800000)), 0);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3A800001)), 1);
|
||||
|
||||
/* rounding up border, denorm */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3b3fffff)), 1);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3b400000)), 2);
|
||||
|
||||
/* rounding down border, denorm */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3ba00000)), 2);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x3ba00001)), 3);
|
||||
|
||||
/* rounding up border, normal */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x4017ffff)), 0x41);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x40180000)), 0x42);
|
||||
|
||||
/* rounding down border, normal */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x40080000)), 0x40);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x40080001)), 0x41);
|
||||
|
||||
/* overflow border */
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x43e80000)), 0x7e);
|
||||
EXPECT_EQ(_mesa_float_to_e4m3fn(uif(0x43e80001)), 0xff);
|
||||
}
|
||||
|
||||
TEST(float8_test, float_to_e5m2_rtne)
|
||||
{
|
||||
/* underflow border */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x37000000)), 0);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x37000001)), 1);
|
||||
|
||||
/* rounding up border, denorm */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x37bfffff)), 1);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x37c00000)), 2);
|
||||
|
||||
/* rounding down border, denorm */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x38200000)), 2);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x38200001)), 3);
|
||||
|
||||
/* rounding up border, normal */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x402fffff)), 0x41);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x40300000)), 0x42);
|
||||
|
||||
/* rounding down border, normal */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x40100000)), 0x40);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x40100001)), 0x41);
|
||||
|
||||
/* overflow border */
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x476fffff)), 0x7b);
|
||||
EXPECT_EQ(_mesa_float_to_e5m2(uif(0x47700000)), 0x7c);
|
||||
}
|
||||
Loading…
Add table
Reference in a new issue