| 1 | // SPDX-License-Identifier: MIT |
| 2 | // |
| 3 | // Copyright 2024 Advanced Micro Devices, Inc. |
| 4 | |
| 5 | #include "spl_debug.h" |
| 6 | #include "spl_custom_float.h" |
| 7 | |
| 8 | static bool spl_build_custom_float(struct spl_fixed31_32 value, |
| 9 | const struct spl_custom_float_format *format, |
| 10 | bool *negative, |
| 11 | uint32_t *mantissa, |
| 12 | uint32_t *exponenta) |
| 13 | { |
| 14 | uint32_t exp_offset = (1 << (format->exponenta_bits - 1)) - 1; |
| 15 | |
| 16 | const struct spl_fixed31_32 mantissa_constant_plus_max_fraction = |
| 17 | spl_fixpt_from_fraction(numerator: (1LL << (format->mantissa_bits + 1)) - 1, |
| 18 | denominator: 1LL << format->mantissa_bits); |
| 19 | |
| 20 | struct spl_fixed31_32 mantiss; |
| 21 | |
| 22 | if (spl_fixpt_eq(arg1: value, arg2: spl_fixpt_zero)) { |
| 23 | *negative = false; |
| 24 | *mantissa = 0; |
| 25 | *exponenta = 0; |
| 26 | return true; |
| 27 | } |
| 28 | |
| 29 | if (spl_fixpt_lt(arg1: value, arg2: spl_fixpt_zero)) { |
| 30 | *negative = format->sign; |
| 31 | value = spl_fixpt_neg(arg: value); |
| 32 | } else { |
| 33 | *negative = false; |
| 34 | } |
| 35 | |
| 36 | if (spl_fixpt_lt(arg1: value, arg2: spl_fixpt_one)) { |
| 37 | uint32_t i = 1; |
| 38 | |
| 39 | do { |
| 40 | value = spl_fixpt_shl(arg: value, shift: 1); |
| 41 | ++i; |
| 42 | } while (spl_fixpt_lt(arg1: value, arg2: spl_fixpt_one)); |
| 43 | |
| 44 | --i; |
| 45 | |
| 46 | if (exp_offset <= i) { |
| 47 | *mantissa = 0; |
| 48 | *exponenta = 0; |
| 49 | return true; |
| 50 | } |
| 51 | |
| 52 | *exponenta = exp_offset - i; |
| 53 | } else if (spl_fixpt_le(arg1: mantissa_constant_plus_max_fraction, arg2: value)) { |
| 54 | uint32_t i = 1; |
| 55 | |
| 56 | do { |
| 57 | value = spl_fixpt_shr(arg: value, shift: 1); |
| 58 | ++i; |
| 59 | } while (spl_fixpt_lt(arg1: mantissa_constant_plus_max_fraction, arg2: value)); |
| 60 | |
| 61 | *exponenta = exp_offset + i - 1; |
| 62 | } else { |
| 63 | *exponenta = exp_offset; |
| 64 | } |
| 65 | |
| 66 | mantiss = spl_fixpt_sub(arg1: value, arg2: spl_fixpt_one); |
| 67 | |
| 68 | if (spl_fixpt_lt(arg1: mantiss, arg2: spl_fixpt_zero) || |
| 69 | spl_fixpt_lt(arg1: spl_fixpt_one, arg2: mantiss)) |
| 70 | mantiss = spl_fixpt_zero; |
| 71 | else |
| 72 | mantiss = spl_fixpt_shl(arg: mantiss, shift: format->mantissa_bits); |
| 73 | |
| 74 | *mantissa = spl_fixpt_floor(arg: mantiss); |
| 75 | |
| 76 | return true; |
| 77 | } |
| 78 | |
| 79 | static bool spl_setup_custom_float(const struct spl_custom_float_format *format, |
| 80 | bool negative, |
| 81 | uint32_t mantissa, |
| 82 | uint32_t exponenta, |
| 83 | uint32_t *result) |
| 84 | { |
| 85 | uint32_t i = 0; |
| 86 | uint32_t j = 0; |
| 87 | uint32_t value = 0; |
| 88 | |
| 89 | /* verification code: |
| 90 | * once calculation is ok we can remove it |
| 91 | */ |
| 92 | |
| 93 | const uint32_t mantissa_mask = |
| 94 | (1 << (format->mantissa_bits + 1)) - 1; |
| 95 | |
| 96 | const uint32_t exponenta_mask = |
| 97 | (1 << (format->exponenta_bits + 1)) - 1; |
| 98 | |
| 99 | if (mantissa & ~mantissa_mask) { |
| 100 | SPL_BREAK_TO_DEBUGGER(); |
| 101 | mantissa = mantissa_mask; |
| 102 | } |
| 103 | |
| 104 | if (exponenta & ~exponenta_mask) { |
| 105 | SPL_BREAK_TO_DEBUGGER(); |
| 106 | exponenta = exponenta_mask; |
| 107 | } |
| 108 | |
| 109 | /* end of verification code */ |
| 110 | |
| 111 | while (i < format->mantissa_bits) { |
| 112 | uint32_t mask = 1 << i; |
| 113 | |
| 114 | if (mantissa & mask) |
| 115 | value |= mask; |
| 116 | |
| 117 | ++i; |
| 118 | } |
| 119 | |
| 120 | while (j < format->exponenta_bits) { |
| 121 | uint32_t mask = 1 << j; |
| 122 | |
| 123 | if (exponenta & mask) |
| 124 | value |= mask << i; |
| 125 | |
| 126 | ++j; |
| 127 | } |
| 128 | |
| 129 | if (negative && format->sign) |
| 130 | value |= 1 << (i + j); |
| 131 | |
| 132 | *result = value; |
| 133 | |
| 134 | return true; |
| 135 | } |
| 136 | |
| 137 | bool spl_convert_to_custom_float_format(struct spl_fixed31_32 value, |
| 138 | const struct spl_custom_float_format *format, |
| 139 | uint32_t *result) |
| 140 | { |
| 141 | uint32_t mantissa; |
| 142 | uint32_t exponenta; |
| 143 | bool negative; |
| 144 | |
| 145 | return spl_build_custom_float(value, format, negative: &negative, mantissa: &mantissa, exponenta: &exponenta) && |
| 146 | spl_setup_custom_float(format, |
| 147 | negative, |
| 148 | mantissa, |
| 149 | exponenta, |
| 150 | result); |
| 151 | } |
| 152 | |