f8_utils.hpp 10.5 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

#pragma once

#include "ck/utility/data_type.hpp"

namespace ck {

// fp8 rounding modes
// use standard for rounding to nearest, the faster one
// use stochastic for stochastic rounding, helps to avoid error accumulation
enum class f8_rounding_mode
{
    standard,
    stochastic
};

19
20
21
__host__ inline int clz(uint32_t x) { return __builtin_clz(x); }
__device__ inline int clz(uint32_t x) { return __clz(x); }

22
23
24
25
26
27
} // namespace ck

namespace ck::utils {

namespace {

28
29
template <typename X, typename Y, bool negative_zero_nan, bool clip, bool stoch>
__host__ __device__ Y run_cast_to_f8(X x, uint32_t rng)
30
{
31
32
33
    // fp8/bf8 exponent/mantissa layout
    constexpr int out_exp  = NumericUtils<Y>::exp;
    constexpr int out_mant = NumericUtils<Y>::mant;
34

35
36
37
    // original type exponent/mantissa layout
    constexpr int in_exp  = NumericUtils<X>::exp;
    constexpr int in_mant = NumericUtils<X>::mant;
38

39
    int exponent, bias;
40
41
    uint32_t head, mantissa, sign;
    // nan code is same for float and half
42
43
    constexpr Y nan_code        = 0x80;
    constexpr uint32_t nan_mask = NumericUtils<X>::nan_mask;
44
45

    // convert to bitwise
46
    using T_bitwise     = typename NumericUtils<X>::bitwise_type;
47
48
49
    T_bitwise x_bitwise = *(reinterpret_cast<T_bitwise*>(&x));

    // unpack the input, depends on datatype
50
51
52
53
    head     = x_bitwise & NumericUtils<X>::head_mask;
    mantissa = x_bitwise & NumericUtils<X>::mant_mask;
    exponent = (head >> in_mant) & NumericUtils<X>::exp_mask;
    sign     = head >> (in_exp + in_mant);
54
    bias     = NumericUtils<X>::bias;
55
56
57
58

    uint32_t signed_inf   = (sign << (in_exp + in_mant)) + (((1 << in_exp) - 1) << in_mant);
    uint32_t drop_mask    = (1 << (in_mant - out_mant)) - 1;
    constexpr int max_exp = (1 << out_exp) - (negative_zero_nan ? 1 : 2);
59
60
61
62
63
64
65
66
67
68
69
70

    if constexpr(negative_zero_nan)
    {
        if((x_bitwise & nan_mask) == nan_mask)
            return nan_code;
    }
    else
    {
        if((x_bitwise & nan_mask) == nan_mask)
            return signed_inf + (mantissa != 0 ? 1 : 0);
    }

71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
    // check if x is 0.0
    if(x_bitwise == 0)
        return 0;

    // First need to check if it is normal or denorm as there is a difference of implict 1
    // Then need to adjust the exponent to align with the F8 exponent, in the meanwhile, shift
    // The mantissa. Then for stochastic rounding, add rng to mantissa and truncate. And for
    // RNE, no need to add rng. Then probably need to check whether there is carry and adjust
    // exponent and mantissa again3

    // For IEEE bias mode, the bias is 2^(k-1)-1 where k is the width of exponent bits
    const int out_bias                  = (1 << (out_exp - 1)) - 1 + (negative_zero_nan ? 1 : 0);
    const int out_denormal_act_exponent = 1 - out_bias; // actual exponent of f8 denormal
    // act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
    // out_exponent is the converted f8 exponent with bias encoding
    // exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
    // the difference needs to be adjusted and mantissa shifted
    int act_exponent, out_exponent, exponent_diff;

    if(exponent == 0)
    { // fp32/fp16 is in denormal.
        /* fp32 denormal is below 2^-127 so it is usually not a concern here, we mostly concern fp16
here. In this case, f8 is usually in denormal. But there could be exceptions. fp16 denormal has
exponent bias 15 while bf8 with NANOO has exponent bias 16. It means that there are some numbers in
fp16 denormal but they are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8 (NANOO) normal.
In this case, the fp16 mantissa should be shift left by 1 */
        act_exponent  = exponent - bias + 1;
        exponent_diff = out_denormal_act_exponent -
                        act_exponent; // actual exponent is exponent-bias+1 as it is denormal
    }
    else
    { // fp32/fp16 is normal with implicit 1
        act_exponent = exponent - bias;
        if(act_exponent <= out_denormal_act_exponent)
106
        {
107
108
109
110
111
112
            /* This is the case where fp32/fp16 is normal but it is in f8 denormal range.
   For example fp8 nanoo mode, denormal exponent is -7, but if the fp32/fp16
   actual exponent is -7, it is actually larger due to the implict 1,
   Therefore it needs to be adjust to -6 and mantissa shift right by 1.
   So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
            exponent_diff = out_denormal_act_exponent - act_exponent;
113
        }
114
115
116
117
118
119
120
        else
        { // both fp32/fp16 and f8 are in normal range
            exponent_diff =
                0; // exponent_diff=0 does not mean there is no difference for this case,
            // act_exponent could be larger. Just that it does not need shift mantissa
        }
        mantissa += (1 << in_mant); // Add the implicit 1 into mantissa
121
122
    }

123
124
125
126
127
128
    bool midpoint = (mantissa & ((1 << (in_mant - out_mant + exponent_diff)) - 1)) ==
                    (1 << (in_mant - out_mant + exponent_diff - 1));
    /* This part is a bit tricky. The judgment of whether it is a tie needs to be done before we
 shift right as shift right could rip off some residual part and make something not midpoint look
 like midpoint. For example, the fp16 number 0x1002 (0 00100 0000000010), it is larger than
 midpoint, but after shift right by 4 bits, it would look like midpoint. */
129

130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
    if(exponent_diff > 0)
        mantissa >>= exponent_diff;
    else if(exponent_diff == -1)
        mantissa <<= -exponent_diff;
    bool implicit_one = mantissa & (1 << in_mant);
    // if there is no implict 1, it  means the f8 is denormal and need to adjust to denorm exponent
    out_exponent =
        (act_exponent + exponent_diff) /*actual f8 exponent*/ + out_bias - (implicit_one ? 0 : 1);

    // Now we have the exponent and mantissa adjusted
    bool odd =
        mantissa &
        (1 << (in_mant - out_mant)); // if the least significant bit that is not truncated is 1
    mantissa += (stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa)) & drop_mask;

    // Now we deal with overflow
    if(out_exponent == 0)
147
    {
148
149
150
151
152
        if((1 << in_mant) & mantissa)
        {
            out_exponent = 1; // denormal overflow to become normal, promote exponent
            // No need to make 1 implicit now as it will be addressed later
        }
153
    }
154
    else
155
    {
156
        if((1 << (in_mant + 1)) & mantissa)
157
        {
158
159
160
            mantissa >>= 1;
            out_exponent++;
            // No need to make 1 implicit now as it will be addressed later
161
162
        }
    }
163
164
165
166

    mantissa >>= (in_mant - out_mant);

    if(out_exponent > max_exp)
167
168
169
    {
        if(clip)
        {
170
171
            mantissa     = (1 << out_mant) - 1;
            out_exponent = max_exp;
172
173
174
175
176
177
178
179
        }
        else
        {
            return signed_inf;
        }
    }

    // check if x is 0.0 or -0.0
180
    if(out_exponent == 0 && mantissa == 0)
181
182
        return negative_zero_nan ? 0 : (sign << (out_exp + out_mant));
    mantissa &= (1 << out_mant) - 1;
183
    return (sign << (out_exp + out_mant)) | (out_exponent << out_mant) | mantissa;
184
185
}

186
187
template <typename X, typename Y, bool negative_zero_nan>
__host__ __device__ Y run_cast_from_f8(X x)
188
{
189
190
191
    // fp8/bf8 exponent/mantissa layout
    constexpr int in_exp  = NumericUtils<X>::exp;
    constexpr int in_mant = NumericUtils<X>::mant;
192
193

    // resulting type exponent/mantissa layout
194
195
    constexpr int out_exp  = NumericUtils<Y>::exp;
    constexpr int out_mant = NumericUtils<Y>::mant;
196
197

    // prepare the codes
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
    constexpr X nan_code = 0x80;
    Y Inf, NegInf, NaN, Neg0;
    using T_bitwise = typename NumericUtils<Y>::bitwise_type;

    constexpr T_bitwise Inf_bitwise    = NumericUtils<Y>::Inf;
    constexpr T_bitwise NegInf_bitwise = NumericUtils<Y>::NegInf;
    constexpr T_bitwise NaN_bitwise    = NumericUtils<Y>::NaN;
    constexpr T_bitwise Neg0_bitwise   = NumericUtils<Y>::Neg0;

    Inf    = *(reinterpret_cast<const Y*>(&Inf_bitwise));
    NegInf = *(reinterpret_cast<const Y*>(&NegInf_bitwise));
    NaN    = *(reinterpret_cast<const Y*>(&NaN_bitwise));
    Neg0   = *(reinterpret_cast<const Y*>(&Neg0_bitwise));

    // check if x is 0.0
    if(x == 0)
        return static_cast<Y>(0);
215
216

    // unpack the input
217
218
219
    uint32_t sign     = x >> (in_exp + in_mant);
    uint32_t mantissa = x & ((1 << in_mant) - 1);
    int exponent      = (x & 0x7F) >> in_mant;
220
221

    constexpr int exp_low_cutoff =
222
223
        (1 << (out_exp - 1)) - (1 << (in_exp - 1)) + 1 - (negative_zero_nan ? 1 : 0);
    T_bitwise retval;
224
225
226
227

    if constexpr(negative_zero_nan)
    {
        if(x == nan_code)
228
            return NaN;
229
230
231
232
    }
    else
    {
        if(x == nan_code)
233
234
235
236
237
238
239
240
241
242
            return Neg0;
        if(exponent == ((1 << in_exp) - 1))
            return (mantissa == 0) ? (sign ? NegInf : Inf) : NaN;
    }

    if((NumericUtils<Y>::mant == 10) && (NumericUtils<X>::mant == 2) && !negative_zero_nan)
    {
        retval = x;
        retval <<= 8;
        return *(reinterpret_cast<const Y*>(&retval));
243
244
245
246
247
248
    }

    // subnormal input
    if(exponent == 0)
    {
        // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
249
250
251
        int sh = 1 + clz(mantissa) - (32 - in_mant);
        mantissa <<= sh;
        exponent += 1 - sh;
252
        mantissa &= ((1 << in_mant) - 1);
253
254
    }
    exponent += exp_low_cutoff - 1;
255
    mantissa <<= out_mant - in_mant;
256
257
258
259

    // subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
    if(exponent <= 0)
    {
260
        mantissa |= 1 << out_mant;
261
262
263
264
        mantissa >>= 1 - exponent;
        exponent = 0;
    }

265
266
    retval = (sign << (out_exp + out_mant)) | (exponent << out_mant) | mantissa;
    return *(reinterpret_cast<const Y*>(&retval));
267
268
269
270
}

} // namespace

271
272
template <typename X, typename Y, bool negative_zero_nan, bool clip, bool stoch>
__host__ __device__ Y cast_to_f8(X x, uint32_t rng)
273
{
274
275
276
277
    // check datatypes
    constexpr bool is_half  = std::is_same<X, half_t>::value;
    constexpr bool is_float = std::is_same<X, float>::value;
    static_assert(is_half || is_float, "Only half and float can be casted.");
278

279
    return run_cast_to_f8<X, Y, negative_zero_nan, clip, stoch>(x, rng);
280
281
}

282
283
template <typename X, typename Y, bool negative_zero_nan>
__host__ __device__ Y cast_from_f8(X x)
284
285
{
    // check datatype
286
287
    constexpr bool is_half  = std::is_same<Y, half_t>::value;
    constexpr bool is_float = std::is_same<Y, float>::value;
288
289
    static_assert(is_half || is_float, "only half and float are supported.");

290
    return run_cast_from_f8<X, Y, negative_zero_nan>(x);
291
292
293
}

} // namespace ck::utils