eunomia 0.8.0

Atlas datatype law: the Single Source of Truth for numeric/scalar datatype vocabulary (scalars, complex, packed formats, conversions)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304

use crate::convert::{widen_finite, widen_finite_high_word, widen_high_word};
use crate::types::{Bf16, Bf4, Bf8, F32, F4, F8};

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_bf8_to_bf16(packed: &[Bf8], unpacked: &mut [Bf16]) {
    use core::arch::x86_64::*;

    let len = packed.len().min(unpacked.len());
    let mut i = 0;

    static TABLE: [i32; 256] = {
        let mut table = [0i32; 256];
        let mut index = 0;
        while index < table.len() {
            table[index] = widen_high_word::<5, 2>(index as u32) as i32;
            index += 1;
        }
        table
    };

    while i + 16 <= len {
        let input = _mm_loadu_si128(packed.as_ptr().add(i) as *const __m128i);
        let lower_indices = _mm256_cvtepu8_epi32(input);
        let upper_indices = _mm256_cvtepu8_epi32(_mm_srli_si128(input, 8));
        let lower = _mm256_i32gather_epi32(TABLE.as_ptr(), lower_indices, 4);
        let upper = _mm256_i32gather_epi32(TABLE.as_ptr(), upper_indices, 4);
        let packed_words = _mm256_packus_epi32(lower, upper);
        let ordered = _mm256_permute4x64_epi64(packed_words, 0xD8);
        _mm256_storeu_si256(unpacked.as_mut_ptr().add(i) as *mut __m256i, ordered);
        i += 16;
    }
    for j in i..len {
        unpacked[j] = Bf16(widen_high_word::<5, 2>(u32::from(packed[j].0)));
    }
}

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_bf4_to_bf16(packed: &[Bf4], unpacked: &mut [Bf16]) {
    use core::arch::x86_64::*;
    let len = packed.len().min(unpacked.len());
    let mut i = 0;

    static TABLE_LO: [u8; 16] = {
        let mut t = [0u8; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = (widen_finite_high_word::<2, 1>(idx as u32) & 0xFF) as u8;
            idx += 1;
        }
        t
    };
    static TABLE_HI: [u8; 16] = {
        let mut t = [0u8; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = (widen_finite_high_word::<2, 1>(idx as u32) >> 8) as u8;
            idx += 1;
        }
        t
    };

    let table_lo = _mm_loadu_si128(TABLE_LO.as_ptr() as *const _);
    let table_hi = _mm_loadu_si128(TABLE_HI.as_ptr() as *const _);
    let mask_0f = _mm_set1_epi8(0x0F);

    while i + 16 <= len {
        let ptr = packed.as_ptr().add(i) as *const __m128i;
        let v_in = _mm_loadu_si128(ptr);
        let indices = _mm_and_si128(v_in, mask_0f);

        let res_lo = _mm_shuffle_epi8(table_lo, indices);
        let res_hi = _mm_shuffle_epi8(table_hi, indices);

        let out_lo = _mm_unpacklo_epi8(res_lo, res_hi);
        let out_hi = _mm_unpackhi_epi8(res_lo, res_hi);

        _mm_storeu_si128(unpacked.as_mut_ptr().add(i) as *mut _, out_lo);
        _mm_storeu_si128(unpacked.as_mut_ptr().add(i + 8) as *mut _, out_hi);

        i += 16;
    }
    for j in i..len {
        let b = packed[j].0;
        unpacked[j] = Bf16(widen_finite_high_word::<2, 1>(b as u32));
    }
}

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_bf4_to_bf16_packed(packed: &[u8], unpacked: &mut [Bf16]) {
    use core::arch::x86_64::*;
    let len = packed.len();
    let n = len.min(unpacked.len() / 2);
    let mut i = 0;

    static TABLE_LO: [u8; 16] = {
        let mut t = [0u8; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = (widen_finite_high_word::<2, 1>(idx as u32) & 0xFF) as u8;
            idx += 1;
        }
        t
    };
    static TABLE_HI: [u8; 16] = {
        let mut t = [0u8; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = (widen_finite_high_word::<2, 1>(idx as u32) >> 8) as u8;
            idx += 1;
        }
        t
    };

    let table_lo = _mm_loadu_si128(TABLE_LO.as_ptr() as *const _);
    let table_hi = _mm_loadu_si128(TABLE_HI.as_ptr() as *const _);
    let mask_0f = _mm_set1_epi8(0x0F);

    while i + 16 <= n {
        let ptr = packed.as_ptr().add(i) as *const __m128i;
        let v = _mm_loadu_si128(ptr);

        let low_nibbles = _mm_and_si128(v, mask_0f);
        let high_nibbles = _mm_and_si128(_mm_srli_epi16(v, 4), mask_0f);

        let res_lo = _mm_unpacklo_epi8(low_nibbles, high_nibbles);
        let res_hi = _mm_unpackhi_epi8(low_nibbles, high_nibbles);

        let res_lo_lo = _mm_shuffle_epi8(table_lo, res_lo);
        let res_lo_hi = _mm_shuffle_epi8(table_hi, res_lo);
        let out_lo0 = _mm_unpacklo_epi8(res_lo_lo, res_lo_hi);
        let out_lo1 = _mm_unpackhi_epi8(res_lo_lo, res_lo_hi);

        let res_hi_lo = _mm_shuffle_epi8(table_lo, res_hi);
        let res_hi_hi = _mm_shuffle_epi8(table_hi, res_hi);
        let out_hi0 = _mm_unpacklo_epi8(res_hi_lo, res_hi_hi);
        let out_hi1 = _mm_unpackhi_epi8(res_hi_lo, res_hi_hi);

        _mm_storeu_si128(unpacked.as_mut_ptr().add(2 * i) as *mut _, out_lo0);
        _mm_storeu_si128(unpacked.as_mut_ptr().add(2 * i + 8) as *mut _, out_lo1);
        _mm_storeu_si128(unpacked.as_mut_ptr().add(2 * i + 16) as *mut _, out_hi0);
        _mm_storeu_si128(unpacked.as_mut_ptr().add(2 * i + 24) as *mut _, out_hi1);

        i += 16;
    }

    for j in i..n {
        let byte = packed[j];
        unpacked[2 * j] = Bf16(widen_finite_high_word::<2, 1>((byte & 0x0F) as u32));
        unpacked[2 * j + 1] = Bf16(widen_finite_high_word::<2, 1>(((byte >> 4) & 0x0F) as u32));
    }
}

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_f4_to_f32(packed: &[F4], unpacked: &mut [F32]) {
    let len = packed.len().min(unpacked.len());
    let mut i = 0;

    static TABLE_BITS: [u32; 16] = {
        let mut t = [0u32; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = widen_finite::<3, 0>(idx as u32);
            idx += 1;
        }
        t
    };

    while i + 8 <= len {
        let ptr = packed.as_ptr().add(i) as *const core::arch::x86_64::__m128i;
        let v_in = core::arch::x86_64::_mm_loadl_epi64(ptr);
        let v_u32 = core::arch::x86_64::_mm256_cvtepu8_epi32(v_in);
        let indices = core::arch::x86_64::_mm256_and_si256(
            v_u32,
            core::arch::x86_64::_mm256_set1_epi32(0x0f),
        );

        let result =
            core::arch::x86_64::_mm256_i32gather_ps(TABLE_BITS.as_ptr() as *const _, indices, 4);

        let out_ptr = unpacked.as_mut_ptr().add(i) as *mut f32;
        core::arch::x86_64::_mm256_storeu_ps(out_ptr, result);
        i += 8;
    }

    for j in i..len {
        unpacked[j] = F32(f32::from_bits(TABLE_BITS[(packed[j].0 & 0x0f) as usize]));
    }
}

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_f4_to_f32_packed(packed: &[u8], unpacked: &mut [F32]) {
    let len = packed.len();
    let n = len.min(unpacked.len() / 2);
    let mut i = 0;

    static TABLE_BITS: [u32; 16] = {
        let mut t = [0u32; 16];
        let mut idx = 0;
        while idx < 16 {
            t[idx] = widen_finite::<3, 0>(idx as u32);
            idx += 1;
        }
        t
    };

    while i + 8 <= n {
        let bytes_ptr = packed.as_ptr().add(i);
        let v_bytes = core::arch::x86_64::_mm_loadl_epi64(bytes_ptr as *const _);
        let v_u32 = core::arch::x86_64::_mm256_cvtepu8_epi32(v_bytes);

        let low_nibbles = core::arch::x86_64::_mm256_and_si256(
            v_u32,
            core::arch::x86_64::_mm256_set1_epi32(0x0f),
        );
        let high_nibbles = core::arch::x86_64::_mm256_and_si256(
            core::arch::x86_64::_mm256_srli_epi32(v_u32, 4),
            core::arch::x86_64::_mm256_set1_epi32(0x0f),
        );

        let val_low = core::arch::x86_64::_mm256_i32gather_ps(
            TABLE_BITS.as_ptr() as *const _,
            low_nibbles,
            4,
        );
        let val_high = core::arch::x86_64::_mm256_i32gather_ps(
            TABLE_BITS.as_ptr() as *const _,
            high_nibbles,
            4,
        );

        let val_lo_u = core::arch::x86_64::_mm256_castps_si256(val_low);
        let val_hi_u = core::arch::x86_64::_mm256_castps_si256(val_high);

        let res0 = core::arch::x86_64::_mm256_unpacklo_epi32(val_lo_u, val_hi_u);
        let res1 = core::arch::x86_64::_mm256_unpackhi_epi32(val_lo_u, val_hi_u);

        let out0 = core::arch::x86_64::_mm256_permute2x128_si256(res0, res1, 0x20);
        let out1 = core::arch::x86_64::_mm256_permute2x128_si256(res0, res1, 0x31);

        let out_ptr0 = unpacked.as_mut_ptr().add(2 * i) as *mut core::arch::x86_64::__m256i;
        core::arch::x86_64::_mm256_storeu_si256(out_ptr0, out0);
        let out_ptr1 = unpacked.as_mut_ptr().add(2 * i + 8) as *mut core::arch::x86_64::__m256i;
        core::arch::x86_64::_mm256_storeu_si256(out_ptr1, out1);

        i += 8;
    }

    for j in i..n {
        let byte = packed[j];
        unpacked[2 * j] = F32(f32::from_bits(TABLE_BITS[(byte & 0x0F) as usize]));
        unpacked[2 * j + 1] = F32(f32::from_bits(TABLE_BITS[((byte >> 4) & 0x0F) as usize]));
    }
}

/// # Safety
/// Callers must ensure the running CPU supports AVX2 (the dispatcher checks
/// `has_avx2()`). Reads and writes stay within the shorter of the two slices.
#[target_feature(enable = "avx2")]
pub unsafe fn unpack_f8_to_f32(packed: &[F8], unpacked: &mut [F32]) {
    let len = packed.len().min(unpacked.len());
    let mut i = 0;

    static TABLE_BITS: [u32; 256] = {
        let mut t = [0u32; 256];
        let mut idx = 0;
        while idx < 256 {
            t[idx] = widen_finite::<4, 3>(idx as u32);
            idx += 1;
        }
        t
    };

    while i + 8 <= len {
        let ptr = packed.as_ptr().add(i) as *const core::arch::x86_64::__m128i;
        let v_in = core::arch::x86_64::_mm_loadl_epi64(ptr);
        let v_u32 = core::arch::x86_64::_mm256_cvtepu8_epi32(v_in);

        let result =
            core::arch::x86_64::_mm256_i32gather_ps(TABLE_BITS.as_ptr() as *const _, v_u32, 4);

        let out_ptr = unpacked.as_mut_ptr().add(i) as *mut f32;
        core::arch::x86_64::_mm256_storeu_ps(out_ptr, result);
        i += 8;
    }

    for j in i..len {
        unpacked[j] = F32(f32::from_bits(TABLE_BITS[packed[j].0 as usize]));
    }
}