poulpy-cpu-avx 0.6.0

A crate providing concrete AVX accelerated CPU implementations of poulpy-hal through its open extension points
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

//! Polynomial convolution AVX2 kernels for [`NTT120Avx`](crate::NTT120Avx).
//!
//! Reuses the block-outer pack-then-multiply structure from the generic NTT120
//! implementation, but commits overwrite-mode results with non-temporal stores
//! so convolution output lines do not evict the packed operand rows from cache.

use std::mem::size_of;

use bytemuck::{cast_slice, cast_slice_mut};
use core::arch::x86_64::_mm_sfence;

use poulpy_cpu_ref::reference::ntt120::{mat_vec::BbcMeta, primes::Primes30, types::Q120bScalar, vec_znx_dft::NttModuleHandle};
use poulpy_hal::layouts::{CnvPVecLBackendRef, CnvPVecRBackendRef, Module, VecZnxDftBackendMut, ZnxView, ZnxViewMut};

use super::{
    arithmetic_avx::{
        pack_left_1blk_x2_avx2, pack_right_1blk_x2_avx2, pairwise_pack_left_1blk_x2_avx2, pairwise_pack_right_1blk_x2_avx2,
    },
    mat_vec_avx::vec_mat1col_product_x2_bbc_avx2,
};
use crate::NTT120Avx;

/// Scratch bytes required by [`cnv_apply_dft_avx`].
pub(crate) fn cnv_apply_dft_avx_tmp_bytes(a_size: usize, b_size: usize) -> usize {
    16 * (a_size + b_size) * size_of::<u32>()
}

/// Scratch bytes required by [`cnv_pairwise_apply_dft_avx`].
pub(crate) fn cnv_pairwise_apply_dft_avx_tmp_bytes(res_size: usize, a_size: usize, b_size: usize) -> usize {
    if a_size == 0 || b_size == 0 || res_size == 0 {
        0
    } else {
        (16 * (a_size + b_size) * size_of::<u32>()).max(cnv_apply_dft_avx_tmp_bytes(a_size, b_size))
    }
}

#[allow(clippy::too_many_arguments)]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn cnv_apply_dft_avx(
    module: &Module<NTT120Avx>,
    res: &mut VecZnxDftBackendMut<'_, NTT120Avx>,
    cnv_offset: usize,
    res_col: usize,
    a: &CnvPVecLBackendRef<'_, NTT120Avx>,
    a_col: usize,
    b: &CnvPVecRBackendRef<'_, NTT120Avx>,
    b_col: usize,
    tmp: &mut [u8],
) {
    let n = res.n();
    let res_size = res.size();
    let a_size = a.size();
    let b_size = b.size();
    if res_size == 0 || a_size == 0 || b_size == 0 {
        for j in 0..res_size {
            cast_slice_mut::<_, u64>(res.at_mut(res_col, j)).fill(0);
        }
        return;
    }

    let bound = a_size + b_size - 1;
    let offset = cnv_offset.min(bound);
    let min_size = res_size.min((bound + 1).saturating_sub(offset));

    let meta: &BbcMeta<Primes30> = module.get_bbc_meta();
    let a_cols = a.cols();
    let b_cols = b.cols();
    let n_blks = n / 2;
    let a_row_stride_u64 = 4 * n * a_cols;
    let b_row_stride_u32 = 8 * n * b_cols;
    let a_col_offset_u64 = 4 * n * a_col;
    let b_col_offset_u32 = 8 * n * b_col;
    let a_raw_u64: &[u64] = cast_slice(a.raw());
    let b_raw_u32: &[u32] = cast_slice(b.raw());

    let (prefix, tmp_u32, suffix) = unsafe { tmp.align_to_mut::<u32>() };
    debug_assert!(prefix.is_empty());
    debug_assert!(suffix.is_empty());
    debug_assert!(tmp_u32.len() >= 16 * (a_size + b_size));
    let (a_tmp, b_tmp) = tmp_u32.split_at_mut(16 * a_size);

    for blk in 0..n_blks {
        unsafe {
            pack_left_1blk_x2_avx2(a_tmp, &a_raw_u64[a_col_offset_u64..], a_size, a_row_stride_u64, blk);
            pack_right_1blk_x2_avx2(b_tmp, &b_raw_u32[b_col_offset_u32..], b_size, b_row_stride_u32, blk);
        }

        for k in 0..min_size {
            let k_abs = k + offset;
            let j_min = k_abs.saturating_sub(a_size - 1);
            let j_max = (k_abs + 1).min(b_size);
            let ell = j_max - j_min;
            let a_start = k_abs + 1 - j_max;
            let b_start = b_size - j_max;

            let res_u64: &mut [u64] = cast_slice_mut(res.at_mut(res_col, k));
            unsafe {
                vec_mat1col_product_x2_bbc_avx2::<true>(
                    meta,
                    ell,
                    &mut res_u64[8 * blk..8 * blk + 8],
                    &a_tmp[16 * a_start..],
                    &b_tmp[16 * b_start..],
                );
            }
        }
    }

    for j in min_size..res_size {
        res.at_mut(res_col, j).fill(Q120bScalar([0; 4]));
    }
    _mm_sfence();
}

#[allow(clippy::too_many_arguments)]
#[target_feature(enable = "avx2")]
pub(crate) unsafe fn cnv_pairwise_apply_dft_avx(
    module: &Module<NTT120Avx>,
    res: &mut VecZnxDftBackendMut<'_, NTT120Avx>,
    cnv_offset: usize,
    res_col: usize,
    a: &CnvPVecLBackendRef<'_, NTT120Avx>,
    b: &CnvPVecRBackendRef<'_, NTT120Avx>,
    col_i: usize,
    col_j: usize,
    tmp: &mut [u8],
) {
    if col_i == col_j {
        unsafe { cnv_apply_dft_avx(module, res, cnv_offset, res_col, a, col_i, b, col_j, tmp) };
        return;
    }

    let n = res.n();
    let res_size = res.size();
    let a_size = a.size();
    let b_size = b.size();
    if res_size == 0 || a_size == 0 || b_size == 0 {
        for j in 0..res_size {
            cast_slice_mut::<_, u64>(res.at_mut(res_col, j)).fill(0);
        }
        return;
    }

    let meta: &BbcMeta<Primes30> = module.get_bbc_meta();
    let a_cols = a.cols();
    let b_cols = b.cols();

    let bound = a_size + b_size - 1;
    let offset = cnv_offset.min(bound);
    let min_size = res_size.min((bound + 1).saturating_sub(offset));
    let n_blks = n / 2;
    let a_row_stride_u64 = 4 * n * a_cols;
    let b_row_stride_u32 = 8 * n * b_cols;
    let a_col_offset_u64_i = 4 * n * col_i;
    let a_col_offset_u64_j = 4 * n * col_j;
    let b_col_offset_u32_i = 8 * n * col_i;
    let b_col_offset_u32_j = 8 * n * col_j;
    let a_raw_u64: &[u64] = cast_slice(a.raw());
    let b_raw_u32: &[u32] = cast_slice(b.raw());

    let (prefix, tmp_u32, suffix) = unsafe { tmp.align_to_mut::<u32>() };
    debug_assert!(prefix.is_empty());
    debug_assert!(suffix.is_empty());
    debug_assert!(tmp_u32.len() >= 16 * (a_size + b_size));
    let (a_tmp, b_tmp) = tmp_u32.split_at_mut(16 * a_size);

    for blk in 0..n_blks {
        unsafe {
            pairwise_pack_left_1blk_x2_avx2(
                a_tmp,
                &a_raw_u64[a_col_offset_u64_i..],
                &a_raw_u64[a_col_offset_u64_j..],
                a_size,
                a_row_stride_u64,
                blk,
            );
            pairwise_pack_right_1blk_x2_avx2(
                b_tmp,
                &b_raw_u32[b_col_offset_u32_i..],
                &b_raw_u32[b_col_offset_u32_j..],
                b_size,
                b_row_stride_u32,
                blk,
            );
        }

        for k in 0..min_size {
            let k_abs = k + offset;
            let j_min = k_abs.saturating_sub(a_size - 1);
            let j_max = (k_abs + 1).min(b_size);
            let ell = j_max - j_min;
            let a_start = k_abs + 1 - j_max;
            let b_start = b_size - j_max;

            let res_u64: &mut [u64] = cast_slice_mut(res.at_mut(res_col, k));
            unsafe {
                vec_mat1col_product_x2_bbc_avx2::<true>(
                    meta,
                    ell,
                    &mut res_u64[8 * blk..8 * blk + 8],
                    &a_tmp[16 * a_start..],
                    &b_tmp[16 * b_start..],
                );
            }
        }
    }

    for j in min_size..res_size {
        res.at_mut(res_col, j).fill(Q120bScalar([0; 4]));
    }
    _mm_sfence();
}