numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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

//! AVX-512 binary operation kernels
//!
//! Processes 16 f32s or 8 f64s per iteration using 512-bit vectors.
//!
//! # Streaming Stores
//!
//! For large arrays (> 1MB), this module uses streaming (non-temporal) stores
//! (`_mm512_stream_ps`) to bypass the cache. This improves performance by
//! avoiding cache pollution when the output won't be read immediately.
//!
//! Streaming stores require 64-byte aligned output pointers. If alignment
//! requirements are not met, the code falls back to regular cached stores.

#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

use super::super::super::math::avx512::{
    exp_f32 as exp_vec_f32, exp_f64 as exp_vec_f64, log_f32 as log_vec_f32, log_f64 as log_vec_f64,
};
use super::super::super::streaming::{is_aligned_avx512, should_stream_f32, should_stream_f64};
use super::super::{binary_scalar_f32, binary_scalar_f64};
use crate::ops::BinaryOp;

const F32_LANES: usize = 16;
const F64_LANES: usize = 8;

// ============================================================================
// Macro to generate binary kernels (eliminates duplication between regular/streaming)
// ============================================================================

/// Generates a binary kernel for f32 with the specified store instruction.
macro_rules! impl_binary_f32_avx512 {
    ($name:ident, $vec_op:ident, $store:ident) => {
        #[target_feature(enable = "avx512f")]
        unsafe fn $name(a: *const f32, b: *const f32, out: *mut f32, chunks: usize) {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = _mm512_loadu_ps(a.add(offset));
                let vb = _mm512_loadu_ps(b.add(offset));
                let vr = $vec_op(va, vb);
                $store(out.add(offset), vr);
            }
        }
    };
}

/// Generates a binary kernel for f64 with the specified store instruction.
macro_rules! impl_binary_f64_avx512 {
    ($name:ident, $vec_op:ident, $store:ident) => {
        #[target_feature(enable = "avx512f")]
        unsafe fn $name(a: *const f64, b: *const f64, out: *mut f64, chunks: usize) {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = _mm512_loadu_pd(a.add(offset));
                let vb = _mm512_loadu_pd(b.add(offset));
                let vr = $vec_op(va, vb);
                $store(out.add(offset), vr);
            }
        }
    };
}

// Generate regular (cached) f32 kernels
impl_binary_f32_avx512!(binary_add_f32, _mm512_add_ps, _mm512_storeu_ps);
impl_binary_f32_avx512!(binary_sub_f32, _mm512_sub_ps, _mm512_storeu_ps);
impl_binary_f32_avx512!(binary_mul_f32, _mm512_mul_ps, _mm512_storeu_ps);
impl_binary_f32_avx512!(binary_div_f32, _mm512_div_ps, _mm512_storeu_ps);
impl_binary_f32_avx512!(binary_max_f32, _mm512_max_ps, _mm512_storeu_ps);
impl_binary_f32_avx512!(binary_min_f32, _mm512_min_ps, _mm512_storeu_ps);

// Generate streaming (non-temporal) f32 kernels
impl_binary_f32_avx512!(binary_add_f32_stream, _mm512_add_ps, _mm512_stream_ps);
impl_binary_f32_avx512!(binary_sub_f32_stream, _mm512_sub_ps, _mm512_stream_ps);
impl_binary_f32_avx512!(binary_mul_f32_stream, _mm512_mul_ps, _mm512_stream_ps);
impl_binary_f32_avx512!(binary_div_f32_stream, _mm512_div_ps, _mm512_stream_ps);
impl_binary_f32_avx512!(binary_max_f32_stream, _mm512_max_ps, _mm512_stream_ps);
impl_binary_f32_avx512!(binary_min_f32_stream, _mm512_min_ps, _mm512_stream_ps);

// Generate regular (cached) f64 kernels
impl_binary_f64_avx512!(binary_add_f64, _mm512_add_pd, _mm512_storeu_pd);
impl_binary_f64_avx512!(binary_sub_f64, _mm512_sub_pd, _mm512_storeu_pd);
impl_binary_f64_avx512!(binary_mul_f64, _mm512_mul_pd, _mm512_storeu_pd);
impl_binary_f64_avx512!(binary_div_f64, _mm512_div_pd, _mm512_storeu_pd);
impl_binary_f64_avx512!(binary_max_f64, _mm512_max_pd, _mm512_storeu_pd);
impl_binary_f64_avx512!(binary_min_f64, _mm512_min_pd, _mm512_storeu_pd);

// Generate streaming (non-temporal) f64 kernels
impl_binary_f64_avx512!(binary_add_f64_stream, _mm512_add_pd, _mm512_stream_pd);
impl_binary_f64_avx512!(binary_sub_f64_stream, _mm512_sub_pd, _mm512_stream_pd);
impl_binary_f64_avx512!(binary_mul_f64_stream, _mm512_mul_pd, _mm512_stream_pd);
impl_binary_f64_avx512!(binary_div_f64_stream, _mm512_div_pd, _mm512_stream_pd);
impl_binary_f64_avx512!(binary_max_f64_stream, _mm512_max_pd, _mm512_stream_pd);
impl_binary_f64_avx512!(binary_min_f64_stream, _mm512_min_pd, _mm512_stream_pd);

// ============================================================================
// Pow kernels (more complex, not macro-generated)
// ============================================================================

/// Generates pow kernel with specified store instruction.
macro_rules! impl_pow_f32_avx512 {
    ($name:ident, $store:ident) => {
        #[target_feature(enable = "avx512f")]
        unsafe fn $name(a: *const f32, b: *const f32, out: *mut f32, chunks: usize) {
            for i in 0..chunks {
                let offset = i * F32_LANES;
                let va = _mm512_loadu_ps(a.add(offset));
                let vb = _mm512_loadu_ps(b.add(offset));
                // pow(a, b) = exp(b * log(a))
                let log_a = log_vec_f32(va);
                let b_log_a = _mm512_mul_ps(vb, log_a);
                let vr = exp_vec_f32(b_log_a);
                $store(out.add(offset), vr);
            }
        }
    };
}

macro_rules! impl_pow_f64_avx512 {
    ($name:ident, $store:ident) => {
        #[target_feature(enable = "avx512f")]
        unsafe fn $name(a: *const f64, b: *const f64, out: *mut f64, chunks: usize) {
            for i in 0..chunks {
                let offset = i * F64_LANES;
                let va = _mm512_loadu_pd(a.add(offset));
                let vb = _mm512_loadu_pd(b.add(offset));
                // pow(a, b) = exp(b * log(a))
                let log_a = log_vec_f64(va);
                let b_log_a = _mm512_mul_pd(vb, log_a);
                let vr = exp_vec_f64(b_log_a);
                $store(out.add(offset), vr);
            }
        }
    };
}

impl_pow_f32_avx512!(binary_pow_f32, _mm512_storeu_ps);
impl_pow_f32_avx512!(binary_pow_f32_stream, _mm512_stream_ps);
impl_pow_f64_avx512!(binary_pow_f64, _mm512_storeu_pd);
impl_pow_f64_avx512!(binary_pow_f64_stream, _mm512_stream_pd);

// ============================================================================
// Public dispatch functions
// ============================================================================

/// AVX-512 binary operation for f32
///
/// Uses streaming (non-temporal) stores for arrays > 1MB when output is 64-byte aligned.
///
/// # Safety
/// - CPU must support AVX-512F
/// - All pointers must be valid for `len` elements
#[target_feature(enable = "avx512f")]
pub unsafe fn binary_f32(op: BinaryOp, a: *const f32, b: *const f32, out: *mut f32, len: usize) {
    let chunks = len / F32_LANES;
    let remainder = len % F32_LANES;

    // Use streaming stores for large aligned arrays
    let use_streaming = should_stream_f32(len) && is_aligned_avx512(out);

    // Atan2 has no SIMD implementation - use scalar fallback
    if op == BinaryOp::Atan2 {
        binary_scalar_f32(op, a, b, out, len);
        return;
    }

    if use_streaming {
        match op {
            BinaryOp::Add => binary_add_f32_stream(a, b, out, chunks),
            BinaryOp::Sub => binary_sub_f32_stream(a, b, out, chunks),
            BinaryOp::Mul => binary_mul_f32_stream(a, b, out, chunks),
            BinaryOp::Div => binary_div_f32_stream(a, b, out, chunks),
            BinaryOp::Max => binary_max_f32_stream(a, b, out, chunks),
            BinaryOp::Min => binary_min_f32_stream(a, b, out, chunks),
            BinaryOp::Pow => binary_pow_f32_stream(a, b, out, chunks),
            BinaryOp::Atan2 => unreachable!(), // Handled above
        }
        // Memory fence ensures streaming stores are globally visible
        _mm_sfence();
    } else {
        match op {
            BinaryOp::Add => binary_add_f32(a, b, out, chunks),
            BinaryOp::Sub => binary_sub_f32(a, b, out, chunks),
            BinaryOp::Mul => binary_mul_f32(a, b, out, chunks),
            BinaryOp::Div => binary_div_f32(a, b, out, chunks),
            BinaryOp::Max => binary_max_f32(a, b, out, chunks),
            BinaryOp::Min => binary_min_f32(a, b, out, chunks),
            BinaryOp::Pow => binary_pow_f32(a, b, out, chunks),
            BinaryOp::Atan2 => unreachable!(), // Handled above
        }
    }

    // Handle tail with scalar
    if remainder > 0 {
        let offset = chunks * F32_LANES;
        binary_scalar_f32(op, a.add(offset), b.add(offset), out.add(offset), remainder);
    }
}

/// AVX-512 binary operation for f64
///
/// Uses streaming (non-temporal) stores for arrays > 1MB when output is 64-byte aligned.
///
/// # Safety
/// - CPU must support AVX-512F
/// - All pointers must be valid for `len` elements
#[target_feature(enable = "avx512f")]
pub unsafe fn binary_f64(op: BinaryOp, a: *const f64, b: *const f64, out: *mut f64, len: usize) {
    let chunks = len / F64_LANES;
    let remainder = len % F64_LANES;

    // Use streaming stores for large aligned arrays
    let use_streaming = should_stream_f64(len) && is_aligned_avx512(out);

    // Atan2 has no SIMD implementation - use scalar fallback
    if op == BinaryOp::Atan2 {
        binary_scalar_f64(op, a, b, out, len);
        return;
    }

    if use_streaming {
        match op {
            BinaryOp::Add => binary_add_f64_stream(a, b, out, chunks),
            BinaryOp::Sub => binary_sub_f64_stream(a, b, out, chunks),
            BinaryOp::Mul => binary_mul_f64_stream(a, b, out, chunks),
            BinaryOp::Div => binary_div_f64_stream(a, b, out, chunks),
            BinaryOp::Max => binary_max_f64_stream(a, b, out, chunks),
            BinaryOp::Min => binary_min_f64_stream(a, b, out, chunks),
            BinaryOp::Pow => binary_pow_f64_stream(a, b, out, chunks),
            BinaryOp::Atan2 => unreachable!(), // Handled above
        }
        // Memory fence ensures streaming stores are globally visible
        _mm_sfence();
    } else {
        match op {
            BinaryOp::Add => binary_add_f64(a, b, out, chunks),
            BinaryOp::Sub => binary_sub_f64(a, b, out, chunks),
            BinaryOp::Mul => binary_mul_f64(a, b, out, chunks),
            BinaryOp::Div => binary_div_f64(a, b, out, chunks),
            BinaryOp::Max => binary_max_f64(a, b, out, chunks),
            BinaryOp::Min => binary_min_f64(a, b, out, chunks),
            BinaryOp::Pow => binary_pow_f64(a, b, out, chunks),
            BinaryOp::Atan2 => unreachable!(), // Handled above
        }
    }

    // Handle tail with scalar
    if remainder > 0 {
        let offset = chunks * F64_LANES;
        binary_scalar_f64(op, a.add(offset), b.add(offset), out.add(offset), remainder);
    }
}