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
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

//! SIMD-accelerated f16/bf16 ↔ f32 conversion utilities
//!
//! These are the building blocks for the block-convert-compute pattern:
//! convert half-precision data to f32 in L1-sized blocks, run existing
//! f32 SIMD kernels, then convert back.
//!
//! # Conversion strategies
//!
//! - **x86 f16**: F16C instructions (`_mm256_cvtph_ps` / `_mm256_cvtps_ph`)
//! - **x86 bf16**: SIMD integer bit-shift (`u32 << 16` for load, rounded `>> 16` for store)
//! - **ARM f16**: NEON `vcvt_f32_f16` / `vcvt_f16_f32`
//! - **ARM bf16**: NEON integer bit-shift
//! - **Fallback**: `half` crate scalar conversion

#[cfg(target_arch = "aarch64")]
mod aarch64;
#[cfg(target_arch = "x86_64")]
mod x86_64;

/// Block size for stack-allocated conversion buffers.
/// 256 f32s = 1024 bytes, fits comfortably in L1 cache.
pub const HALF_BLOCK: usize = 256;

// ---------------------------------------------------------------------------
// Public API
// ---------------------------------------------------------------------------

/// Convert f16 values to f32 using SIMD when available.
///
/// # Safety
/// - `src` must be valid for reads of `len` u16 values (f16 bit patterns)
/// - `dst` must be valid for writes of `len` f32 values
#[inline]
pub unsafe fn convert_f16_to_f32(src: *const u16, dst: *mut f32, len: usize) {
    if len == 0 {
        return;
    }

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("f16c") {
            return x86_64::convert_f16_to_f32_f16c(src, dst, len);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        return aarch64::convert_f16_to_f32_neon(src, dst, len);
    }

    #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
    convert_f16_to_f32_scalar(src, dst, len);
}

/// Convert f32 values to f16 using SIMD when available.
///
/// # Safety
/// - `src` must be valid for reads of `len` f32 values
/// - `dst` must be valid for writes of `len` u16 values (f16 bit patterns)
#[inline]
pub unsafe fn convert_f32_to_f16(src: *const f32, dst: *mut u16, len: usize) {
    if len == 0 {
        return;
    }

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("f16c") {
            return x86_64::convert_f32_to_f16_f16c(src, dst, len);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        return aarch64::convert_f32_to_f16_neon(src, dst, len);
    }

    #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
    convert_f32_to_f16_scalar(src, dst, len);
}

/// Convert bf16 values to f32 using SIMD when available.
///
/// # Safety
/// - `src` must be valid for reads of `len` u16 values (bf16 bit patterns)
/// - `dst` must be valid for writes of `len` f32 values
#[inline]
pub unsafe fn convert_bf16_to_f32(src: *const u16, dst: *mut f32, len: usize) {
    if len == 0 {
        return;
    }

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("avx2") {
            return x86_64::convert_bf16_to_f32_avx2(src, dst, len);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        return aarch64::convert_bf16_to_f32_neon(src, dst, len);
    }

    #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
    convert_bf16_to_f32_scalar(src, dst, len);
}

/// Convert f32 values to bf16 using SIMD when available (with round-to-nearest-even).
///
/// # Safety
/// - `src` must be valid for reads of `len` f32 values
/// - `dst` must be valid for writes of `len` u16 values (bf16 bit patterns)
#[inline]
pub unsafe fn convert_f32_to_bf16(src: *const f32, dst: *mut u16, len: usize) {
    if len == 0 {
        return;
    }

    #[cfg(target_arch = "x86_64")]
    {
        if is_x86_feature_detected!("avx2") {
            return x86_64::convert_f32_to_bf16_avx2(src, dst, len);
        }
    }

    #[cfg(target_arch = "aarch64")]
    {
        return aarch64::convert_f32_to_bf16_neon(src, dst, len);
    }

    #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
    convert_f32_to_bf16_scalar(src, dst, len);
}

// ---------------------------------------------------------------------------
// Scalar fallbacks
// ---------------------------------------------------------------------------

#[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
#[inline]
unsafe fn convert_f16_to_f32_scalar(src: *const u16, dst: *mut f32, len: usize) {
    for i in 0..len {
        *dst.add(i) = half::f16::from_bits(*src.add(i)).to_f32();
    }
}

#[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
#[inline]
unsafe fn convert_f32_to_f16_scalar(src: *const f32, dst: *mut u16, len: usize) {
    for i in 0..len {
        *dst.add(i) = half::f16::from_f32(*src.add(i)).to_bits();
    }
}

#[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
#[inline]
unsafe fn convert_bf16_to_f32_scalar(src: *const u16, dst: *mut f32, len: usize) {
    for i in 0..len {
        *dst.add(i) = half::bf16::from_bits(*src.add(i)).to_f32();
    }
}

#[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
#[inline]
unsafe fn convert_f32_to_bf16_scalar(src: *const f32, dst: *mut u16, len: usize) {
    for i in 0..len {
        let bits = (*src.add(i)).to_bits();
        let rounded = bits.wrapping_add(0x7FFF + ((bits >> 16) & 1));
        *dst.add(i) = (rounded >> 16) as u16;
    }
}

// ---------------------------------------------------------------------------
// Tests
// ---------------------------------------------------------------------------

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_f16_roundtrip() {
        let values: Vec<f32> = vec![
            0.0,
            1.0,
            -1.0,
            0.5,
            -0.5,
            65504.0,
            -65504.0,
            0.000061035156,
            3.15,
        ];
        let f16_bits: Vec<u16> = values
            .iter()
            .map(|&v| half::f16::from_f32(v).to_bits())
            .collect();
        let mut f32_out = vec![0.0f32; values.len()];
        let mut f16_out = vec![0u16; values.len()];

        unsafe {
            convert_f16_to_f32(f16_bits.as_ptr(), f32_out.as_mut_ptr(), values.len());
            convert_f32_to_f16(f32_out.as_ptr(), f16_out.as_mut_ptr(), f32_out.len());
        }

        for i in 0..values.len() {
            assert_eq!(
                f16_bits[i], f16_out[i],
                "f16 roundtrip failed at index {}: input bits {:04x}, output bits {:04x}",
                i, f16_bits[i], f16_out[i]
            );
        }
    }

    #[test]
    fn test_bf16_roundtrip() {
        let values: Vec<f32> = vec![0.0, 1.0, -1.0, 0.5, -0.5, 100.0, -100.0, 3.15];
        let bf16_bits: Vec<u16> = values
            .iter()
            .map(|&v| half::bf16::from_f32(v).to_bits())
            .collect();
        let mut f32_out = vec![0.0f32; values.len()];
        let mut bf16_out = vec![0u16; values.len()];

        unsafe {
            convert_bf16_to_f32(bf16_bits.as_ptr(), f32_out.as_mut_ptr(), values.len());
            convert_f32_to_bf16(f32_out.as_ptr(), bf16_out.as_mut_ptr(), f32_out.len());
        }

        for i in 0..values.len() {
            assert_eq!(
                bf16_bits[i], bf16_out[i],
                "bf16 roundtrip failed at index {}: input bits {:04x}, output bits {:04x}",
                i, bf16_bits[i], bf16_out[i]
            );
        }
    }

    #[test]
    fn test_f16_conversion_accuracy() {
        let f16_bits: Vec<u16> = (0..512)
            .map(|i| half::f16::from_f32((i as f32 - 256.0) * 0.1).to_bits())
            .collect();
        let mut f32_out = vec![0.0f32; f16_bits.len()];
        unsafe { convert_f16_to_f32(f16_bits.as_ptr(), f32_out.as_mut_ptr(), f16_bits.len()) }

        for i in 0..f16_bits.len() {
            let expected = half::f16::from_bits(f16_bits[i]).to_f32();
            assert_eq!(f32_out[i], expected, "f16→f32 mismatch at index {}", i);
        }
    }

    #[test]
    fn test_bf16_conversion_accuracy() {
        let bf16_bits: Vec<u16> = (0..512)
            .map(|i| half::bf16::from_f32((i as f32 - 256.0) * 0.1).to_bits())
            .collect();
        let mut f32_out = vec![0.0f32; bf16_bits.len()];
        unsafe { convert_bf16_to_f32(bf16_bits.as_ptr(), f32_out.as_mut_ptr(), bf16_bits.len()) }

        for i in 0..bf16_bits.len() {
            let expected = half::bf16::from_bits(bf16_bits[i]).to_f32();
            assert_eq!(f32_out[i], expected, "bf16→f32 mismatch at index {}", i);
        }
    }

    #[test]
    fn test_empty_conversion() {
        unsafe {
            convert_f16_to_f32(std::ptr::null(), std::ptr::null_mut(), 0);
            convert_f32_to_f16(std::ptr::null(), std::ptr::null_mut(), 0);
            convert_bf16_to_f32(std::ptr::null(), std::ptr::null_mut(), 0);
            convert_f32_to_bf16(std::ptr::null(), std::ptr::null_mut(), 0);
        }
    }

    #[test]
    fn test_unaligned_lengths() {
        for len in [1, 3, 5, 7, 9, 15, 17, 31, 33] {
            let f16_bits: Vec<u16> = (0..len)
                .map(|i| half::f16::from_f32(i as f32).to_bits())
                .collect();
            let mut f32_out = vec![0.0f32; len];

            unsafe { convert_f16_to_f32(f16_bits.as_ptr(), f32_out.as_mut_ptr(), len) }

            for i in 0..len {
                let expected = half::f16::from_bits(f16_bits[i]).to_f32();
                assert_eq!(f32_out[i], expected, "mismatch at len={}, index={}", len, i);
            }
        }
    }
}