treeboost 0.1.0

High-performance Gradient Boosted Decision Tree engine for large-scale tabular data
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

//! SIMD-optimized histogram merge operations
//!
//! # Key Optimization
//!
//! When reducing partial histograms from parallel blocks, we need to merge
//! 256 bins worth of (gradient, hessian, count) tuples. With AVX2, we can
//! process 8 floats / 8 u32s at a time, giving ~8x speedup over scalar.
//!
//! # Memory Layout
//!
//! Histograms use Array-of-Structs: [BinEntry; 256] where BinEntry is:
//!   - sum_gradients: f32
//!   - sum_hessians: f32
//!   - count: u32
//!
//! Total: 12 bytes per bin, 3072 bytes per histogram.

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

/// SIMD merge of two SoA histogram arrays (gradients)
///
/// Adds `other_grads` into `self_grads` using AVX2.
/// Processes 8 floats (32 bytes) at a time.
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn merge_histogram_grads_avx2(self_grads: &mut [f32; 256], other_grads: &[f32; 256]) {
    // 256 floats = 32 AVX2 iterations (8 floats each)
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_ps(self_grads.as_ptr().add(i));
        let other_vec = _mm256_loadu_ps(other_grads.as_ptr().add(i));
        let sum = _mm256_add_ps(self_vec, other_vec);
        _mm256_storeu_ps(self_grads.as_mut_ptr().add(i), sum);
    }
}

/// SIMD merge of two SoA histogram arrays (hessians)
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn merge_histogram_hess_avx2(self_hess: &mut [f32; 256], other_hess: &[f32; 256]) {
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_ps(self_hess.as_ptr().add(i));
        let other_vec = _mm256_loadu_ps(other_hess.as_ptr().add(i));
        let sum = _mm256_add_ps(self_vec, other_vec);
        _mm256_storeu_ps(self_hess.as_mut_ptr().add(i), sum);
    }
}

/// SIMD merge of two SoA histogram arrays (counts)
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn merge_histogram_counts_avx2(self_counts: &mut [u32; 256], other_counts: &[u32; 256]) {
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_si256(self_counts.as_ptr().add(i) as *const __m256i);
        let other_vec = _mm256_loadu_si256(other_counts.as_ptr().add(i) as *const __m256i);
        let sum = _mm256_add_epi32(self_vec, other_vec);
        _mm256_storeu_si256(self_counts.as_mut_ptr().add(i) as *mut __m256i, sum);
    }
}

/// SIMD subtract of two SoA histogram arrays (gradients)
///
/// Subtracts `other_grads` from `self_grads` using AVX2.
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn subtract_histogram_grads_avx2(self_grads: &mut [f32; 256], other_grads: &[f32; 256]) {
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_ps(self_grads.as_ptr().add(i));
        let other_vec = _mm256_loadu_ps(other_grads.as_ptr().add(i));
        let diff = _mm256_sub_ps(self_vec, other_vec);
        _mm256_storeu_ps(self_grads.as_mut_ptr().add(i), diff);
    }
}

/// SIMD subtract of two SoA histogram arrays (hessians)
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn subtract_histogram_hess_avx2(self_hess: &mut [f32; 256], other_hess: &[f32; 256]) {
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_ps(self_hess.as_ptr().add(i));
        let other_vec = _mm256_loadu_ps(other_hess.as_ptr().add(i));
        let diff = _mm256_sub_ps(self_vec, other_vec);
        _mm256_storeu_ps(self_hess.as_mut_ptr().add(i), diff);
    }
}

/// SIMD subtract of two SoA histogram arrays (counts)
///
/// # Safety
/// - Requires AVX2 support
/// - Arrays must be exactly 256 elements
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn subtract_histogram_counts_avx2(
    self_counts: &mut [u32; 256],
    other_counts: &[u32; 256],
) {
    for i in (0..256).step_by(8) {
        let self_vec = _mm256_loadu_si256(self_counts.as_ptr().add(i) as *const __m256i);
        let other_vec = _mm256_loadu_si256(other_counts.as_ptr().add(i) as *const __m256i);
        let diff = _mm256_sub_epi32(self_vec, other_vec);
        _mm256_storeu_si256(self_counts.as_mut_ptr().add(i) as *mut __m256i, diff);
    }
}

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

    #[test]
    fn test_merge_grads_avx2() {
        if !std::arch::is_x86_feature_detected!("avx2") {
            println!("AVX2 not available, skipping test");
            return;
        }

        let mut self_grads = [0.0f32; 256];
        let mut other_grads = [0.0f32; 256];

        // Set up test data
        for i in 0..256 {
            self_grads[i] = i as f32;
            other_grads[i] = (256 - i) as f32;
        }

        unsafe {
            merge_histogram_grads_avx2(&mut self_grads, &other_grads);
        }

        // All should be 256
        for (i, &val) in self_grads.iter().enumerate() {
            assert!(
                (val - 256.0).abs() < 1e-5,
                "Mismatch at index {}: expected 256, got {}",
                i,
                val
            );
        }
    }

    #[test]
    fn test_merge_counts_avx2() {
        if !std::arch::is_x86_feature_detected!("avx2") {
            println!("AVX2 not available, skipping test");
            return;
        }

        let mut self_counts = [0u32; 256];
        let mut other_counts = [0u32; 256];

        for i in 0..256 {
            self_counts[i] = i as u32;
            other_counts[i] = 1;
        }

        unsafe {
            merge_histogram_counts_avx2(&mut self_counts, &other_counts);
        }

        for i in 0..256 {
            assert_eq!(self_counts[i], i as u32 + 1, "Mismatch at index {}", i);
        }
    }

    #[test]
    fn test_subtract_grads_avx2() {
        if !std::arch::is_x86_feature_detected!("avx2") {
            println!("AVX2 not available, skipping test");
            return;
        }

        let mut self_grads = [100.0f32; 256];
        let mut other_grads = [0.0f32; 256];

        for i in 0..256 {
            other_grads[i] = i as f32;
        }

        unsafe {
            subtract_histogram_grads_avx2(&mut self_grads, &other_grads);
        }

        for i in 0..256 {
            let expected = 100.0 - i as f32;
            assert!(
                (self_grads[i] - expected).abs() < 1e-5,
                "Mismatch at index {}: expected {}, got {}",
                i,
                expected,
                self_grads[i]
            );
        }
    }
}