ldsc 0.3.1

LD Score Regression — fast Rust reimplementation of Bulik-Sullivan et al. LDSC
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

#![cfg(feature = "gpu")]

use anyhow::{Result, anyhow};
use cubecl::client::ComputeClient;
use cubecl::cuda::{CudaDevice, CudaRuntime};
use cubecl::prelude::*;
use cubecl::std::tensor::TensorHandle;
use cubek_matmul::definition::MatmulElems;
use cubek_matmul::launch::{MatmulInputHandleRef, Strategy};

type R = CudaRuntime;

/// GPU context holding a CUDA device and compute client.
pub struct GpuContext {
    client: ComputeClient<R>,
    strategy: Strategy,
}

/// Extract a contiguous row-major (m × n) result from a potentially pitched GPU output buffer.
fn extract_result(raw: &[f32], m: usize, n: usize, row_stride: usize) -> Vec<f32> {
    let mut out = Vec::with_capacity(m * n);
    for i in 0..m {
        let start = i * row_stride;
        out.extend_from_slice(&raw[start..start + n]);
    }
    out
}

impl GpuContext {
    /// Create a new GPU context on CUDA device 0.
    pub fn new() -> Result<Self> {
        let device = CudaDevice::new(0);
        let client = R::client(&device);
        Ok(Self {
            client,
            strategy: Strategy::Auto,
        })
    }

    /// Perform A^T × B on GPU.
    ///
    /// `a_data` is column-major f32 (n_rows × a_cols), `b_data` is column-major f32 (n_rows × b_cols).
    /// Returns the result matrix (a_cols × b_cols) as a flat row-major f32 vec.
    pub fn matmul_tn(
        &self,
        a_data: &[f32],
        n_rows: usize,
        a_cols: usize,
        b_data: &[f32],
        b_cols: usize,
    ) -> Result<Vec<f32>> {
        let m = a_cols;
        let k = n_rows;
        let n = b_cols;
        let f32_storage = f32::as_type_native_unchecked();

        // Upload A and B to GPU
        let a_handle = self.client.create_from_slice(f32::as_bytes(a_data));
        let b_handle = self.client.create_from_slice(f32::as_bytes(b_data));

        // A is column-major (k × m): element (i,j) at offset j*k + i.
        // For A^T (m × k): element (j,i) maps to j*k + i, so shape [m,k], strides [k,1].
        let a_tensor = TensorHandle::<R>::new(a_handle, vec![m, k], vec![k, 1], f32_storage);

        // B is column-major (k × n): element (i,j) at offset j*k + i, strides [1, k].
        let b_tensor = TensorHandle::<R>::new(b_handle, vec![k, n], vec![1, k], f32_storage);

        // Output: (m × n) — may be pitched/padded by the runtime
        let out_tensor = TensorHandle::<R>::empty(&self.client, vec![m, n], f32_storage);
        let out_row_stride = out_tensor.strides()[0];

        let mut dtypes = MatmulElems::from_single_dtype(f32_storage);

        cubek_matmul::launch::launch_ref::<R>(
            &self.strategy,
            &self.client,
            &MatmulInputHandleRef::new(a_tensor.as_ref(), f32_storage),
            &MatmulInputHandleRef::new(b_tensor.as_ref(), f32_storage),
            &out_tensor.as_ref(),
            &mut dtypes,
        )
        .map_err(|e| anyhow!("GPU matmul failed: {:?}", e))?;

        // Read result back, accounting for potential row padding
        let out_bytes = self.client.read_one(out_tensor.handle);
        let raw: &[f32] = f32::from_bytes(&out_bytes);
        Ok(extract_result(raw, m, n, out_row_stride))
    }

    /// Perform A^T × B on GPU with tiled A for large windows that don't fit VRAM.
    #[allow(dead_code)]
    pub fn matmul_tn_tiled(
        &self,
        a_data: &[f32],
        n_rows: usize,
        a_cols: usize,
        b_data: &[f32],
        b_cols: usize,
        tile_cols: usize,
    ) -> Result<Vec<f32>> {
        let n = b_cols;
        let f32_storage = f32::as_type_native_unchecked();
        let mut result = vec![0.0f32; a_cols * n];

        for tile_start in (0..a_cols).step_by(tile_cols) {
            let tile_end = (tile_start + tile_cols).min(a_cols);
            let tile_m = tile_end - tile_start;

            // Extract tile columns from column-major A
            let tile_data: Vec<f32> = (tile_start..tile_end)
                .flat_map(|col| {
                    let offset = col * n_rows;
                    a_data[offset..offset + n_rows].iter().copied()
                })
                .collect();

            // Re-upload B each tile (handle is consumed by read_one)
            let b_handle = self.client.create_from_slice(f32::as_bytes(b_data));
            let b_tensor =
                TensorHandle::<R>::new(b_handle, vec![n_rows, n], vec![1, n_rows], f32_storage);

            let a_handle = self.client.create_from_slice(f32::as_bytes(&tile_data));
            let a_tensor = TensorHandle::<R>::new(
                a_handle,
                vec![tile_m, n_rows],
                vec![n_rows, 1],
                f32_storage,
            );

            let out_tensor = TensorHandle::<R>::empty(&self.client, vec![tile_m, n], f32_storage);
            let out_row_stride = out_tensor.strides()[0];

            let mut dtypes = MatmulElems::from_single_dtype(f32_storage);

            cubek_matmul::launch::launch_ref::<R>(
                &self.strategy,
                &self.client,
                &MatmulInputHandleRef::new(a_tensor.as_ref(), f32_storage),
                &MatmulInputHandleRef::new(b_tensor.as_ref(), f32_storage),
                &out_tensor.as_ref(),
                &mut dtypes,
            )
            .map_err(|e| anyhow!("GPU matmul tile failed: {:?}", e))?;

            let tile_bytes = self.client.read_one(out_tensor.handle);
            let raw: &[f32] = f32::from_bytes(&tile_bytes);

            for i in 0..tile_m {
                for j in 0..n {
                    result[(tile_start + i) * n + j] = raw[i * out_row_stride + j];
                }
            }
        }

        Ok(result)
    }

    /// Perform A^T × B on GPU using flex32 precision (f32 upload, f16 in GPU shared memory,
    /// f32 accumulation and output). Zero CPU-side conversion overhead.
    pub fn matmul_tn_flex32(
        &self,
        a_data: &[f32],
        n_rows: usize,
        a_cols: usize,
        b_data: &[f32],
        b_cols: usize,
    ) -> Result<Vec<f32>> {
        let m = a_cols;
        let k = n_rows;
        let n = b_cols;
        let f32_storage = f32::as_type_native_unchecked();

        // Upload f32 directly — GPU converts to f16 in shared memory
        let a_handle = self.client.create_from_slice(f32::as_bytes(a_data));
        let b_handle = self.client.create_from_slice(f32::as_bytes(b_data));

        let a_tensor = TensorHandle::<R>::new(a_handle, vec![m, k], vec![k, 1], f32_storage);
        let b_tensor = TensorHandle::<R>::new(b_handle, vec![k, n], vec![1, k], f32_storage);

        let out_tensor = TensorHandle::<R>::empty(&self.client, vec![m, n], f32_storage);
        let out_row_stride = out_tensor.strides()[0];

        // flex32: global f32, stage/register f16, accumulator f32
        let mut dtypes = MatmulElems::new_deprecated::<flex32>();

        cubek_matmul::launch::launch_ref::<R>(
            &self.strategy,
            &self.client,
            &MatmulInputHandleRef::new(a_tensor.as_ref(), f32_storage),
            &MatmulInputHandleRef::new(b_tensor.as_ref(), f32_storage),
            &out_tensor.as_ref(),
            &mut dtypes,
        )
        .map_err(|e| anyhow!("GPU flex32 matmul failed: {:?}", e))?;

        let out_bytes = self.client.read_one(out_tensor.handle);
        let raw: &[f32] = f32::from_bytes(&out_bytes);
        Ok(extract_result(raw, m, n, out_row_stride))
    }

    /// Perform A^T × B on GPU using flex32 precision with tiled A.
    #[allow(dead_code)]
    pub fn matmul_tn_tiled_flex32(
        &self,
        a_data: &[f32],
        n_rows: usize,
        a_cols: usize,
        b_data: &[f32],
        b_cols: usize,
        tile_cols: usize,
    ) -> Result<Vec<f32>> {
        let n = b_cols;
        let f32_storage = f32::as_type_native_unchecked();
        let mut result = vec![0.0f32; a_cols * n];

        for tile_start in (0..a_cols).step_by(tile_cols) {
            let tile_end = (tile_start + tile_cols).min(a_cols);
            let tile_m = tile_end - tile_start;

            let tile_data: Vec<f32> = (tile_start..tile_end)
                .flat_map(|col| {
                    let offset = col * n_rows;
                    a_data[offset..offset + n_rows].iter().copied()
                })
                .collect();

            let b_handle = self.client.create_from_slice(f32::as_bytes(b_data));
            let b_tensor =
                TensorHandle::<R>::new(b_handle, vec![n_rows, n], vec![1, n_rows], f32_storage);

            let a_handle = self.client.create_from_slice(f32::as_bytes(&tile_data));
            let a_tensor = TensorHandle::<R>::new(
                a_handle,
                vec![tile_m, n_rows],
                vec![n_rows, 1],
                f32_storage,
            );

            let out_tensor = TensorHandle::<R>::empty(&self.client, vec![tile_m, n], f32_storage);
            let out_row_stride = out_tensor.strides()[0];

            let mut dtypes = MatmulElems::new_deprecated::<flex32>();

            cubek_matmul::launch::launch_ref::<R>(
                &self.strategy,
                &self.client,
                &MatmulInputHandleRef::new(a_tensor.as_ref(), f32_storage),
                &MatmulInputHandleRef::new(b_tensor.as_ref(), f32_storage),
                &out_tensor.as_ref(),
                &mut dtypes,
            )
            .map_err(|e| anyhow!("GPU flex32 matmul tile failed: {:?}", e))?;

            let tile_bytes = self.client.read_one(out_tensor.handle);
            let raw: &[f32] = f32::from_bytes(&tile_bytes);

            for i in 0..tile_m {
                for j in 0..n {
                    result[(tile_start + i) * n + j] = raw[i * out_row_stride + j];
                }
            }
        }

        Ok(result)
    }
}