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

//! Xoshiro256++ PRNG WGSL shaders and launchers
//!
//! Rust `rand` crate default. Uses emulated 64-bit arithmetic.
//! Reference: Blackman & Vigna "Scrambled Linear Pseudorandom Number Generators"

use wgpu::{Buffer, Queue};

use super::check_float_dtype;
use crate::dtype::DType;
use crate::error::Result;
use crate::runtime::wgpu::shaders::pipeline::{LayoutKey, PipelineCache, workgroup_count};

const XOSHIRO256_UNIFORM_WGSL: &str = r#"
struct Xoshiro256Params {
    numel: u32,
    seed_lo: u32,
    seed_hi: u32,
    _pad0: u32,
}

@group(0) @binding(0) var<storage, read_write> output: array<f32>;
@group(0) @binding(1) var<uniform> params: Xoshiro256Params;

fn rotl64(lo: u32, hi: u32, k: u32) -> vec2<u32> {
    if (k == 0u) { return vec2<u32>(lo, hi); }
    if (k < 32u) {
        return vec2<u32>((lo << k) | (hi >> (32u - k)), (hi << k) | (lo >> (32u - k)));
    }
    let k2 = k - 32u;
    return vec2<u32>((hi << k2) | (lo >> (32u - k2)), (lo << k2) | (hi >> (32u - k2)));
}

fn add64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    let lo = a_lo + b_lo;
    return vec2<u32>(lo, a_hi + b_hi + select(0u, 1u, lo < a_lo));
}

fn xor64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    return vec2<u32>(a_lo ^ b_lo, a_hi ^ b_hi);
}

fn shr64(lo: u32, hi: u32, k: u32) -> vec2<u32> {
    if (k == 0u) { return vec2<u32>(lo, hi); }
    if (k < 32u) { return vec2<u32>((lo >> k) | (hi << (32u - k)), hi >> k); }
    return vec2<u32>(hi >> (k - 32u), 0u);
}

fn mul64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    let a0 = a_lo & 0xFFFFu; let a1 = a_lo >> 16u;
    let b0 = b_lo & 0xFFFFu; let b1 = b_lo >> 16u;
    let p00 = a0 * b0; let p01 = a0 * b1; let p10 = a1 * b0; let p11 = a1 * b1;
    let mid = p01 + p10 + (p00 >> 16u);
    let lo = (p00 & 0xFFFFu) | ((mid & 0xFFFFu) << 16u);
    let hi = p11 + (mid >> 16u) + a_lo * b_hi + a_hi * b_lo;
    return vec2<u32>(lo, hi);
}

fn splitmix64(state: ptr<function, vec2<u32>>) -> vec2<u32> {
    *state = add64((*state).x, (*state).y, 0x7F4A7C15u, 0x9E3779B9u);
    var z = *state;
    let s30 = shr64(z.x, z.y, 30u); z = xor64(z.x, z.y, s30.x, s30.y);
    z = mul64(z.x, z.y, 0x1CE4E5B9u, 0xBF58476Du);
    let s27 = shr64(z.x, z.y, 27u); z = xor64(z.x, z.y, s27.x, s27.y);
    z = mul64(z.x, z.y, 0x133111EBu, 0x94D049BBu);
    let s31 = shr64(z.x, z.y, 31u); z = xor64(z.x, z.y, s31.x, s31.y);
    return z;
}

@compute @workgroup_size(256)
fn xoshiro256_uniform_f32(@builtin(global_invocation_id) gid: vec3<u32>) {
    let idx = gid.x;
    if (idx >= params.numel) { return; }

    var sm_state = add64(params.seed_lo, params.seed_hi, idx, 0u);
    let s0 = splitmix64(&sm_state);
    let s1 = splitmix64(&sm_state);
    let s2 = splitmix64(&sm_state);
    let s3 = splitmix64(&sm_state);

    let sum03 = add64(s0.x, s0.y, s3.x, s3.y);
    let rotated = rotl64(sum03.x, sum03.y, 23u);
    let result = add64(rotated.x, rotated.y, s0.x, s0.y);

    output[idx] = f32(result.y) / 4294967296.0;
}
"#;

const XOSHIRO256_RANDN_WGSL: &str = r#"
const PI: f32 = 3.14159265359;

struct Xoshiro256Params {
    numel: u32,
    seed_lo: u32,
    seed_hi: u32,
    _pad0: u32,
}

@group(0) @binding(0) var<storage, read_write> output: array<f32>;
@group(0) @binding(1) var<uniform> params: Xoshiro256Params;

fn rotl64(lo: u32, hi: u32, k: u32) -> vec2<u32> {
    if (k == 0u) { return vec2<u32>(lo, hi); }
    if (k < 32u) {
        return vec2<u32>((lo << k) | (hi >> (32u - k)), (hi << k) | (lo >> (32u - k)));
    }
    let k2 = k - 32u;
    return vec2<u32>((hi << k2) | (lo >> (32u - k2)), (lo << k2) | (hi >> (32u - k2)));
}

fn add64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    let lo = a_lo + b_lo;
    return vec2<u32>(lo, a_hi + b_hi + select(0u, 1u, lo < a_lo));
}

fn xor64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    return vec2<u32>(a_lo ^ b_lo, a_hi ^ b_hi);
}

fn shl64(lo: u32, hi: u32, k: u32) -> vec2<u32> {
    if (k == 0u) { return vec2<u32>(lo, hi); }
    if (k < 32u) { return vec2<u32>(lo << k, (hi << k) | (lo >> (32u - k))); }
    return vec2<u32>(0u, lo << (k - 32u));
}

fn shr64(lo: u32, hi: u32, k: u32) -> vec2<u32> {
    if (k == 0u) { return vec2<u32>(lo, hi); }
    if (k < 32u) { return vec2<u32>((lo >> k) | (hi << (32u - k)), hi >> k); }
    return vec2<u32>(hi >> (k - 32u), 0u);
}

fn mul64(a_lo: u32, a_hi: u32, b_lo: u32, b_hi: u32) -> vec2<u32> {
    let a0 = a_lo & 0xFFFFu; let a1 = a_lo >> 16u;
    let b0 = b_lo & 0xFFFFu; let b1 = b_lo >> 16u;
    let p00 = a0 * b0; let p01 = a0 * b1; let p10 = a1 * b0; let p11 = a1 * b1;
    let mid = p01 + p10 + (p00 >> 16u);
    let lo = (p00 & 0xFFFFu) | ((mid & 0xFFFFu) << 16u);
    let hi = p11 + (mid >> 16u) + a_lo * b_hi + a_hi * b_lo;
    return vec2<u32>(lo, hi);
}

fn splitmix64(state: ptr<function, vec2<u32>>) -> vec2<u32> {
    *state = add64((*state).x, (*state).y, 0x7F4A7C15u, 0x9E3779B9u);
    var z = *state;
    let s30 = shr64(z.x, z.y, 30u); z = xor64(z.x, z.y, s30.x, s30.y);
    z = mul64(z.x, z.y, 0x1CE4E5B9u, 0xBF58476Du);
    let s27 = shr64(z.x, z.y, 27u); z = xor64(z.x, z.y, s27.x, s27.y);
    z = mul64(z.x, z.y, 0x133111EBu, 0x94D049BBu);
    let s31 = shr64(z.x, z.y, 31u); z = xor64(z.x, z.y, s31.x, s31.y);
    return z;
}

fn xoshiro_next(s: ptr<function, array<vec2<u32>, 4>>) -> u32 {
    let sum03 = add64((*s)[0].x, (*s)[0].y, (*s)[3].x, (*s)[3].y);
    let rotated = rotl64(sum03.x, sum03.y, 23u);
    let result = add64(rotated.x, rotated.y, (*s)[0].x, (*s)[0].y);

    let t = shl64((*s)[1].x, (*s)[1].y, 17u);
    (*s)[2] = xor64((*s)[2].x, (*s)[2].y, (*s)[0].x, (*s)[0].y);
    (*s)[3] = xor64((*s)[3].x, (*s)[3].y, (*s)[1].x, (*s)[1].y);
    (*s)[1] = xor64((*s)[1].x, (*s)[1].y, (*s)[2].x, (*s)[2].y);
    (*s)[0] = xor64((*s)[0].x, (*s)[0].y, (*s)[3].x, (*s)[3].y);
    (*s)[2] = xor64((*s)[2].x, (*s)[2].y, t.x, t.y);
    (*s)[3] = rotl64((*s)[3].x, (*s)[3].y, 45u);

    return result.y;
}

fn box_muller(u1: f32, u2: f32) -> vec2<f32> {
    let r = sqrt(-2.0 * log(max(u1, 1e-10)));
    let theta = 2.0 * PI * u2;
    return vec2<f32>(r * cos(theta), r * sin(theta));
}

@compute @workgroup_size(256)
fn xoshiro256_randn_f32(@builtin(global_invocation_id) gid: vec3<u32>) {
    let base_idx = gid.x * 2u;
    if (base_idx >= params.numel) { return; }

    var sm_state = add64(params.seed_lo, params.seed_hi, gid.x, 0u);
    var state: array<vec2<u32>, 4>;
    state[0] = splitmix64(&sm_state);
    state[1] = splitmix64(&sm_state);
    state[2] = splitmix64(&sm_state);
    state[3] = splitmix64(&sm_state);

    let r1 = xoshiro_next(&state);
    let r2 = xoshiro_next(&state);

    let u1 = f32(r1) / 4294967296.0;
    let u2 = f32(r2) / 4294967296.0;
    let normals = box_muller(u1, u2);

    if (base_idx < params.numel) { output[base_idx] = normals.x; }
    if (base_idx + 1u < params.numel) { output[base_idx + 1u] = normals.y; }
}
"#;

/// Launch Xoshiro256++ uniform kernel
pub fn launch_xoshiro256_uniform(
    cache: &PipelineCache,
    queue: &Queue,
    out: &Buffer,
    params: &Buffer,
    numel: usize,
    dtype: DType,
) -> Result<()> {
    if numel == 0 {
        return Ok(());
    }
    check_float_dtype(dtype, "xoshiro256_uniform")?;

    let name = "xoshiro256_uniform_f32";
    let module = cache.get_or_create_module(name, XOSHIRO256_UNIFORM_WGSL);
    let layout = cache.get_or_create_layout(LayoutKey {
        num_storage_buffers: 1,
        num_uniform_buffers: 1,
        num_readonly_storage: 0,
    });
    let pipeline = cache.get_or_create_pipeline(name, name, &module, &layout);
    let bind_group = cache.create_bind_group(&layout, &[out, params]);

    let mut encoder = cache
        .device()
        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("xoshiro256_uniform"),
        });
    {
        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("xoshiro256_uniform"),
            timestamp_writes: None,
        });
        pass.set_pipeline(&pipeline);
        pass.set_bind_group(0, Some(&bind_group), &[]);
        pass.dispatch_workgroups(workgroup_count(numel), 1, 1);
    }
    queue.submit(std::iter::once(encoder.finish()));
    Ok(())
}

/// Launch Xoshiro256++ randn kernel (Box-Muller transform)
pub fn launch_xoshiro256_randn(
    cache: &PipelineCache,
    queue: &Queue,
    out: &Buffer,
    params: &Buffer,
    numel: usize,
    dtype: DType,
) -> Result<()> {
    if numel == 0 {
        return Ok(());
    }
    check_float_dtype(dtype, "xoshiro256_randn")?;

    let name = "xoshiro256_randn_f32";
    let module = cache.get_or_create_module(name, XOSHIRO256_RANDN_WGSL);
    let layout = cache.get_or_create_layout(LayoutKey {
        num_storage_buffers: 1,
        num_uniform_buffers: 1,
        num_readonly_storage: 0,
    });
    let pipeline = cache.get_or_create_pipeline(name, name, &module, &layout);
    let bind_group = cache.create_bind_group(&layout, &[out, params]);

    let mut encoder = cache
        .device()
        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("xoshiro256_randn"),
        });
    {
        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("xoshiro256_randn"),
            timestamp_writes: None,
        });
        pass.set_pipeline(&pipeline);
        pass.set_bind_group(0, Some(&bind_group), &[]);
        pass.dispatch_workgroups(workgroup_count((numel + 1) / 2), 1, 1);
    }
    queue.submit(std::iter::once(encoder.finish()));
    Ok(())
}