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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
// SPDX-License-Identifier: AGPL-3.0-or-later
//! `NORM_CDF` - Normal distribution CDF and PDF - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Hardware-agnostic implementation
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: Production-ready, no mocks
//! - Hardware-agnostic: Pure WGSL for universal compute
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Normal distribution CDF Φ(x) and PDF φ(x)
pub struct NormCdf {
input: Tensor,
mu: f32,
sigma: f32,
compute_pdf: bool,
}
impl NormCdf {
/// Create standard normal CDF operation (μ=0, σ=1)
#[must_use]
pub fn standard_cdf(input: Tensor) -> Self {
Self {
input,
mu: 0.0,
sigma: 1.0,
compute_pdf: false,
}
}
/// Create standard normal PDF operation (μ=0, σ=1)
#[must_use]
pub fn standard_pdf(input: Tensor) -> Self {
Self {
input,
mu: 0.0,
sigma: 1.0,
compute_pdf: true,
}
}
/// Create general normal CDF operation with custom μ, σ
#[must_use]
pub fn cdf(input: Tensor, mu: f32, sigma: f32) -> Self {
Self {
input,
mu,
sigma,
compute_pdf: false,
}
}
/// Create general normal PDF operation with custom μ, σ
#[must_use]
pub fn pdf(input: Tensor, mu: f32, sigma: f32) -> Self {
Self {
input,
mu,
sigma,
compute_pdf: true,
}
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/special/norm_cdf_f64.wgsl").to_string()
});
&SHADER
}
/// Execute normal CDF or PDF on the input tensor.
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, the
/// device is lost, or compute submission fails.
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let size: usize = self.input.shape().iter().product();
let output_buffer = device.create_buffer_f32(size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
size: u32,
mu: f32,
sigma: f32,
mode: u32, // 0 = CDF, 1 = PDF
}
let params = Params {
size: size as u32,
mu: self.mu,
sigma: self.sigma,
mode: u32::from(self.compute_pdf),
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("NormCdf Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("NormCdf Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 1,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 2,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("NormCdf Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("NormCdf"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("NormCdf Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("NormCdf Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("NormCdf Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("NormCdf Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.input.shape().to_vec(),
device.clone(),
))
}
}
impl Tensor {
/// Compute standard normal CDF Φ(x) for each element
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, the
/// device is lost, or compute submission fails.
pub fn norm_cdf(self) -> Result<Self> {
NormCdf::standard_cdf(self).execute()
}
/// Compute standard normal PDF φ(x) for each element
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, the
/// device is lost, or compute submission fails.
pub fn norm_pdf(self) -> Result<Self> {
NormCdf::standard_pdf(self).execute()
}
/// Compute general normal CDF Φ(x; μ, σ) for each element
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, the
/// device is lost, or compute submission fails.
pub fn norm_cdf_params(self, mu: f32, sigma: f32) -> Result<Self> {
NormCdf::cdf(self, mu, sigma).execute()
}
/// Compute general normal PDF φ(x; μ, σ) for each element
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, the
/// device is lost, or compute submission fails.
pub fn norm_pdf_params(self, mu: f32, sigma: f32) -> Result<Self> {
NormCdf::pdf(self, mu, sigma).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_norm_cdf_zero() {
let device = crate::device::test_pool::get_test_device().await;
// Φ(0) = 0.5
let input = Tensor::new(vec![0.0], vec![1], device);
let output = input.norm_cdf().unwrap();
let result = output.to_vec().unwrap();
assert!(
(result[0] - 0.5).abs() < 0.001,
"Φ(0) = {}, expected 0.5",
result[0]
);
}
#[tokio::test]
async fn test_norm_cdf_critical() {
let device = crate::device::test_pool::get_test_device().await;
// Φ(-1.96) ≈ 0.025, Φ(1.96) ≈ 0.975
let input = Tensor::new(vec![-1.96, 1.96], vec![2], device);
let output = input.norm_cdf().unwrap();
let result = output.to_vec().unwrap();
assert!(
(result[0] - 0.025).abs() < 0.01,
"Φ(-1.96) = {}, expected ~0.025",
result[0]
);
assert!(
(result[1] - 0.975).abs() < 0.01,
"Φ(1.96) = {}, expected ~0.975",
result[1]
);
}
#[tokio::test]
async fn test_norm_pdf_peak() {
let device = crate::device::test_pool::get_test_device().await;
// φ(0) = 1/√(2π) ≈ 0.3989
let input = Tensor::new(vec![0.0], vec![1], device);
let output = input.norm_pdf().unwrap();
let result = output.to_vec().unwrap();
let expected = 1.0 / (2.0 * std::f32::consts::PI).sqrt();
assert!(
(result[0] - expected).abs() < 0.001,
"φ(0) = {}, expected {}",
result[0],
expected
);
}
#[tokio::test]
async fn test_norm_cdf_general() {
let device = crate::device::test_pool::get_test_device().await;
// Φ(μ; μ, σ) = 0.5 for any σ > 0
let input = Tensor::new(vec![5.0], vec![1], device);
let output = input.norm_cdf_params(5.0, 2.0).unwrap();
let result = output.to_vec().unwrap();
assert!(
(result[0] - 0.5).abs() < 0.001,
"Φ(5; 5, 2) = {}, expected 0.5",
result[0]
);
}
}