primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Graph Batch Normalization - Batch normalization adapted for graph data (Pure WGSL)
//!
//! Normalizes node features across the batch and feature dimensions
//! Similar to standard batch norm, but operates on graph nodes
//!
//! **Deep Debt Principles**:
//! - Pure WGSL implementation (no CPU code)
//! - Safe Rust wrapper (no unsafe code)
//! - Hardware-agnostic via WebGPU
//! - Complete implementation (production-ready)

use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;

/// Graph Batch Normalization
pub struct GraphBatchNorm {
    input: Tensor,
    gamma: Tensor,
    beta: Tensor,
    num_nodes: usize,
    num_features: usize,
    epsilon: f32,
}

impl GraphBatchNorm {
    /// Create graph batch normalization.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn new(input: Tensor, gamma: Tensor, beta: Tensor, epsilon: f32) -> Result<Self> {
        let input_shape = input.shape();
        if input_shape.len() != 2 {
            return Err(BarracudaError::invalid_op(
                "graph_batch_norm",
                "input must be 2D [num_nodes, num_features]",
            ));
        }

        let num_nodes = input_shape[0];
        let num_features = input_shape[1];

        let gamma_shape = gamma.shape();
        let gamma_size = gamma_shape.iter().product::<usize>();
        if gamma_size != num_features {
            return Err(BarracudaError::invalid_op(
                "graph_batch_norm",
                "gamma must have num_features elements",
            ));
        }

        let beta_shape = beta.shape();
        let beta_size = beta_shape.iter().product::<usize>();
        if beta_size != num_features {
            return Err(BarracudaError::invalid_op(
                "graph_batch_norm",
                "beta must have num_features elements",
            ));
        }

        if epsilon <= 0.0 {
            return Err(BarracudaError::invalid_op(
                "graph_batch_norm",
                "epsilon must be positive",
            ));
        }

        Ok(Self {
            input,
            gamma,
            beta,
            num_nodes,
            num_features,
            epsilon,
        })
    }

    fn wgsl_shader() -> &'static str {
        {
            static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
                include_str!("../shaders/norm/graph_batch_norm_f64.wgsl").to_string()
            });
            std::sync::LazyLock::force(&SHADER).as_str()
        }
    }

    /// Execute graph batch normalization.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn execute(self) -> Result<Tensor> {
        let device = self.input.device();
        // Create intermediate buffers for mean and variance
        let mean_buffer = device.create_buffer_f32(self.num_features)?;
        let variance_buffer = device.create_buffer_f32(self.num_features)?;

        // Create output buffer
        let output_size = self.num_nodes * self.num_features;
        let output_buffer = device.create_buffer_f32(output_size)?;

        // Create uniform buffer
        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct Params {
            num_nodes: u32,
            num_features: u32,
            epsilon: f32,
            _pad1: u32,
        }

        let params = Params {
            num_nodes: self.num_nodes as u32,
            num_features: self.num_features as u32,
            epsilon: self.epsilon,
            _pad1: 0,
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("GraphBatchNorm Params"),
                contents: bytemuck::cast_slice(&[params]),
                usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            });

        // Compile shader
        let shader_module =
            device.compile_shader(Self::wgsl_shader(), Some("GraphBatchNorm Shader"));

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("GraphBatchNorm Bind Group Layout"),
                    entries: &[
                        wgpu::BindGroupLayoutEntry {
                            binding: 0,
                            visibility: wgpu::ShaderStages::COMPUTE,
                            ty: wgpu::BindingType::Buffer {
                                ty: wgpu::BufferBindingType::Uniform,
                                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: true },
                                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::Storage { read_only: true },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                        wgpu::BindGroupLayoutEntry {
                            binding: 3,
                            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: 4,
                            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: 5,
                            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: 6,
                            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,
                        },
                    ],
                });

        // Create bind group
        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("GraphBatchNorm Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: params_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.input.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: self.gamma.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: self.beta.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: mean_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 6,
                    resource: variance_buffer.as_entire_binding(),
                },
            ],
        });

        // Create pipeline layout
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("GraphBatchNorm Pipeline Layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        // Create encoder
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("GraphBatchNorm Encoder"),
        });

        // Step 1: Compute mean
        let mean_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("GraphBatchNorm Mean Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("compute_mean"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("GraphBatchNorm Mean Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&mean_pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            // Deep Debt Evolution: Capability-based dispatch
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
            let workgroups = (self.num_features as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        // Step 2: Compute variance
        let variance_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("GraphBatchNorm Variance Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("compute_variance"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("GraphBatchNorm Variance Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&variance_pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            // Deep Debt Evolution: Capability-based dispatch
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
            let workgroups = (self.num_features as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        // Step 3: Normalize, scale, and shift
        let normalize_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("GraphBatchNorm Normalize Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("normalize"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("GraphBatchNorm Normalize Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&normalize_pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);
            // Deep Debt Evolution: Capability-based dispatch
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
            let workgroups = (output_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,
            vec![self.num_nodes, self.num_features],
            device.clone(),
        ))
    }
}

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

    #[tokio::test]
    async fn test_graph_batch_norm_basic() {
        let device = crate::device::test_pool::get_test_device().await;
        let num_nodes = 4;
        let num_features = 8;

        let input = Tensor::from_vec_on(
            vec![1.0; num_nodes * num_features],
            vec![num_nodes, num_features],
            device.clone(),
        )
        .await
        .unwrap();

        let gamma =
            Tensor::from_vec_on(vec![1.0; num_features], vec![num_features], device.clone())
                .await
                .unwrap();

        let beta = Tensor::from_vec_on(vec![0.0; num_features], vec![num_features], device.clone())
            .await
            .unwrap();

        let batch_norm = GraphBatchNorm::new(input, gamma, beta, 1e-5).unwrap();
        let output = batch_norm.execute().unwrap();

        assert_eq!(output.shape(), &[num_nodes, num_features]);
    }

    #[tokio::test]
    async fn test_graph_batch_norm_large_batch() {
        let device = crate::device::test_pool::get_test_device().await;
        let num_nodes = 100;
        let num_features = 128;

        let input = Tensor::from_vec_on(
            vec![1.0; num_nodes * num_features],
            vec![num_nodes, num_features],
            device.clone(),
        )
        .await
        .unwrap();

        let gamma =
            Tensor::from_vec_on(vec![1.0; num_features], vec![num_features], device.clone())
                .await
                .unwrap();

        let beta = Tensor::from_vec_on(vec![0.0; num_features], vec![num_features], device.clone())
            .await
            .unwrap();

        let batch_norm = GraphBatchNorm::new(input, gamma, beta, 1e-5).unwrap();
        let output = batch_norm.execute().unwrap();

        assert_eq!(output.shape(), &[num_nodes, num_features]);
    }
}