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
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Mosaic - Mosaic augmentation (YOLO-style)
//!
//! Combines 4 images into one mosaic.
//! Used in object detection for multi-scale training.
//!
//! Deep Debt Principles:
//! - Pure GPU/WGSL execution
//! - Safe Rust wrappers
//! - Hardware-agnostic via WebGPU
//! - Runtime device discovery
//! - Zero CPU fallbacks in execution

use crate::device::DeviceCapabilities;
use crate::error::Result;
use crate::tensor::Tensor;

const SHADER_F64: &str = include_str!("../shaders/augmentation/mosaic_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
    std::sync::LazyLock::new(|| SHADER_F64.to_string());

/// Mosaic operation
pub struct Mosaic {
    images: [Tensor; 4],
    seed: u64,
}

impl Mosaic {
    /// Create a new mosaic operation
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn new(images: [Tensor; 4], seed: u64) -> Result<Self> {
        // Validate all images have same shape
        let shape = images[0].shape();
        for (i, img) in images.iter().enumerate() {
            if img.shape() != shape {
                return Err(crate::error::BarracudaError::invalid_op(
                    "Mosaic",
                    format!("All images must have same shape, image {i} differs"),
                ));
            }
        }

        if shape.len() != 3 {
            return Err(crate::error::BarracudaError::invalid_op(
                "Mosaic",
                format!("Expected 3D tensor (C, H, W), got {}D", shape.len()),
            ));
        }

        Ok(Self { images, seed })
    }

    /// Get the WGSL shader source
    fn wgsl_shader() -> &'static str {
        &SHADER_F32
    }

    /// Execute the mosaic operation
    ///
    /// # 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.images[0].device();
        let shape = self.images[0].shape();

        let channels = shape[0];
        let height = shape[1];
        let width = shape[2];

        // Compute random split point from seed
        let split_x = ((self.seed * 1_103_515_245) % width as u64) as usize;
        let split_y = ((self.seed * 22_695_477) % height as u64) as usize;

        let output_size = channels * height * width;

        // Create buffers
        let image0_buffer = self.images[0].buffer();
        let image1_buffer = self.images[1].buffer();
        let image2_buffer = self.images[2].buffer();
        let image3_buffer = self.images[3].buffer();

        let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("Mosaic Output"),
            size: (output_size * std::mem::size_of::<f32>()) as u64,
            usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
            mapped_at_creation: false,
        });

        // Create uniform buffer for parameters
        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct Params {
            channels: u32,
            height: u32,
            width: u32,
            split_x: u32,
            split_y: u32,
        }

        let params = Params {
            channels: channels as u32,
            height: height as u32,
            width: width as u32,
            split_x: split_x as u32,
            split_y: split_y as u32,
        };

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

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("Mosaic 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: 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::Uniform,
                                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("Mosaic Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: image0_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: image1_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: image2_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: image3_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

        // Create compute pipeline
        let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Mosaic Shader"));

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

        let compute_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("Mosaic Pipeline"),
                    layout: Some(&pipeline_layout),
                    module: &shader_module,
                    entry_point: Some("main"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        // Execute compute shader
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("Mosaic Encoder"),
        });

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("Mosaic Pass"),
                timestamp_writes: None,
            });
            compute_pass.set_pipeline(&compute_pipeline);
            compute_pass.set_bind_group(0, Some(&bind_group), &[]);

            // Dispatch using standard 2D shader workgroup size (16, 16)
            let caps = DeviceCapabilities::from_device(device);
            let (workgroups_x, workgroups_y) = caps.dispatch_2d(width as u32, height as u32);
            compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
        }

        device.submit_commands(Some(encoder.finish()));

        // Read back results
        let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;

        Ok(Tensor::new(
            output_data,
            vec![channels, height, width],
            device.clone(),
        ))
    }
}

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

    #[tokio::test]
    async fn test_mosaic_basic() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let t1 = Tensor::from_vec_on(vec![1.0; 3 * 640 * 640], vec![3, 640, 640], device.clone())
            .await
            .unwrap();
        let t2 = Tensor::from_vec_on(vec![0.8; 3 * 640 * 640], vec![3, 640, 640], device.clone())
            .await
            .unwrap();
        let t3 = Tensor::from_vec_on(vec![0.6; 3 * 640 * 640], vec![3, 640, 640], device.clone())
            .await
            .unwrap();
        let t4 = Tensor::from_vec_on(vec![0.4; 3 * 640 * 640], vec![3, 640, 640], device)
            .await
            .unwrap();
        let mosaic_tensor = Mosaic::new([t1, t2, t3, t4], 77_777)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic_img = mosaic_tensor.to_vec().unwrap();
        assert_eq!(mosaic_img.len(), 3 * 640 * 640);
        assert!(mosaic_img.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_mosaic_edge_cases() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        // Small images
        let t1 = Tensor::from_vec_on(vec![1.0; 3 * 32 * 32], vec![3, 32, 32], device.clone())
            .await
            .unwrap();
        let t2 = Tensor::from_vec_on(vec![2.0; 3 * 32 * 32], vec![3, 32, 32], device.clone())
            .await
            .unwrap();
        let t3 = Tensor::from_vec_on(vec![3.0; 3 * 32 * 32], vec![3, 32, 32], device.clone())
            .await
            .unwrap();
        let t4 = Tensor::from_vec_on(vec![4.0; 3 * 32 * 32], vec![3, 32, 32], device.clone())
            .await
            .unwrap();
        let mosaic_tensor = Mosaic::new([t1, t2, t3, t4], 12_345)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic_img = mosaic_tensor.to_vec().unwrap();
        assert_eq!(mosaic_img.len(), 3 * 32 * 32);

        // Single channel (grayscale)
        let t1 = Tensor::from_vec_on(vec![1.0; 64 * 64], vec![1, 64, 64], device.clone())
            .await
            .unwrap();
        let t2 = Tensor::from_vec_on(vec![2.0; 64 * 64], vec![1, 64, 64], device.clone())
            .await
            .unwrap();
        let t3 = Tensor::from_vec_on(vec![3.0; 64 * 64], vec![1, 64, 64], device.clone())
            .await
            .unwrap();
        let t4 = Tensor::from_vec_on(vec![4.0; 64 * 64], vec![1, 64, 64], device)
            .await
            .unwrap();
        let mosaic_tensor = Mosaic::new([t1, t2, t3, t4], 99_999)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic_img = mosaic_tensor.to_vec().unwrap();
        assert_eq!(mosaic_img.len(), 64 * 64);
    }

    #[tokio::test]
    async fn test_mosaic_boundary() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        // Different seeds produce different mosaics
        let t1 = Tensor::from_vec_on(vec![1.0; 3 * 128 * 128], vec![3, 128, 128], device.clone())
            .await
            .unwrap();
        let t2 = Tensor::from_vec_on(vec![2.0; 3 * 128 * 128], vec![3, 128, 128], device.clone())
            .await
            .unwrap();
        let t3 = Tensor::from_vec_on(vec![3.0; 3 * 128 * 128], vec![3, 128, 128], device.clone())
            .await
            .unwrap();
        let t4 = Tensor::from_vec_on(vec![4.0; 3 * 128 * 128], vec![3, 128, 128], device)
            .await
            .unwrap();
        let mosaic_tensor1 = Mosaic::new([t1.clone(), t2.clone(), t3.clone(), t4.clone()], 111)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic1 = mosaic_tensor1.to_vec().unwrap();

        let mosaic_tensor2 = Mosaic::new([t1, t2, t3, t4], 222)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic2 = mosaic_tensor2.to_vec().unwrap();
        assert_eq!(mosaic1.len(), mosaic2.len());
    }

    #[tokio::test]
    async fn test_mosaic_large_images() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        // HD images
        let t1 = Tensor::from_vec_on(
            vec![1.0; 3 * 1024 * 1024],
            vec![3, 1024, 1024],
            device.clone(),
        )
        .await
        .unwrap();
        let t2 = Tensor::from_vec_on(
            vec![0.5; 3 * 1024 * 1024],
            vec![3, 1024, 1024],
            device.clone(),
        )
        .await
        .unwrap();
        let t3 = Tensor::from_vec_on(
            vec![0.25; 3 * 1024 * 1024],
            vec![3, 1024, 1024],
            device.clone(),
        )
        .await
        .unwrap();
        let t4 = Tensor::from_vec_on(vec![0.0; 3 * 1024 * 1024], vec![3, 1024, 1024], device)
            .await
            .unwrap();
        let mosaic_tensor = Mosaic::new([t1, t2, t3, t4], 42)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic_img = mosaic_tensor.to_vec().unwrap();
        assert_eq!(mosaic_img.len(), 3 * 1024 * 1024);
    }

    #[tokio::test]
    async fn test_mosaic_precision() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        // Test that all 4 quadrants are represented
        let t1 = Tensor::from_vec_on(vec![1.0; 3 * 100 * 100], vec![3, 100, 100], device.clone())
            .await
            .unwrap();
        let t2 = Tensor::from_vec_on(vec![2.0; 3 * 100 * 100], vec![3, 100, 100], device.clone())
            .await
            .unwrap();
        let t3 = Tensor::from_vec_on(vec![3.0; 3 * 100 * 100], vec![3, 100, 100], device.clone())
            .await
            .unwrap();
        let t4 = Tensor::from_vec_on(vec![4.0; 3 * 100 * 100], vec![3, 100, 100], device)
            .await
            .unwrap();
        let mosaic_tensor = Mosaic::new([t1, t2, t3, t4], 50_505)
            .unwrap()
            .execute()
            .unwrap();
        let mosaic_img = mosaic_tensor.to_vec().unwrap();

        // Should contain values from all 4 images
        assert_eq!(mosaic_img.len(), 3 * 100 * 100);
        assert!(mosaic_img.iter().all(|&x| (1.0..=4.0).contains(&x)));
    }
}