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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Adaptive Instance Normalization (`AdaIN`) - Style transfer
//!
//! Transfers style from one image to another.
//! Used in neural style transfer, GANs.
//!
//! Deep Debt Principles:
//! - Pure GPU/WGSL execution
//! - Safe Rust wrappers
//! - Hardware-agnostic via WebGPU
//! - Runtime device discovery
//! - Zero CPU fallbacks in execution

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

/// f64 is the canonical source — math is universal, precision is silicon.
const SHADER_F64: &str = include_str!("../shaders/norm/adaptive_instance_norm_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
    std::sync::LazyLock::new(|| SHADER_F64.to_string());

/// `AdaptiveInstanceNorm` operation
pub struct AdaptiveInstanceNorm {
    content: Tensor,
    style_mean: Tensor,
    style_std: Tensor,
}

impl AdaptiveInstanceNorm {
    /// Create a new adaptive instance norm operation
    /// # Errors
    /// Returns [`Err`] if content is not 4D (NCHW), style mean/std are not 1D, or style statistics
    /// do not match the number of channels.
    pub fn new(content: Tensor, style_mean: Tensor, style_std: Tensor) -> Result<Self> {
        let content_shape = content.shape();
        let style_mean_shape = style_mean.shape();
        let style_std_shape = style_std.shape();

        if content_shape.len() != 4 {
            return Err(crate::error::BarracudaError::invalid_op(
                "AdaptiveInstanceNorm",
                format!("Content must be 4D (NCHW), got {}D", content_shape.len()),
            ));
        }

        if style_mean_shape.len() != 1 || style_std_shape.len() != 1 {
            return Err(crate::error::BarracudaError::invalid_op(
                "AdaptiveInstanceNorm",
                "Style mean and std must be 1D tensors",
            ));
        }

        if style_mean_shape[0] != content_shape[1] || style_std_shape[0] != content_shape[1] {
            return Err(crate::error::BarracudaError::invalid_op(
                "AdaptiveInstanceNorm",
                "Style statistics must match number of channels",
            ));
        }

        Ok(Self {
            content,
            style_mean,
            style_std,
        })
    }

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

    /// Execute the adaptive instance norm operation
    /// # Errors
    /// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, buffer readback fails,
    /// or the device is lost.
    pub fn execute(self) -> Result<Tensor> {
        let device = self.content.device();
        let shape = self.content.shape();

        let batch = shape[0];
        let channels = shape[1];
        let height = shape[2];
        let width = shape[3];
        let spatial_size = height * width;
        let output_size = batch * channels * spatial_size;

        // Create buffers
        let content_buffer = self.content.buffer();
        let style_mean_buffer = self.style_mean.buffer();
        let style_std_buffer = self.style_std.buffer();

        let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("AdaptiveInstanceNorm 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 {
            batch: u32,
            channels: u32,
            height: u32,
            width: u32,
            spatial_size: u32,
        }

        let params = Params {
            batch: batch as u32,
            channels: channels as u32,
            height: height as u32,
            width: width as u32,
            spatial_size: spatial_size as u32,
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("AdaptiveInstanceNorm 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("AdaptiveInstanceNorm 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: false },
                                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::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("AdaptiveInstanceNorm Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: content_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: style_mean_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: style_std_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

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

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

        let compute_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("AdaptiveInstanceNorm 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("AdaptiveInstanceNorm Encoder"),
        });

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

            // Deep Debt Evolution: Capability-based dispatch
            use crate::device::{DeviceCapabilities, WorkloadType};
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
            let workgroups = (output_size as u32).div_ceil(optimal_wg_size);
            compute_pass.dispatch_workgroups(workgroups, 1, 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![batch, channels, height, width],
            device.clone(),
        ))
    }
}

impl Tensor {
    /// Apply adaptive instance normalization (`AdaIN`) for style transfer
    /// # Arguments
    /// * `style_mean` - Style mean tensor [C]
    /// * `style_std` - Style std tensor [C]
    /// # Errors
    /// Returns [`Err`] if content is not 4D, style mean/std shapes are invalid, channel count
    /// mismatches, buffer allocation fails, GPU dispatch fails, buffer readback fails, or the device is lost.
    pub fn adaptive_instance_norm(self, style_mean: Self, style_std: Self) -> Result<Self> {
        AdaptiveInstanceNorm::new(self, style_mean, style_std)?.execute()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    // No longer needed - using Tensor method API

    #[tokio::test]
    async fn test_adaptive_instance_norm_basic() {
        let dev = crate::device::test_pool::get_test_device().await;
        let content_data = vec![1.0; 3 * 4 * 4];
        let content = Tensor::new(content_data.clone(), vec![1, 3, 4, 4], dev.clone());
        let style_mean = Tensor::new(vec![0.5, 0.5, 0.5], vec![3], dev.clone());
        let style_std = Tensor::new(vec![0.2, 0.2, 0.2], vec![3], dev);
        let output_tensor = content
            .adaptive_instance_norm(style_mean, style_std)
            .unwrap();
        let output = output_tensor.to_vec().unwrap();
        assert_eq!(output.len(), content_data.len());
        assert!(output.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_adaptive_instance_norm_edge_cases() {
        let dev = crate::device::test_pool::get_test_device().await;
        // Test with zero style std (should clamp)
        let content = Tensor::new(vec![1.0, 2.0, 3.0, 4.0], vec![1, 1, 2, 2], dev.clone());
        let style_mean = Tensor::new(vec![0.0], vec![1], dev.clone());
        let style_std = Tensor::new(vec![0.0], vec![1], dev.clone());
        let output_tensor = content
            .adaptive_instance_norm(style_mean, style_std)
            .unwrap();
        let output = output_tensor.to_vec().unwrap();
        assert!(output.iter().all(|&x| x.is_finite()));

        // Test with single channel, single pixel
        let content = Tensor::new(vec![5.0], vec![1, 1, 1, 1], dev.clone());
        let style_mean = Tensor::new(vec![1.0], vec![1], dev.clone());
        let style_std = Tensor::new(vec![2.0], vec![1], dev);
        let output_tensor = content
            .adaptive_instance_norm(style_mean, style_std)
            .unwrap();
        let output = output_tensor.to_vec().unwrap();
        assert_eq!(output.len(), 1);
        assert!(output[0].is_finite());
    }

    #[tokio::test]
    async fn test_adaptive_instance_norm_boundary() {
        let dev = crate::device::test_pool::get_test_device().await;
        // Test with different style statistics
        let content_data = vec![0.0, 1.0, 2.0, 3.0];

        // Style 1: mean=0, std=1
        let content1 = Tensor::new(content_data.clone(), vec![1, 1, 2, 2], dev.clone());
        let style_mean1 = Tensor::new(vec![0.0], vec![1], dev.clone());
        let style_std1 = Tensor::new(vec![1.0], vec![1], dev.clone());
        let result1 = content1
            .adaptive_instance_norm(style_mean1, style_std1)
            .unwrap();
        let output1 = result1.to_vec().unwrap();

        // Style 2: mean=10, std=5
        let content2 = Tensor::new(content_data, vec![1, 1, 2, 2], dev.clone());
        let style_mean2 = Tensor::new(vec![10.0], vec![1], dev.clone());
        let style_std2 = Tensor::new(vec![5.0], vec![1], dev);
        let result2 = content2
            .adaptive_instance_norm(style_mean2, style_std2)
            .unwrap();
        let output2 = result2.to_vec().unwrap();

        assert!(output1.iter().all(|&x| x.is_finite()));
        assert!(output2.iter().all(|&x| x.is_finite()));
        // Different style should produce different output
        assert_ne!(output1, output2);
        // Output2 should have higher values (mean=10)
        assert!(output2.iter().sum::<f32>() > output1.iter().sum::<f32>());
    }

    #[tokio::test]
    async fn test_adaptive_instance_norm_large_batch() {
        let dev = crate::device::test_pool::get_test_device().await;
        // Multiple batches and channels
        let batch_size = 2;
        let channels = 4;
        let height = 8;
        let width = 8;

        let content_data: Vec<f32> = (0..batch_size * channels * height * width)
            .map(|i| (i % 10) as f32)
            .collect();
        let content = Tensor::new(
            content_data.clone(),
            vec![batch_size, channels, height, width],
            dev.clone(),
        );
        let style_mean = Tensor::new(vec![0.5, 1.0, 1.5, 2.0], vec![channels], dev.clone());
        let style_std = Tensor::new(vec![0.1, 0.2, 0.3, 0.4], vec![channels], dev);

        let output_tensor = content
            .adaptive_instance_norm(style_mean, style_std)
            .unwrap();
        let output = output_tensor.to_vec().unwrap();

        assert_eq!(output.len(), content_data.len());
        assert!(output.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_adaptive_instance_norm_precision() {
        let dev = crate::device::test_pool::get_test_device().await;
        // Test with known values for style transfer
        let content_data = vec![
            0.0, 1.0, 2.0, 3.0, // Mean = 1.5
        ];
        let content = Tensor::new(content_data, vec![1, 1, 2, 2], dev.clone());
        let style_mean = Tensor::new(vec![5.0], vec![1], dev.clone()); // Target mean
        let style_std = Tensor::new(vec![2.0], vec![1], dev); // Target std

        let result = content
            .adaptive_instance_norm(style_mean, style_std)
            .unwrap();
        let output = result.to_vec().unwrap();

        // After AdaIN, output should have approximately the target mean
        let out_mean = output.iter().sum::<f32>() / output.len() as f32;
        assert!((out_mean - 5.0).abs() < 0.1);

        // Output should preserve relative relationships (normalized)
        assert!(output[0] < output[1]);
        assert!(output[1] < output[2]);
        assert!(output[2] < output[3]);
    }
}