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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! RBF Kernel Evaluation - Radial Basis Function Kernels - Pure WGSL
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (GPU-optimized)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//! - ✅ Runtime-configured kernel type
//! - ✅ Fused distance + kernel computation
//!
//! ## Algorithm
//!
//! Computes RBF kernel matrix:
//! ```text
//! Input:  X [N, d], Y [M, d] - point clouds
//! Output: K [N, M] where K[i,j] = φ(‖xᵢ - yⱼ‖)
//!
//! Fused computation: distance + kernel in single pass
//! Avoids N×M intermediate distance matrix
//! ```
//!
//! ## Kernel Functions
//!
//! - **Thin Plate Spline**: φ(r) = r² · log(r) [default, best for physics]
//! - **Gaussian**: φ(r) = exp(-ε²r²)
//! - **Multiquadric**: φ(r) = sqrt(1 + ε²r²)
//! - **Inverse Multiquadric**: φ(r) = 1/sqrt(1 + ε²r²)
//! - **Cubic**: φ(r) = r³
//! - **Quintic**: φ(r) = r⁵
//! - **Linear**: φ(r) = r
//!
//! ## Use Case
//!
//! **RBF Surrogate Learning** (hotSpring physics integration):
//! - Training: K = `rbf_kernel(X_train`, `X_train`) → Cholesky → solve for weights
//! - Prediction: K = `rbf_kernel(X_new`, `X_train`) → matmul with weights
//!
//! ## References
//!
//! - Fasshauer, "Meshfree Approximation Methods with MATLAB"
//! - scipy.interpolate.RBFInterpolator
//! - Used in surrogate-based optimization

use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};

/// f64 is the canonical source — math is universal, precision is silicon.
static SHADER_F64: &str = include_str!("../../shaders/interpolation/rbf_kernel_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
    std::sync::LazyLock::new(|| SHADER_F64.to_string());

/// RBF kernel types
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u32)]
#[derive(Default)]
pub enum RbfKernelType {
    /// Thin Plate Spline: r² · log(r) [default, best for physics]
    #[default]
    ThinPlateSpline = 0,
    /// Gaussian: exp(-ε²r²)
    Gaussian = 1,
    /// Multiquadric: sqrt(1 + ε²r²)
    Multiquadric = 2,
    /// Inverse Multiquadric: 1/sqrt(1 + ε²r²)
    InverseMultiquadric = 3,
    /// Cubic: r³
    Cubic = 4,
    /// Quintic: r⁵
    Quintic = 5,
    /// Linear: r
    Linear = 6,
}

/// RBF kernel evaluation operation
///
/// Computes K[i,j] = φ(‖xᵢ - yⱼ‖) for radial basis function φ
pub struct RbfKernel {
    x: Tensor, // Points X [N, d]
    y: Tensor, // Points Y [M, d]
    kernel_type: RbfKernelType,
    epsilon: f32, // Shape parameter (for Gaussian, MQ, IMQ)
}

impl RbfKernel {
    /// Create new RBF kernel operation
    /// # Arguments
    /// * `x` - First point cloud [N, d]
    /// * `y` - Second point cloud [M, d]
    /// * `kernel_type` - Type of RBF kernel
    /// * `epsilon` - Shape parameter (default 1.0)
    /// # Deep Debt Compliance
    /// - No hardcoded sizes (runtime N, M, d)
    /// - No unsafe blocks
    /// - Agnostic design (works with any point cloud)
    /// - Composable (fuses distance + kernel)
    #[must_use]
    pub fn new(x: Tensor, y: Tensor, kernel_type: RbfKernelType, epsilon: f32) -> Self {
        Self {
            x,
            y,
            kernel_type,
            epsilon,
        }
    }

    fn wgsl_shader() -> &'static str {
        &SHADER_F32
    }

    /// Execute RBF kernel evaluation on GPU
    /// # Returns
    /// Kernel matrix K [N, M] where K[i,j] = φ(‖xᵢ - yⱼ‖)
    /// # Errors
    /// - Returns error if x and y have different dimensions
    /// - Returns error if x or y are not 2D
    /// # Deep Debt Compliance
    /// - Pure WGSL execution (no CPU fallback)
    /// - Fused distance + kernel (single pass)
    /// - Safe buffer management
    pub fn execute(self) -> Result<Tensor> {
        let device = self.x.device();
        let x_shape = self.x.shape();
        let y_shape = self.y.shape();

        // Validate 2D tensors
        if x_shape.len() != 2 || y_shape.len() != 2 {
            return Err(BarracudaError::InvalidShape {
                expected: vec![0, 0],
                actual: x_shape.to_vec(),
            });
        }

        let n_rows = x_shape[0]; // N
        let n_cols = y_shape[0]; // M
        let x_dims = x_shape[1]; // d
        let y_dims = y_shape[1]; // d

        // Validate same dimensionality
        if x_dims != y_dims {
            return Err(BarracudaError::InvalidShape {
                expected: vec![n_rows, x_dims],
                actual: vec![n_cols, y_dims],
            });
        }

        let n_dims = x_dims;

        // Create output buffer for kernel matrix K [N, M]
        let output_size = n_rows * n_cols;
        let output_buffer = device.create_buffer_f32(output_size)?;

        // Create params struct matching WGSL layout
        #[repr(C)]
        #[derive(Copy, Clone, Debug, Pod, Zeroable)]
        struct RbfKernelParams {
            n_rows: u32,
            n_cols: u32,
            n_dims: u32,
            kernel_type: u32,
            epsilon: f32,
            _pad0: u32,
            _pad1: u32,
            _pad2: u32,
        }

        let params = RbfKernelParams {
            n_rows: n_rows as u32,
            n_cols: n_cols as u32,
            n_dims: n_dims as u32,
            kernel_type: self.kernel_type as u32,
            epsilon: self.epsilon,
            _pad0: 0,
            _pad1: 0,
            _pad2: 0,
        };

        let params_buffer = device.create_uniform_buffer("RbfKernel Params", &params);

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("RbfKernel BGL"),
                    entries: &[
                        // Points X [N, d]
                        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,
                        },
                        // Points Y [M, d]
                        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,
                        },
                        // Output kernel matrix K [N, M]
                        wgpu::BindGroupLayoutEntry {
                            binding: 2,
                            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,
                        },
                        // Parameters (n_rows, n_cols, n_dims, kernel_type, epsilon)
                        wgpu::BindGroupLayoutEntry {
                            binding: 3,
                            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("RbfKernel BG"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.x.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.y.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

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

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

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

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

        // Execute compute pass
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("RbfKernel Pass"),
                timestamp_writes: None,
            });

            pass.set_pipeline(&pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);

            // Deep Debt Evolution: 2D parallel workgroup dispatch
            // Each (i,j) pair is independent
            let workgroup_size = 16;
            let workgroups_x = (n_rows as u32).div_ceil(workgroup_size);
            let workgroups_y = (n_cols as u32).div_ceil(workgroup_size);
            pass.dispatch_workgroups(workgroups_x.max(1), workgroups_y.max(1), 1);
        }

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

        let output_size = n_rows * n_cols;
        let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
        Ok(Tensor::new(
            output_data,
            vec![n_rows, n_cols],
            device.clone(),
        ))
    }
}

/// Tensor extension for RBF kernel evaluation
impl Tensor {
    /// Compute RBF kernel matrix between two point clouds
    /// # Arguments
    /// * `other` - Second point cloud Y [M, d]
    /// * `kernel_type` - Type of RBF kernel
    /// * `epsilon` - Shape parameter (default 1.0)
    /// # Returns
    /// Kernel matrix K [N, M] where K[i,j] = φ(‖xᵢ - yⱼ‖)
    /// # Example
    /// ```ignore
    /// let x = Tensor::from_vec(vec![...], vec![10, 3], device)?;  // 10 points, 3D
    /// let y = Tensor::from_vec(vec![...], vec![20, 3], device)?;  // 20 points, 3D
    /// let k = x.rbf_kernel(&y, RbfKernelType::ThinPlateSpline, 1.0)?;
    /// // k is [10, 20] kernel matrix
    /// ```
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn rbf_kernel(
        &self,
        other: &Self,
        kernel_type: RbfKernelType,
        epsilon: f32,
    ) -> Result<Self> {
        RbfKernel::new(self.clone(), other.clone(), kernel_type, epsilon).execute()
    }
}

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

    #[tokio::test]
    async fn test_rbf_kernel_same_points() {
        let device = crate::device::test_pool::get_test_device().await;
        // Two identical points should have kernel value = 0 for TPS (r=0)
        let x_data = vec![1.0, 2.0, 3.0, 1.0, 2.0, 3.0];
        let x = Tensor::from_vec_on(x_data, vec![2, 3], device)
            .await
            .unwrap();

        let k = x
            .rbf_kernel(&x, RbfKernelType::ThinPlateSpline, 1.0)
            .unwrap();
        let kernel = k.to_vec().unwrap();

        // K should be [2, 2]
        assert_eq!(kernel.len(), 4);

        // Diagonal should be 0 for TPS (r=0)
        assert!(
            kernel[0].abs() < 1e-5,
            "K[0,0] should be 0, got {}",
            kernel[0]
        );
        assert!(
            kernel[3].abs() < 1e-5,
            "K[1,1] should be 0, got {}",
            kernel[3]
        );
    }

    #[tokio::test]
    async fn test_rbf_kernel_gaussian() {
        let device = crate::device::test_pool::get_test_device().await;
        // Test Gaussian kernel at origin: should be 1.0
        let x_data = vec![0.0, 0.0, 0.0];
        let x = Tensor::from_vec_on(x_data, vec![1, 3], device)
            .await
            .unwrap();

        let k = x.rbf_kernel(&x, RbfKernelType::Gaussian, 1.0).unwrap();
        let kernel = k.to_vec().unwrap();

        // Gaussian at r=0 should be exp(0) = 1.0
        assert!(
            (kernel[0] - 1.0).abs() < 1e-5,
            "Gaussian at r=0 should be 1.0, got {}",
            kernel[0]
        );
    }

    #[tokio::test]
    async fn test_rbf_kernel_dimensions() {
        let device = crate::device::test_pool::get_test_device().await;
        // Test that output has correct dimensions
        let x_data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]; // 2 points, 3D
        let y_data = vec![7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0]; // 3 points, 3D

        let x = Tensor::from_vec_on(x_data, vec![2, 3], device.clone())
            .await
            .unwrap();
        let y = Tensor::from_vec_on(y_data, vec![3, 3], device)
            .await
            .unwrap();

        let k = x.rbf_kernel(&y, RbfKernelType::Cubic, 1.0).unwrap();
        let shape = k.shape();

        // Should be [2, 3]
        assert_eq!(shape, &[2, 3]);

        let kernel = k.to_vec().unwrap();
        assert_eq!(kernel.len(), 6);
    }
}