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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! `SearchSorted` - GPU parallel binary search
//!
//! **Deep Debt Principles**:
//! - Complete GPU implementation: Parallel binary search for each value
//! - No CPU fallbacks: All computation on GPU
//! - Self-knowledge: Validates sorted array
//! - Modern idiomatic Rust: Result<T, E>

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

#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct SearchSortedParams {
    sorted_size: u32,
    values_size: u32,
    side_right: u32, // 0 = left (default), 1 = right
    _pad1: u32,
}

/// GPU parallel binary search for insertion indices in a sorted array.
pub struct SearchSorted {
    sorted: Tensor,
    values: Tensor,
    side_right: bool,
}

impl SearchSorted {
    /// Creates a new searchsorted operation. `side_right` selects left (false) or right (true) insertion.
    /// # Errors
    /// Returns [`Err`] if sorted or values are empty, or if either is not 1D.
    pub fn new(sorted: Tensor, values: Tensor, side_right: bool) -> Result<Self> {
        if sorted.is_empty() {
            return Err(BarracudaError::invalid_op(
                "searchsorted",
                "Sorted array cannot be empty",
            ));
        }

        if values.is_empty() {
            return Err(BarracudaError::invalid_op(
                "searchsorted",
                "Values array cannot be empty",
            ));
        }

        // Validate sorted array is 1D
        if sorted.shape().len() != 1 {
            return Err(BarracudaError::invalid_op(
                "searchsorted",
                "Sorted array must be 1D",
            ));
        }

        if values.shape().len() != 1 {
            return Err(BarracudaError::invalid_op(
                "searchsorted",
                "Values array must be 1D",
            ));
        }

        Ok(Self {
            sorted,
            values,
            side_right,
        })
    }

    fn wgsl_shader() -> &'static str {
        {
            static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
                include_str!("../shaders/misc/searchsorted_f64.wgsl").to_string()
            });
            std::sync::LazyLock::force(&SHADER).as_str()
        }
    }

    fn u32_to_f32_shader() -> &'static str {
        static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
            include_str!("../shaders/misc/u32_to_f32_f64.wgsl").to_string()
        });
        std::sync::LazyLock::force(&SHADER).as_str()
    }

    /// Executes the binary search and returns insertion indices as f32 tensor.
    /// # 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.sorted.device();
        let sorted_size = self.sorted.len();
        let values_size = self.values.len();

        let output_buffer = device.create_buffer_u32(values_size)?;

        let params = SearchSortedParams {
            sorted_size: sorted_size as u32,
            values_size: values_size as u32,
            side_right: u32::from(self.side_right),
            _pad1: 0,
        };

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

        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("SearchSorted 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: false },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                    ],
                });

        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("SearchSorted Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: params_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.sorted.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: self.values.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: output_buffer.as_entire_binding(),
                },
            ],
        });

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

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

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("SearchSorted Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&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::ElementWise);
            let workgroups = (values_size as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

        // Convert u32 indices to f32 on GPU (for Tensor compatibility)
        let indices_f32_buffer = device.create_buffer_f32(values_size)?;

        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct ConvertParams {
            size: u32,
            _pad1: u32,
            _pad2: u32,
            _pad3: u32,
        }

        let convert_params = ConvertParams {
            size: values_size as u32,
            _pad1: 0,
            _pad2: 0,
            _pad3: 0,
        };

        let convert_params_buffer =
            device
                .device
                .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                    label: Some("SearchSorted Convert Params"),
                    contents: bytemuck::bytes_of(&convert_params),
                    usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
                });

        let convert_bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("SearchSorted Convert 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: false },
                                has_dynamic_offset: false,
                                min_binding_size: None,
                            },
                            count: None,
                        },
                    ],
                });

        let convert_bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("SearchSorted Convert Bind Group"),
            layout: &convert_bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: convert_params_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: indices_f32_buffer.as_entire_binding(),
                },
            ],
        });

        let convert_shader =
            device.compile_shader(Self::u32_to_f32_shader(), Some("SearchSorted Convert"));
        let convert_pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("SearchSorted Convert Pipeline Layout"),
                    bind_group_layouts: &[&convert_bind_group_layout],
                    immediate_size: 0,
                });

        let convert_pipeline =
            device
                .device
                .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                    label: Some("SearchSorted Convert Pipeline"),
                    layout: Some(&convert_pipeline_layout),
                    module: &convert_shader,
                    entry_point: Some("main"),
                    cache: None,
                    compilation_options: Default::default(),
                });

        let mut convert_encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("SearchSorted Convert Encoder"),
        });

        {
            let mut pass = convert_encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("SearchSorted Convert Pass"),
                timestamp_writes: None,
            });
            pass.set_pipeline(&convert_pipeline);
            pass.set_bind_group(0, Some(&convert_bind_group), &[]);
            // Deep Debt Evolution: Capability-based dispatch
            let caps = DeviceCapabilities::from_device(device);
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
            let workgroups = (values_size as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

        Ok(Tensor::from_buffer(
            indices_f32_buffer,
            vec![values_size],
            device.clone(),
        ))
    }
}

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

    #[tokio::test]
    async fn test_searchsorted_basic() {
        let device = crate::device::test_pool::get_test_device().await;
        let sorted = Tensor::from_vec_on(vec![1.0, 3.0, 5.0, 7.0], vec![4], device.clone())
            .await
            .unwrap();
        let values = Tensor::from_vec_on(vec![2.0, 4.0, 6.0], vec![3], device.clone())
            .await
            .unwrap();

        let result = SearchSorted::new(sorted, values, false)
            .unwrap()
            .execute()
            .unwrap();
        let indices = result.to_vec().unwrap();
        assert_eq!(indices.len(), 3);
        // Should be [1, 2, 3] (insertion points)
    }

    #[tokio::test]
    async fn test_searchsorted_right() {
        let device = crate::device::test_pool::get_test_device().await;
        let sorted = Tensor::from_vec_on(vec![1.0, 3.0, 5.0], vec![3], device.clone())
            .await
            .unwrap();
        let values = Tensor::from_vec_on(vec![3.0], vec![1], device.clone())
            .await
            .unwrap();

        let result = SearchSorted::new(sorted, values, true)
            .unwrap()
            .execute()
            .unwrap();
        let indices = result.to_vec().unwrap();
        assert_eq!(indices.len(), 1);
    }
}