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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Sparse matrix multiplication with quantization
//!
//! Performs matrix multiplication with sparse matrices and quantized values.
//! Critical for neuromorphic and edge computing where memory/power are constrained.
//!
//! # Neuromorphic Computing
//!
//! Sparse quantized operations are essential for NPU efficiency:
//! - Sparse matrices: Only store non-zero values
//! - Quantization: Use int8 instead of fp32 (4x memory savings)
//! - Combined: Massive efficiency gains for SNNs
//!
//! # Storage Format
//!
//! **COO (Coordinate) Format**:
//! - Values: Non-zero elements
//! - Rows: Row indices
//! - Cols: Column indices
//!
//! # Example
//!
//! ```no_run
//! use barracuda::sparse_matmul_quantized;
//! use barracuda::prelude::WgpuDevice;
//!
//! # async fn example() -> Result<(), Box<dyn std::error::Error>> {
//! let device = WgpuDevice::new().await?;
//!
//! // Sparse matrix: only non-zero values
//! let values = vec![127, -128, 64];  // int8 values
//! let rows = vec![0, 1, 2];
//! let cols = vec![0, 1, 0];
//! let dense_vec = vec![100, -50];     // int8 input
//!
//! let result = sparse_matmul_quantized(
//!     &device,
//!     &values,
//!     &rows,
//!     &cols,
//!     &dense_vec,
//!     3, // output size
//!     127.0, // scale factor
//! )?;
//! # Ok(())
//! # }
//! ```

use std::borrow::Cow;
use wgpu::util::DeviceExt;

use crate::device::WgpuDevice;
use crate::device::capabilities::{
    WORKGROUP_SIZE_1D, WORKGROUP_SIZE_COMPACT, WORKGROUP_SIZE_MEDIUM,
};
use crate::error::{BarracudaError, Result as BarracudaResult};

/// Sparse matrix multiply with quantized int8 values
///
/// # Arguments
///
/// * `device` - The `WgpuDevice` (provides device, queue, and readback)
/// * `sparse_values` - Non-zero values (int8)
/// * `sparse_rows` - Row indices
/// * `sparse_cols` - Column indices
/// * `dense_vector` - Dense input vector (int8)
/// * `output_size` - Size of output vector
/// * `scale` - Quantization scale factor
///
/// # Returns
///
/// Dense output vector (fp32, dequantized)
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn sparse_matmul_quantized(
    device: &WgpuDevice,
    sparse_values: &[i8],
    sparse_rows: &[u32],
    sparse_cols: &[u32],
    dense_vector: &[i8],
    output_size: u32,
    scale: f32,
) -> BarracudaResult<Vec<f32>> {
    if sparse_values.is_empty() {
        return Err(BarracudaError::InvalidInput {
            message: "Sparse values cannot be empty".to_string(),
        });
    }

    if sparse_values.len() != sparse_rows.len() || sparse_values.len() != sparse_cols.len() {
        return Err(BarracudaError::InvalidInput {
            message: "Sparse arrays must have same length".to_string(),
        });
    }

    let nnz = crate::utils::checked_u32(sparse_values.len(), "sparse_matmul nnz")?;
    let d = device.device();
    let q = device.queue();

    let shader = d.create_shader_module(wgpu::ShaderModuleDescriptor {
        label: Some("Sparse MatMul Quantized Shader"),
        source: wgpu::ShaderSource::Wgsl(Cow::Borrowed(include_str!(
            "sparse_matmul_quantized.wgsl"
        ))),
    });

    // Convert i8 to i32 for GPU (WGSL doesn't have i8)
    let values_i32: Vec<i32> = sparse_values.iter().map(|&x| x as i32).collect();
    let dense_i32: Vec<i32> = dense_vector.iter().map(|&x| x as i32).collect();

    let values_buffer = d.create_buffer_init(&wgpu::util::BufferInitDescriptor {
        label: Some("Sparse Values"),
        contents: bytemuck::cast_slice(&values_i32),
        usage: wgpu::BufferUsages::STORAGE,
    });

    let rows_buffer = d.create_buffer_init(&wgpu::util::BufferInitDescriptor {
        label: Some("Sparse Rows"),
        contents: bytemuck::cast_slice(sparse_rows),
        usage: wgpu::BufferUsages::STORAGE,
    });

    let cols_buffer = d.create_buffer_init(&wgpu::util::BufferInitDescriptor {
        label: Some("Sparse Cols"),
        contents: bytemuck::cast_slice(sparse_cols),
        usage: wgpu::BufferUsages::STORAGE,
    });

    let dense_buffer = d.create_buffer_init(&wgpu::util::BufferInitDescriptor {
        label: Some("Dense Vector"),
        contents: bytemuck::cast_slice(&dense_i32),
        usage: wgpu::BufferUsages::STORAGE,
    });

    let output_buffer = d.create_buffer(&wgpu::BufferDescriptor {
        label: Some("Output"),
        size: (output_size * std::mem::size_of::<f32>() as u32) as u64,
        usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
        mapped_at_creation: false,
    });

    #[repr(C)]
    #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
    struct Params {
        nnz: u32,
        output_size: u32,
        scale: f32,
        _padding: u32,
    }

    let params = Params {
        nnz,
        output_size,
        scale,
        _padding: 0,
    };

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

    let bind_group_layout = d.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
        label: Some("Sparse MatMul 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,
            },
        ],
    });

    let bind_group = d.create_bind_group(&wgpu::BindGroupDescriptor {
        label: Some("Sparse MatMul Bind Group"),
        layout: &bind_group_layout,
        entries: &[
            wgpu::BindGroupEntry {
                binding: 0,
                resource: values_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 1,
                resource: rows_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 2,
                resource: cols_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 3,
                resource: dense_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 4,
                resource: output_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 5,
                resource: params_buffer.as_entire_binding(),
            },
        ],
    });

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

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

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

    {
        let mut cpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("Sparse MatMul Pass"),
            timestamp_writes: None,
        });
        cpass.set_pipeline(&pipeline);
        cpass.set_bind_group(0, Some(&bind_group), &[]);
        // Deep Debt Evolution: Capability-based dispatch
        // Note: This function uses raw wgpu::Device, so we use device limits for capability awareness
        let limits = d.limits();
        let max_invocations = limits.max_compute_invocations_per_workgroup;
        // Use capability-aware workgroup size, respecting device limits
        // Choose optimal size based on device capabilities: prefer 256 for discrete GPUs,
        // 128 for integrated GPUs, but always respect device limits
        let optimal_wg_size = if max_invocations >= WORKGROUP_SIZE_1D {
            WORKGROUP_SIZE_1D
        } else if max_invocations >= WORKGROUP_SIZE_MEDIUM {
            WORKGROUP_SIZE_MEDIUM
        } else {
            max_invocations.max(WORKGROUP_SIZE_COMPACT)
        };
        let workgroups = output_size.div_ceil(optimal_wg_size);
        cpass.dispatch_workgroups(workgroups.max(1), 1, 1);
    }

    let staging_buffer = d.create_buffer(&wgpu::BufferDescriptor {
        label: Some("Staging"),
        size: (output_size * std::mem::size_of::<f32>() as u32) as u64,
        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
        mapped_at_creation: false,
    });

    encoder.copy_buffer_to_buffer(
        &output_buffer,
        0,
        &staging_buffer,
        0,
        (output_size * std::mem::size_of::<f32>() as u32) as u64,
    );
    q.submit(Some(encoder.finish()));

    device.map_staging_buffer::<f32>(&staging_buffer, output_size as usize)
}

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

    #[tokio::test]
    async fn test_sparse_matmul_quantized_basic() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let values = vec![127, -64, 32];
        let rows = vec![0, 1, 2];
        let cols = vec![0, 0, 1];
        let dense = vec![10, 20];
        let result =
            sparse_matmul_quantized(device.as_ref(), &values, &rows, &cols, &dense, 3, 1.0)
                .unwrap();
        assert_eq!(result.len(), 3);
        assert!((result[0] - 1270.0).abs() < 1.0);
        assert!((result[1] - -640.0).abs() < 1.0);
        assert!((result[2] - 640.0).abs() < 1.0);
    }

    #[tokio::test]
    async fn test_sparse_matmul_quantized_edge_cases() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let values = vec![0];
        let rows = vec![0];
        let cols = vec![0];
        let dense = vec![100];
        let result =
            sparse_matmul_quantized(device.as_ref(), &values, &rows, &cols, &dense, 1, 1.0)
                .unwrap();
        assert!(result[0].abs() < 0.1);
    }

    #[tokio::test]
    async fn test_sparse_matmul_quantized_boundary() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let empty: Vec<i8> = vec![];
        assert!(sparse_matmul_quantized(device.as_ref(), &empty, &[], &[], &[1], 1, 1.0).is_err());
    }

    #[tokio::test]
    async fn test_sparse_matmul_quantized_large_tensor() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let values: Vec<i8> = (0..1000).map(|i| (i % 128) as i8).collect();
        let rows: Vec<u32> = (0..1000).map(|i| i % 100).collect();
        let cols: Vec<u32> = (0..1000).map(|i| i % 50).collect();
        let dense: Vec<i8> = (0..50).map(|i| (i % 10) as i8).collect();
        let result =
            sparse_matmul_quantized(device.as_ref(), &values, &rows, &cols, &dense, 100, 1.0)
                .unwrap();
        assert_eq!(result.len(), 100);
        assert!(result.iter().all(|&x| x.is_finite()));
    }

    #[tokio::test]
    async fn test_sparse_matmul_quantized_precision() {
        let Some(device) = crate::device::test_pool::get_test_device_if_gpu_available().await
        else {
            return;
        };
        let values = vec![127, 127];
        let rows = vec![0, 0];
        let cols = vec![0, 1];
        let dense = vec![1, 1];
        let result =
            sparse_matmul_quantized(device.as_ref(), &values, &rows, &cols, &dense, 1, 1.0)
                .unwrap();
        assert!((result[0] - 254.0).abs() < 1.0);
    }
}