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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Min - Reduction operation finding minimum values - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Hardware-agnostic implementation
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: Production-ready, no mocks
//! - Hardware-agnostic: Pure WGSL for universal compute
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Simple min reduction variant (f64 canonical).
const WGSL_MIN_SIMPLE_F64: &str = include_str!("../shaders/math/min_simple_f64.wgsl");
/// Simple min reduction variant (f32 derived from f64).
pub static WGSL_MIN_SIMPLE: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| WGSL_MIN_SIMPLE_F64.to_string());
/// f64 is the canonical source — math is universal, precision is silicon.
const WGSL_MIN_BASIC_F64: &str = include_str!("../shaders/math/min_f64.wgsl");
/// Basic min reduction shader (f32 derived from f64).
pub static WGSL_MIN_BASIC: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| WGSL_MIN_BASIC_F64.to_string());
/// Min reduction operation
pub struct Min {
input: Tensor,
dim: Option<usize>, // None = global min, Some(d) = min along dimension d
keepdim: bool, // Whether to keep dimension with size 1
}
impl Min {
/// Create a new min operation
#[must_use]
pub fn new(input: Tensor, dim: Option<usize>, keepdim: bool) -> Self {
Self {
input,
dim,
keepdim,
}
}
/// Get the WGSL shader source for global reduction
fn wgsl_shader_reduce() -> &'static str {
static SHADER_REDUCE: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/math/min_reduce_f64.wgsl").to_string()
});
&SHADER_REDUCE
}
/// Get the WGSL shader source for dimension-wise reduction
fn wgsl_shader_dim() -> &'static str {
static SHADER_DIM: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/math/min_dim_f64.wgsl").to_string()
});
&SHADER_DIM
}
/// Execute the min operation
/// # Errors
/// Returns [`Err`] if `dim` is out of range for the tensor shape, buffer
/// allocation fails, shader compilation fails, GPU dispatch fails, or buffer
/// readback fails (e.g. device lost).
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let shape = self.input.shape();
let input_buffer = self.input.buffer();
match self.dim {
None => {
// Global min reduction
let size: usize = shape.iter().product();
// Deep Debt Evolution: Capability-based dispatch
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
let num_workgroups = (size as u32).div_ceil(optimal_wg_size);
// Create output buffer for partial results
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Min Reduce Output"),
size: (num_workgroups as usize * 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 {
size: u32,
}
let params = Params { size: size as u32 };
let params_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Min Reduce Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
// Compile shader
let shader_module =
device.compile_shader(Self::wgsl_shader_reduce(), Some("Min Reduce Shader"));
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Min Reduce 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: false },
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::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("Min Reduce Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
// Create compute pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Min Reduce Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Min Reduce 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("Min Reduce Encoder"),
});
{
let mut compute_pass =
encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Min Reduce Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
compute_pass.dispatch_workgroups(num_workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
// CPU final reduction of partial workgroup results.
// A two-pass GPU reduce is worthwhile at >1M elements;
// below that threshold, readback + CPU reduce is faster.
let partial_results =
device.read_buffer_f32(&output_buffer, num_workgroups as usize)?;
let global_min = partial_results.iter().fold(f32::INFINITY, |a, &b| a.min(b));
// Return scalar tensor
Ok(Tensor::new(vec![global_min], vec![], device.clone()))
}
Some(dim) => {
// Dimension-wise min reduction
if dim >= shape.len() {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!("Dimension {dim} out of range for shape {shape:?}"),
});
}
let dim_size = shape[dim];
let outer_size: usize = shape[..dim].iter().product();
let inner_size: usize = shape[dim + 1..].iter().product();
let output_size = outer_size * inner_size;
// Create output buffer
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Min Dim 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 {
dim_size: u32,
outer_size: u32,
inner_size: u32,
}
let params = Params {
dim_size: dim_size as u32,
outer_size: outer_size as u32,
inner_size: inner_size as u32,
};
let params_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Min Dim Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
// Compile shader
let shader_module =
device.compile_shader(Self::wgsl_shader_dim(), Some("Min Dim Shader"));
// Create bind group layout
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Min Dim 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: false },
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::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("Min Dim Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
// Create compute pipeline
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Min Dim Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Min Dim 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("Min Dim Encoder"),
});
{
let mut compute_pass =
encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Min Dim 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
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
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 = device.read_buffer_f32(&output_buffer, output_size)?;
// Calculate output shape
let mut output_shape = shape.to_vec();
if self.keepdim {
output_shape[dim] = 1;
} else {
output_shape.remove(dim);
}
Ok(Tensor::new(output_data, output_shape, device.clone()))
}
}
}
}
impl Tensor {
/// Find minimum value (global reduction)
/// # Errors
/// Returns [`Err`] if buffer allocation fails, shader compilation fails, GPU
/// dispatch fails, or buffer readback fails (e.g. device lost).
pub fn min(&self) -> Result<Self> {
Min::new(self.clone(), None, false).execute()
}
/// Find minimum value along a dimension
/// # Arguments
/// * `dim` - Dimension to find min along
/// * `keepdim` - Whether to keep the reduced dimension with size 1
/// # Errors
/// Returns [`Err`] if `dim` is out of range for the tensor shape, buffer
/// allocation fails, shader compilation fails, GPU dispatch fails, or buffer
/// readback fails (e.g. device lost).
pub fn min_dim(&self, dim: usize, keepdim: bool) -> Result<Self> {
Min::new(self.clone(), Some(dim), keepdim).execute()
}
/// Find minimum value (legacy method for backward compatibility)
/// # Errors
/// Returns [`Err`] if `dim` is out of range for the tensor shape, buffer
/// allocation fails, shader compilation fails, GPU dispatch fails, or buffer
/// readback fails (e.g. device lost).
pub fn min_wgsl(self, dim: Option<usize>) -> Result<Self> {
match dim {
None => Min::new(self, None, false).execute(),
Some(d) => Min::new(self, Some(d), false).execute(),
}
}
}