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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Max - Reduction operation finding maximum 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::compute_pipeline::ComputeDispatch;
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
/// Simple max reduction variant (f64 canonical).
const WGSL_MAX_SIMPLE_F64: &str = include_str!("../shaders/math/max_simple_f64.wgsl");
/// Simple max reduction variant (f32 derived from f64).
pub static WGSL_MAX_SIMPLE: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| WGSL_MAX_SIMPLE_F64.to_string());
/// f64 is the canonical source — math is universal, precision is silicon.
const WGSL_MAX_BASIC_F64: &str = include_str!("../shaders/math/max_f64.wgsl");
/// Basic max reduction shader (f32 derived from f64).
pub static WGSL_MAX_BASIC: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| WGSL_MAX_BASIC_F64.to_string());
/// Max reduction operation
pub struct Max {
input: Tensor,
dim: Option<usize>, // None = global max, Some(d) = max along dimension d
keepdim: bool, // Whether to keep dimension with size 1
}
impl Max {
/// Create a new max 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/max_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/max_dim_f64.wgsl").to_string()
});
&SHADER_DIM
}
/// Execute the max operation
/// # Errors
/// Returns [`Err`] if `dim` is out of range for the tensor shape, buffer
/// allocation 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 max 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("Max 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.create_uniform_buffer("Max Reduce Params", ¶ms);
ComputeDispatch::new(device, "max_reduce")
.shader(Self::wgsl_shader_reduce(), "main")
.storage_read(0, input_buffer)
.storage_rw(1, &output_buffer)
.uniform(2, ¶ms_buffer)
.dispatch(num_workgroups, 1, 1)
.submit()?;
// 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_max = partial_results
.iter()
.fold(f32::NEG_INFINITY, |a: f32, &b| a.max(b));
// Return scalar tensor
Ok(Tensor::new(vec![global_max], vec![], device.clone()))
}
Some(dim) => {
// Dimension-wise max 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("Max 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.create_uniform_buffer("Max Dim Params", ¶ms);
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);
ComputeDispatch::new(device, "max_dim")
.shader(Self::wgsl_shader_dim(), "main")
.storage_read(0, input_buffer)
.storage_rw(1, &output_buffer)
.uniform(2, ¶ms_buffer)
.dispatch(workgroups, 1, 1)
.submit()?;
// 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 maximum value (global reduction)
/// # Errors
/// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, or buffer
/// readback fails (e.g. device lost).
pub fn max(&self) -> Result<Self> {
Max::new(self.clone(), None, false).execute()
}
/// Find maximum value along a dimension
/// # Arguments
/// * `dim` - Dimension to find max 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, GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn max_dim(&self, dim: usize, keepdim: bool) -> Result<Self> {
Max::new(self.clone(), Some(dim), keepdim).execute()
}
/// Find maximum value (legacy method for backward compatibility)
/// # Errors
/// Returns [`Err`] if `dim` is out of range for the tensor shape, buffer
/// allocation fails, GPU dispatch fails, or buffer readback fails (e.g. device lost).
pub fn max_wgsl(self, dim: Option<usize>) -> Result<Self> {
match dim {
None => Max::new(self, None, false).execute(),
Some(d) => Max::new(self, Some(d), false).execute(),
}
}
}