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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Argmax - Find indices of maximum values - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its dimension
//! - Zero hardcoding: All parameters passed at runtime
//! - 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;
/// Generic sort shader.
#[must_use]
pub fn wgsl_sort() -> &'static str {
static SHADER: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| include_str!("../shaders/misc/sort_f64.wgsl").to_string());
std::sync::LazyLock::force(&SHADER).as_str()
}
/// Argmax operation - Find indices of maximum values
pub struct Argmax {
input: Tensor,
dim: Option<usize>, // None = global argmax, Some(d) = argmax along dimension d
keepdim: bool, // Whether to keep dimension with size 1
}
impl Argmax {
/// Create a new argmax 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/argmax_reduce_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER_REDUCE).as_str()
}
/// 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/argmax_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER_DIM).as_str()
}
/// Execute the argmax 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 argmax 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 (indices)
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Argmax Reduce Output"),
size: (num_workgroups as usize * std::mem::size_of::<u32>()) 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("Argmax Reduce Params", ¶ms);
ComputeDispatch::new(device, "argmax_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()?;
// Read back partial results and find the global argmax on CPU
// We need to compare values at the partial indices to find the true global argmax
let partial_indices =
crate::utils::read_buffer_u32(device, &output_buffer, num_workgroups as usize)?;
// Read the input values at these indices to find the maximum
let input_data = device.read_buffer_f32(input_buffer, size)?;
let mut global_max_idx = 0u32;
let mut global_max_val = f32::NEG_INFINITY;
for &idx in &partial_indices {
if (idx as usize) < size {
let val = input_data[idx as usize];
if val > global_max_val {
global_max_val = val;
global_max_idx = idx;
}
}
}
// Return scalar tensor (single index)
Ok(Tensor::new(
vec![global_max_idx as f32],
vec![],
device.clone(),
))
}
Some(dim) => {
// Dimension-wise argmax 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("Argmax Dim Output"),
size: (output_size * std::mem::size_of::<u32>()) 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("Argmax 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, "argmax_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 =
crate::utils::read_buffer_u32(device, &output_buffer, output_size)?;
// Convert u32 to f32 for tensor
let output_f32: Vec<f32> = output_data.iter().map(|&x| x as f32).collect();
// 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_f32, output_shape, device.clone()))
}
}
}
}
impl Tensor {
/// Find index of maximum value (global reduction)
/// # Errors
/// Returns [`Err`] if buffer allocation fails, GPU dispatch fails, or buffer
/// readback fails (e.g. device lost).
pub fn argmax(&self) -> Result<Self> {
Argmax::new(self.clone(), None, false).execute()
}
/// Find indices of maximum values along a dimension (GPU/WGSL).
/// # 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 argmax_dim_keepdim(&self, dim: usize, keepdim: bool) -> Result<Self> {
Argmax::new(self.clone(), Some(dim), keepdim).execute()
}
/// Find indices of maximum values along a dimension (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 argmax_wgsl(self, dim: usize) -> Result<Self> {
Argmax::new(self, Some(dim), false).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_argmax_1d() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 5.0, 3.0, 2.0];
let input = Tensor::new(data, vec![4], device);
let output = input.argmax_wgsl(0).unwrap();
assert_eq!(output.shape(), &[] as &[usize]); // Scalar output
let result = output.to_vec().unwrap();
assert_eq!(result[0] as u32, 1); // Index of max value (5.0)
}
#[tokio::test]
async fn test_argmax_2d_dim0() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 5.0, 3.0, 2.0, 4.0, 6.0];
let input = Tensor::new(data, vec![3, 2], device);
let output = input.argmax_wgsl(0).unwrap();
assert_eq!(output.shape(), &[2]);
let result = output.to_vec().unwrap();
assert_eq!(result[0] as u32, 2); // Max in column 0 is at index 2 (value 4.0)
assert_eq!(result[1] as u32, 2); // Max in column 1 is at index 2 (value 6.0)
}
#[tokio::test]
async fn test_argmax_global() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 5.0, 3.0, 2.0];
let input = Tensor::new(data, vec![4], device);
let output = input.argmax().unwrap();
assert_eq!(output.shape(), &[] as &[usize]); // Scalar output
let result = output.to_vec().unwrap();
assert_eq!(result[0] as u32, 1); // Index of max value (5.0)
}
#[tokio::test]
async fn test_argmax_dim_keepdim() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 5.0, 3.0, 2.0, 4.0, 6.0];
let input = Tensor::new(data, vec![3, 2], device);
let output = input.argmax_dim_keepdim(0, true).unwrap();
assert_eq!(output.shape(), &[1, 2]); // Dimension kept
let result = output.to_vec().unwrap();
assert_eq!(result[0] as u32, 2); // Max in column 0 is at index 2
assert_eq!(result[1] as u32, 2); // Max in column 1 is at index 2
}
}