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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! GPU compute operations for Adam Optimizer
//!
//! This module contains the single-pass GPU execution for Adam optimizer
//! with bias correction.
use super::{Adam, AdamParams};
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
impl Adam {
/// Execute Adam optimizer step (GPU single-pass with bias correction)
///
/// **Deep Debt**: Efficient single-pass update with bias correction
///
/// Returns: (`new_params`, `new_m`, `new_v`)
///
/// # 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, Tensor, Tensor)> {
let device = self.params().device();
let size = self.params().shape().iter().product::<usize>();
let adam_params = AdamParams {
num_params: size as u32,
learning_rate: self.learning_rate(),
beta1: self.beta1(),
beta2: self.beta2(),
epsilon: 1e-8,
weight_decay: 0.0,
step: self.step() as u32,
};
// Create writable buffers (shader does in-place updates)
let zeros = vec![0.0f32; size];
// Copy params to writable buffer
let params_data = self.params().to_vec()?;
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("adam_params"),
contents: bytemuck::cast_slice(¶ms_data),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
});
// Copy or create m buffer
let m_data = if let Some(m_tensor) = self.m() {
m_tensor.to_vec()?
} else {
zeros.clone()
};
let m_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("adam_m"),
contents: bytemuck::cast_slice(&m_data),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
});
// Copy or create v buffer
let v_data = if let Some(v_tensor) = self.v() {
v_tensor.to_vec()?
} else {
zeros
};
let v_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("adam_v"),
contents: bytemuck::cast_slice(&v_data),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
});
let adam_params_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("adam_params"),
contents: bytemuck::cast_slice(&[adam_params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::shader(), Some("adam_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("adam_bind_group_layout"),
entries: &[
// binding 0: gradients (read)
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,
},
// binding 1: params (read_write)
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,
},
// binding 2: m (read_write)
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,
},
// binding 3: v (read_write)
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,
},
// binding 4: adam_params (uniform)
wgpu::BindGroupLayoutEntry {
binding: 4,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("adam_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("adam_pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("adam_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.gradients().buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: m_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: v_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: adam_params_buffer.as_entire_binding(),
},
],
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("adam_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("adam_pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&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::ElementWise);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
let updated_params = Tensor::from_buffer(
params_buffer,
self.params().shape().to_vec(),
device.clone(),
);
let updated_m =
Tensor::from_buffer(m_buffer, self.params().shape().to_vec(), device.clone());
let updated_v =
Tensor::from_buffer(v_buffer, self.params().shape().to_vec(), device.clone());
Ok((updated_params, updated_m, updated_v))
}
}