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
// SPDX-License-Identifier: AGPL-3.0-or-later
//! GELU — GPU-resident, pipeline-cached, batchable
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Capability-based workgroup dispatch
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: No mocks or stubs
//! - Batchable: routes through `TensorContext::record_operation()`
//! - Zero-copy output: buffer pool, no GPU→CPU→GPU round-trip
//! - Pipeline cached: `GLOBAL_CACHE` eliminates recompilation overhead
/// GELU f64 WGSL shader source — canonical.
///
/// Uses `array<f64>` storage; requires GPU f64 support. Springs building
/// custom f64 pipelines via `ComputeDispatch` can use this directly.
pub const WGSL_GELU_F64: &str = include_str!("../shaders/activation/gelu_f64.wgsl");
pub(crate) static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| WGSL_GELU_F64.to_string());
use crate::device::pipeline_cache::{BindGroupLayoutSignature, GLOBAL_CACHE};
use crate::device::tensor_context::get_device_context;
use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};
#[repr(C)]
#[derive(Copy, Clone, Pod, Zeroable)]
struct Params {
size: u32,
}
/// GELU activation: `x · Φ(x)` where Φ is the Gaussian CDF.
pub struct GELU {
input: Tensor,
}
impl GELU {
/// Creates a GELU operation for the given input tensor.
#[must_use]
pub fn new(input: Tensor) -> Self {
Self { input }
}
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
/// Execute GELU.
///
/// - Output stays GPU-resident (no readback).
/// - Pipeline compiled once, cached globally.
/// - Dispatch batched when inside `TensorSession`.
///
/// # 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> {
let device = self.input.device();
let size: usize = self.input.shape().iter().product();
let ctx = get_device_context(device);
let caps = DeviceCapabilities::from_device(device);
let wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = (size as u32).div_ceil(wg_size);
// Pooled output — zero allocation in steady state.
let output_buffer = ctx.acquire_pooled_output(size);
let params_buf = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GELU Params"),
contents: bytemuck::bytes_of(&Params { size: size as u32 }),
usage: wgpu::BufferUsages::UNIFORM,
});
// Cached BGL: reduction() = (1 read-only, 1 read-write, 1 uniform)
let layout_sig = BindGroupLayoutSignature::reduction();
let adapter_info = device.adapter_info();
let bgl = GLOBAL_CACHE.get_or_create_layout(
device.device(),
adapter_info,
layout_sig,
Some("GELU BGL"),
);
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("GELU BG"),
layout: &bgl,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buf.as_entire_binding(),
},
],
});
let bind_group = std::sync::Arc::new(bind_group);
// Cached pipeline.
let pipeline = GLOBAL_CACHE.get_or_create_pipeline(
device.device(),
adapter_info,
Self::wgsl_shader(),
layout_sig,
"main",
Some("GELU Pipeline"),
);
// Record into TensorContext — immediate when not batching, deferred otherwise.
ctx.record_operation(move |encoder| {
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GELU Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&*bind_group), &[]);
pass.dispatch_workgroups(workgroups, 1, 1);
drop(params_buf); // keep alive until dispatch
})?;
Ok(Tensor::from_pooled_buffer(
output_buffer,
self.input.shape().to_vec(),
device.clone(),
))
}
}
impl Tensor {
/// Compute GELU element-wise (GPU-resident, pipeline-cached, batchable).
///
/// # Errors
///
/// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
/// readback fails (e.g. device lost or out of memory).
pub fn gelu_wgsl(self) -> Result<Self> {
GELU::new(self).execute()
}
}