rumus 0.3.0

A native-Rust deep learning framework with explicit memory safety and hardware acceleration
Documentation 
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

// SPDX-License-Identifier: Apache-2.0 OR MIT
//! JIT compiler for kernel fusion.
//!
//! Fuses element-wise operations (Add, Relu, Scale, etc.) into a single
//! dynamically generated WGSL kernel.  Autograd tape recording proceeds
//! normally — only the forward GPU dispatch is intercepted.
//!
//! Feature-gated behind `--features jit`.

pub mod cache;
pub mod codegen;
pub mod tracer;

pub use tracer::compile;

use std::sync::OnceLock;

use crate::backend::gpu::context::{GpuContext, STORAGE_USAGE};
use crate::jit::cache::{FusionKey, JitCache};
use crate::jit::tracer::FusionBlock;

use wgpu::util::{BufferInitDescriptor, DeviceExt};

static JIT_CACHE: OnceLock<JitCache> = OnceLock::new();

fn get_cache() -> &'static JitCache {
    JIT_CACHE.get_or_init(JitCache::new)
}

/// Uniform params for fused kernels: just numel.
#[repr(C)]
#[derive(Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
struct FusedParams {
    numel: u32,
    _p0: u32,
    _p1: u32,
    _p2: u32,
}

/// Flush a fusion block: codegen → cache lookup → dispatch → materialize outputs.
pub(crate) fn flush_block(block: FusionBlock) {
    if block.ops.is_empty() || block.num_outputs == 0 {
        return;
    }

    let ctx = GpuContext::get().expect("GPU required for JIT fusion");
    let cache = get_cache();

    // Build cache key and look up / compile.
    let key = FusionKey::from_block(&block);
    let cached = cache.get_or_compile(&key, &block, &ctx.device);

    // Allocate output GPU buffers.
    let mut output_buffers = Vec::with_capacity(block.num_outputs);
    for _ in 0..block.num_outputs {
        let buf = ctx.pool.acquire(
            &ctx.device,
            block.dtype.gpu_buf_size(block.numel),
            STORAGE_USAGE,
        );
        output_buffers.push(buf);
    }

    // Build uniform buffer.
    let params = FusedParams {
        numel: block.numel as u32,
        _p0: 0, _p1: 0, _p2: 0,
    };
    let params_buf = ctx.device.create_buffer_init(&BufferInitDescriptor {
        label: Some("jit_fused_params"),
        contents: bytemuck::bytes_of(&params),
        usage: wgpu::BufferUsages::UNIFORM,
    });

    // Build bind group entries.
    let mut bg_entries: Vec<wgpu::BindGroupEntry<'_>> = Vec::with_capacity(cached.total_bindings);

    // Input bindings.
    let input_guards: Vec<_> = block
        .input_storages
        .iter()
        .map(|s| {
            s.ensure_gpu();
            s.gpu_buffer()
        })
        .collect();
    for (i, guard) in input_guards.iter().enumerate() {
        bg_entries.push(wgpu::BindGroupEntry {
            binding: i as u32,
            resource: guard.as_entire_binding(),
        });
    }

    // Output bindings.
    for (i, buf) in output_buffers.iter().enumerate() {
        let binding = (block.num_inputs + i) as u32;
        bg_entries.push(wgpu::BindGroupEntry {
            binding,
            resource: buf.as_entire_binding(),
        });
    }

    // Uniform binding.
    let uniform_binding = (block.num_inputs + block.num_outputs) as u32;
    bg_entries.push(wgpu::BindGroupEntry {
        binding: uniform_binding,
        resource: params_buf.as_entire_binding(),
    });

    let bind_group = ctx.device.create_bind_group(&wgpu::BindGroupDescriptor {
        layout: &cached.layout,
        entries: &bg_entries,
        label: Some("jit_fused_bg"),
    });

    // Dispatch.
    let workgroups = (block.numel as u32 + 255) / 256;
    let mut encoder = ctx.device.create_command_encoder(&Default::default());
    {
        let mut pass = encoder.begin_compute_pass(&Default::default());
        pass.set_pipeline(&cached.pipeline);
        pass.set_bind_group(0, &bind_group, &[]);
        pass.dispatch_workgroups(workgroups, 1, 1);
    }
    ctx.queue.submit(std::iter::once(encoder.finish()));

    // Drop input guards before materializing outputs (release read locks).
    drop(input_guards);

    // Materialize deferred output storages with real GPU buffers.
    for (storage, buffer) in block.output_storages.iter().zip(output_buffers.into_iter()) {
        storage.materialize_gpu(buffer);
    }
}