rumus 0.2.0

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
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150

// Convolution helper kernels: im2col, col2im, channel-wise bias.
//
// All operate on a SINGLE batch element.  The Rust dispatch layer
// loops over the batch dimension.

struct Im2ColParams {
    c_in: u32,
    h: u32,
    w: u32,
    k: u32,
    stride: u32,
    pad: u32,
    out_h: u32,
    out_w: u32,
}
// 8 * 4 = 32 bytes — multiple of 16 ✓

// --- im2col -----------------------------------------------------------------
// Input:  [C_in, H, W]          (binding 0)
// Output: [col_height, num_patches]  (binding 1)
//   col_height  = C_in * K * K
//   num_patches = out_h * out_w
// Each thread writes one element of the output matrix.

@group(0) @binding(0) var<storage, read>       im2col_input: array<f32>;
@group(0) @binding(1) var<storage, read_write> im2col_output: array<f32>;
@group(0) @binding(2) var<uniform>             im2col_params: Im2ColParams;

@compute @workgroup_size(64)
fn im2col_kernel(@builtin(global_invocation_id) gid: vec3<u32>) {
    let p = im2col_params;
    let col_height = p.c_in * p.k * p.k;
    let num_patches = p.out_h * p.out_w;
    let idx = gid.x;
    if (idx >= col_height * num_patches) { return; }

    let row = idx / num_patches;
    let col = idx % num_patches;

    let kw_val = row % p.k;
    let kh_val = (row / p.k) % p.k;
    let c      = row / (p.k * p.k);

    let oh = col / p.out_w;
    let ow = col % p.out_w;

    let ih_signed: i32 = i32(oh * p.stride + kh_val) - i32(p.pad);
    let iw_signed: i32 = i32(ow * p.stride + kw_val) - i32(p.pad);

    if (ih_signed >= 0 && ih_signed < i32(p.h) && iw_signed >= 0 && iw_signed < i32(p.w)) {
        let ih = u32(ih_signed);
        let iw = u32(iw_signed);
        im2col_output[idx] = im2col_input[c * p.h * p.w + ih * p.w + iw];
    } else {
        im2col_output[idx] = 0.0;
    }
}

// --- col2im -----------------------------------------------------------------
// Input:  [col_height, num_patches]  (binding 0)
// Output: [C_in, H, W]              (binding 1, pre-zeroed)
// Each thread handles one pixel of the output, accumulating from all
// overlapping patches.

@group(0) @binding(0) var<storage, read>       col2im_input: array<f32>;
@group(0) @binding(1) var<storage, read_write> col2im_output: array<f32>;
@group(0) @binding(2) var<uniform>             col2im_params: Im2ColParams;

@compute @workgroup_size(64)
fn col2im_kernel(@builtin(global_invocation_id) gid: vec3<u32>) {
    let p = col2im_params;
    let total = p.c_in * p.h * p.w;
    let idx = gid.x;
    if (idx >= total) { return; }

    let c  = idx / (p.h * p.w);
    let hw = idx % (p.h * p.w);
    let ih = hw / p.w;
    let iw = hw % p.w;

    let num_patches = p.out_h * p.out_w;
    var sum: f32 = 0.0;

    for (var kh: u32 = 0u; kh < p.k; kh++) {
        for (var kw: u32 = 0u; kw < p.k; kw++) {
            let ih_shifted_signed: i32 = i32(ih) + i32(p.pad) - i32(kh);
            let iw_shifted_signed: i32 = i32(iw) + i32(p.pad) - i32(kw);
            if (ih_shifted_signed < 0 || iw_shifted_signed < 0) { continue; }
            let ih_shifted = u32(ih_shifted_signed);
            let iw_shifted = u32(iw_shifted_signed);
            if (ih_shifted % p.stride != 0u || iw_shifted % p.stride != 0u) { continue; }
            let oh = ih_shifted / p.stride;
            let ow = iw_shifted / p.stride;
            if (oh >= p.out_h || ow >= p.out_w) { continue; }

            let row = c * p.k * p.k + kh * p.k + kw;
            let col_idx = oh * p.out_w + ow;
            sum += col2im_input[row * num_patches + col_idx];
        }
    }
    col2im_output[idx] = sum;
}

// --- add_channel_bias -------------------------------------------------------
// src:  [C, spatial]  (binding 0)   where spatial = H*W
// bias: [C]           (binding 1)
// out:  [C, spatial]  (binding 2)
// out[c * spatial + s] = src[c * spatial + s] + bias[c]

struct ChannelBiasParams {
    channels: u32,
    spatial: u32,
    _pad0: u32,
    _pad1: u32,
}
// 4 * 4 = 16 bytes ✓

@group(0) @binding(0) var<storage, read>       chb_src:  array<f32>;
@group(0) @binding(1) var<storage, read>       chb_bias: array<f32>;
@group(0) @binding(2) var<storage, read_write> chb_out:  array<f32>;
@group(0) @binding(3) var<uniform>             chb_params: ChannelBiasParams;

@compute @workgroup_size(64)
fn add_channel_bias_kernel(@builtin(global_invocation_id) gid: vec3<u32>) {
    let total = chb_params.channels * chb_params.spatial;
    let idx = gid.x;
    if (idx >= total) { return; }
    let c = idx / chb_params.spatial;
    chb_out[idx] = chb_src[idx] + chb_bias[c];
}

// --- sum_channel_bias_grad --------------------------------------------------
// Reduces [C, spatial] → [C] by summing over the spatial dimension.
// Each thread handles one channel.

@group(0) @binding(0) var<storage, read>       scbg_src: array<f32>;
@group(0) @binding(1) var<storage, read>       scbg_dummy: array<f32>;
@group(0) @binding(2) var<storage, read_write> scbg_out: array<f32>;
@group(0) @binding(3) var<uniform>             scbg_params: ChannelBiasParams;

@compute @workgroup_size(64)
fn sum_channel_bias_grad_kernel(@builtin(global_invocation_id) gid: vec3<u32>) {
    let c = gid.x;
    if (c >= scbg_params.channels) { return; }
    var sum: f32 = 0.0;
    for (var s: u32 = 0u; s < scbg_params.spatial; s++) {
        sum += scbg_src[c * scbg_params.spatial + s];
    }
    scbg_out[c] = sum;
}