hanzo-ml 0.10.2

Minimalist ML framework.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

#version 450
// Fill a contiguous run of `n` 4-byte words with a constant bit pattern `val`:
// out[i] = val for i in [0, n). Powers const_set / Tensor::full / ones on a contiguous,
// offset-0, whole-buffer view, replacing a GPU->CPU->GPU read-modify-write with one
// on-GPU dispatch. The value is passed as raw bits (uint) so the SAME kernel serves f32
// and u32 storage bit-exactly (the host packs f32::to_bits() or the u32 directly), with
// no float canonicalization. Partial / strided / offset views still use the CPU path,
// which must preserve unaddressed elements.
layout(local_size_x = 64) in;

layout(set = 0, binding = 0) writeonly buffer Out { uint outp[]; };
layout(push_constant) uniform Pc {
    uint n;
    uint val; // raw 32-bit pattern to store
};

void main() {
    uint i = gl_GlobalInvocationID.x;
    if (i < n) {
        outp[i] = val;
    }
}