luboslenco
22 lines
705 B
Plaintext
22 lines
705 B
Plaintext
#version 450
|
|
/*
|
|
* Gated residual add: out[s,h] += gate[h] * proj[s,h]
|
|
* gate is a per-channel f32 vector (the modulation gate, broadcast over seq).
|
|
* One invocation per (s, h).
|
|
*/
|
|
layout(local_size_x = 16, local_size_y = 16) in;
|
|
|
|
layout(std430, binding = 0) buffer Out { float outp[]; };
|
|
layout(std430, binding = 1) readonly buffer Gate { float gate[]; };
|
|
layout(std430, binding = 2) readonly buffer Proj { float proj[]; };
|
|
|
|
layout(push_constant) uniform PC { uint seq, hidden; } pc;
|
|
|
|
void main() {
|
|
uint hh = gl_GlobalInvocationID.x;
|
|
uint s = gl_GlobalInvocationID.y;
|
|
if (s >= pc.seq || hh >= pc.hidden) return;
|
|
uint idx = s * pc.hidden + hh;
|
|
outp[idx] += gate[hh] * proj[idx];
|
|
}
|