22 lines
707 B
Plaintext
22 lines
707 B
Plaintext
|
#version 450
|
||
|
|
/* Nearest-neighbour 2x upsample, batch=1: [C,H,W] -> [C,2H,2W]. */
|
||
|
|
layout(local_size_x = 256) in;
|
||
|
|
|
||
|
|
layout(std430, binding = 0) readonly buffer In { float x[]; };
|
||
|
|
layout(std430, binding = 1) writeonly buffer Out { float outp[]; };
|
||
|
|
|
||
|
|
layout(push_constant) uniform PC { uint channels, H, W; } pc;
|
||
|
|
|
||
|
|
void main() {
|
||
|
|
uint OW = pc.W * 2u, OH = pc.H * 2u;
|
||
|
|
uint total = pc.channels * OH * OW;
|
||
|
|
uint stride_g = gl_NumWorkGroups.x * 256u;
|
||
|
|
for (uint idx = gl_GlobalInvocationID.x; idx < total; idx += stride_g) {
|
||
|
|
uint ox = idx % OW;
|
||
|
|
uint t = idx / OW;
|
||
|
|
uint oy = t % OH;
|
||
|
|
uint c = t / OH;
|
||
|
|
outp[idx] = x[(c * pc.H + oy / 2u) * pc.W + ox / 2u];
|
||
|
|
}
|
||
|
|
}
|