layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; | |
layout (binding = 0) readonly buffer A {float data_a[];}; | |
layout (binding = 0) readonly buffer A4 {vec4 data_a4[];}; | |
layout (binding = 1) writeonly buffer D {float data_d[];}; | |
layout (binding = 1) writeonly buffer D4 {vec4 data_d4[];}; | |
layout (push_constant) uniform parameter { | |
uint ne; | |
uint k_num; | |
} p; | |
void main() { | |
// Each invocation handles four consecutive components | |
const uint idx = gl_GlobalInvocationID.x * 4; | |
if (idx >= p.ne) { | |
return; | |
} | |
// Check if all four components are in bounds and aligned, | |
// then use vector loads | |
if (idx + 3 < p.ne && (p.ne % 4) == 0) { | |
vec4 result = vec4(0.0f); | |
[[unroll]] for (uint i = 0; i < p.k_num; i++) { | |
result += data_a4[(i * p.ne + idx) / 4]; | |
} | |
data_d4[idx / 4] = result; | |
} else { | |
[[unroll]] for (uint j = 0; j < 4; ++j) { | |
if (idx + j < p.ne) { | |
float result = 0.0f; | |
[[unroll]] for (uint i = 0; i < p.k_num; i++) { | |
result += data_a[i * p.ne + idx + j]; | |
} | |
data_d[idx + j] = result; | |
} | |
} | |
} | |
} | |