mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-12-27 03:44:35 +00:00
43 lines
1.2 KiB
Plaintext
43 lines
1.2 KiB
Plaintext
|
#version 450
|
||
|
|
||
|
#include "generic_head.comp"
|
||
|
#include "types.comp"
|
||
|
|
||
|
#extension GL_EXT_control_flow_attributes : enable
|
||
|
#define BLOCK_SIZE 512
|
||
|
|
||
|
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
|
||
|
|
||
|
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
|
||
|
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
|
||
|
|
||
|
shared FLOAT_TYPE sum[BLOCK_SIZE];
|
||
|
|
||
|
void main() {
|
||
|
const uint row = gl_WorkGroupID.x;
|
||
|
const uint tid = gl_LocalInvocationID.x;
|
||
|
|
||
|
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
|
||
|
|
||
|
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
|
||
|
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
|
||
|
sum[tid] += xi * xi;
|
||
|
}
|
||
|
|
||
|
// sum up partial sums and write back result
|
||
|
barrier();
|
||
|
[[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
|
||
|
if (tid < s) {
|
||
|
sum[tid] += sum[tid + s];
|
||
|
}
|
||
|
barrier();
|
||
|
}
|
||
|
|
||
|
const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
|
||
|
const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
|
||
|
|
||
|
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
|
||
|
data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
|
||
|
}
|
||
|
}
|