KBaba7/llama.cpp

1#version 4502 3#include "generic_head.comp"4#include "types.comp"5 6#extension GL_EXT_control_flow_attributes : enable7#define BLOCK_SIZE 5128 9layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;10 11layout (binding = 0) readonly buffer X {A_TYPE data_a[];};12layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};13 14shared FLOAT_TYPE sum[BLOCK_SIZE];15 16void main() {17    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;18    const uint tid = gl_LocalInvocationID.x;19 20    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp21 22    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {23        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);24        sum[tid] += xi * xi;25    }26 27    // sum up partial sums and write back result28    barrier();29    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {30        if (tid < s) {31            sum[tid] += sum[tid + s];32        }33        barrier();34    }35 36    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);37    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));38 39    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {40        data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));41    }42}43