char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
80f19b418
char/aliceai-t5-moe
1#pragma OPENCL EXTENSION cl_khr_fp16 : enable 2 3//------------------------------------------------------------------------------ 4// silu 5//------------------------------------------------------------------------------ 6kernel void kernel_silu( 7 global float * src0, 8 ulong offset0, 9 global float * dst, 10 ulong offsetd 11) { 12 src0 = (global float*)((global char*)src0 + offset0); 13 dst = (global float*)((global char*)dst + offsetd); 14 15 float x = src0[get_global_id(0)]; 16 dst[get_global_id(0)] = x / (1.0f + exp(-x)); 17} 18 19kernel void kernel_silu_4( 20 global float4 * src0, 21 ulong offset0, 22 global float4 * dst, 23 ulong offsetd 24) { 25 src0 = (global float4*)((global char*)src0 + offset0); 26 dst = (global float4*)((global char*)dst + offsetd); 27 28 float4 x = src0[get_global_id(0)]; 29 dst[get_global_id(0)] = x / (1.0f + exp(-x)); 30}