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// clamp 5//------------------------------------------------------------------------------ 6kernel void kernel_clamp( 7 global float * src0, 8 ulong offset0, 9 global float * dst, 10 ulong offsetd, 11 float min, 12 float max 13) { 14 src0 = (global float*)((global char*)src0 + offset0); 15 dst = (global float*)((global char*)dst + offsetd); 16 17 dst[get_global_id(0)] = src0[get_global_id(0)] < min ? 18 min : 19 (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]); 20}