char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

lhezopencl: refactor some ops, concat, repeat, tanh and scale (#19226)91ea44e89

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
774 B27 linesraw
1#pragma OPENCL EXTENSION cl_khr_fp16 : enable
2
3kernel void kernel_scale_f32(
4        global float * src0,
5        ulong offset0,
6        global float * dst,
7        ulong offsetd,
8        float scale,
9        float bias
10) {
11    src0 = (global float*)((global char*)src0 + offset0);
12    dst = (global float*)((global char*)dst + offsetd);
13    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
14}
15
16kernel void kernel_scale_f32_4(
17        global float4 * src0,
18        ulong offset0,
19        global float4 * dst,
20        ulong offsetd,
21        float scale,
22        float bias
23) {
24    src0 = (global float4*)((global char*)src0 + offset0);
25    dst = (global float4*)((global char*)dst + offsetd);
26    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
27}