add gelu vulkan operator

[FhqTreap]

add gelu vulkan operator

[codecov-commenter]

Codecov Report

Merging #5001 (2afe627) into master (9ecf6a6) will increase coverage by 0.00%.
Report is 2 commits behind head on master.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #5001   +/-   ##
=======================================
  Coverage   94.71%   94.72%           
=======================================
  Files         767      768    +1     
  Lines      228515   228610   +95     
=======================================
+ Hits       216445   216540   +95     
  Misses      12070    12070           

Files Changed	Coverage Δ
src/layer/vulkan/gelu_vulkan.cpp	100.00% <100.00%> (ø)

[nihui]

lgtm

[nihui]

Thanks for your contribution !

[DickyQi]

for MacOS with vulkan fp16 build failed.

[FhqTreap]

for MacOS with vulkan fp16 build failed.

more details on this?😮

[nihui]

for MacOS with vulkan fp16 build failed.

more details on this?😮

tanh fp16 on moltenvk has precision issue, so it is banned on half values

workaround follows (ref tanh comp)

#if NCNN_moltenvk
    v = afpvec4(tanh(vec4(v)));
#else
    v = tanh(v);
#endif

[FhqTreap]

for MacOS with vulkan fp16 build failed.

more details on this?😮

tanh fp16 on moltenvk has precision issue, so it is banned on half values

workaround follows (ref tanh comp)

#if NCNN_moltenvk
    v = afpvec4(tanh(vec4(v)));
#else
    v = tanh(v);
#endif

so it shuold look like this?

#if NCNN_moltenvk
    v = 0.5f * v * (1.0f + afp(tanh(float(0.79788452f * (v + 0.044715f * v * v * v)))));
#else
    v = 0.5f * v * (1.0f + tanh(0.79788452f * (v + 0.044715f * v * v * v)));
#endif

[nihui]

for MacOS with vulkan fp16 build failed.

more details on this?😮

tanh fp16 on moltenvk has precision issue, so it is banned on half values
workaround follows (ref tanh comp)

#if NCNN_moltenvk
    v = afpvec4(tanh(vec4(v)));
#else
    v = tanh(v);
#endif

so it shuold look like this?

#if NCNN_moltenvk
    v = 0.5f * v * (1.0f + afp(tanh(float(0.79788452f * (v + 0.044715f * v * v * v)))));
#else
    v = 0.5f * v * (1.0f + tanh(0.79788452f * (v + 0.044715f * v * v * v)));
#endif

yeah, exactly