Fix ARM Neon compilation and MulAdd implementation #3
Conversation
|
Thanks, did you run the tests on MacOS ARM? I don't have a Mac device to test it myself |
|
Yes, I ran FastNoise2 (NewFastSIMD branch) on a MacBook with the M1 CPU. |
|
The enable_if on the mask register is to avoid ambiguous implicit conversions. The AVX512 m32 also uses it, without it there are compiler errors. What error are you getting with the enable_if? |
|
Hey guys. I've confirmed this indeed fixes Auburn/FastNoise2#161 I've confirmed this works with latest NewFastSIMD branch on Apple Silicon. Could this be merged? Thanks to all involved. |
|
|
||
| template<typename T = NativeType> | ||
| FS_FORCEINLINE Register( std::enable_if_t<OPTIMISE_FLOAT, T> v ) : native( v ) { } | ||
| FS_FORCEINLINE Register( NativeType v ) : native( v ) { } |
There was a problem hiding this comment.
This is possibly outdated. I was able to confirm the fix without this change.
|
Could you remake the change in a new PR without the change to enable_if. I don't have any apple silcone to test on, thanks |
3 participants
This fixes an incorrect implementation for FMA on NEON.
I also had issues compiling the library due to the the enable_if constraint on NEON f32 register constructor. Looking at the other implementations, they don't have the enable_if, but I wasn't entirely sure what it's purpose was.
Here's a before an after for comparison:
