jblas causes UnsatisfiedLinkError on aarch64 (m1) mac

[sirno]

jblas doesn't support m1 (jblas-project/jblas#134 )as @tgvaughan mentioned in CompEvol/beast2#1047

This results in the following errror.

java.lang.reflect.InvocationTargetException
	at java.base/jdk.internal.reflect.DirectMethodHandleAccessor.invoke(DirectMethodHandleAccessor.java:119)
	at java.base/java.lang.reflect.Method.invoke(Method.java:577)
	at beast.app.beastapp.BeastLauncher.run(Unknown Source)
	at beast.app.beastapp.BeastLauncher.main(Unknown Source)
Caused by: java.lang.UnsatisfiedLinkError: 'void org.jblas.NativeBlas.dgemm(char, char, int, int, int, double, double[], int, int, double[], int, int, double, double[], int, int)'
	at org.jblas.NativeBlas.dgemm(Native Method)
	at org.jblas.SimpleBlas.gemm(SimpleBlas.java:247)
	at org.jblas.DoubleMatrix.mmuli(DoubleMatrix.java:1781)
	at org.jblas.DoubleMatrix.mmul(DoubleMatrix.java:3138)
	at org.jblas.MatrixFunctions.expm(MatrixFunctions.java:428)
	at multitypetree.operators.UniformizationRetypeOperator.getBranchTypeProb(Unknown Source)
	at multitypetree.operators.NodeShiftRetype.nonRootProposal(Unknown Source)
	at multitypetree.operators.NodeShiftRetype.proposal(Unknown Source)
	at beast.core.Operator.proposal(Unknown Source)
	at beast.core.MCMC.propagateState(Unknown Source)
	at beast.core.MCMC.doLoop(Unknown Source)
	at beast.core.MCMC.run(Unknown Source)
	at beast.app.BeastMCMC.run(Unknown Source)
	at beast.app.beastapp.BeastMain.<init>(Unknown Source)
	at beast.app.beastapp.BeastMain.main(Unknown Source)
	at java.base/jdk.internal.reflect.DirectMethodHandleAccessor.invoke(DirectMethodHandleAccessor.java:104)
	... 3 more
-- org.jblas INFO Deleting /var/folders/sl/b99vml7s45q3zx3wb1fqvq240000gp/T/jblas15640803248391520753

[tgvaughan]

Thanks Nicolas :)

[sirno]

Actually I just checked again, and it looks like version 1.2.5 should work on arm64, which is one minor version higher than what is packaged with mtt. jblas-project/jblas#118

[triple-jay]

Actually I just checked again, and it looks like version 1.2.5 should work on arm64, which is one minor version higher than what is packaged with mtt. jblas-project/jblas#118

Hi, I'm running into the same issue—were you able to solve this? I tried replacing the jblas 1.2.4 jar file with jblas 1.2.5 and still ran into the same issue.

[tgvaughan]

Hi @triple-jay, I've just released a new version of MTT which includes (and is compiled against) the new jblas. Maybe this helps get things working on your system?

[triple-jay]

Thank you for your quick response and help! Unfortunately, it still is giving the same error 🥲

[tgvaughan]

Sorry to hear this. :-/ I think we're a bit stuck in this case then as far as apple silicon goes. Unless somebody has time to port MTT to a new linear algebra library which is supported on that hardware, or to include a Java-only implementation of the matrix computations. (This latter approach would probably incur quite a speed penalty.)

For now it seems that MTT will have to remain an intel-only package :-(

[triple-jay]

Got it thank you so much for your help :)

[jugne]

Hi, you can solve this issue by using Java JDK or JRE for x86 64-bit architecture. It's the wrong architecture, but you'll be able to use it due to rosetta 2. Rosetta should be preinstalled, and in my experience, it's picked up automatically. Simply download JAVA for MacOS x86 64-bit as .zip, extract it to some non-default java location and point to it when running MTT.

This may be slower than using proper Java architecture, but will work for testing/developing purposes. I do not advise to use x86 64-bit Java as system default (so still have JAVA_HOME environment variable pointing to some proper Java architecture installation).

[rbouckaert]

Curiously enough, java for ARM is in my experience slower than Java for x86 translated with rosetta. Therefore, the recommended JRE for BEAST2 (and the one packaged in releases) is x86 based. It has been a while since I looked at it, so things may have changed.