Problem about Hexagon HVX support

#1

Recently, I want to utilize HVX instructions when generating assemblys for Hexagon.

While I noticed the description in the TVM target, that floating point support in HVX requires LLVM 14+. image

My LLVM version is 12.0.1, and HVX support for int is really no problem, as I can see instructions like vmpyie in assembly.

However, even I update my LLVM to 14.0.6, HVX support for floating points still seems not implemented? I still don’t see any HVX instructions in the generated assemblys.

Was it because I missed sth. for the target?

#2

The cpu_ver in the target must be at least v68. Could you show how you construct the target (i.e. the parameters)?

#3

I construct the target by the following:

target = tvm.target.hexagon('v68', link_params=True, hvx=128)

I have also tried to set use_qfloat and use_ieee to true, but it doesn’t seem to bring any change.

BTW, do you have any documents on using hvx for reference?

#4

There shouldn’t be anything additional required. Are you using vectorize in your schedule, or are you relying on auto-vectorization in LLVM?

#5

I’m relying on auto-vectorization in LLVM.

#6

Could you try compiling for ‘v69’ instead, just to see if HVX code is generated?

#7

Sorry. Forgot to reply.

I think the problem is because there are no instructions for floating point in HVX yet.

For example, there is sfmpy in Hexagon but no vfmpy in HVX.

So after I quantize the floating point to fixed point, everything is ok.

I don’t know if this is the way you were intended, but it works.

BTW, I’ve also tried with target ‘v69’, and no HVX instruction was generated.