Compile for gpu use failed at GPU Objectives

[hsukiang]

Hello there,

I m trying to compile megahit for GPU usage. However it failed at this stage. I am using POWER8 server and CUDA 8.0 on Ubuntu16.04.
Am I missing a library and NVCC could not locate?

~/project/megahit# make use_gpu=1
"/usr/local/cuda-8.0/bin/nvcc" -arch=sm_35 lv2_gpu_functions.cu -Xptxas -v -Xcudafe -# -cuda --ptxas-options=-v -Xptxas -abi=no -mpowerpc64 -I"/usr/local/cuda-8.0/bin/../include" -I. -O3 -DTUNE_ARCH=350 -DTUNE_SIZE=4 -o .lv2_gpu_functions_4B_sm350_nvvm_8.0_abi_ppc64.cpp
nvcc fatal : 'powerpc64': expected a number
Makefile:253: recipe for target '.lv2_gpu_functions_4B_sm350_nvvm_8.0_abi_ppc64.cpp' failed
make: *** [.lv2_gpu_functions_4B_sm350_nvvm_8.0_abi_ppc64.cpp] Error 1

Thank you!

Best regards,
James

[aquaskyline]

Hi James, could you please use the CPU version? The CPU version using 8 cores is in fact faster than the GPU version.

[hsukiang]

Hello there!
Absolutely, I did use CPU alone and run an assembly examples. It picked up 192 threads and took 9.3 secs to complete. The question beckons:

1. for a newer GPU with more cores, does it still hold true?
2. Does it scale with larger sequencing data?

This exercise is to investigate if GPU acceleration still have speedup.
How much effort do you think is needed to get Megahit to compile for CUDA 8?

Thanks!

Best regards,
James

[aquaskyline]

1. Yes, for GTX 1080, it still hold true, the fact is that the bottleneck is still at the memory bandwidth.
2. Yes.

We stopped supporting the GPU version officially because the CPU version is indeed faster, and it has a larger user-base. I don't have a good estimation on how much effort you need to make it work with CUDA 8. From your error message I think you can make it work by removing the PPC related codes from the Makefile.

Good luck.