AccelerateHS/accelerate

[BUG] CUDA-10 library doesn't support the Turing-based RTX 2060?

Closed this issue · 8 comments

sedwards-lab commented

Looks like the Turing-based RTX 2060 is too new for the cuda-10 library. It's apparently "sm_75" but the cuda-10 library only supports up to sm_72.

Trying to run the fwaccel-gpu example from Simon Marlow's book, which I patched to start with

{-# LANGUAGE CPP, BangPatterns #-}
{-# OPTIONS_GHC -Wall -fno-warn-name-shadowing #-}

module Main ( main, test {-, maxDistances -} ) where

import Prelude
import System.Environment
import Data.Array.Accelerate as A
import Data.Array.Accelerate.LLVM.PTX

(I removed the import of AccelerateCompat)

Runtime error:

$ stack run fwaccel-gpu -- 200
fwaccel-gpu:
*** Internal error in package accelerate ***
*** Please submit a bug report at https://github.com/AccelerateHS/accelerate/issues

ptxas - -o /home/sedwards/.cache/accelerate/accelerate-llvm-1.3.0.0/accelerate-llvm-ptx-1.3.0.0/llvm-hs-9.0.1/nvptx64-nvidia-cuda/sm75/rel/morp9202ae4c7dc569aa0ea6fb2c02b0efca3bfd980b6e440b3de34a094604d51ee0.sass -arch=sm_75 (exit 255)
ptxas fatal : Value 'sm_75' is not defined for option 'gpu-name'

CallStack (from HasCallStack):
internalError: Data.Array.Accelerate.LLVM.PTX.Compile:185:24
compileCUBIN: Data.Array.Accelerate.LLVM.PTX.Compile:123:20
compile: Data.Array.Accelerate.LLVM.PTX.Compile:88:22

If the bug is with the accelerate-llvm-ptx GPU backend, include the output of nvidia-device-query

resolver: lts-16.25

extra-deps:

  • accelerate-1.3.0.0
  • accelerate-llvm-ptx-1.3.0.0
  • accelerate-llvm-1.3.0.0
  • cuda-0.10.2.0
  • nvvm-0.10.0.0

resolver: lts-16.25

Ubuntu 18.04.5 LTS

GeForce RTX 2060

cuda-10

tmcdonell commented

I have a 2080ti so this definitely should work! What version of llvm are you using? I have llvm-9.

tmcdonell commented

Oh, I think you need cuda-10.1 installed (that's what I have) (or later, but I haven't updated the bindings for cuda-11* yet, sorry...). It seems the error is coming from ptxas, which is part of the CUDA toolchain.

sedwards-lab commented

I think I have cuda-10 installed, although I first started by installing cuda-11, which gave link errors with the Haskell libraries. And I'm also using LLVM 9 (standard under Ubuntu 18.04)

How does the system choose to send sm_75 to ptxas? I couldn't quickly find it in the code.

tmcdonell commented

In the file Target.hs. Your system is very similar to mine so it should work.

CUDA-10 indeed added support for sm_75: https://docs.nvidia.com/cuda/archive/10.0/cuda-toolkit-release-notes/index.html#cuda-general-new-features

Oh, if you upgraded your CUDA Toolkit after installing my CUDA bindings package, you might need to unregister (stack exec ghc-pkg -- unregister --force cuda) that package and reinstall it.

sedwards-lab commented
tmcdonell commented

(sorry for the late response). I'm running ubuntu 20.04 and CUDA 10.1.

$ lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description:    Ubuntu 20.04.1 LTS
Release:        20.04
Codename:       focal
sedwards-lab commented
tmcdonell commented

okay, closing for now then. let us know if you want any help and I'll do my best.