"access violation" problem

Reporting a bug

• I am using the latest released version of Numba
• I have included below a minimal working reproducer

Thank you all for developing numba! I’m now using numba 0.49.1 with cuda 10.2.89 on python 3.7.5, windows 10. I have a problem in my code that it… sometimes fail to execute due to OSError: exception: access violation reading 0x<non-zero-non-0xFF...address>.

cuda-memcheck.exe shows nothing.

The memory access violation occurs every time in the numba\cuda\cudadrv\nvvm.py in compile(self, **options), line 232, err = self.driver.nvvmCompileProgram(self._handle, len(opts), c_opts)

---------------------------------------------------------------------------
OSError                                   Traceback (most recent call last)
 in 
     39 arr2 = np.arange(16,dtype=np.int8)
     40 arr5 = np.zeros(16*16,dtype=np.int8)
---> 41 foo[2, 16](arr1,arr2,arr5)   # more threads than array elements

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in __call__(self, *args)
    756         Specialize and invoke this kernel with *args*.
    757         '''
--> 758         kernel = self.specialize(*args)
    759         cfg = kernel[self.griddim, self.blockdim, self.stream, self.sharedmem]
    760         cfg(*args)

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in specialize(self, *args)
    767         argtypes = tuple(
    768             [self.typingctx.resolve_argument_type(a) for a in args])
--> 769         kernel = self.compile(argtypes)
    770         return kernel
    771 

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in compile(self, sig)
    786             self.definitions[(cc, argtypes)] = kernel
    787             if self.bind:
--> 788                 kernel.bind()
    789         return kernel
    790 

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in bind(self)
    528         Force binding to current CUDA context
    529         """
--> 530         self._func.get()
    531 
    532     @property

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in get(self)
    405         cufunc = self.cache.get(device.id)
    406         if cufunc is None:
--> 407             ptx = self.ptx.get()
    408 
    409             # Link

V:\Program Files\Python\lib\site-packages\numba\cuda\compiler.py in get(self)
    376             arch = nvvm.get_arch_option(*cc)
    377             ptx = nvvm.llvm_to_ptx(self.llvmir, opt=3, arch=arch,
--> 378                                    **self._extra_options)
    379             self.cache[cc] = ptx
    380             if config.DUMP_ASSEMBLY:

V:\Program Files\Python\lib\site-packages\numba\cuda\cudadrv\nvvm.py in llvm_to_ptx(llvmir, **opts)
    493     cu.add_module(libdevice.get())
    494 
--> 495     ptx = cu.compile(**opts)
    496     # XXX remove debug_pubnames seems to be necessary sometimes
    497     return patch_ptx_debug_pubnames(ptx)

V:\Program Files\Python\lib\site-packages\numba\cuda\cudadrv\nvvm.py in compile(self, **options)
    229         c_opts = (c_char_p * len(opts))(*[c_char_p(x.encode('utf8'))
    230                                           for x in opts])
--> 231         err = self.driver.nvvmCompileProgram(self._handle, len(opts), c_opts)
    232         self._try_error(err, 'Failed to compile\n')
    233 

OSError: exception: access violation reading 0x000002087936AB58

If run the program again, it shows CudaAPIError: [500] Call to cuModuleGetFunction results in CUDA_ERROR_NOT_FOUND.

Here’s a reproducer works on my machine:

import numpy as np
import numba
from numba import cuda

sbox = np.array([0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],dtype=np.int8)
pbox = np.array([15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0],dtype=np.int8)

@cuda.jit(device=True)
def bar(txt,el,r):
  cudas_box=cuda.const.array_like(sbox)
  cudap_box=cuda.const.array_like(pbox)
  if r!=0:
    temp=cuda.local.array(16,numba.int8)

    #error do not occur if either for loop is removed
    for i in range(16):
      temp[i]=cudas_box[txt[cudap_box[i]]]
    for i in range(16):
      txt[i]=temp[i]
    
  return txt[10]+el

@cuda.jit(debug=True)
def foo(arr1,arr2,arr5):
  tx = cuda.threadIdx.x
  ty = cuda.blockIdx.x
  bw = cuda.blockDim.x
  pos = tx + ty * bw

  pos1=pos%16
  pos2=(pos//16)%16
  if pos<len(arr5):
    roundtext=cuda.local.array(16,numba.int8)
    for i in range(16):
      roundtext[i]=arr1[pos1][i]
    arr5[pos] = bar(roundtext,arr2[pos2],pos%2)

arr1 = np.arange(16,dtype=np.int8).repeat(16).reshape(16,16)
arr2 = np.arange(16,dtype=np.int8)
arr5 = np.zeros(16*16,dtype=np.int8)
foo[2, 16](arr1,arr2,arr5)

[gmarkall]

Thanks for linking to the answer. This issue was caused by some transformations that were done to the IR that made it invalid when passing it to NVVM. However, for errors in user code using the simulator can also be a good option. If the simulator is too slow, then one can also add debug=True,opt=False kwargs to the @cuda.jit decorator and run under compute-sanitizer (e.g. compute-sanitizer python bug.py - if there are any source lines with invalid accesses these will be reported. It’s not as interactive / deep a tool for CUDA Python code as the CUDA simulator, but it is a lot faster.

[gmarkall]

Thanks for the report and nice reproducer! I can reproduce this.