[BUG] Incorrect behavior of asynchronous D2D copy

cupy.copyto misses to apply user specified stream. Please see following mini repro.

• Conditions
  • CuPy 5.4.0 and 6.0.0rc1
  • Ubuntu 16.04
  • CUDA 10.1
• Code to reproduce

import cupy as cp

def main():
  SIZE = 250000000
  a = cp.empty((1, SIZE), cp.float32)
  b = cp.empty((1, SIZE), cp.float32)
  my_stream = cp.cuda.stream.Stream()

  my_stream.use()
  cp.copyto(a, b) # <- it will use default stream but it should use `my_stream`
  my_stream.synchronize()

  my_stream.use()
  a[:,:] = b # <- use `my_stream`
  my_stream.synchronize()

if __name__ == '__main__':
  main()

• Error messages, stack traces, or logs nvvp log

[niboshi]

@takagi Thanks. If the synchronization only occurs once at the first dot, we can ignore it. The situation is similar to other functions that trigger kernel compilation on the fly (loading a CUDA module incurs synchronization too.).

[niboshi]

Closing, as cupy.copyto has been fixed in #2942 . @jeng1220 @takagi Thanks!