ops_cuda fix race condition on cubin file read when testing with multiple cores (#1172)

tinygrad/runtime/ops_cuda.py

@@ -1,7 +1,5 @@
-import subprocess
+import subprocess, time, re, hashlib, tempfile
 from typing import Optional
-import time
-import re
 import numpy as np
 from pycuda.compiler import compile as cuda_compile # type: ignore
 from tinygrad.helpers import DEBUG, getenv, fromimport, colored
@@ -56,9 +54,10 @@ class CUDAProgram:
   def __init__(self, name:str, prg:str, binary=False):
     try:
       if DEBUG >= 6:
-        with open("/tmp/cubin", "wb") as f:
+        fn = f"{tempfile.gettempdir()}/tinycuda_{hashlib.md5(prg.encode('utf-8')).hexdigest()}"
+        with open(fn, "wb") as f:
           f.write(cuda_compile(prg, target="cubin", no_extern_c=True))
-        sass = subprocess.check_output(['nvdisasm', '/tmp/cubin']).decode('utf-8')
+        sass = subprocess.check_output(['nvdisasm', fn]).decode('utf-8')
         print(sass)
       if not binary: prg = cuda_compile(prg, target="ptx", no_extern_c=True, options=['-Wno-deprecated-gpu-targets']).decode('utf-8')
     except cuda.CompileError as e: