Switch to using the per-thread default stream instead of a stream per…

… task.

parla/cuda.py

@@ -3,6 +3,7 @@
 
 import logging
 from functools import wraps, lru_cache
+import os, sys
 from typing import Dict, List, Optional, Collection
 
 from parla import array
@@ -15,6 +16,14 @@
 
 logger = logging.getLogger(__name__)
 
+if 'cupy_backends' in sys.modules:
+    # TODO: This should be dynamically configurable. That needs to be fixed upstream though.
+    raise ImportError("cupy must be imported after parla.gpu for per-thread default stream configuration to work properly.")
+
+os.env["CUPY_CUDA_PER_THREAD_DEFAULT_STREAM"] = "1"
+# numba responds to this env var even if it has already been imported.
+os.env["NUMBA_CUDA_PER_THREAD_DEFAULT_STREAM"] = "1"
+
 try:
     import cupy
     import cupy.cuda
@@ -167,33 +176,19 @@ def get_array_module(self, a):
 # Integration with parla.environments
 
 class _GPUStacksLocal(threading.local):
-    _stream_stack: List[cupy.cuda.Stream]
     _device_stack: List[cupy.cuda.Device]
 
     def __init__(self):
         super(_GPUStacksLocal, self).__init__()
-        self._stream_stack = []
         self._device_stack = []
 
-    def push_stream(self, stream):
-        self._stream_stack.append(stream)
-
-    def pop_stream(self) -> cupy.cuda.Stream:
-        return self._stream_stack.pop()
-
     def push_device(self, dev):
         self._device_stack.append(dev)
 
     def pop_device(self) -> cupy.cuda.Device:
         return self._device_stack.pop()
 
     @property
-    def stream(self):
-        if self._stream_stack:
-            return self._stream_stack[-1]
-        else:
-            return None
-    @property
     def device(self):
         if self._device_stack:
             return self._device_stack[-1]
@@ -213,30 +208,20 @@ def __init__(self, descriptor: "GPUComponent", env: TaskEnvironment):
         # Use a stack per thread per GPU component just in case.
         self._stack = _GPUStacksLocal()
 
-    def _make_stream(self):
-        with self.gpu.cupy_device:
-            return cupy.cuda.Stream(null=False, non_blocking=True)
-
     def __enter__(self):
         _gpu_locals._gpus = self.gpus
         dev = self.gpu.cupy_device
         dev.__enter__()
         self._stack.push_device(dev)
-        stream = self._make_stream()
-        stream.__enter__()
-        self._stack.push_stream(stream)
         return self
 
     def __exit__(self, exc_type, exc_val, exc_tb):
         dev = self._stack.device
-        stream = self._stack.stream
         try:
-            stream.synchronize()
-            stream.__exit__(exc_type, exc_val, exc_tb)
+            cupy.cuda.get_current_stream().synchronize()
             _gpu_locals._gpus = None
             ret = dev.__exit__(exc_type, exc_val, exc_tb)
         finally:
-            self._stack.pop_stream()
             self._stack.pop_device()
         return ret
 
@@ -245,15 +230,15 @@ def initialize_thread(self) -> None:
             # Trigger cuBLAS/etc. initialization for this GPU in this thread.
             with cupy.cuda.Device(gpu.index) as device:
                 a = cupy.asarray([2.])
-                cupy.cuda.get_current_stream().synchronize()
-                with cupy.cuda.Stream(False, True) as stream:
-                    cupy.asnumpy(cupy.sqrt(a))
-                    device.cublas_handle
-                    device.cusolver_handle
-                    device.cusolver_sp_handle
-                    device.cusparse_handle
-                    stream.synchronize()
-                    device.synchronize()
+                stream = cupy.cuda.get_current_stream()
+                stream.synchronize()
+                cupy.asnumpy(cupy.sqrt(a))
+                device.cublas_handle
+                device.cusolver_handle
+                device.cusolver_sp_handle
+                device.cusparse_handle
+                stream.synchronize()
+                device.synchronize()
 
 class GPUComponent(EnvironmentComponentDescriptor):
     """A single GPU CUDA component which configures the environment to use the specific GPU using a single
@@ -312,15 +297,15 @@ def initialize_thread(self) -> None:
             # Trigger cuBLAS/etc. initialization for this GPU in this thread.
             with cupy.cuda.Device(gpu.index) as device:
                 a = cupy.asarray([2.])
-                cupy.cuda.get_current_stream().synchronize()
-                with cupy.cuda.Stream(False, True) as stream:
-                    cupy.asnumpy(cupy.sqrt(a))
-                    device.cublas_handle
-                    device.cusolver_handle
-                    device.cusolver_sp_handle
-                    device.cusparse_handle
-                    stream.synchronize()
-                    device.synchronize()
+                stream = cupy.cuda.get_current_stream()
+                stream.synchronize()
+                cupy.asnumpy(cupy.sqrt(a))
+                device.cublas_handle
+                device.cusolver_handle
+                device.cusolver_sp_handle
+                device.cusparse_handle
+                stream.synchronize()
+                device.synchronize()
 
 
 class MultiGPUComponent(EnvironmentComponentDescriptor):