Python bindings for cuda_async_memory_resource

include/rmm/mr/device/cuda_async_memory_resource.hpp

@@ -50,7 +50,8 @@ class cuda_async_memory_resource final : public device_memory_resource {
     RMM_EXPECTS(e == cudaSuccess && v == 1,
                 "cudaMallocAsync not supported with this CUDA driver/runtime version");
 #else
-    RMM_FAIL("cudaMallocAsync not supported");
+    RMM_FAIL(
+      "cudaMallocAsync not supported by the version of the CUDA Toolkit used for this build");
 #endif
   }
 

python/rmm/_cuda/11.0/gpu.pxi → python/rmm/_cuda/11.x/gpu.pxi

@@ -1,5 +1,5 @@
 # Copyright (c) 2020, NVIDIA CORPORATION.
-# For CUDA 11.0
+# For CUDA 11.0, 11.1 and 11.2
 
 cdef extern from "cuda.h" nogil:
     cpdef enum cudaDeviceAttr:

python/rmm/_lib/memory_resource.pxd

@@ -27,6 +27,9 @@ cdef class CudaMemoryResource(DeviceMemoryResource):
 cdef class ManagedMemoryResource(DeviceMemoryResource):
     pass
 
+cdef class CudaAsyncMemoryResource(DeviceMemoryResource):
+    pass
+
 cdef class PoolMemoryResource(UpstreamResourceAdaptor):
     pass
 

python/rmm/_lib/memory_resource.pyx

@@ -38,6 +38,11 @@ cdef extern from "rmm/mr/device/managed_memory_resource.hpp" \
     cdef cppclass managed_memory_resource(device_memory_resource):
         managed_memory_resource() except +
 
+cdef extern from "rmm/mr/device/cuda_async_memory_resource.hpp" \
+        namespace "rmm::mr" nogil:
+    cdef cppclass cuda_async_memory_resource(device_memory_resource):
+        cuda_async_memory_resource() except +
+
 cdef extern from "rmm/mr/device/pool_memory_resource.hpp" \
         namespace "rmm::mr" nogil:
     cdef cppclass pool_memory_resource[Upstream](device_memory_resource):
@@ -99,6 +104,7 @@ cdef extern from "rmm/mr/device/per_device_resource.hpp" namespace "rmm" nogil:
     cdef device_memory_resource* _get_per_device_resource \
         "rmm::mr::get_per_device_resource"(cuda_device_id id)
 
+
 cdef class DeviceMemoryResource:
 
     cdef device_memory_resource* get_mr(self):
@@ -135,6 +141,17 @@ cdef class CudaMemoryResource(DeviceMemoryResource):
         pass
 
 
+cdef class CudaAsyncMemoryResource(DeviceMemoryResource):
+    """
+    Memory resource that uses cudaMallocAsync/Free for
+    allocation/deallocation
+    """
+    def __cinit__(self, device=None):
+        self.c_obj.reset(
+            new cuda_async_memory_resource()
+        )
+
+
 cdef class ManagedMemoryResource(DeviceMemoryResource):
     def __cinit__(self):
         self.c_obj.reset(

python/rmm/mr.py

@@ -1,6 +1,7 @@
 # Copyright (c) 2020, NVIDIA CORPORATION.
 from rmm._lib.memory_resource import (
     BinningMemoryResource,
+    CudaAsyncMemoryResource,
     CudaMemoryResource,
     DeviceMemoryResource,
     FixedSizeMemoryResource,
@@ -22,6 +23,7 @@
 
 __all__ = [
     "BinningMemoryResource",
+    "CudaAsyncMemoryResource",
     "CudaMemoryResource",
     "DeviceMemoryResource",
     "FixedSizeMemoryResource",

python/rmm/tests/test_rmm.py

@@ -21,6 +21,9 @@
 
 cuda.set_memory_manager(rmm.RMMNumbaManager)
 
+_driver_version = rmm._cuda.gpu.driverGetVersion()
+_runtime_version = rmm._cuda.gpu.runtimeGetVersion()
+
 
 @pytest.fixture(scope="function", autouse=True)
 def rmm_auto_reinitialize():
@@ -444,3 +447,34 @@ def test_mr_upstream_lifetime():
     # Delete cuda_mr first. Should be kept alive by pool_mr
     del cuda_mr
     del pool_mr
+
+
+@pytest.mark.skipif(
+    (_driver_version, _runtime_version) < (11020, 11020),
+    reason="cudaMallocAsync not supported",
+)
+@pytest.mark.parametrize("dtype", _dtypes)
+@pytest.mark.parametrize("nelem", _nelems)
+@pytest.mark.parametrize("alloc", _allocs)
+def test_cuda_async_memory_resource(dtype, nelem, alloc):
+    mr = rmm.mr.CudaAsyncMemoryResource()
+    rmm.mr.set_current_device_resource(mr)
+    assert rmm.mr.get_current_device_resource_type() is type(mr)
+    array_tester(dtype, nelem, alloc)
+
+
+@pytest.mark.skipif(
+    (_driver_version, _runtime_version) < (11020, 11020),
+    reason="cudaMallocAsync not supported",
+)
+@pytest.mark.parametrize("nelems", _nelems)
+def test_cuda_async_memory_resource_stream(nelems):
+    # test that using CudaAsyncMemoryResource
+    # with a non-default stream works
+    mr = rmm.mr.CudaAsyncMemoryResource()
+    rmm.mr.set_current_device_resource(mr)
+    stream = rmm._cuda.stream.Stream()
+    expected = np.full(nelems, 5, dtype="u1")
+    dbuf = rmm.DeviceBuffer.to_device(expected, stream=stream)
+    result = np.asarray(dbuf.copy_to_host())
+    np.testing.assert_equal(expected, result)

python/setup.py

@@ -1,4 +1,5 @@
 # Copyright (c) 2019-2020, NVIDIA CORPORATION.
+import filecmp
 import glob
 import os
 import re
@@ -63,22 +64,40 @@ def get_cuda_version_from_header(cuda_include_dir):
 # valid symbols for specific version of CUDA.
 
 cwd = os.getcwd()
-preprocess_files = ["gpu.pxd"]
-supported_cuda_versions = {"10.1", "10.2", "11.0"}
-
-for file_p in preprocess_files:
-    pxi_file = ".".join(file_p.split(".")[:-1])
-    pxi_file = pxi_file + ".pxi"
-
-    if CUDA_VERSION in supported_cuda_versions:
-        shutil.copyfile(
-            os.path.join(cwd, "rmm/_cuda", CUDA_VERSION, pxi_file),
-            os.path.join(cwd, "rmm/_cuda", file_p),
+files_to_preprocess = ["gpu.pxd"]
+
+# The .pxi file is unchanged between some CUDA versions
+# (e.g., 11.0 & 11.1), so we keep only a single copy
+# of it
+cuda_version_to_pxi_dir = {
+    "10.1": "10.1",
+    "10.2": "10.2",
+    "11.0": "11.x",
+    "11.1": "11.x",
+    "11.2": "11.x",
+}
+
+for pxd_basename in files_to_preprocess:
+    pxi_basename = os.path.splitext(pxd_basename)[0] + ".pxi"
+    if CUDA_VERSION in cuda_version_to_pxi_dir:
+        pxi_pathname = os.path.join(
+            cwd,
+            "rmm/_cuda",
+            cuda_version_to_pxi_dir[CUDA_VERSION],
+            pxi_basename,
         )
+        pxd_pathname = os.path.join(cwd, "rmm/_cuda", pxd_basename)
+        try:
+            if filecmp.cmp(pxi_pathname, pxd_pathname):
+                # files are the same, no need to copy
+                continue
+        except FileNotFoundError:
+            # pxd_pathname doesn't exist yet
+            pass
+        shutil.copyfile(pxi_pathname, pxd_pathname)
     else:
         raise TypeError(f"{CUDA_VERSION} is not supported.")
 
-
 try:
     nthreads = int(os.environ.get("PARALLEL_LEVEL", "0") or "0")
 except Exception: