refactor: Remove C shared memory shim. Refine shared memory utils beh…

…avior. Add unit test for shared memory (#797)

* test: add unit test for shared memory

* refactor: reflect latest comment on create / destroy behavior

* fix: fix variable shadowing

* chore: update function comment. address comment

* refactor: remove C shared memory shim

* style: add commnet and fix copyright

* fix: address comment

src/python/library/CMakeLists.txt

@@ -96,7 +96,6 @@ add_custom_target(
 if (NOT WIN32)
   # Can generate linux specific wheel file on linux systems only.
   set(LINUX_WHEEL_DEPENDS
-        cshm
         ${WHEEL_DEPENDS}
   )
 

src/python/library/build_wheel.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-# Copyright 2021-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -174,10 +174,6 @@ def sed(pattern, replace, source, dest=None):
             "tritonclient/utils/shared_memory",
             os.path.join(FLAGS.whl_dir, "tritonclient/utils/shared_memory"),
         )
-        shutil.copyfile(
-            "tritonclient/utils/libcshm.so",
-            os.path.join(FLAGS.whl_dir, "tritonclient/utils/shared_memory/libcshm.so"),
-        )
         cpdir(
             "tritonclient/utils/cuda_shared_memory",
             os.path.join(FLAGS.whl_dir, "tritonclient/utils/cuda_shared_memory"),

src/python/library/setup.py

@@ -76,8 +76,6 @@ def req_file(filename, folder="requirements"):
 extras_require["all"] = list(chain(extras_require.values()))
 
 platform_package_data = []
-if PLATFORM_FLAG != "any":
-    platform_package_data += ["libcshm.so"]
 
 data_files = [
     ("", ["LICENSE.txt"]),

src/python/library/tests/test_shared_memory.py

@@ -0,0 +1,183 @@
+# Copyright 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import unittest
+
+import numpy
+import tritonclient.utils as utils
+import tritonclient.utils.shared_memory as shm
+
+
+class SharedMemoryTest(unittest.TestCase):
+    """
+    Testing shared memory utilities
+    """
+
+    def setUp(self):
+        self.shm_handles = []
+
+    def tearDown(self):
+        for shm_handle in self.shm_handles:
+            shm.destroy_shared_memory_region(shm_handle)
+
+    def test_lifecycle(self):
+        cpu_tensor = numpy.ones([4, 4], dtype=numpy.float32)
+        byte_size = 64
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", byte_size)
+        )
+
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 1)
+
+        # Set data from Numpy array
+        shm.set_shared_memory_region(self.shm_handles[0], [cpu_tensor])
+        shm_tensor = shm.get_contents_as_numpy(
+            self.shm_handles[0], numpy.float32, [4, 4]
+        )
+
+        self.assertTrue(numpy.allclose(cpu_tensor, shm_tensor))
+
+        shm.destroy_shared_memory_region(self.shm_handles.pop(0))
+
+    def test_invalid_create_shm(self):
+        # Raises error since tried to create invalid system shared memory region
+        with self.assertRaisesRegex(
+            shm.SharedMemoryException, "unable to create the shared memory region"
+        ):
+            self.shm_handles.append(
+                shm.create_shared_memory_region("dummy_data", "/dummy_data", -1)
+            )
+
+    def test_set_region_offset(self):
+        large_tensor = numpy.ones([4, 4], dtype=numpy.float32)
+        large_size = 64
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", large_size)
+        )
+        shm.set_shared_memory_region(self.shm_handles[0], [large_tensor])
+        small_tensor = numpy.zeros([2, 4], dtype=numpy.float32)
+        small_size = 32
+        shm.set_shared_memory_region(
+            self.shm_handles[0], [small_tensor], offset=large_size - small_size
+        )
+        shm_tensor = shm.get_contents_as_numpy(
+            self.shm_handles[0], numpy.float32, [2, 4], offset=large_size - small_size
+        )
+
+        self.assertTrue(numpy.allclose(small_tensor, shm_tensor))
+
+    def test_set_region_oversize(self):
+        large_tensor = numpy.ones([4, 4], dtype=numpy.float32)
+        small_size = 32
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", small_size)
+        )
+        with self.assertRaisesRegex(
+            shm.SharedMemoryException, "unable to set the shared memory region"
+        ):
+            shm.set_shared_memory_region(self.shm_handles[0], [large_tensor])
+
+    def test_duplicate_key(self):
+        # by default, return the same handle if existed, warning will be print
+        # if size is different
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", 32)
+        )
+        with self.assertRaisesRegex(
+            shm.SharedMemoryException,
+            "unable to create the shared memory region",
+        ):
+            self.shm_handles.append(
+                shm.create_shared_memory_region(
+                    "shm_name", "shm_key", 32, create_only=True
+                )
+            )
+
+        # Get handle to the same shared memory region but with larger size requested,
+        # check if actual size is checked
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", 64)
+        )
+
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 1)
+
+        large_tensor = numpy.ones([4, 4], dtype=numpy.float32)
+        with self.assertRaisesRegex(
+            shm.SharedMemoryException, "unable to set the shared memory region"
+        ):
+            shm.set_shared_memory_region(self.shm_handles[-1], [large_tensor])
+
+    def test_destroy_duplicate(self):
+        # destruction of duplicate shared memory region will occur when the last
+        # managed handle is destroyed
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 0)
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", 64)
+        )
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", 32)
+        )
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", 32)
+        )
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 1)
+
+        shm.destroy_shared_memory_region(self.shm_handles.pop(0))
+        shm.destroy_shared_memory_region(self.shm_handles.pop(0))
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 1)
+
+        shm.destroy_shared_memory_region(self.shm_handles.pop(0))
+        self.assertEqual(len(shm.mapped_shared_memory_regions()), 0)
+
+    def test_numpy_bytes(self):
+        int_tensor = numpy.arange(start=0, stop=16, dtype=numpy.int32)
+        bytes_tensor = numpy.array(
+            [str(x).encode("utf-8") for x in int_tensor.flatten()], dtype=object
+        )
+        bytes_tensor = bytes_tensor.reshape(int_tensor.shape)
+        bytes_tensor_serialized = utils.serialize_byte_tensor(bytes_tensor)
+        byte_size = utils.serialized_byte_size(bytes_tensor_serialized)
+
+        self.shm_handles.append(
+            shm.create_shared_memory_region("shm_name", "shm_key", byte_size)
+        )
+
+        # Set data from Numpy array
+        shm.set_shared_memory_region(self.shm_handles[0], [bytes_tensor_serialized])
+
+        shm_tensor = shm.get_contents_as_numpy(
+            self.shm_handles[0],
+            numpy.object_,
+            [
+                16,
+            ],
+        )
+
+        self.assertTrue(numpy.array_equal(bytes_tensor, shm_tensor))
+
+
+if __name__ == "__main__":
+    unittest.main()

src/python/library/tritonclient/utils/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright 2020-2023, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2020-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -28,20 +28,6 @@ configure_file(__init__.py __init__.py COPYONLY)
 configure_file(_dlpack.py _dlpack.py COPYONLY)
 configure_file(_shared_memory_tensor.py _shared_memory_tensor.py COPYONLY)
 
-if(NOT WIN32)
-  file(COPY shared_memory DESTINATION .)
-
-  #
-  # libcshm.so
-  #
-  add_library(cshm SHARED shared_memory/shared_memory.cc)
-  if(${TRITON_ENABLE_GPU})
-    target_compile_definitions(cshm PUBLIC TRITON_ENABLE_GPU=1)
-    target_link_libraries(cshm PUBLIC CUDA::cudart)
-  endif() # TRITON_ENABLE_GPU
-  target_link_libraries(cshm PRIVATE rt)
-endif() # WIN32
-
 if(NOT WIN32)
   configure_file(shared_memory/__init__.py shared_memory/__init__.py COPYONLY)
   configure_file(cuda_shared_memory/__init__.py cuda_shared_memory/__init__.py COPYONLY)