Initial commit

Author: fpetrini15

CMakeLists.txt

@@ -64,6 +64,13 @@ if(NOT TRITON_ENABLE_PYTHON_GRPC)
   set(TRITON_COMMON_ENABLE_PROTOBUF_PYTHON OFF)
 endif()
 
+# FIXME: Windows client currently does not support GPU tensors.
+# For simplicity, we will override this option here.
+if(WIN32 AND TRITON_ENABLE_GPU)
+  message("GPU shared memory is not currently supported by the Windows client. Forcing TRITON_ENABLE_GPU to false.")
+  set(TRITON_ENABLE_GPU OFF CACHE BOOL "GPU disabled" FORCE)
+endif()
+
 #
 # Dependencies
 #

src/python/library/CMakeLists.txt

@@ -38,9 +38,7 @@ if(${TRITON_ENABLE_PYTHON_HTTP})
   file(COPY tritonhttpclient DESTINATION .)
 endif() # TRITON_ENABLE_PYTHON_HTTP
 file(COPY tritonclientutils DESTINATION .)
-if (NOT WIN32)
-  file(COPY tritonshmutils DESTINATION .)
-endif() # NOT WIN32
+file(COPY tritonshmutils DESTINATION .)
 ####################################
 
 file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/TRITON_VERSION ${TRITON_VERSION})
@@ -90,6 +88,35 @@ add_custom_target(
     "${wheel_stamp_file}"
 )
 
+#
+# Windows-specific Wheel file.
+#
+if(WIN32)
+  set(WINDOWS_WHEEL_DEPENDS
+        cshm
+        ${WHEEL_DEPENDS}
+  )
+
+  if (${TRITON_ENABLE_PERF_ANALYZER})
+    set(perf_analyzer_arg --perf-analyzer ${CMAKE_INSTALL_PREFIX}/bin/perf_analyzer)
+  endif()
+  set(windows_wheel_stamp_file "windows_stamp.whl")
+  add_custom_command(
+    OUTPUT "${windows_wheel_stamp_file}"
+    COMMAND python3
+    ARGS
+      "${CMAKE_CURRENT_SOURCE_DIR}/build_wheel.py"
+      --dest-dir "${CMAKE_CURRENT_BINARY_DIR}/windows"
+      --windows
+      ${perf_analyzer_arg}
+    DEPENDS ${LINUX_WHEEL_DEPENDS}
+  )
+
+  add_custom_target(
+    windows-client-wheel ALL
+    DEPENDS
+      "${windows_wheel_stamp_file}"
+  )
 #
 # Linux specific Wheel file. Compatible with x86, x64 and aarch64
 #
@@ -147,14 +174,20 @@ if(${TRITON_ENABLE_PYTHON_GRPC})
   )
 endif() # TRITON_ENABLE_PYTHON_GRPC
 
-install(
-  CODE "file(GLOB _Wheel \"${CMAKE_CURRENT_BINARY_DIR}/generic/triton*.whl\")"
-  CODE "file(INSTALL \${_Wheel} DESTINATION \"${CMAKE_INSTALL_PREFIX}/python\")"
-)
 
-if (NOT WIN32)
+if(WIN32)
+  install(
+    CODE "file(GLOB _Wheel \"${CMAKE_CURRENT_BINARY_DIR}/windows/triton*.whl\")"
+    CODE "file(INSTALL \${_Wheel} DESTINATION \"${CMAKE_INSTALL_PREFIX}/python\")"
+  )
+elseif(NOT WIN32)
   install(
     CODE "file(GLOB _Wheel \"${CMAKE_CURRENT_BINARY_DIR}/linux/triton*.whl\")"
     CODE "file(INSTALL \${_Wheel} DESTINATION \"${CMAKE_INSTALL_PREFIX}/python\")"
   )
+else()
+  install(
+    CODE "file(GLOB _Wheel \"${CMAKE_CURRENT_BINARY_DIR}/generic/triton*.whl\")"
+    CODE "file(INSTALL \${_Wheel} DESTINATION \"${CMAKE_INSTALL_PREFIX}/python\")"
+  )
 endif() # NOT WIN32

src/python/library/build_wheel.py

@@ -28,6 +28,7 @@
 import argparse
 import os
 import pathlib
+import platform
 import re
 import shutil
 import subprocess
@@ -83,6 +84,12 @@ def sed(pattern, replace, source, dest=None):
         required=False,
         help="Include linux specific artifacts.",
     )
+    parser.add_argument(
+        "--windows",
+        action="store_true",
+        required=False,
+        help="Include windows specific artifacts.",
+    )
     parser.add_argument(
         "--perf-analyzer",
         type=str,
@@ -118,7 +125,7 @@ def sed(pattern, replace, source, dest=None):
         cpdir("tritonhttpclient", os.path.join(FLAGS.whl_dir, "tritonhttpclient"))
     if os.path.isdir("tritongrpcclient"):
         cpdir("tritongrpcclient", os.path.join(FLAGS.whl_dir, "tritongrpcclient"))
-    if FLAGS.linux:
+    if FLAGS.linux or FLAGS.windows:
         if os.path.isdir("tritonshmutils"):
             cpdir("tritonshmutils", os.path.join(FLAGS.whl_dir, "tritonshmutils"))
 
@@ -194,6 +201,21 @@ def sed(pattern, replace, source, dest=None):
             if not os.path.exists(os.path.join(FLAGS.whl_dir, "perf_client")):
                 os.symlink("perf_analyzer", os.path.join(FLAGS.whl_dir, "perf_client"))
 
+    if FLAGS.windows:
+        cpdir(
+            "tritonclient/utils/shared_memory",
+            os.path.join(FLAGS.whl_dir, "tritonclient/utils/shared_memory"),
+        )
+        shutil.copyfile(
+            "tritonclient/utils/Release/cshm.dll",
+            os.path.join(FLAGS.whl_dir, "tritonclient/utils/shared_memory/cshm.dll"),
+        )
+        # FIXME: Enable when Windows supports GPU tensors
+        # cpdir(
+        #     "tritonclient/utils/cuda_shared_memory",
+        #     os.path.join(FLAGS.whl_dir, "tritonclient/utils/cuda_shared_memory"),
+        # )
+
     shutil.copyfile("LICENSE.txt", os.path.join(FLAGS.whl_dir, "LICENSE.txt"))
     shutil.copyfile("setup.py", os.path.join(FLAGS.whl_dir, "setup.py"))
     cpdir("requirements", os.path.join(FLAGS.whl_dir, "requirements"))
@@ -208,6 +230,9 @@ def sed(pattern, replace, source, dest=None):
         else:
             platform_name = "manylinux1_x86_64"
         args = ["python3", "setup.py", "bdist_wheel", "--plat-name", platform_name]
+    elif FLAGS.windows and platform.uname().machine == "AMD64":
+        platform_name = "win_amd64"
+        args = ["python3", "setup.py", "bdist_wheel", "--plat-name", platform_name]
     else:
         args = ["python3", "setup.py", "bdist_wheel"]
 

src/python/library/setup.py

@@ -76,8 +76,10 @@ def req_file(filename, folder="requirements"):
 extras_require["all"] = list(chain(extras_require.values()))
 
 platform_package_data = []
-if PLATFORM_FLAG != "any":
+if "linux" in PLATFORM_FLAG:
     platform_package_data += ["libcshm.so"]
+elif PLATFORM_FLAG == "win_amd64":
+    platform_package_data += ["cshm.dll"]
 
 data_files = [
     ("", ["LICENSE.txt"]),

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

@@ -28,22 +28,23 @@ 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 .)
+file(COPY shared_memory DESTINATION .)
+#
+# libcshm.so / cshm.dll
+#
+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
 
-  #
-  # 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
+if(NOT WIN32)
   target_link_libraries(cshm PRIVATE rt)
 endif() # WIN32
 
-if(NOT WIN32)
-  configure_file(shared_memory/__init__.py shared_memory/__init__.py COPYONLY)
+configure_file(shared_memory/__init__.py shared_memory/__init__.py COPYONLY)
+
+if(${TRITON_ENABLE_GPU})
   configure_file(cuda_shared_memory/__init__.py cuda_shared_memory/__init__.py COPYONLY)
   configure_file(cuda_shared_memory/_utils.py cuda_shared_memory/_utils.py COPYONLY)
-endif() # NOT WIN32
+endif() # TRITON_ENABLE_GPU

src/python/library/tritonclient/utils/shared_memory/__init__.py

@@ -326,6 +326,7 @@ def __init__(self, err):
             -4: "unable to read/mmap the shared memory region",
             -5: "unable to unlink the shared memory region",
             -6: "unable to munmap the shared memory region",
+            -7: "unable to create file mapping",
         }
         self._msg = None
         if type(err) == str: