[PHI]Fix cub bugs when use PHI shared lib in CUDA12 (#57820)

YuanRisheng · web-flow · commit 8044ca4f71f5 · 2023-10-11T13:41:31.000+08:00
* fix cuda12 bugs

* fix windows bugs

* move cccl to third_party

* perfect code

* fix cuda bugs
diff --git a/.gitmodules b/.gitmodules
@@ -106,3 +106,7 @@
 	path = third_party/jitify
 	url = https://github.com/NVIDIA/jitify.git
 	ignore = dirty
+[submodule "third_party/cccl"]
+	path = third_party/cccl
+	url = https://github.com/NVIDIA/cccl.git
+	ignore = dirty
diff --git a/cmake/external/cccl.cmake b/cmake/external/cccl.cmake
@@ -0,0 +1,31 @@
+include(ExternalProject)
+
+set(CCCL_PATH
+    "${THIRD_PARTY_PATH}/cccl"
+    CACHE STRING "A path setting for external_cccl path.")
+set(CCCL_PREFIX_DIR ${CCCL_PATH})
+set(CCCL_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/cccl)
+
+# The latest commit has bugs in windows, so we set a fix commit.
+set(CCCL_TAG 1f6e4bcae0fbf1bbed87f88544d8d2161c490fc1)
+execute_process(COMMAND git --git-dir=${CCCL_SOURCE_DIR}/.git
+                        --work-tree=${CCCL_SOURCE_DIR} checkout ${CCCL_TAG})
+
+set(CCCL_INCLUDE_DIR ${CCCL_SOURCE_DIR})
+message("CCCL_INCLUDE_DIR is ${CCCL_INCLUDE_DIR}")
+include_directories(${CCCL_INCLUDE_DIR})
+
+ExternalProject_Add(
+  extern_cccl
+  ${EXTERNAL_PROJECT_LOG_ARGS}
+  SOURCE_DIR ${CCCL_SOURCE_DIR}
+  PREFIX ${CCCL_PREFIX_DIR}
+  UPDATE_COMMAND ""
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND ""
+  INSTALL_COMMAND ""
+  TEST_COMMAND "")
+
+add_library(cccl INTERFACE)
+
+add_dependencies(cccl extern_cccl)
diff --git a/cmake/generic.cmake b/cmake/generic.cmake
@@ -1345,6 +1345,9 @@ function(math_library TARGET)
   if(WITH_GPU)
     if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
       list(APPEND math_common_deps cub)
+    elseif(${CMAKE_CUDA_COMPILER_VERSION} EQUAL 12.0
+           OR ${CMAKE_CUDA_COMPILER_VERSION} GREATER 12.0)
+      list(APPEND math_common_deps cccl)
     else()
       list(APPEND math_common_deps)
     endif()
diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
@@ -375,6 +375,10 @@ if(WITH_GPU)
   if(${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0)
     include(external/cub) # download cub
     list(APPEND third_party_deps extern_cub)
+  elseif(${CMAKE_CUDA_COMPILER_VERSION} EQUAL 12.0
+         OR ${CMAKE_CUDA_COMPILER_VERSION} GREATER 12.0)
+    include(external/cccl)
+    list(APPEND third_party_deps extern_cccl)
   endif()
   set(URL
       "https://paddlepaddledeps.bj.bcebos.com/externalErrorMsg_20210928.tar.gz"
diff --git a/python/paddle/tensor/creation.py b/python/paddle/tensor/creation.py
@@ -2264,22 +2264,16 @@ def convert_scalar(x):
             )
             dtype = core.DataType.FLOAT32
 
-        if dtype == core.VarDesc.VarType.BOOL or dtype == core.DataType.BOOL:
+        if dtype in [core.VarDesc.VarType.BOOL, core.DataType.BOOL]:
             value_name = "bool_values"
             values = [int(v) for v in input.flat]
-        elif (
-            dtype == core.VarDesc.VarType.FP32 or dtype == core.DataType.FLOAT32
-        ):
+        elif dtype in [core.VarDesc.VarType.FP32, core.DataType.FLOAT32]:
             value_name = "fp32_values"
             values = [float(v) for v in input.flat]
-        elif (
-            dtype == core.VarDesc.VarType.INT32 or dtype == core.DataType.INT32
-        ):
+        elif dtype in [core.VarDesc.VarType.INT32, core.DataType.INT32]:
             value_name = "int32_values"
             values = [int(v) for v in input.flat]
-        elif (
-            dtype == core.VarDesc.VarType.INT64 or dtype == core.DataType.INT64
-        ):
+        elif dtype in [core.VarDesc.VarType.INT64, core.DataType.INT64]:
             value_name = "int64_values"
             values = [int(v) for v in input.flat]
         else:
diff --git a/third_party/cccl b/third_party/cccl
@@ -0,0 +1 @@
+Subproject commit 1f6e4bcae0fbf1bbed87f88544d8d2161c490fc1
