Support build with gcc12 for CUDA less than 12.0 (#50106)

cmake/experiments/cuda_module_loading_lazy.cmake

@@ -31,6 +31,11 @@ if(LINUX)
     message("cuda 11.7+ already support lazy module loading")
     return()
   endif()
+  if(${CUDA_VERSION} VERSION_LESS "11.2" AND ${CMAKE_CXX_COMPILER_VERSION}
+                                             VERSION_GREATER_EQUAL 12.0)
+    message("cuda less than 11.2 doesn't support gcc12")
+    return()
+  endif()
 
   message(
     "for cuda before 11.7, libcudart.so must be used for the lazy module loading trick to work, instead of libcudart_static.a"

cmake/external/gloo.cmake

@@ -31,6 +31,17 @@ set(GLOO_LIBRARIES
     "${GLOO_INSTALL_DIR}/lib/libgloo.a"
     CACHE FILEPATH "gloo library." FORCE)
 
+set(GLOO_PATCH_COMMAND "")
+if(WITH_GPU)
+  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0 AND ${CMAKE_CXX_COMPILER_VERSION}
+                                                  VERSION_GREATER 12.0)
+    file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/gloo/device.cc.patch
+         native_dst)
+    set(GLOO_PATCH_COMMAND patch -d ${GLOO_SOURCE_DIR}/gloo/transport/tcp <
+                           ${native_dst})
+  endif()
+endif()
+
 include_directories(${GLOO_INCLUDE_DIR})
 
 if(WITH_ASCEND OR WITH_ASCEND_CL)
@@ -59,6 +70,7 @@ else()
     GIT_TAG ${GLOO_TAG}
     PREFIX "${GLOO_PREFIX_DIR}"
     UPDATE_COMMAND ""
+    PATCH_COMMAND ${GLOO_PATCH_COMMAND}
     CONFIGURE_COMMAND ""
     BUILD_COMMAND
       mkdir -p ${GLOO_SOURCE_DIR}/build && cd ${GLOO_SOURCE_DIR}/build && cmake

cmake/external/protobuf.cmake

@@ -250,6 +250,12 @@ function(build_protobuf TARGET_NAME BUILD_FOR_HOST)
   else()
     set(PROTOBUF_REPOSITORY ${GIT_URL}/protocolbuffers/protobuf.git)
     set(PROTOBUF_TAG 9f75c5aa851cd877fb0d93ccc31b8567a6706546)
+    if(WITH_GPU)
+      if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0
+         AND ${CMAKE_CXX_COMPILER_VERSION} VERSION_GREATER 12.0)
+        set(PROTOBUF_TAG 2dc747c574b68a808ea4699d26942c8132fe2b09)
+      endif()
+    endif()
   endif()
   if(WITH_ARM_BRPC)
     set(ARM_PROTOBUF_URL
@@ -322,6 +328,12 @@ elseif(WITH_ARM_BRPC)
   set(PROTOBUF_VERSION 3.7.1-baidu-ee-common)
 else()
   set(PROTOBUF_VERSION 3.1.0)
+  if(WITH_GPU)
+    if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0
+       AND ${CMAKE_CXX_COMPILER_VERSION} VERSION_GREATER 12.0)
+      set(PROTOBUF_VERSION 3.16.0)
+    endif()
+  endif()
 endif()
 
 if(NOT PROTOBUF_FOUND)

cmake/external/warpctc.cmake

@@ -25,6 +25,19 @@ set(WARPCTC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/warpctc)
 set(WARPCTC_REPOSITORY ${GIT_URL}/baidu-research/warp-ctc.git)
 set(WARPCTC_TAG bdc2b4550453e0ef2d3b5190f9c6103a84eff184)
 
+set(WARPCTC_SOURCE_DIR ${THIRD_PARTY_PATH}/warpctc/src/extern_warpctc)
+set(WARPCTC_PATCH_COMMAND "")
+set(WARPCTC_CCBIN_OPTION "")
+if(NOT WIN32 AND WITH_GPU)
+  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0 AND ${CMAKE_CXX_COMPILER_VERSION}
+                                                  VERSION_GREATER 12.0)
+    file(TO_NATIVE_PATH
+         ${PADDLE_SOURCE_DIR}/patches/warpctc/CMakeLists.txt.patch native_src)
+    set(WARPCTC_PATCH_COMMAND patch -d ${WARPCTC_SOURCE_DIR} < ${native_src})
+    set(WARPCTC_CCBIN_OPTION -DCCBIN_COMPILER=${CCBIN_COMPILER})
+  endif()
+endif()
+
 set(WARPCTC_INCLUDE_DIR
     "${WARPCTC_INSTALL_DIR}/include"
     CACHE PATH "Warp-ctc Directory" FORCE)
@@ -112,7 +125,7 @@ else()
     GIT_TAG ${WARPCTC_TAG}
     PREFIX ${WARPCTC_PREFIX_DIR}
     UPDATE_COMMAND ""
-    PATCH_COMMAND ""
+    PATCH_COMMAND ${WARPCTC_PATCH_COMMAND}
     #BUILD_ALWAYS    1
     CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
                -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
@@ -132,7 +145,9 @@ else()
                -DBUILD_TESTS=OFF
                -DCMAKE_POSITION_INDEPENDENT_CODE=ON
                -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
+               -DCUDA_TOOLKIT_ROOT_DIR=${CUDA_TOOLKIT_ROOT_DIR}
                ${EXTERNAL_OPTIONAL_ARGS}
+               ${WARPCTC_CCBIN_OPTION}
     CMAKE_CACHE_ARGS
       -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
       -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON

paddle/fluid/framework/fleet/heter_ps/gpu_graph_utils.h

@@ -62,7 +62,7 @@ inline std::vector<int> shuffle_int_vector(int n) {
     e = i++;
   }
   std::shuffle(ret.begin(), ret.end(), random_engine_wrapper.engine);
-  return std::move(ret);
+  return ret;
 }
 
 #define CUDA_CHECK(cmd)                                                       \

paddle/fluid/operators/jit/CMakeLists.txt

@@ -39,14 +39,23 @@ cc_test(
   SRCS test.cc
   DEPS jit_kernel_helper)
 if(NOT WIN32)
-  cc_binary(
-    jit_kernel_benchmark
-    SRCS
-    benchmark.cc
-    DEPS
-    jit_kernel_helper
-    device_tracer
-    tensor)
+  set(cuda_less12_and_gcc_greater12 false)
+  if(DEFINED CMAKE_CUDA_COMPILER_VERSION)
+    if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0
+       AND ${CMAKE_CXX_COMPILER_VERSION} VERSION_GREATER 12.0)
+      set(cuda_less12_and_gcc_greater12 true)
+    endif()
+  endif()
+  if(NOT cuda_less12_and_gcc_greater12)
+    cc_binary(
+      jit_kernel_benchmark
+      SRCS
+      benchmark.cc
+      DEPS
+      jit_kernel_helper
+      device_tracer
+      tensor)
+  endif()
 endif()
 if(WITH_TESTING AND TEST jit_kernel_test)
   set_tests_properties(jit_kernel_test PROPERTIES TIMEOUT 120)

paddle/phi/tools/CMakeLists.txt

+if(WITH_GPU)
+  if(${CMAKE_CUDA_COMPILER_VERSION} LESS 12.0 AND ${CMAKE_CXX_COMPILER_VERSION}
+                                                  VERSION_GREATER 12.0)
+    return()
+  endif()
+endif()
+
 add_executable(print_phi_kernels print_phi_kernels.cc)
 target_link_libraries(print_phi_kernels phi phi_api_utils)
 if(WIN32)

patches/gloo/device.cc.patch

+--- a/device.cc
++++ b/device.cc
+@@ -12,6 +12,7 @@
+ #include <netdb.h>
+ #include <netinet/in.h>
+ #include <string.h>
++#include <array>
+
+ #include "gloo/common/linux.h"
+ #include "gloo/common/logging.h"

patches/warpctc/CMakeLists.txt.patch

+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -23,6 +23,10 @@
+ option(BUILD_SHARED "build warp-ctc shared library."  ON)
+ option(WITH_ROCM     "Compile PaddlePaddle with ROCM platform"             OFF)
+ 
++set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -ccbin ${CCBIN_COMPILER} -allow-unsupported-compiler")
++set(NVCC_PREPEND_FLAGS "${NVCC_PREPEND_FLAGS} -ccbin ${CCBIN_COMPILER} -allow-unsupported-compiler")
++set(CCBIN "-ccbin ${CCBIN_COMPILER} -allow-unsupported-compiler")
++
+ if(WITH_ROCM)
+     add_definitions(-DWARPCTC_WITH_HIP)
+     include(hip)