update cmake for max blocksize and device judge

CMakeLists.txt

@@ -33,6 +33,7 @@ dgl_option(EXTERNAL_METIS_PATH "Path to external metis" OFF)
 dgl_option(EXTERNAL_METIS_LIB_PATH "Path to external metis library" OFF)
 dgl_option(EXTERNAL_GKLIB_PATH "Path to external gklib" OFF)
 
+
 # Options for building DGL features: "none," "dev," "dogfood," "release," and
 # "all."
 #    "none"  - The feature is OFF for all build types. This is used when
@@ -275,9 +276,6 @@ file(GLOB_RECURSE DGL_SRC_1
 )
 
 list(APPEND DGL_SRC ${DGL_SRC_1})
-if(NOT USE_HIP AND NOT USE_CUDA)
-  add_library(dgl SHARED ${DGL_SRC})
-endif()
 
 if (NOT MSVC)
   file(GLOB_RECURSE DGL_RPC_SRC src/rpc/*.cc)
@@ -296,6 +294,7 @@ if(USE_HIP)
   set(HIP_HIPCC_FLAGS "-std=c++17")
   add_library(dgl SHARED ${DGL_SRC})
   target_link_options(dgl PRIVATE "-Wl,--allow-multiple-definition")
+  target_compile_options(dgl PUBLIC "--gpu-max-threads-per-block=1024")
   # set_target_properties(dgl PROPERTIES LINKER_LANGUAGE hip)
   target_link_libraries(dgl ${DGL_LINKER_LIBS})
   target_include_directories(dgl PRIVATE "${CMAKE_SOURCE_DIR}/include/dgl")

graphbolt/CMakeLists.txt

@@ -77,18 +77,13 @@ target_link_libraries(${LIB_GRAPHBOLT_NAME} "${TORCH_LIBRARIES}")
 if(USE_HIP)
   # set_target_properties(${LIB_GRAPHBOLT_NAME} PROPERTIES CUDA_STANDARD 17)
   message(STATUS "Use external CCCL library for a consistent API and performance for graphbolt.")
+  target_compile_options(${LIB_GRAPHBOLT_NAME} PRIVATE "--gpu-max-threads-per-block=1024")
   target_include_directories(${LIB_GRAPHBOLT_NAME} PRIVATE
-  #                           #  "/opt/dgl_dep/hipcub-install-0915/include/"
-  #                           #  "/opt/dgl_dep/rocprim-install-0915/include/"
   #                            "${ROCM_PATH}/include/thrust"
                              "${ROCM_PATH}/include/hipcub"
                              "${ROCM_PATH}/include/rocprim"
   )
 
-  # target_include_directories(${LIB_GRAPHBOLT_NAME} PRIVATE
-  #                            "../third_party/cccl/thrust"
-  #                            "../third_party/cccl/cub"
-  #                            "../third_party/cccl/libcudacxx/include")
 
   message(STATUS "Use HugeCTR gpu_cache for graphbolt with INCLUDE_DIRS $ENV{GPU_CACHE_INCLUDE_DIRS}.")
   target_include_directories(${LIB_GRAPHBOLT_NAME} PRIVATE $ENV{GPU_CACHE_INCLUDE_DIRS})

src/partition/ndarray_partition.cc

@@ -107,7 +107,7 @@ class RangePartition : public NDArrayPartition {
         // we have only one CPU context, and can safely copy the array to that.
         range_cpu_(range.CopyTo(DGLContext{kDGLCPU, 0})) {
     auto ctx = range->ctx;
-    if (ctx.device_type != kDGLCUDA) {
+    if (ctx.device_type != kDGLCUDA || ctx.device_type != kDGLROCM) {
       LOG(FATAL) << "The range for an NDArrayPartition is only supported "
                     " on GPUs. Transfer the range to the target device before "
                     "creating the partition.";

src/runtime/cuda/cuda_device_api.cc

@@ -74,7 +74,6 @@ class CUDADeviceAPI final : public DeviceAPI {
         hipDeviceProp_t props;
         CUDA_CALL(hipGetDeviceProperties(&props, ctx.device_id));
         *rv = std::string(props.name);
-        printf("******* debug: device.name:%s\n ",std::string(props.name).c_str());
         return;
       }
       case kMaxClockRate: {