[ILUVATAR_GPU] Update for compiling CUDAExtension for iluvatar (#73762)

Author: tianyuzhou668

paddle/common/backend_header.h

@@ -18,7 +18,7 @@
 #include <cuda.h>
 #endif
 
-#if defined(__CUDACC__) && CUDA_VERSION >= 11000
+#if defined(__CUDACC__) && (CUDA_VERSION >= 11000 || defined(PADDLE_WITH_COREX))
 #define PADDLE_CUDA_BF16
 #include <cuda_bf16.h>
 #endif

paddle/utils/flat_hash_map.h

@@ -684,7 +684,8 @@ class sherwood_v3_table : private EntryAlloc, private Hasher, private Equal {
 
   size_t num_buckets_for_reserve(size_t num_elements) const {
     return static_cast<size_t>(std::ceil(
-        num_elements / std::min(0.5, static_cast<double>(_max_load_factor))));
+        num_elements /
+        std::min((double)0.5, static_cast<double>(_max_load_factor))));
   }
   void rehash_for_other_container(const sherwood_v3_table &other) {
     rehash(

python/paddle/utils/cpp_extension/cpp_extension.py

@@ -467,6 +467,20 @@ def unix_custom_compile_single_file(
                         # {'nvcc': {}, 'cxx: {}}
                         if isinstance(cflags, dict):
                             cflags = cflags['hipcc']
+                    elif core.is_compiled_with_custom_device("iluvatar_gpu"):
+                        ixcc_cmd = os.path.join(
+                            os.getenv("COREX_HOME", "/usr/local/corex/"),
+                            'bin',
+                            'clang++',
+                        )
+                        if not os.path.isfile(ixcc_cmd):
+                            raise ValueError(
+                                "Corex compiler is unavailable, please set `COREX_HOME` to specify it."
+                            )
+                        self.set_executable('compiler_so', ixcc_cmd)
+                        # {'nvcc': {}, 'cxx: {}}
+                        if isinstance(cflags, dict):
+                            cflags = cflags['nvcc']
                     else:
                         assert (
                             CUDA_HOME is not None

python/paddle/utils/cpp_extension/extension_utils.py

@@ -394,6 +394,8 @@ def prepare_unix_cudaflags(cflags):
             *cflags,
             *get_rocm_arch_flags(cflags),
         ]
+    elif core.is_compiled_with_custom_device("iluvatar_gpu"):
+        cflags = [*COMMON_NVCC_FLAGS, '-fPIC', '-DPADDLE_WITH_COREX', *cflags]
     else:
         cflags = [
             *COMMON_NVCC_FLAGS,