arch: get shm max size on device

Author: mloubout

devito/arch/archinfo.py

@@ -1069,6 +1069,32 @@ def march(self):
                 return 'tesla'
         return None
 
+    @cached_property
+    def max_shm_per_block(self):
+        """
+        Get the maximum amount of shared memory per thread block
+        """
+        # Load libcudart
+        libname = ctypes.util.find_library("cudart")
+        if not libname:
+            raise RuntimeError("cudart library not found")
+        lib = ctypes.CDLL(libname)
+
+        cudaDevAttrMaxSharedMemoryPerBlockOptin = 97
+        # get current device
+        dev = ctypes.c_int()
+        lib.cudaGetDevice(ctypes.byref(dev))
+
+        # query attribute
+        value = ctypes.c_int()
+        lib.cudaDeviceGetAttribute(
+            ctypes.byref(value),
+            ctypes.c_int(cudaDevAttrMaxSharedMemoryPerBlockOptin),
+            dev
+        )
+
+        return value.value
+
     def supports(self, query, language=None):
         if language != 'cuda':
             return False
@@ -1125,6 +1151,8 @@ class AmdDevice(Device):
 
     max_mem_trans_nbytes = 256
 
+    max_shm_per_block = 64*1024  # 64 KB
+
     @cached_property
     def march(cls):
         # TODO: this corresponds to Vega, which acts as the fallback `march`