Change the block size to 1024 during compilation. (#2460)

* The default block size of rocm is 256.The designed kenel has a block size of 512.Change the block size to 1024 during compilation.

* The default block size of rocm is 256.The designed kenel has a block size of 512.Change the block size to 1024 during compilation.

* The default block size of rocm is 256.The designed kenel has a block size of 512.Change the block size to 1024 during compilation.

* The default block size of rocm is 256.The designed kenel has a block size of 512.Change the block size to 1024 during compilation.

setup.py

@@ -284,6 +284,9 @@ def get_extensions():
             define_macros += [('MMCV_WITH_CUDA', None)]
             cuda_args = os.getenv('MMCV_CUDA_ARGS')
             extra_compile_args['nvcc'] = [cuda_args] if cuda_args else []
+            if is_rocm_pytorch and platform.system() != 'Windows':
+                extra_compile_args['nvcc'] += \
+                    ['--gpu-max-threads-per-block=1024']
             op_files = glob.glob('./mmcv/ops/csrc/pytorch/*.cpp') + \
                 glob.glob('./mmcv/ops/csrc/pytorch/cpu/*.cpp') + \
                 glob.glob('./mmcv/ops/csrc/pytorch/cuda/*.cu') + \