fix custom op makefile (apache#17516)

Makefile

@@ -667,18 +667,21 @@ pylint:
 	python3 -m pylint --rcfile=$(ROOTDIR)/ci/other/pylintrc --ignore-patterns=".*\.so$$,.*\.dll$$,.*\.dylib$$" python/mxnet
 
 # MXNet extension dynamically loading libraries
-EXT_LIBS = custom_op_lib subgraph_lib
+EXT_LIBS = build/libcustomop_lib.so build/libsubgraph_lib.so
 ifeq ($(USE_CUDA), 1)
-	EXT_LIBS += custom_op_gpu_lib
+        EXT_LIBS += build/libcustomop_gpu_lib.so
 endif
 extension_libs: $(EXT_LIBS)
 
-custom_op_lib:
-	$(CXX) -shared -fPIC -std=c++11 example/extensions/lib_custom_op/gemm_lib.cc -o build/libcustomop_lib.so -I include/mxnet
-custom_op_gpu_lib:
-	$(NVCC) -shared -std=c++11 -Xcompiler -fPIC example/extensions/lib_custom_op/relu_lib.cu -o build/libcustomop_gpu_lib.so -I include/mxnet
-subgraph_lib:
-	$(CXX) -shared -fPIC -std=c++11 example/extensions/lib_subgraph/subgraph_lib.cc -o build/libsubgraph_lib.so -I include/mxnet
+build/libcustomop_lib.so:
+	@mkdir -p $(@D)
+	$(CXX) -shared -fPIC -std=c++11 example/extensions/lib_custom_op/gemm_lib.cc -o $@ -I include/mxnet
+build/libcustomop_gpu_lib.so:
+	@mkdir -p $(@D)
+	$(NVCC) -shared -std=c++11 -Xcompiler -fPIC example/extensions/lib_custom_op/relu_lib.cu -o $@ -I include/mxnet
+build/libsubgraph_lib.so:
+	@mkdir -p $(@D)
+	$(CXX) -shared -fPIC -std=c++11 example/extensions/lib_subgraph/subgraph_lib.cc -o $@ -I include/mxnet
 
 # Cython build
 cython:

example/extensions/lib_custom_op/test_relu.py

@@ -65,6 +65,7 @@
 print("--------start testing larger ndarray---------")
 a = mx.nd.uniform(shape=(100,100,100), ctx=mx.cpu())
 b = mx.nd.uniform(shape=(100,100,100), ctx=mx.gpu())
+mx.nd.waitall()
 t1 = time.time()
 r1 = mx.nd.my_relu(a)
 mx.nd.waitall()