Fixing avx compilation issue

llvm/projects/hpvm-tensor-rt/CMakeLists.txt

@@ -8,7 +8,7 @@ if (CMAKE_BUILD_TYPE STREQUAL "Debug")
   message("Debug mode")
     set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=compute_60;-std=c++11;-g;-lineinfo;-Xcompiler;-ggdb;-lcurand)
 else()
-   set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=compute_60;-std=c++11;-O3;-DNDEBUG;-Xcompiler;-DNDEBUG;-lcurand)
+   set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_60,code=compute_60;-std=c++11;-DNDEBUG;-Xcompiler;-DNDEBUG;-lcurand)
 endif()
 
 set(CUDA_PROPAGATE_HOST_FLAGS OFF)
@@ -17,6 +17,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 " )
 
 add_definitions(-DNO_INJECTION)
+add_definitions(-DPROMISE_TUNER_ENABLED)
 if(USE_GFLAGS)
   add_definitions(-DUSE_GFLAGS)
 endif()
@@ -68,6 +69,15 @@ target_link_libraries(lenet_keras  tensor_runtime)
 add_executable(cifar_keras  dnn_sources/src/cifar_keras.cc)
 target_link_libraries(cifar_keras  tensor_runtime)
 
+add_executable(cifar_keras2  dnn_sources/src/cifar_keras2.cc)
+target_link_libraries(cifar_keras2  tensor_runtime)
+
+add_executable(cifar_keras3  dnn_sources/src/cifar_keras3.cc)
+target_link_libraries(cifar_keras3  tensor_runtime)
+
+add_executable(test_keras  dnn_sources/src/test_keras.cc)
+target_link_libraries(test_keras  tensor_runtime)
+
 
 # Half precision networks
 add_executable(fc2_half  dnn_sources/src/half/fc2_half.cc)
@@ -86,9 +96,33 @@ add_executable(lenet_keras_half  dnn_sources/src/half/lenet_keras_half.cc)
 target_link_libraries(lenet_keras_half  tensor_runtime)
 
 
+# GPU Layer sources
+add_executable(lenet_layers  dnn_sources/src/layers/lenet_layers.cc)
+target_link_libraries(lenet_layers  tensor_runtime)
+
+add_executable(lenet_layers2  dnn_sources/src/layers/lenet_layers2.cc)
+target_link_libraries(lenet_layers2  tensor_runtime)
+
+add_executable(test_layers  dnn_sources/src/layers/test_layers.cc)
+target_link_libraries(test_layers  tensor_runtime)
+
+add_executable(test_layers2  dnn_sources/src/layers/test_layers2.cc)
+target_link_libraries(test_layers2  tensor_runtime)
+
+add_executable(cifar10_layers  dnn_sources/src/layers/cifar10_layers.cc)
+target_link_libraries(cifar10_layers  tensor_runtime)
+
+
 # Promise API sources
 add_executable(lenet_keras_promise  dnn_sources/src/promise/lenet_keras_promise.cc)
 target_link_libraries(lenet_keras_promise  tensor_runtime)
 
 add_executable(fc4_clipped_promise  dnn_sources/src/promise/fc4_clipped_promise.cc)
 target_link_libraries(fc4_clipped_promise  tensor_runtime)
+
+add_executable(fc3_clipped_promise  dnn_sources/src/promise/fc3_clipped_promise.cc)
+target_link_libraries(fc3_clipped_promise  tensor_runtime)
+
+add_executable(fc2_clipped_promise  dnn_sources/src/promise/fc2_clipped_promise.cc)
+target_link_libraries(fc2_clipped_promise  tensor_runtime)
+