Try fixing potential runtime errors when running DNN benchmarks

hpvm/CMakeLists.txt

@@ -12,7 +12,6 @@ include_directories(${CMAKE_CURRENT_BINARY_DIR})
 add_subdirectory(lib)
 add_subdirectory(projects)
 add_subdirectory(tools)
-
 add_subdirectory(test)
 
 # Add a global check rule now that all subdirectories have been traversed

hpvm/lib/Transforms/CMakeLists.txt

@@ -8,3 +8,6 @@ add_subdirectory(DFG2LLVM_WrapperAPI)
 add_subdirectory(DFG2LLVM_CUDNN)
 add_subdirectory(FuseHPVMTensorNodes)
 add_subdirectory(InPlaceDFG)
+
+set(TENSOR_RT_PREFIX ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})
+set(TENSOR_RT_LL ${TENSOR_RT_PREFIX}/tensor_runtime.ll)

hpvm/test/dnn_benchmarks/benchmarks/alexnet/data/profile_info_0.txt

+Iteration 0
+tensorConvApprox 11.4549 0
+tensorAdd 0.987733 0
+tensorTanh 0.89782 0
+tensorPooling 3.92176 0
+tensorConvApprox 30.4695 0
+tensorAdd 0.756682 0
+tensorTanh 0.659855 0
+tensorPooling 3.87498 0
+tensorConvApprox 12.9027 0
+tensorAdd 0.529304 0
+tensorTanh 0.360859 0
+tensorConvApprox 14.1188 0
+tensorAdd 0.572512 0
+tensorTanh 0.269676 0
+tensorConvApprox 12.5122 0
+tensorAdd 0.37547 0
+tensorTanh 0.260915 0
+tensorPooling 2.68986 0
+tensorGemmGPU 0.215392 0
+tensorAdd 0.120663 0
+tensorSoftmax 0.134412 0
+
+Iteration Compute Time   : 98.0861
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.025326
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 1
+tensorConvApprox 14.367 0
+tensorAdd 1.01869 0
+tensorTanh 0.88308 0
+tensorPooling 3.88747 0
+tensorConvApprox 31.0697 0
+tensorAdd 0.775118 0
+tensorTanh 0.672938 0
+tensorPooling 3.17431 0
+tensorConvApprox 19.3971 0
+tensorAdd 0.528325 0
+tensorTanh 0.361839 0
+tensorConvApprox 14.727 0
+tensorAdd 0.367463 0
+tensorTanh 0.258956 0
+tensorConvApprox 15.0059 0
+tensorAdd 0.361187 0
+tensorTanh 0.258366 0
+tensorPooling 2.70526 0
+tensorGemmGPU 0.18939 0
+tensorAdd 0.068016 0
+tensorSoftmax 0.536969 0
+
+Iteration Compute Time   : 110.614
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.021939
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 2
+tensorConvApprox 10.204 0
+tensorAdd 1.0115 0
+tensorTanh 0.900257 0
+tensorPooling 3.90068 0
+tensorConvApprox 27.5914 0
+tensorAdd 0.822439 0
+tensorTanh 0.849694 0
+tensorPooling 2.97768 0
+tensorConvApprox 17.8808 0
+tensorAdd 0.55192 0
+tensorTanh 0.378285 0
+tensorConvApprox 13.1066 0
+tensorAdd 0.621008 0
+tensorTanh 0.258276 0
+tensorConvApprox 18.631 0
+tensorAdd 0.422421 0
+tensorTanh 0.274397 0
+tensorPooling 2.65083 0
+tensorGemmGPU 0.409805 0
+tensorAdd 0.068461 0
+tensorSoftmax 0.104962 0
+
+Iteration Compute Time   : 103.616
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.152392
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 3
+tensorConvApprox 16.3712 0
+tensorAdd 1.03894 0
+tensorTanh 0.891796 0
+tensorPooling 4.02378 0
+tensorConvApprox 25.7412 0
+tensorAdd 0.820657 0
+tensorTanh 0.702908 0
+tensorPooling 3.12982 0
+tensorConvApprox 17.68 0
+tensorAdd 0.543514 0
+tensorTanh 0.378761 0
+tensorConvApprox 12.42 0
+tensorAdd 0.408593 0
+tensorTanh 0.403677 0
+tensorConvApprox 14.9018 0
+tensorAdd 0.408336 0
+tensorTanh 0.269063 0
+tensorPooling 2.6564 0
+tensorGemmGPU 0.169787 0
+tensorAdd 0.057522 0
+tensorSoftmax 0.083927 0
+
+Iteration Compute Time   : 103.102
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.022148
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 4
+tensorConvApprox 9.93468 0
+tensorAdd 1.02456 0
+tensorTanh 0.899216 0
+tensorPooling 3.89126 0
+tensorConvApprox 25.9977 0
+tensorAdd 1.24468 0
+tensorTanh 0.704036 0
+tensorPooling 3.74854 0
+tensorConvApprox 17.644 0
+tensorAdd 0.55182 0
+tensorTanh 0.37049 0
+tensorConvApprox 12.5094 0
+tensorAdd 0.407149 0
+tensorTanh 0.248609 0
+tensorConvApprox 13.9976 0
+tensorAdd 0.403548 0
+tensorTanh 0.26787 0
+tensorPooling 2.65781 0
+tensorGemmGPU 0.163417 0
+tensorAdd 0.055534 0
+tensorSoftmax 0.079422 0
+
+Iteration Compute Time   : 96.8013
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.021165
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 5
+tensorConvApprox 9.80512 0
+tensorAdd 1.00169 0
+tensorTanh 0.921429 0
+tensorPooling 3.91348 0
+tensorConvApprox 25.9371 0
+tensorAdd 0.815127 0
+tensorTanh 0.673571 0
+tensorPooling 7.36816 0
+tensorConvApprox 16.8525 0
+tensorAdd 0.525325 0
+tensorTanh 0.375484 0
+tensorConvApprox 13.4568 0
+tensorAdd 0.413802 0
+tensorTanh 0.250907 0
+tensorConvApprox 14.0005 0
+tensorAdd 0.408961 0
+tensorTanh 0.248077 0
+tensorPooling 2.66095 0
+tensorGemmGPU 0.165503 0
+tensorAdd 0.056493 0
+tensorSoftmax 0.077052 0
+
+Iteration Compute Time   : 99.9281
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.021977
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 6
+tensorConvApprox 9.80471 0
+tensorAdd 0.999942 0
+tensorTanh 0.915084 0
+tensorPooling 3.91451 0
+tensorConvApprox 25.834 0
+tensorAdd 1.14345 0
+tensorTanh 0.74676 0
+tensorPooling 3.82146 0
+tensorConvApprox 17.6392 0
+tensorAdd 0.556454 0
+tensorTanh 0.370676 0
+tensorConvApprox 12.4922 0
+tensorAdd 0.404391 0
+tensorTanh 0.250653 0
+tensorConvApprox 14.002 0
+tensorAdd 0.403829 0
+tensorTanh 0.25152 0
+tensorPooling 2.66016 0
+tensorGemmGPU 0.161451 0
+tensorAdd 0.059232 0
+tensorSoftmax 0.076951 0
+
+Iteration Compute Time   : 96.5086
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.021092
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 7
+tensorConvApprox 9.79061 0
+tensorAdd 1.00411 0
+tensorTanh 0.923385 0
+tensorPooling 3.89759 0
+tensorConvApprox 25.887 0
+tensorAdd 0.821996 0
+tensorTanh 0.685319 0
+tensorPooling 3.12557 0
+tensorConvApprox 17.6648 0
+tensorAdd 0.555787 0
+tensorTanh 0.385425 0
+tensorConvApprox 12.5025 0
+tensorAdd 0.404708 0
+tensorTanh 0.25099 0
+tensorConvApprox 13.9948 0
+tensorAdd 0.403679 0
+tensorTanh 0.252795 0
+tensorPooling 2.65976 0
+tensorGemmGPU 0.154435 0
+tensorAdd 0.058066 0
+tensorSoftmax 0.102014 0
+
+Iteration Compute Time   : 95.5253
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.03922
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 8
+tensorConvApprox 10.3697 0
+tensorAdd 1.02741 0
+tensorTanh 0.912612 0
+tensorPooling 3.87946 0
+tensorConvApprox 26.2823 0
+tensorAdd 0.816944 0
+tensorTanh 0.686638 0
+tensorPooling 3.12319 0
+tensorConvApprox 17.918 0
+tensorAdd 0.530344 0
+tensorTanh 0.362426 0
+tensorConvApprox 12.8083 0
+tensorAdd 0.360747 0
+tensorTanh 0.257041 0
+tensorConvApprox 14.2165 0
+tensorAdd 0.528494 0
+tensorTanh 0.257368 0
+tensorPooling 2.53903 0
+tensorGemmGPU 0.170898 0
+tensorAdd 0.100506 0
+tensorSoftmax 0.119107 0
+
+Iteration Compute Time   : 97.2669
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.041931
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+Iteration 9
+tensorConvApprox 10.2616 0
+tensorAdd 0.98462 0
+tensorTanh 0.895782 0
+tensorPooling 3.9356 0
+tensorConvApprox 26.2674 0
+tensorAdd 0.786113 0
+tensorTanh 0.676243 0
+tensorPooling 3.163 0
+tensorConvApprox 23.1817 0
+tensorAdd 0.847734 0
+tensorTanh 0.367746 0
+tensorConvApprox 12.9463 0
+tensorAdd 0.371015 0
+tensorTanh 0.263012 0
+tensorConvApprox 15.1519 0
+tensorAdd 0.428313 0
+tensorTanh 0.28493 0
+tensorPooling 2.61769 0
+tensorGemmGPU 0.167254 0
+tensorAdd 0.056753 0
+tensorSoftmax 0.082013 0
+
+Iteration Compute Time   : 103.737
+Iteration Compute Energy : 0
+Iteration Control Time   : 0.022577
+Iteration Control Energy : 0
+Iteration Config Time   : 0
+Iteration Config Energy : 0
+Iteration End Frequency : 0
+
+
+
+
+Total Compute Time  : 1005.19
+Total Compute Energy: 6.91275e-310
+
+Total Control Time  : 0.389767
+Total Control Energy: 3.51633e-315
+
+Total Config Time  : 3.51633e-315
+Total Config Energy: 3.51633e-315
+
+Total Time  : 1005.57
+Total Energy: 0