From 52ccb25dcfecd26d983e9c9477718a347a62ac48 Mon Sep 17 00:00:00 2001 From: Elizabeth <hashim.sharif91@gmail.com> Date: Tue, 19 Nov 2019 00:44:20 -0600 Subject: [PATCH] Generated tuner confs files for federated autotuning --- .../alexnet/data/soc_data/alexnet3.txt | 91 ++ .../alexnet2/data/soc_data/alexnet2_3.txt | 100 +++ .../mobilenet/data/soc_data/mobilenet3.txt | 784 ++++++++++++++++++ .../resnet18/data/soc_data/resnet3.txt | 406 +++++++++ .../data/soc_data/vgg16_10_3.txt | 172 ++++ .../data/soc_data/vgg16_100_3.txt | 172 ++++ 6 files changed, 1725 insertions(+) create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/data/soc_data/alexnet3.txt create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet2/data/soc_data/alexnet2_3.txt create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/mobilenet/data/soc_data/mobilenet3.txt create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/data/soc_data/resnet3.txt create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/data/soc_data/vgg16_10_3.txt create mode 100644 llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar100/data/soc_data/vgg16_100_3.txt diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/data/soc_data/alexnet3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/data/soc_data/alexnet3.txt new file mode 100644 index 0000000000..e9118452b0 --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/data/soc_data/alexnet3.txt @@ -0,0 +1,91 @@ +4650.778334 ++++++ +conf1 1 1 78.051000325 1.048499512499994 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 7 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf2 1.0 1.000537467642168 78.113499925 0.9547501125000011 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 4 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf3 1.0 1.001729400037653 77.941999725 1.2120004124999966 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf4 1.0 1.001729400037653 77.92550035 1.2367494750000105 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf5 1.0 1.0 77.981501175 1.152748237499992 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 7 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf6 1.0 1.001729400037653 77.9790006 1.1564990999999907 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf7 1.0 0.9988207866576465 78.032500275 1.0762495874999942 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 4 +3 promise swing_level 6 +4 promise swing_level 4 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf8 1.0 1.001729400037653 78.0065002 1.1152496999999926 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- ++++++ +conf9 1.0 1.001729400037653 78.0029999 1.120500149999991 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 pool_max fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 promise swing_level 6 +6 gpu mul fp16 1 add fp16 1 +7 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet2/data/soc_data/alexnet2_3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet2/data/soc_data/alexnet2_3.txt new file mode 100644 index 0000000000..51ec86de02 --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet2/data/soc_data/alexnet2_3.txt @@ -0,0 +1,100 @@ +1485.1935210000001 ++++++ +conf1 1 1 83.475999725 1.9260004125000165 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 3 +3 promise swing_level 3 +4 promise swing_level 4 +5 gpu conv samp 33 add fp16 1 tanh fp16 1 +6 promise swing_level 6 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf2 1.5484556890138508 1.5615566950038826 82.863499425 2.844750862500007 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 6 +3 promise swing_level 5 +4 promise swing_level 3 +5 promise swing_level 6 +6 promise swing_level 6 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf3 1.0 1.0030044219908405 83.26149955 2.247750675000013 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 3 +3 promise swing_level 3 +4 promise swing_level 3 +5 gpu conv samp 33 add fp16 1 tanh fp16 1 +6 promise swing_level 3 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf4 1.0 1.0023992912963957 83.302000575 2.186999137500017 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 3 +3 promise swing_level 5 +4 promise swing_level 3 +5 gpu conv samp 33 add fp16 1 tanh fp16 1 +6 promise swing_level 3 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf5 0.9082116763340733 0.9025816179600147 83.179499625 2.3707505624999996 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 6 +3 gpu conv samp 31 add fp16 1 tanh fp16 1 +4 promise swing_level 3 +5 promise swing_level 7 +6 promise swing_level 3 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf6 1.5484556890138508 1.5603544259848947 82.779999175 2.970001237500007 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 5 +3 promise swing_level 5 +4 promise swing_level 5 +5 promise swing_level 6 +6 promise swing_level 7 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf7 0.9987465178894045 1.0048879316617951 83.6874994 1.6087509000000182 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 4 +3 promise swing_level 3 +4 promise swing_level 3 +5 gpu conv samp 35 add fp16 1 tanh fp16 1 +6 promise swing_level 4 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf8 0.9987465178894045 1.0053076710120923 83.55850015 1.802249775000007 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 3 +3 promise swing_level 4 +4 promise swing_level 3 +5 gpu conv samp 35 add fp16 1 tanh fp16 1 +6 promise swing_level 4 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- ++++++ +conf9 0.9987465178894045 1.0063585541989162 83.308 2.1779999999999973 +1 gpu conv fp16 1 add fp16 1 tanh fp16 1 +2 promise swing_level 3 +3 promise swing_level 3 +4 promise swing_level 3 +5 gpu conv samp 35 add fp16 1 tanh fp16 1 +6 promise swing_level 3 +7 gpu mul fp16 1 add fp16 1 +8 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/mobilenet/data/soc_data/mobilenet3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/mobilenet/data/soc_data/mobilenet3.txt new file mode 100644 index 0000000000..a840a581e8 --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/mobilenet/data/soc_data/mobilenet3.txt @@ -0,0 +1,784 @@ +10583.745193000002 ++++++ +conf1 1 1 81.9308327 2.3537509500000056 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 4 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 3 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 5 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 3 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 3 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 23 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 4 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 gpu conv perf 21 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 3 +84 gpu softmax fp32 1 +----- ++++++ +conf2 0.956346411385161 0.9523144131507536 81.851666825 2.472499762500007 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 gpu conv fp16 1 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 3 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 3 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 gpu conv perf 21 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 promise swing_level 5 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 promise swing_level 6 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 4 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 gpu conv samp 36 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 3 +84 gpu softmax fp32 1 +----- ++++++ +conf3 1.0240762074005454 1.022260386702397 81.89250005 2.411249925000007 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 4 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 3 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 5 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 4 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 3 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 22 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 7 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 5 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 4 +84 gpu softmax fp32 1 +----- ++++++ +conf4 0.9779579760207721 0.9717937012841189 81.76500005 2.6024999250000036 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 gpu conv fp16 1 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 5 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 3 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 4 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 5 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 promise swing_level 3 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 21 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 7 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 3 +84 gpu softmax fp32 1 +----- ++++++ +conf5 1.0240762074005454 1.0222006357142792 81.764166375 2.603750437499997 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 5 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 6 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 3 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 4 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 22 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 5 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 4 +84 gpu softmax fp32 1 +----- ++++++ +conf6 1.0067921101791497 1.0067619136195214 81.904999375 2.392500937499996 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 4 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 4 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 3 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 4 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 3 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 22 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 gpu conv perf 24 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 4 +84 gpu softmax fp32 1 +----- ++++++ +conf7 1.0592736051015887 1.0554868785054459 81.810832725 2.533750912500004 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 6 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 6 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 5 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 4 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 promise swing_level 6 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 5 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 4 +84 gpu softmax fp32 1 +----- ++++++ +conf8 1.0240762074005454 1.0222007380529154 81.839167875 2.4912481874999983 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 5 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 6 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 3 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 4 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 22 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 5 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 3 +84 gpu softmax fp32 1 +----- ++++++ +conf9 1.0240762074005454 1.0223933395037377 81.8899998 2.4150003000000027 +1 gpu conv fp16 1 +2 gpu batchnorm fp16 1 +3 gpu relu fp16 1 +4 gpu group_conv fp16 1 +5 gpu batchnorm fp16 1 +6 gpu relu fp16 1 +7 gpu conv fp16 1 +8 gpu batchnorm fp16 1 +9 gpu relu fp16 1 +10 gpu group_conv fp16 1 +11 gpu batchnorm fp16 1 +12 gpu relu fp16 1 +13 promise swing_level 5 +14 gpu batchnorm fp16 1 +15 gpu relu fp16 1 +16 gpu group_conv fp16 1 +17 gpu batchnorm fp16 1 +18 gpu relu fp16 1 +19 gpu conv fp16 1 +20 gpu batchnorm fp16 1 +21 gpu relu fp16 1 +22 gpu group_conv fp16 1 +23 gpu batchnorm fp16 1 +24 gpu relu fp16 1 +25 promise swing_level 3 +26 gpu batchnorm fp16 1 +27 gpu relu fp16 1 +28 gpu group_conv fp16 1 +29 gpu batchnorm fp16 1 +30 gpu relu fp16 1 +31 promise swing_level 5 +32 gpu batchnorm fp16 1 +33 gpu relu fp16 1 +34 gpu group_conv fp16 1 +35 gpu batchnorm fp16 1 +36 gpu relu fp16 1 +37 promise swing_level 3 +38 gpu batchnorm fp16 1 +39 gpu relu fp16 1 +40 gpu group_conv fp16 1 +41 gpu batchnorm fp16 1 +42 gpu relu fp16 1 +43 gpu conv fp16 1 +44 gpu batchnorm fp16 1 +45 gpu relu fp16 1 +46 gpu group_conv fp16 1 +47 gpu batchnorm fp16 1 +48 gpu relu fp16 1 +49 promise swing_level 3 +50 gpu batchnorm fp16 1 +51 gpu relu fp16 1 +52 gpu group_conv fp16 1 +53 gpu batchnorm fp16 1 +54 gpu relu fp16 1 +55 gpu conv perf 25 +56 gpu batchnorm fp16 1 +57 gpu relu fp16 1 +58 gpu group_conv fp16 1 +59 gpu batchnorm fp16 1 +60 gpu relu fp16 1 +61 promise swing_level 4 +62 gpu batchnorm fp16 1 +63 gpu relu fp16 1 +64 gpu group_conv fp16 1 +65 gpu batchnorm fp16 1 +66 gpu relu fp16 1 +67 gpu conv perf 22 +68 gpu batchnorm fp16 1 +69 gpu relu fp16 1 +70 gpu group_conv fp16 1 +71 gpu batchnorm fp16 1 +72 gpu relu fp16 1 +73 promise swing_level 3 +74 gpu batchnorm fp16 1 +75 gpu relu fp16 1 +76 gpu group_conv fp16 1 +77 gpu batchnorm fp16 1 +78 gpu relu fp16 1 +79 promise swing_level 3 +80 gpu batchnorm fp16 1 +81 gpu relu fp16 1 +82 gpu pool_mean fp16 1 +83 promise swing_level 4 +84 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/data/soc_data/resnet3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/data/soc_data/resnet3.txt new file mode 100644 index 0000000000..fc0f3c9932 --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/data/soc_data/resnet3.txt @@ -0,0 +1,406 @@ +5554.477625 ++++++ +conf1 1 1 87.516000375 2.825999437500002 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv fp16 1 add fp16 1 relu fp16 1 +3 promise swing_level 7 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 promise swing_level 6 +7 promise swing_level 6 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 7 +11 gpu conv perf 22 add fp16 1 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 7 +16 promise swing_level 3 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 5 +20 promise swing_level 7 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 promise swing_level 6 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 gpu conv fp16 1 add fp16 1 +29 promise swing_level 6 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 promise swing_level 5 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 5 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf2 0.8417170952921683 0.8361318137313652 87.5435004 2.7847494000000097 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv fp16 1 add fp16 1 relu fp16 1 +3 promise swing_level 7 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 promise swing_level 5 +7 promise swing_level 3 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 gpu conv perf 30 add fp16 1 relu fp16 1 +11 gpu conv perf 26 add fp16 1 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 7 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 gpu conv samp 32 add fp16 1 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 promise swing_level 6 +24 promise swing_level 5 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 5 +29 promise swing_level 3 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 promise swing_level 5 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 7 +37 promise swing_level 7 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf3 0.6622673984003916 0.6577405535611336 87.82049955 2.3692506750000177 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 7 +3 gpu conv fp16 1 add fp16 1 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 gpu conv perf 23 add fp16 1 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 7 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 6 +16 promise swing_level 6 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 gpu conv perf 25 add fp16 1 relu fp16 1 +20 gpu conv samp 32 add fp16 1 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 33 add fp16 1 relu fp16 1 +24 promise swing_level 6 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 promise swing_level 5 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 7 +37 promise swing_level 5 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf4 0.7382626993097082 0.7367062530412197 87.4914997 2.86275045 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv fp16 1 add fp16 1 relu fp16 1 +3 promise swing_level 5 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 gpu conv perf 22 add fp16 1 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 7 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 3 +16 gpu conv perf 23 add fp16 1 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 gpu conv samp 32 add fp16 1 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 36 add fp16 1 relu fp16 1 +24 promise swing_level 5 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 promise swing_level 6 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 promise swing_level 7 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 5 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf5 0.8907182071298466 0.8848096818830573 87.56749975 2.7487503750000144 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 7 +3 gpu conv fp16 1 add fp16 1 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 promise swing_level 3 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 3 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 6 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 promise swing_level 5 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 33 add fp16 1 relu fp16 1 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 promise swing_level 3 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 7 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf6 0.8907182071298466 0.8848884916645441 87.5164995 2.8252507500000164 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 7 +3 gpu conv fp16 1 add fp16 1 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 promise swing_level 3 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 3 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 6 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 promise swing_level 5 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 33 add fp16 1 relu fp16 1 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 promise swing_level 3 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 6 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf7 0.8742929289382473 0.8687435210865424 87.554999875 2.767500187500019 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 7 +3 gpu conv fp16 1 add fp16 1 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 promise swing_level 3 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 3 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 5 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 promise swing_level 5 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 33 add fp16 1 relu fp16 1 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 gpu conv fp16 1 add fp16 1 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 7 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf8 0.8877370342726504 0.8880914474537416 87.6829996 2.575500600000005 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv fp16 1 add fp16 1 relu fp16 1 +3 promise swing_level 7 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 26 add fp16 1 relu fp16 1 +7 promise swing_level 5 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 5 +11 promise swing_level 3 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 6 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 5 +20 promise swing_level 5 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 gpu conv samp 33 add fp16 1 relu fp16 1 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 5 +29 promise swing_level 3 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 7 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- ++++++ +conf9 0.964679225085488 0.9627569589243773 87.68000035 2.579999475000008 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv fp16 1 add fp16 1 relu fp16 1 +3 promise swing_level 7 +4 gpu add fp16 1 +5 gpu relu fp16 1 +6 gpu conv perf 25 add fp16 1 relu fp16 1 +7 promise swing_level 3 +8 gpu add fp16 1 +9 gpu relu fp16 1 +10 promise swing_level 6 +11 promise swing_level 3 +12 gpu add fp16 1 +13 gpu relu fp16 1 +14 gpu conv fp16 1 add fp16 1 relu fp16 1 +15 promise swing_level 6 +16 promise swing_level 5 +17 gpu add fp16 1 +18 gpu relu fp16 1 +19 promise swing_level 7 +20 promise swing_level 5 +21 gpu add fp16 1 +22 gpu relu fp16 1 +23 promise swing_level 5 +24 promise swing_level 3 +25 gpu add fp16 1 +26 gpu relu fp16 1 +27 gpu conv fp16 1 add fp16 1 relu fp16 1 +28 promise swing_level 7 +29 promise swing_level 3 +30 gpu add fp16 1 +31 gpu relu fp16 1 +32 gpu conv fp16 1 add fp16 1 relu fp16 1 +33 gpu conv fp16 1 add fp16 1 +34 gpu add fp16 1 +35 gpu relu fp16 1 +36 promise swing_level 5 +37 promise swing_level 6 +38 gpu add fp16 1 +39 gpu relu fp16 1 +40 gpu pool_max fp16 1 +41 gpu mul fp16 1 add fp16 1 +42 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/data/soc_data/vgg16_10_3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/data/soc_data/vgg16_10_3.txt new file mode 100644 index 0000000000..23921e98ea --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/data/soc_data/vgg16_10_3.txt @@ -0,0 +1,172 @@ +3849.973075 ++++++ +conf1 1 1 87.38999925 2.7450011249999946 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 5 +3 promise swing_level 6 +4 promise swing_level 3 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 5 +7 promise swing_level 6 +8 promise swing_level 3 +9 promise swing_level 3 +10 promise swing_level 3 +11 gpu conv samp 31 add fp16 1 relu fp16 1 +12 gpu conv perf 24 add fp16 1 relu fp16 1 +13 promise swing_level 7 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf2 1.0261976440908955 1.0308363738497694 87.3341672 2.8287492000000043 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 6 +3 promise swing_level 6 +4 promise swing_level 6 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 promise swing_level 3 +11 promise swing_level 3 +12 promise swing_level 5 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf3 0.8435853258572804 0.8555700473010334 87.302499475 2.876250787499991 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 3 +3 promise swing_level 6 +4 promise swing_level 3 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 3 +8 promise swing_level 3 +9 promise swing_level 3 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 gpu conv samp 31 add fp16 1 relu fp16 1 +12 promise swing_level 3 +13 promise swing_level 3 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf4 0.9311907031224316 0.9322568025003101 87.4466671 2.6599993499999997 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 7 +3 promise swing_level 7 +4 promise swing_level 5 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 5 +7 promise swing_level 3 +8 gpu conv samp 32 add fp16 1 relu fp16 1 +9 promise swing_level 5 +10 promise swing_level 3 +11 gpu conv perf 21 add fp16 1 relu fp16 1 +12 promise swing_level 5 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf5 0.5134548746824162 0.5167533689561125 87.4116671 2.7124993499999945 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 6 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 promise swing_level 6 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 5 +7 promise swing_level 5 +8 promise swing_level 3 +9 gpu conv samp 31 add fp16 1 relu fp16 1 +10 promise swing_level 5 +11 gpu conv fp16 1 add fp16 1 relu fp16 1 +12 gpu conv perf 28 add fp16 1 relu fp16 1 +13 promise swing_level 3 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf6 0.8901327291165676 0.8932467207725764 87.360832475 2.7887512875000056 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 5 +3 promise swing_level 7 +4 promise swing_level 5 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 5 +8 gpu conv samp 32 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 promise swing_level 3 +11 promise swing_level 3 +12 gpu conv perf 28 add fp16 1 relu fp16 1 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf7 0.5911190709711305 0.5972249232748029 87.390833475 2.7437497875000076 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 5 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 promise swing_level 3 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 5 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 promise swing_level 3 +11 promise swing_level 3 +12 gpu conv perf 28 add fp16 1 relu fp16 1 +13 promise swing_level 3 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf8 1.0261976440908955 1.0343635986793345 87.2600004 2.939999400000005 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 promise swing_level 5 +3 promise swing_level 6 +4 promise swing_level 3 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 5 +10 promise swing_level 3 +11 promise swing_level 6 +12 promise swing_level 6 +13 promise swing_level 3 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf9 0.447943678904132 0.453836321656264 87.305834375 2.871248437499993 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 28 add fp16 1 relu fp16 1 pool_max fp16 1 +3 promise swing_level 6 +4 promise swing_level 5 +5 gpu conv fp16 1 add fp16 1 relu fp16 1 +6 promise swing_level 3 +7 promise swing_level 6 +8 promise swing_level 3 +9 promise swing_level 3 +10 promise swing_level 3 +11 gpu conv fp16 1 add fp16 1 relu fp16 1 +12 gpu conv perf 28 add fp16 1 relu fp16 1 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar100/data/soc_data/vgg16_100_3.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar100/data/soc_data/vgg16_100_3.txt new file mode 100644 index 0000000000..da0f430bf1 --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar100/data/soc_data/vgg16_100_3.txt @@ -0,0 +1,172 @@ +11447.231556999996 ++++++ +conf1 1 1 66.988333075 2.132500387499995 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 5 +6 promise swing_level 3 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 6 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 promise swing_level 5 +12 promise swing_level 3 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf2 0.8982889313390673 0.8976137592411119 66.743334225 2.4999986624999977 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 26 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv samp 32 add fp16 1 relu fp16 1 +4 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 7 +6 promise swing_level 6 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 7 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 gpu conv perf 24 add fp16 1 relu fp16 1 +12 promise swing_level 6 +13 promise swing_level 3 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf3 0.8266384933930085 0.8228075750844537 66.798332725 2.417500912500003 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv samp 32 add fp16 1 relu fp16 1 +4 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 5 +6 promise swing_level 5 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 gpu conv perf 25 add fp16 1 relu fp16 1 +10 promise swing_level 3 +11 gpu conv fp16 1 add fp16 1 relu fp16 1 +12 gpu conv perf 25 add fp16 1 relu fp16 1 +13 gpu conv perf 23 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf4 0.8942472732473843 0.8909756722188954 66.579166775 2.746249837499988 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 26 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv samp 32 add fp16 1 relu fp16 1 +4 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 7 +6 promise swing_level 7 +7 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 5 +10 promise swing_level 3 +11 promise swing_level 6 +12 gpu conv perf 21 add fp16 1 relu fp16 1 +13 gpu conv samp 31 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf5 0.9444162598549138 0.9421190267448466 66.78249995 2.441250074999992 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 promise swing_level 5 +5 promise swing_level 3 +6 promise swing_level 6 +7 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 gpu conv perf 21 add fp16 1 relu fp16 1 pool_max fp16 1 +11 promise swing_level 5 +12 promise swing_level 5 +13 gpu conv samp 31 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf6 0.7434423618227001 0.7449596141840523 66.5350008 2.812498799999986 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 5 +6 promise swing_level 5 +7 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +8 gpu conv samp 32 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 promise swing_level 7 +11 gpu conv perf 24 add fp16 1 relu fp16 1 +12 promise swing_level 5 +13 gpu conv samp 31 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf7 0.898242268940609 0.8998966686330817 66.81000065 2.3999990249999854 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 26 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv samp 31 add fp16 1 relu fp16 1 +4 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 3 +6 promise swing_level 3 +7 promise swing_level 3 +8 gpu conv perf 25 add fp16 1 relu fp16 1 +9 promise swing_level 7 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 gpu conv perf 21 add fp16 1 relu fp16 1 +12 promise swing_level 3 +13 promise swing_level 5 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf8 0.9382354893566366 0.9379730375509677 66.6666662 2.615000700000003 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 5 +6 promise swing_level 6 +7 promise swing_level 3 +8 gpu conv samp 32 add fp16 1 relu fp16 1 +9 promise swing_level 3 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 gpu conv perf 21 add fp16 1 relu fp16 1 +12 promise swing_level 5 +13 gpu conv samp 31 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- ++++++ +conf9 0.7619679421937245 0.7656675022185832 67.00583275 2.1062508750000006 +1 gpu conv fp16 1 add fp16 1 relu fp16 1 +2 gpu conv perf 27 add fp16 1 relu fp16 1 pool_max fp16 1 +3 gpu conv fp16 1 add fp16 1 relu fp16 1 +4 gpu conv fp16 1 add fp16 1 relu fp16 1 pool_max fp16 1 +5 promise swing_level 3 +6 promise swing_level 5 +7 promise swing_level 5 +8 gpu conv samp 32 add fp16 1 relu fp16 1 +9 gpu conv samp 31 add fp16 1 relu fp16 1 +10 gpu conv samp 32 add fp16 1 relu fp16 1 pool_max fp16 1 +11 promise swing_level 5 +12 gpu conv perf 25 add fp16 1 relu fp16 1 +13 gpu conv samp 31 add fp16 1 relu fp16 1 pool_max fp16 1 +14 gpu mul fp16 1 add fp16 1 relu fp16 1 +15 gpu mul fp16 1 add fp16 1 +16 gpu softmax fp32 1 +----- -- GitLab