diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/data/tuner_confs_base.txt b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/data/tuner_confs_base.txt new file mode 100644 index 0000000000000000000000000000000000000000..cf93cd1286cb6f1358a46cde5991d19ab451c78a --- /dev/null +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/data/tuner_confs_base.txt @@ -0,0 +1,21 @@ +19194.623482 ++++++ +conf1 1 1 72.84 0.0 +1 gpu conv fp32 11 add fp32 1 relu fp32 1 +2 gpu conv fp32 11 add fp32 1 relu fp32 1 pool_max fp32 1 +3 gpu conv fp32 11 add fp32 1 relu fp32 1 +4 gpu conv fp32 11 add fp32 1 relu fp32 1 pool_max fp32 1 +5 gpu conv fp32 11 add fp32 1 relu fp32 1 +6 gpu conv fp32 11 add fp32 1 relu fp32 1 +7 gpu conv fp32 11 add fp32 1 relu fp32 1 pool_max fp32 1 +8 gpu conv fp32 11 add fp32 1 relu fp32 1 +9 gpu conv fp32 11 add fp32 1 relu fp32 1 +10 gpu conv fp32 11 add fp32 1 relu fp32 1 pool_max fp32 1 +11 gpu conv fp32 11 add fp32 1 relu fp32 1 +12 gpu conv fp32 11 add fp32 1 relu fp32 1 +13 gpu conv fp32 11 add fp32 1 relu fp32 1 pool_max fp32 1 +14 gpu mul fp32 11 add fp32 1 relu fp32 1 +15 gpu mul fp32 11 add fp32 1 relu fp32 1 +16 gpu mul fp32 11 add fp32 1 +17 gpu softmax fp32 1 +----- diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/src/vgg16_imagenet_loop.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/src/vgg16_imagenet_loop.cpp index 164eba72adb4eca8e1a325d5bb6f61e9d5c01011..44f46b96afc828db2b163dc09d0192ad57a85f1a 100644 --- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/src/vgg16_imagenet_loop.cpp +++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_imagenet/src/vgg16_imagenet_loop.cpp @@ -1027,7 +1027,7 @@ int main(){ startMemTracking(); startProfiling(); - unsigned int batch_size = 100; + unsigned int batch_size = 20; unsigned int test_input_size = 500; unsigned int batch_count = test_input_size / batch_size; @@ -1037,11 +1037,10 @@ int main(){ int start = i * batch_size; int end = (i + 1) * batch_size; - void* input = readInputBatch(input_path.c_str(), 0, - start, end, - 3, 224, 224); + start, end, + 3, 224, 224); args->input = input; args->input_bytes = 0; @@ -1053,9 +1052,7 @@ int main(){ void *result = static_cast<RootIn*>(args)->input; hpvm_request_tensor(result, 0); - __visc__cleanup(); - - + llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end); freeBatchMemory(); @@ -1064,5 +1061,9 @@ int main(){ } + stopProfiling(); + __visc__cleanup(); + + return 0; }