HPVM Tensor Runtime

Dependencies

• CUDA >= 9.1

• cuDNN >= 7

Building Tensor Runtime

Tensor Runtime and the DNN sources using the Tensor runtime are built with the unified HPVM build system. These can also be separately built. HPVM Tensor Runtime can be built under the build directory as:

make -j ${NUM_THREADS} tensor_runtime

The tensor runtime is built as a static library under build/lib/liibtensor_runtime.a

TensorRT DNN Benchmarks

To assist development of tensor-based programs using only the tensor runtime, we include sources under dnn_sources that directly invoke the HPVM Tensor Runtime API calls for tensor operations, e.g., convolution, matrix multiplication, add, relu, among others.

Each benchmark can be build under your build directory as:

make -j ${NUM_THREADS} ${BENCHMARK}

Currently, 17 Benchmarks included:

_fp32 suffix denotes fp32 binaries - these use the FP32 API calls

_fp_16 suffix denotes fp16 binaries - these use FP16 (half precision) calls.