The repository contains a set of convolution tests and benchmarks for NVIDIA's cuDNN library.
This is not an officially supported Google product.
Install bazel (instructions).
Install the CUDA Toolkit (CUDA 8 is the minimal supported version)
Install the cuDNN SDK (cuDNN 6 is the minimal supported version).
Bazel parameters:
-
-c opt
Build with optimizations. Recommended for benchmarks. -
--action_env=CUDA_PATH=<path>
: Path to the CUDA SDK directory. Default is/usr/local/cuda
. -
--action_env=CUDNN_PATH=<path>
: Path to the CUDNN SDK directory. Default isCUDA_PATH
. -
--action_env=CC=<compiler>
: Name of (or path to) the compiler. Examples:clang
,gcc-6
.
Executable parameters:
-
--cuda_device=<device_id>
: CUDA device to use. Default is 0. -
--device_memory_limit_mb=<size>
: Upper limit of device memory (in megabytes) to use for cuDNN workspace after tensors have been allocated. Negative values specify an offset from the memory available at startup. Default is 4096.
bazel run [bazel parameters] //:cudnn_test -- [test parameters]
Bazel can run tests in a sandbox (which allows reporting crashes as failures). To run in a sandbox, replace 'bazel run ...' with 'bazel test ...' and prefix each test parameter with '--test_arg='
bazel test [bazel parameters] //:cudnn_test --test_arg=[test parameter 1] ...
Test parameters:
-
--gtest_filter=<pattern>
: Only run tests that match the given pattern. Example pattern:'*FWD*'
. -
--proto_path=<path>
: Path to textproto file that contains additional convolutions to run. Default is'cudnn_tests.textproto'
. -
--gtest_random_seed=<value>
: Seed for random generator. Changing the value produces tests with a different mix of tensor shapes, filter sizes, etc. Default is 0. -
--gtest_also_run_disabled_tests
: Include disabled tests (i.e. tests with names that start withDISABLED_
). -
--help
for more options.
bazel run [bazel parameters] //:cudnn_benchmark -- [benchmark parameters]
Benchmark parameters:
-
--benchmark_filter=<regex>
: Only run tests that match the given regex pattern. Example:'BWD_DATA'
. -
--proto_path=<path>
: Path to textproto file that contains additional convolutions to run. Default is'cudnn_benchmarks.textproto'
. -
--timing=<method>
: How to measure execution time. One of'kernel-duration'
(default),'kernel-cycles'
, or'host-duration'
. -
--help
and--helpfull
for more options.