cuda_gtest_plugin

Hacks into googletest to provide the possibility to test small code fragments in CUDA kernels (and in host code).

When do you need it?

Suppose you want to test a class that uses different code on host and device. With this plugin you need to write only a single test. In the test summary two tests will appear, the one with suffix "_CUDA" is run in a kernel.

How does it work

Googletest uses preprocessor macros for defining TESTs. We define a new macro CUDA_TEST. The body of the test is duplicated into original TESTs and, the one for CUDA, is surrounded with CUDA code. The ASSERTs are manipulated as well, to work in CUDA kernels. The result of a CUDA_ASSERT is copied back to host and evaluated there with the original ASSERT.

Limitations

• Only ASSERT_EQ and ASSERT_FLOAT_EQ is supported.
• The number of ASSERTs in one test is limited to 10 for each type.