Function run_test

tests/python/codegen/test_target_codegen_cuda.py:676–700

(tvm_intrin, np_func, dtype)

    ]

    def run_test(tvm_intrin, np_func, dtype):
        if dtype == "float16" and not have_fp16(tvm.cuda(0).compute_version):
            print("Skip because gpu does not have fp16 support")
            return
        # set of intrinsics does not support fp16 yet.
        skip_set = {
            tvm.tirx.abs,
            tvm.tirx.round,
            tvm.tirx.tan,
            tvm.tirx.atan,
            tvm.tirx.tanh,
            tvm.tirx.cosh,
            tvm.tirx.sinh,
        }
        if dtype == "float16" and tvm_intrin in skip_set:
            print(f"Skip because '{tvm_intrin.__name__}' does not support fp16 yet")
            return

        n = 128
        f = sched(tvm_intrin, dtype, n)
        dev = tvm.cuda(0)
        a = tvm.runtime.tensor(np.random.uniform(0, 1, size=n).astype(dtype), dev)
        b = tvm.runtime.tensor(np.zeros(shape=(n,)).astype(dtype), dev)
        f(a, b)
        tvm.testing.assert_allclose(b.numpy(), np_func(a.numpy()), atol=1e-3, rtol=1e-3)

    for func in test_funcs:
        run_test(*func, "float32")