test fail on intel gpu

[meteor-lsw]

When I run the script 'run-checks.sh all' using Intel's integrated graphics card, I got the following error.

failures:

---- tests::ad_cos::tests::should_diff_cos stdout ----
thread 'tests::ad_cos::tests::should_diff_cos' panicked at burn-wgpu/src/lib.rs:48:5:
Tensors are not approx eq:
=> Position 1: -39.12511444091797 != -39.123374938964844 | difference 0.001739501953125 > tolerance 0.0010000000000000002
=> Position 3: 49.75035095214844 != 49.748355865478516 | difference 0.001995086669921875 > tolerance 0.0010000000000000002
note: run with RUST_BACKTRACE=1 environment variable to display a backtrace

---- tests::ad_sin::tests::should_diff_sin stdout ----
thread 'tests::ad_sin::tests::should_diff_sin' panicked at burn-wgpu/src/lib.rs:48:5:
Tensors are not approx eq:
=> Position 1: 44.195899963378906 != 44.19477462768555 | difference 0.001125335693359375 > tolerance 0.0010000000000000002
=> Position 3: -80.46344757080078 != -80.4609375 | difference 0.00251007080078125 > tolerance 0.0010000000000000002

failures:
tests::ad_cos::tests::should_diff_cos
tests::ad_sin::tests::should_diff_sin

test result: FAILED. 640 passed; 2 failed; 0 ignored; 0 measured; 0 filtered out; finished in 11.82s

error: test failed, to rerun pass -p burn-wgpu --lib

[nathanielsimard]

I seems to be a precision problem, we should probably aims to set the right precision for those tests.

[meteor-lsw]

The current precision set for assert_approx_eq in these two tests is 3. Should we change the precision to 2 now?

[nathanielsimard]

We could do that, or add another function similar to assert_approx_eq with a custom difference not forced to be a multiple of 10.

assert_approx_eq_diff(actual, expected, 2.5e-3)

[louisfd]

Fixed in #954