Separate the tests for coverage versus non-coverage

[DilumAluthge]

No description provided.

[DilumAluthge]

If this works well, we can reproduce this pattern in other repos where the test suite performance really slows down when code coverage is enabled.

[codecov]

Codecov Report

Merging #8 (4e85e7e) into master (cf3b37a) will decrease coverage by 5.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master       #8      +/-   ##
==========================================
- Coverage   82.92%   77.87%   -5.06%     
==========================================
  Files           7        7              
  Lines         123      113      -10     
==========================================
- Hits          102       88      -14     
- Misses         21       25       +4     

Impacted Files	Coverage Δ
src/Octavian.jl	100.00% <ø> (ø)
src/macrokernel.jl	54.54% <ø> (-37.77%)	⬇️
src/matmul.jl	100.00% <ø> (ø)
src/utils.jl	71.42% <ø> (-1.91%)	⬇️
src/block_sizes.jl	93.75% <100.00%> (-0.37%)	⬇️
src/types.jl	44.44% <100.00%> (ø)
src/macros.jl	100.00% <0.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update cf3b37a...4e85e7e. Read the comment docs.

[DilumAluthge]

Wow, look how fast!

[chriselrod]

Note that different matrix sizes hit different code paths.
But I think we're actually okay here, because being transposed also changes the codepaths, meaning it should be hitting them all anyway.

[chriselrod]

Inside matmul! we have:

if M * K ≤ _Mc * _Kc && A isa DenseArray && C isa StridedArray && B isa StridedArray && #
        (stride(A,2) ≤ 72 || (iszero(stride(A,2) & (VectorizationBase.pick_vector_width(eltype(A))-1)) && iszero(reinterpret(Int,pointer(A)) & 63)))
    macrokernel!(C, A, B, _α, _β)
    return C
end

a check for whether we should just call macrokernel! directly.
And then similarly:

do_not_pack_B = (B isa DenseArray && (K * N ≤ _Kc * _Nc)) || N ≤ LoopVectorization.nᵣ

To check if we should skip packing B.
Because of the DenseArray checks, we'll still take both sides of each.

[DilumAluthge]

I wonder if this kind of pattern could be useful in other test suites (LoopVectorization.jl, VectorizationBase.jl, PaddedMatrices.jl, Gaius.jl, etc.).

Of course, it's only useful in cases where the test suite time greatly increases between coverage disabled and coverage enabled.

[chriselrod]

The beta != 0 code path actually isn't getting hit anymore.
It gets set to 1 for the second+ iteration of the K loop, but we could also just call matmul!(C, A, B, true, true), for example.

[DilumAluthge]

The beta != 0 code path actually isn't getting hit anymore.
It gets set to 1 for the second+ iteration of the K loop, but we could also just call matmul!(C, A, B, true, true), for example.

Yeah worst case we can just add unit tests to make up the coverage. I'll open a PR in a few minutes.