Add benchmark for BLAS operations

[upsj]

This PR adds benchmarks for Ginkgo's BLAS-like operations (copy, axpy, scale, dot, norm, gemm) as well as a Dense::convert_to(Dense) implementation that doesn't modify the stride of the result vector if unnecessary (thus partially addressing #748). This benchmark can be used later to check whether #733 provides decent implementations for Dense operations.

The BLAS benchmarks are not the most sophisticated thing out there, since they don't flush the cache in any way, or check for correctness/precision (though the current design allows for that), but that should not matter too much as soon as we scale up.

A benchmark test case consists entirely of problem dimensions and strides.
We have input vectors x, y: nxr, a, b: 1xr or 1x1, A: nxk, B: kxm, C:nxm, for which we can specify individual strides stride_x/y/A/B/C with stride being a shorthand for stride_x/y.

The benchmark uses the amount of FLOPs and memory movement to compute FLOP/s and B/s numbers in addition to the runtime. Please double-check those (found in get_memory() and get_flops(), I just claimed that copy does one FLOP per entry), though.

Example output on my GPU:

[
    {
        "n": 1000000,
        "r": 10,
        "stride_x": 20,
        "stride_y": 15,
        "blas": {
            "copy": {
                "time": 0.019050621900000003,
                "flops": 524917246.92725118,
                "bandwidth": 8398675950.836019,
                "completed": true
            },
            "axpy": {
                "time": 0.0243559142,
                "flops": 821155791.3929588,
                "bandwidth": 9853869496.715506,
                "completed": true
            },
            "scal": {
                "time": 0.018665861,
                "flops": 535737408.5235072,
                "bandwidth": 8571798536.376115,
                "completed": true
            }
        }
    }
]

[yhmtsai]

LGTM

[codecov]

Codecov Report

Merging #772 (aca6830) into develop (a1b57a0) will increase coverage by 0.01%.
The diff coverage is n/a.

@@             Coverage Diff             @@
##           develop     #772      +/-   ##
===========================================
+ Coverage    94.17%   94.18%   +0.01%     
===========================================
  Files          400      400              
  Lines        31049    31049              
===========================================
+ Hits         29241    29245       +4     
+ Misses        1808     1804       -4     

Impacted Files	Coverage Δ
include/ginkgo/core/base/range.hpp	95.48% <ø> (ø)
omp/reorder/rcm_kernels.cpp	98.13% <0.00%> (+2.48%)	⬆️

---

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 a1b57a0...aca6830. Read the comment docs.

[sonarcloud]

Kudos, SonarCloud Quality Gate passed!

0 Bugs
0 Vulnerabilities
0 Security Hotspots
13 Code Smells

0.0% Coverage
0.0% Duplication

[tcojean]

LGTM.

I guess you don't want to advertise it which is why you didn't add it to neither run_all_benchmarks.sh or BENCHMARKING.md?

[upsj]

@tcojean Yes, I would like to finish some interface extensions (non-inplace or non-object-creating operations) and implementation changes (convert_to and move_to, mostly) first, so we can also represent conversions, transpose, permute etc. correctly.