Improved profiler result printing.

[mcourteaux]

This 50-cents improvement makes my profiling reports more readable, as I have long function names. Additionally, I cheaply aligned the timing by printing spacing depending on how many order of magnitude the number spans. Example output below:

neonraw_denoiser_f8_k7_p1-x86-64-linux-avx2-fma-profile_by_timer
 total time: 4193.547363 ms  samples: 3593  runs: 11  time/run: 381.231567 ms
 average threads used: 15.640412
 heap allocations: 284680  peak heap usage: 2949120 bytes
  halide_malloc:                 0.575ms (0%)    threads: 16.000
  halide_free:                   0.095ms (0%)    threads: 15.000
  denoised:                     17.235ms (4%)    threads: 15.000
  luminance:                     5.551ms (1%)    threads: 15.446 stack: 13456
  blur_pass_h:                   0.385ms (0%)    threads: 16.000 stack: 12528
  blur_pass_h_sum:               0.767ms (0%)    threads: 15.500 stack: 32
  filter_input:                  0.296ms (0%)    threads: 16.000 stack: 11664
  blur_pass_v:                   0.000ms (0%)    threads: 0.000  stack: 32
  blur_pass_v_sum:               0.191ms (0%)    threads: 16.000 stack: 32
  mask_activation:               1.057ms (0%)    threads: 15.727 peak: 1105920  num: 56936     avg: 73728
  mask_activation_sum:          56.829ms (14%)   threads: 15.670 stack: 32
  exp_logits:                   17.358ms (4%)    threads: 15.521 peak: 1179648  num: 56936     avg: 73728
  logit:                         1.251ms (0%)    threads: 15.923 stack: 1760
  max_logit:                     7.621ms (1%)    threads: 15.688 stack: 32
  blur_conv_pass_0:              5.247ms (1%)    threads: 15.722 peak: 1769472  num: 170808    avg: 110592
  blur_conv_pass_0_sum:        242.229ms (63%)   threads: 15.688 stack: 32
  sum_exp_logits:                1.483ms (0%)    threads: 15.857 stack: 192
  sum_exp_logits$1:              6.088ms (1%)    threads: 15.237 stack: 32
  softmax:                       6.316ms (1%)    threads: 15.677 stack: 2304
  denoise_accum_pass_0:          4.716ms (1%)    threads: 15.833 stack: 4
  denoise_accum_pass_0_intm:     5.929ms (1%)    threads: 15.545 stack: 48

[steven-johnson]

All "if" statements in Halide should be in {braces}, even trivial single-line ones like these

[mcourteaux]

Oh, hmm. Then why doesn't the clang-format CI-step complain?

[steven-johnson]

clang-format didn't, but clang-tidy did

[mcourteaux]

Ah, what, clang-tidy checks this separately. Clang-format has an option for this too, I believe.

[mcourteaux]

Dang this clang-tidy fix makes the code really ugly... ☹️

[abadams]

Yeah, clang-tidy and clang-format make things better on average. There are definitely cases where it makes things worse, but it's worth it in expectation.

[mcourteaux]

I can surround it with // clang-format off/on in this case. Maybe clang-tidy will still complain then... Let me know what you prefer here.

[steven-johnson]

I can surround it with // clang-format off/on in this case. Maybe clang-tidy will still complain then... Let me know what you prefer here.

We should fix our clang-format setings to enforce this as well (if possible). In the meantime, please just manually format in a way that makes clang-tidy happy; we prefer to reserve clang-format off for pathological cases.

[mcourteaux]

Okay, then this is it. 😄

[steven-johnson]

Thanks for the fix!