Save frequencies when bulk-saving of times is enabled

[oleksandr-pavlyk]

Closes #318

This PR:

• applies changes started by @gevtushenko to collect bulk values of frequency estimates and enable saving them alongside bulk of time-stamps
• the same batched writeout_data utility is used to save duration and frequency data faster

[oleksandr-pavlyk]

Questions to be answered:

• should dumping of frequencies require a CLI option separate from the one controlling dumping of time-stamps?
  • Frequencies are saved when throttling control is enabled.
    Q: what happens when throttling control is turned off and bulk data is requested?

[oleksandr-pavlyk]

Setting -throttle-threshold 0 disables collection of GPU frequency data:

$ rg 'm_check_throttling\(' ~/repos/nvbench/nvbench/
~/repos/nvbench/nvbench/detail/measure_cold.cu
47:    , m_check_throttling(!exec_state.get_run_once() && exec_state.get_throttle_threshold() > 0.f)

I executed the benchmark 2 times:

./nvbench.example.cpp17.axes -d 0 --jsonbin axes-d0 --stopping-criterion entropy
./nvbench.example.cpp17.axes -d 0 --jsonbin axes-d1 --throttle-threshold 0.0 --stopping-criterion entropy

Executing ./nvbench.example.cpp17.axes -d 0 --jsonbin axes-d1 --throttle-threshold 0.0 --stopping-criterion entropy thus exercises configuration flagged as possibly problematic in the earlier comment and produces empty BIN files for frequency data.

Sizes of BIN files for each run

Inspecting the JSON file corresponding to the run with default value of throttling threshold, the GPU frequencies are collected and non-empty binary files are generated:

 jq '.benchmarks[].states[] | select(.is_skipped == false ) | .summaries[] | select(has("hint")) | select(.hint | test("file/sample_freqs")) | .data[1].value' axes-d0 | uniq -c | awk '
  { buf[NR]=$0 }
  END {
    if (NR > 7) {
      for (i=1;i<=3;i++) print buf[i]
      printf "... (%d lines omitted) ...\n", NR-6
      for (i=NR-2;i<=NR;i++) print buf[i]
    } else {
      for (i=1;i<=NR;i++) print buf[i]
    }
  }'
      1 "312"
      1 "304"
      1 "204"
... (74 lines omitted) ...
      1 "320"
      1 "310"
      1 "344"

Inspecting the JSON file corresponding to the run with zero throttling threshold which turns off the collection of GPU frequency data, the sample frequencies binary files are all empty:

$ jq '.benchmarks[].states[] | select(.is_skipped == false ) | .summaries[] | select(has("hint")) | select(.hint | test("file/sample_freqs")) | .data[1].value' axes-d1 | uniq -c | awk '
  { buf[NR]=$0 }
  END {
    if (NR > 7) {
      for (i=1;i<=3;i++) print buf[i]
      printf "... (%d lines omitted) ...\n", NR-6
      for (i=NR-2;i<=NR;i++) print buf[i]
    } else {
      for (i=1;i<=NR;i++) print buf[i]
    }
  }'
     80 "0"

[oleksandr-pavlyk]

After discussing the issue internally, we should always collect GPU frequencies even if throttling threshold has been set to zero.

[oleksandr-pavlyk]

**Q**: what happens when throttling control is turned off and bulk data is requested?

With GPU frequency bulk data now collected for any setting of throttling threshold, use of --jsonbin would always produce frequency data of the same size as the kernel execution duration data.