Benchmarks: Revision - Support flexible warmup and non-random data initialization in cublas-benchmark

[yukirora]

Description
revise cublas-benchmark for flexible warmup and fill data with fixed number for perf test to improve the running efficiency.

Major Revision

• remove num_in_steps for warmup to support more flexible warmup setting for users
• Add support to generate input with fixed number for perf test

[codecov]

Codecov Report

Merging #479 (4363c5e) into main (0292366) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##             main     #479   +/-   ##
=======================================
  Coverage   87.82%   87.82%           
=======================================
  Files          87       87           
  Lines        5625     5627    +2     
=======================================
+ Hits         4940     4942    +2     
  Misses        685      685           

Flag	Coverage Δ
cpu-python3.6-unit-test	74.98% <50.00%> (-0.01%)	⬇️
cpu-python3.7-unit-test	74.98% <50.00%> (-0.01%)	⬇️
cpu-python3.8-unit-test	75.45% <50.00%> (-0.01%)	⬇️
cuda-unit-test	87.75% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...nch/benchmarks/micro_benchmarks/cublas_function.py	84.78% <100.00%> (+0.33%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[abuccts]

do you need to update cudnn_function as well?

[rhewett]

I think the random fix looks good. I have a few more thoughts here:

• The random initialization can be accelerated by adding -O3 to the host compiler flags. After some local testing, this cuts about 35% of the run time.
• The random initialization can be further accelerated using OMP.
• The random initialization can be accelerated by moving it to the device.

The issue with the for loops in the warm-up still remains.

[cp5555]

I think the random fix looks good. I have a few more thoughts here:

• The random initialization can be accelerated by adding -O3 to the host compiler flags. After some local testing, this cuts about 35% of the run time.
• The random initialization can be further accelerated using OMP.
• The random initialization can be accelerated by moving it to the device.

The issue with the for loops in the warm-up still remains.

Very good comments. I think we can check it and optimize it in next PR