[rocsparse] Fix csric0 dispatch when rows have many nonzeros

[jsandham]

Motivation

Removes unneeded WF_SIZE template parameter from csrsm kernels. Fixes bug in csric0 where we could return rocsparse_status_invalid_value when the maximum number of non-zeros was between 513 and 1024.

Technical Details

The hash algorithm is only usable up to 512 but the dispatch assumed it could be used up to 1024. This means that if there was a matrix with 512 < max_nnz <= 1024 we would dispatch to the hash table algorithm and then in the hash table algorithm we would throw an error as not supported. This is why I added an additional test to check 512 < max_nnz <= 1024 as without my change this test fails.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 45.45455% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
.../library/src/level3/rocsparse_csrsm_solve_impl.cpp	50.00%	5 Missing ⚠️
.../precond/csric0/rocsparse_csric0_kernel_launch.cpp	0.00%	0 Missing and 1 partial ⚠️

❌ Your project status has failed because the head coverage (77.21%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #5360      +/-   ##
===========================================
+ Coverage    66.43%   66.45%   +0.02%     
===========================================
  Files         1793     1814      +21     
  Lines       277419   279234    +1815     
  Branches     38806    39217     +411     
===========================================
+ Hits        184301   185562    +1261     
- Misses       77117    77441     +324     
- Partials     16001    16231     +230     

Flag	Coverage Δ		*Carryforward flag
hipBLAS	90.67% <ø> (ø)		Carriedforward from d36db75
hipBLASLt	43.99% <ø> (ø)		Carriedforward from d36db75
hipCUB	82.38% <ø> (ø)		Carriedforward from d36db75
hipDNN	83.94% <ø> (ø)		Carriedforward from d36db75
hipFFT	56.38% <ø> (ø)		Carriedforward from d36db75
hipRAND	76.12% <ø> (ø)		Carriedforward from d36db75
hipSOLVER	68.81% <ø> (ø)		Carriedforward from d36db75
hipSPARSE	84.70% <ø> (ø)		Carriedforward from d36db75
rocBLAS	47.97% <ø> (ø)		Carriedforward from d36db75
rocFFT	49.77% <ø> (ø)		Carriedforward from d36db75
rocRAND	57.07% <ø> (ø)		Carriedforward from d36db75
rocSOLVER	77.21% <ø> (ø)		Carriedforward from d36db75
rocSPARSE	71.48% <45.45%> (-0.05%)	⬇️

*This pull request uses carry forward flags. Click here to find out more.

Files with missing lines	Coverage Δ
.../precond/csric0/rocsparse_csric0_kernel_launch.cpp	75.00% <0.00%> (ø)
.../library/src/level3/rocsparse_csrsm_solve_impl.cpp	74.09% <50.00%> (ø)

... and 86 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[YvanMokwinski]

There is change about the use of binsearch, you lowered the size.
Please clarify.

[jsandham]

There is change about the use of binsearch, you lowered the size. Please clarify.

The hash algorithm is only usable up to 512 but the dispatch assumed it could be used up to 1024. This means that if there was a matrix with 512 < max_nnz <= 1024 we would dispatch to the hash table algorithm and then in the hash table algorithm we would throw an error as not supported. This is why I added an additional test to check 512 < max_nnz <= 1024 as without my change this test fails.

[YvanMokwinski]

There is change about the use of binsearch, you lowered the size. Please clarify.

The hash algorithm is only usable up to 512 but the dispatch assumed it could be used up to 1024. This means that if there was a matrix with 512 < max_nnz <= 1024 we would dispatch to the hash table algorithm and then in the hash table algorithm we would throw an error as not supported. This is why I added an additional test to check 512 < max_nnz <= 1024 as without my change this test fails.

Thanks for the clarification, @jsandham

[amd-jnovotny]

Changelog OK.