[Fix](bangc-ops): replace __bang_atomic_add with __bang_atomic_reduce…

[Unireverse]

…_add for better perf.

Thanks for your contribution and we appreciate it a lot. 🚀🚀

1. Motivation

Replace __bang_atomic_add with __bang_atomic_reduce_add for better performance.

2. Modification

modified: bangc-ops/kernels/carafe/carafe_block.mlu
modified: bangc-ops/kernels/deform_roi_pool/deform_roi_pool_union1.mlu
modified: bangc-ops/kernels/psroipool/psroipool_block.mlu
modified: bangc-ops/kernels/roi_align_rotated/roi_align_rotated_block.mlu
modified: bangc-ops/kernels/roi_crop/roi_crop_block.mlu
modified: bangc-ops/kernels/rotated_feature_align/rotated_feature_align_block.mlu

3. Test Report

3.1 Modification Details

3.1.1 Accuracy Acceptance Standard

For static threshold standard details, see: MLU-OPS Accuracy Acceptance Standard.

• static threshold
  • diff1
    • float32 mlu diff1 <= 1e-5
    • float32 mlu diff1 <= 3e-3
    • float16 mlu diff1 <= 3e-3
  • diff2
    • float32 mlu diff2 <= 1e-5
    • float32 mlu diff2 <= 3e-3
    • float16 mlu diff2 <= 3e-3
  • diff3
    • mlu diff3 == 0
    • mlu diff3_1 == 0
    • mlu diff3_2 == 0
• dynamic threshold
  • diff1: mlu diff1 <= max(baseline diff1 * 10, static threshold)
  • diff2: mlu diff2 <= max(baseline diff2 * 10, static threshold)
  • diff3: mlu diff3 <= max(baseline diff3 * 10, static threshold)
    • float32, threshold = 1e-5
    • float16, threshold = 1e-3

3.1.2 Operator Scheme checklist

• Supported hardware
  • MLU370
  • MLU590
• Job types
  • BLOCK
  • UNION1
  • UNION2
  • UNION4
  • The operator will dynamically select the most suitable task type, for example, UNION8

3.2 Accuracy Test

3.2.1 Accuracy Test

Regression test all passed.

[ OK ] copy/TestSuite.mluOp/3 (3 ms)
[----------] 4 tests from copy/TestSuite (14 ms total)

[----------] Global test environment tear-down
[ SUMMARY ] Total 104 cases of 2 op(s).
ALL PASSED.
[==========] 104 test cases from 2 test suites ran. (165101 ms total)
[ PASSED ] 104 test cases.

3.2.2 Parameter Check

No update.

3.3 Performance Test

3.4 Summary Analysis

Replace __bang_atomic_add with __bang_atomic_reduce_add for better performance. According to accuracy test, accuracy is not changed.