opt commonir, change scalar compute to vectorize compute

[hellozmz]

测试数据

【原始测试数据】
Average execution time over 100 runs:
  TileLang: 3.3752 ms
  Triton:   0.1431 ms
  PyTorch:  0.0164 ms

Throughput Analysis:
  Triton:   87.90 GB/s
  TileLang: 3.73 GB/s
  PyTorch:  765.88 GB/s

Performance comparison relative to PyTorch:
  Triton is 8.71x SLOWER than PyTorch
  TileLang is 205.44x SLOWER than PyTorch
  Triton is 23.58x FASTER than TileLang
==================================================
【优化后测试数据】
Average execution time over 100 runs:
  TileLang: 0.1119 ms
  Triton:   0.1828 ms
  PyTorch:  0.0152 ms

Throughput Analysis:
  Triton:   68.82 GB/s
  TileLang: 112.48 GB/s
  PyTorch:  829.82 GB/s

Performance comparison relative to PyTorch:
  Triton is 12.06x SLOWER than PyTorch
  TileLang is 7.38x SLOWER than PyTorch
  Triton is 1.63x SLOWER than TileLang

性能变化

性能提升 3.3752/0.1119= 30 倍

mlir 变化