[Mlir-commits] [mlir] [mlir][nvgpu] make TmaCreateDescriptorOp can use static box and add folder function to it. (PR #135497)

[lonely eagle]

linuxlonelyeagle wrote:

```
 python matmul.py
===-- Running GEMM Multistage f32 += f16 * f16, Size 128x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 128x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 128x128x256, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 128x128x256, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 512x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 512x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 512x128x256, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 512x128x256, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 128x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 128x128x64, Tile 128x128x64, stages 1 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 128x128x256, Tile 128x128x64, stages 3 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 128x128x256, Tile 128x128x64, stages 3 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 512x128x64, Tile 128x128x64, stages 2 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 512x128x64, Tile 128x128x64, stages 2 --===
PASS 
===-- Running GEMM Multistage f32 += f16 * f16, Size 512x128x256, Tile 128x128x64, stages 3 --===
PASS 
===-- Running GEMM Warp specialization f32 += f16 * f16, Size 512x128x256, Tile 128x128x64, stages 3 --===
PASS 
```

But nvgpu toy is reporting errors for other reasons.
```
python Ch2.py
...
ValueError: Operand 3 of operation "nvgpu.tma.async.load" must be a Sequence of Values (std::bad_cast)
```

https://github.com/llvm/llvm-project/pull/135497