Add a Conversion from tt.load to genx.matrix.2Dblockload for 2D blocked tensor pointer cases

[LiyangLingIntel]

This issue is the first stage of task [DPAS]: Use 2d-loads instruction to load the operand of tt.dot #146

Triton CUDA pipeline lower 2d-load to TMALoadTiledOp which does not have limitation on the last dim strides.
For Triton XPU pipeline, if we want to leverage GenISA_LSC2DBlockRead, it requires last dim (stride[-1]=1) continuous for each block pointer.

So, in the first stage, in the pass of 2d load conversion lowering, we will check the stride attr type. If the type is a constant, and meets the 2d block load case, we can leverage genx.matrix.2Dblockload. Otherwise, it will fallback to regular loads.

[LiyangLingIntel]

Besides the stride limitation, there are still restrictions because of HW limitation.:

• The base_pitch must be a multiple of QW (8 bytes)
• Should pay attention to the size of tile_width, tile_height, they are mulitiplied by the element size (bytes) must be a multiple of 4 bytes, no greater than 63.

[LiyangLingIntel]

For cases in 08-experimental-block-pointer and test_block_pointer.py, they produce IR with constant last dim stride.
So these cases can be covered in the situation talked in this issue.

[LiyangLingIntel]

Based on the previous discussion, submitted #941, we have 2 ways of implementation. The idea is similar, differences would be

• Lower tt.load to TritonGEN::Matrix2DBlockLoadOp for tt.dot operands #865
  Using intermediate op triton_intel_gpu.load_2d to work with Triton common passes, like RewriteTensorPointer, with minor code changes.
• 2D block load lowering for tt.dot operands with no intermediate op #941
  Do not need a extra op in Intel GPU dialect, but this way rewrites the common RewriteTensorPointer pass like previously NV did. So that we can keep tt.load pattern with tensor pointer for the later lowering in to-llvm stage.

[LiyangLingIntel]

Split #941 to 2 PRs for better review:

• Rewrite RewriteTensorPointer pass to support 2D block load #958
• Lower block pointer tt.load to 2DBlockRead #959

[etiotto]

Helping with refactoring & code review.