[TIR] VNNI and ARM dot product intrinsic for tensorization

[masahi]

Introduces a new directory python/tvm/tir/tensor_intrin where we put intrinsic descriptions written in TVMScript for various HW targets. They can be used by manual tensorized TIR schedules or auto-tensorized ones. More intrinsics, such as tensorcore ones, DP4A etc will be added later.

@junrushao1994 @vinx13 @shingjan @Hzfengsy @spectrometerHBH

[masahi]

This is equivalent to the TE one in

tvm/python/tvm/topi/arm_cpu/tensor_intrin.py

Line 536 in 912993f

def dot_int8_int8_int32_neon():

cc @tkonolige

[masahi]

This is equivalent to the TE one in

tvm/python/tvm/topi/arm_cpu/tensor_intrin.py

Line 431 in 912993f

def dot_int8_int8_int32_neon_82(int32_lanes, dtype="uint"):

[masahi]

Equivalent to the TE one in

tvm/python/tvm/topi/x86/tensor_intrin.py

Line 244 in 912993f

def dot_16x1x16_uint8_int8_int32_cascadelake():

[masahi]

cc @junrushao1994 @yelite, this is one of the common needs for meta programming support in TVMScript. I think shape parameterization is possible via specialize, but not sure if I can use that with T.Buffer syntax sugar.

A similar need arises for tensorcore (different mma shape x data type)

[masahi]

cc @junrushao1994 @yelite, I want to define and call a convenience function like

tvm/python/tvm/topi/arm_cpu/tensor_intrin.py

Line 625 in 912993f

def pairwise_add_mul(extract_half):

to remove duplication in IR generation. I think this is a low-hanging fruit that can be supported without more foundational work on meta-programming? (just need to tweak the python parser?)

[junrushao]

Yes, it is likely doable and @Hzfengsy probably already has something ready

[junrushao]

CC @vinx13 would you like to review? Thanks a lot!