[Bug] partion_tensor_view在高维tensor场景下会触发pto-isa assert

[werewb]

Component

PTO Dialect / ODS (include/PTO/IR)

Description

一个4维tensor，shape为[B,S,N,D]，现在想从S维和D维取一个切片，则partion_tensor_view的shape为[1, tile_s, 1, tile_d], strides为[stride[0], stride[1], stride[2], stride[3]], 这在表示上没啥问题，但是pto-isa要求tload/tstore时，tensor除了低2维以外的其它shape必须为1，所以上述的partion_tensor虽然表述上没有问题，但是无法使用，会触发pto-isa的assert，正确的做法应该是生成shape为[1, 1, tile_s, tile_d], strides为[stride[0], stride[2], stride[1], stride[3]]的partion_tensor

Reproduction (minimal)

见附件的pto文件

Expected behavior

能生成正确的pto-isa代码，并可以编译成功

Actual behavior / error logs

pto/npu/a2a3/TLoad.hpp:379:5: error: static assertion failed due to requirement 'GlobalTensor<half, pto::Shape<1, 1, 128, 1, 128>, pto::Stride<-1, -1, -1, -1, 1>, pto::Layout::ND>::staticShape[0] == 1 && GlobalTensor<half, pto::Shape<1, 1, 128, 1, 128>, pto::Stride<-1, -1, -1, -1, 1>, pto::Layout::ND>::staticShape[1] == 1 && GlobalTensor<half, pto::Shape<1, 1, 128, 1, 128>, pto::Stride<-1, -1, -1, -1, 1>, pto::Layout::ND>::staticShape[2] == 1': Fix: GlobalTensor ony support 2 dim when ND2NZ!

Git commit

eb419fa

Host platform

None

Target Ascend arch (if relevant)

None

kernel_pto.txt

PTOAS build level (if relevant)

None

[zhangstevenunity]

Suggest use loadgm2l1ND2ND type tload，and use tmov that change layout between l1 and l0