Consider enable SLPVectorization back

[anmyachev]

https://github.com/intel/intel-xpu-backend-for-triton/blob/0b4311746264fb80a6420332c57dfb5014fb8c71/third_party/intel/triton_xpu.cc#L201C4-L201C44

[anmyachev]

The test from the comment now works, but there are other problems: #4166

One example:

****** Running 06-fused-attention test ******

L0 build module failed. Log: 
error: undefined reference to `_Z16__spirv_ocl_fmaxDv32_fS_'
in function: '__spirv_ocl_fmax(float vector[32], float vector[32])' called by kernel: '_attn_fwd'

error: backend compiler failed build.

error: undefined reference to `_Z16__spirv_ocl_exp2Dv32_f'
in function: '__spirv_ocl_exp2(float vector[32])' called by kernel: '_attn_fwd'

error: backend compiler failed build.

Error during Intel loadBinary: Triton Error [ZE]: 0x70000004

@whitneywhtsang WDYT?

[whitneywhtsang]

IGC team suggested to perform vectorization in IGC, and not enable SLPVectorization pass in LLVM opt.

[anmyachev]

IGC team suggested to perform vectorization in IGC, and not enable SLPVectorization pass in LLVM opt.

We could at least tell IGC team about new problems we found. I guess it's more common to use library-side vectorization rather than driver-side vectorization? (other developers may face similar problems)

Or is it a problem on our side that we are not taking something into account?

[whitneywhtsang]

Or is it a problem on our side that we are not taking something into account?

Possible, maybe _Z16__spirv_ocl_exp2Dv32_f is not a valid function?