Intel Advisor/ MLIR

Hi Everyone,

I am wondering if anyone has used before intel Adivsor to analyse the performance with MLIR.
I am trying to use it to analyse my code. But for example, it cannot always detecte correctly vectorized loops. For example, I have this loop that is vectorized

  scf.for %arg14 = %c0 to %c128 step %c8 {
                  %67 = affine.apply #map4(%arg14, %arg12)
                  %68 = vector.transfer_read %61[%66, %67], %cst_0 {in_bounds = [true, true]} : memref<?x?xf64, #map6>, vector<1x8xf64>
                  %69 = vector.transfer_read %65[%arg13, %arg14], %cst_0 {in_bounds = [true, true]} : memref<?x128xf64, #map0>, vector<1x8xf64>
                  %70 = arith.addf %68, %69 : vector<1x8xf64>
                  vector.transfer_write %70, %65[%arg13, %arg14] {in_bounds = [true, true]} : vector<1x8xf64>, memref<?x128xf64, #map0>

I checked also its generated assembly and I see that it is using vector registers but in intel/Advisor it says that it is a scalar loop.

probably I am doing something wrong with my compilation pipline to pass. I used evrywhere with mlir-opt and mlir-translated the option --mlir-print-debuginfo, then to compile to object file

opt -O3 --debug-entry-values -S | llc --debug-entry-values -O3 -mattr=+avx2 -filetype=obj -o heat_transfert.o