Failure running conv2d i32 with objectFifo

[newling]

Checkpoints in IR:

IR Dump Before AssignTargetDevicesPass
IR Dump Before FoldMemRefAliasOps
IR Dump Before AMDAIENormalizeLoopBounds
IR Dump Before AMDAIEInsertCores

Core dumped in AMDAIEInsertCores.

first_run_conv.txt

[newling]

@jtuyls something new introduced here. Our conv tiling creates interleaved L2->L1 copies with compute:

...
 scf.forall (%arg2, %arg3) in (2, 2) {   // the threads
    scf.for %arg4 = %c0 to %c16 step %c1 {
        %7 = amdaie.dma_cpy_nd ...
        ...  
        %13 = vector.contract ... 
     }
}

This is new compared to current matmul, where amdaie.dma_cpy_nd is not in scf.for.

Question: should the insert-cores op should turn this into code like

 scf.forall (%arg2, %arg3) in (2, 2) {
    scf.for %arg4 = %c0 to %c16 step %c1 {
        ...
        %7 = amdaie.dma_cpy_nd
        ...
     }
     ... 
     amd.core(...) {
        scf.for %arg4 = %c0 to %c16 step %c1 {
         vector.contract etc.
        }
     }
}

i.e. should we duplicate loop structure for dma_cpy_nps and the core ? (with the consumer/producer stuff inserted as usual too)

[newling]

Update: probably need some sort of loop subsumption. Hoisting the dma_cpy_nd ops out of the scf.for loops