Skip leading size-1 dims when locating innermost contiguous run

verifyBDSizesStrides previously treated dim 0 as the innermost
contiguous run, but a leading size-1 dim is a DMA no-op and does not
constrain alignment. Walk past leading size-1 entries to find the real
innermost dim before applying the granularity and stride-alignment
checks.

Also drop the degenerate dimensionsToStream=[(1,2),(1,1)] from the
vector_reduce_max memtile designs; it lowered to a zero-effect layout
that only existed to work around the old verifier and is rejected by
the unified helper.

Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>

Author: hunhoffe

lib/Dialect/AIE/IR/AIEDialect.cpp

@@ -2210,14 +2210,21 @@ mlir::LogicalResult xilinx::AIE::verifyBDSizesStrides(
       return forOp->emitOpError("Size ") << i << " must be a positive integer.";
   }
 
+  // The "real" innermost contiguous run is the deepest dim with size > 1; a
+  // leading size==1 entry (innermost in the original op) is a DMA no-op and
+  // does not constrain alignment.
+  int innermost = 0;
+  while (innermost < n - 1 && inputSizes[innermost] == 1)
+    ++innermost;
+
   // Innermost contiguous run must be a multiple of address granularity --
   // hardware moves whole words; a sub-word innermost run is unrealizable.
-  if (inputSizes[0] * elemWidthBits % addressGranularityBits != 0) {
+  if (inputSizes[innermost] * elemWidthBits % addressGranularityBits != 0) {
     std::stringstream msg;
     msg << "Transfer sizes must be multiples of "
-        << (addressGranularityBits / 8) << " bytes. " << inputSizes[0]
+        << (addressGranularityBits / 8) << " bytes. " << inputSizes[innermost]
         << " elements at " << (elemWidthBits / 8) << " bytes each equal "
-        << (inputSizes[0] * elemWidthBits / 8)
+        << (inputSizes[innermost] * elemWidthBits / 8)
         << " bytes, which is not divisible by " << (addressGranularityBits / 8)
         << ". ";
     return forOp->emitOpError(msg.str());
@@ -2235,11 +2242,11 @@ mlir::LogicalResult xilinx::AIE::verifyBDSizesStrides(
     }
   }
 
-  // Stride byte-alignment: innermost stride==1 is always allowed (sub-word
-  // packed contiguous run, paired with the granularity check above); any other
-  // stride must be a granularity multiple in bytes.
+  // Stride byte-alignment: the innermost non-degenerate dim's stride==1 is
+  // always allowed (sub-word packed contiguous run, paired with the granularity
+  // check above); any other stride must be a granularity multiple in bytes.
   for (int i = 0; i < n; ++i) {
-    if (i == 0 && inputStrides[i] == 1)
+    if (i == innermost && inputStrides[i] == 1)
       continue;
     if (inputSizes[i] == 1)
       continue;
@@ -2257,7 +2264,7 @@ mlir::LogicalResult xilinx::AIE::verifyBDSizesStrides(
   // For element widths larger than the granularity (e.g. bfp blocks, i64),
   // the hardware cannot encode a non-1 innermost stride; getHardwareStrides-
   // Wraps would silently drop the stride. Force innermost stride == 1.
-  if (elemWidthBits > addressGranularityBits && inputStrides[0] != 1)
+  if (elemWidthBits > addressGranularityBits && inputStrides[innermost] != 1)
     return forOp->emitOpError(
                "For element widths larger than the address granularity (")
            << (addressGranularityBits / 8)

programming_examples/basic/vector_reduce_max/single_column_designs/vector_reduce_max_memtile.py

@@ -50,7 +50,7 @@ def my_reduce_max(dev, in1_size, out_size, dtype_str, trace_size):
     out_fifos = []
 
     of_in = ObjectFifo(mem_ty, name="of_in")
-    outC = ObjectFifo(int_ty, name="outC", dims_to_stream=[(1, 2), (1, 1)])
+    outC = ObjectFifo(int_ty, name="outC")
     of_out = ObjectFifo(out_ty, name="of_out")
 
     if n_cores > 1:

programming_examples/basic/vector_reduce_max/single_column_designs/vector_reduce_max_memtile_placed.py

@@ -111,7 +111,6 @@ def device_body():
             cores[0],
             buffer_depth,
             int_ty,
-            dimensionsToStream=[(1, O), (1, 1)],
         )
         object_fifo_link(out_fifos, outC, of_c_offsets, [])