[Mlir-commits] [mlir] [mlir][SCF] Deprecate `linalg::tileToForallOp` and `linalg::tileToForallOpUsingTileSizes` (PR #91878)

[llvmlistbot at llvm.org]

================
@@ -241,17 +251,25 @@ getTileOffsetAndSizes(RewriterBase &rewriter, Location loc, ValueRange ivs,
   SmallVector<OpFoldResult> offsets, sizes;
   int materializedLoopNum = 0;
 
+  assert(
+      llvm::all_of(iterationDomain,
+                   [](Range r) { return isConstantIntValue(r.stride, 1); }) &&
+      "the offset and tile size computation assumes stride 1 for all "
+      "dimensions of the iteration domain");
+
   if (!numThreads.empty()) {
-    AffineExpr d0, d1, s0, s1, s2;
+    AffineExpr d0, d1, s0, s1;
     AffineExpr offsetExpr, residualTileSizeExpr;
     bindDims(rewriter.getContext(), d0, d1);
-    bindSymbols(rewriter.getContext(), s0, s1, s2);
-    offsetExpr = d0 + d1 * s0 * s1;
-    residualTileSizeExpr = s2 - (d0 + d1 * s0 * s1);
+    bindSymbols(rewriter.getContext(), s0, s1);
+    offsetExpr = d0 + d1 * s0;
+    residualTileSizeExpr = s1 - (d0 + d1 * s0);
----------------
MaheshRavishankar wrote:

There was some fly-by comment that we need to support non-unit strides in tiling, but I dont have a full context on it. Ill keep an eye out for this.

https://github.com/llvm/llvm-project/pull/91878