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

[Nicolas Vasilache]

================
@@ -96,7 +216,113 @@ static OpFoldResult getBoundedTileSize(OpBuilder &b, Location loc,
   AffineMap minMap = AffineMap::get(1, 2, {s0, s1 - d0}, b.getContext());
   Value size = getValueOrCreateConstantIndexOp(b, loc, loopRange.size);
   return affine::makeComposedFoldedAffineMin(
-      b, loc, minMap, SmallVector<OpFoldResult>{iv, tileSize, size});
+      b, loc, minMap, SmallVector<OpFoldResult>{offset, tileSize, size});
+}
+
+/// Returns true if the maximum tile offset `tileSize * numThreads-1` is less
+/// than `iterationSize`.
+static bool canOmitTileOffsetInBoundsCheck(OpFoldResult tileSize,
+                                           OpFoldResult numThreads,
+                                           OpFoldResult iterationSize) {
+  std::optional<int64_t> tileSizeConst = getConstantIntValue(tileSize);
+  std::optional<int64_t> numThreadsConst = getConstantIntValue(numThreads);
+  std::optional<int64_t> iterSizeConst = getConstantIntValue(iterationSize);
+  if (!tileSizeConst || !numThreadsConst || !iterSizeConst)
+    return false;
+  return *tileSizeConst * (*numThreadsConst - 1) < *iterSizeConst;
+}
+
+/// Compute the tile offsets and sizes.
+static std::tuple<SmallVector<OpFoldResult>, SmallVector<OpFoldResult>>
+getTileOffsetAndSizes(RewriterBase &rewriter, Location loc, ValueRange ivs,
+                      ArrayRef<Range> iterationDomain,
+                      ArrayRef<OpFoldResult> tileSizes,
+                      ArrayRef<OpFoldResult> numThreads) {
+  SmallVector<OpFoldResult> offsets, sizes;
+  int materializedLoopNum = 0;
+
+  if (!numThreads.empty()) {
+    AffineExpr d0, d1, s0, s1, s2;
+    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);
----------------
nicolasvasilache wrote:

Is the extra multiplicative factor here coming from the fact that you added support for strides (which was introduced here IIRC https://discourse.llvm.org/t/rfc-parallel-loops-on-tensors-in-mlir/68332) ?

If so, we also need tests for such cases.


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