[Mlir-commits] [mlir] [mlir][memref][spirv] Add SPIR-V Image Lowering (PR #150978)

[Jack Frankland]

================
@@ -661,6 +682,79 @@ LoadOpPattern::matchAndRewrite(memref::LoadOp loadOp, OpAdaptor adaptor,
   return success();
 }
 
+LogicalResult
+ImageLoadOpPattern::matchAndRewrite(memref::LoadOp loadOp, OpAdaptor adaptor,
+                                    ConversionPatternRewriter &rewriter) const {
+  auto memrefType = cast<MemRefType>(loadOp.getMemref().getType());
+  if (memrefType.getMemorySpace() !=
+      spirv::StorageClassAttr::get(rewriter.getContext(),
+                                   spirv::StorageClass::Image))
+    return failure();
+
+  auto loadPtr = adaptor.getMemref();
+  auto memoryRequirements = calculateMemoryRequirements(loadPtr, loadOp);
+  if (failed(memoryRequirements))
+    return rewriter.notifyMatchFailure(
+        loadOp, "failed to determine memory requirements");
+
+  const auto [memoryAccess, alignment] = *memoryRequirements;
+
+  if (!loadOp.getMemRefType().hasRank())
+    return rewriter.notifyMatchFailure(
+        loadOp, "cannot lower unranked memrefs to SPIR-V images");
+
+  // We currently only support lowering of scalar memref elements to texels in
+  // the R[16|32][f|i|ui] formats. Future work will enable lowering of vector
+  // elements to texels in richer formats.
+  if (!loadOp.getMemRefType().getElementType().isIntOrFloat())
+    return rewriter.notifyMatchFailure(
+        loadOp, "cannot lower memrefs who's element type is not int or float "
+                "to SPIR-V images");
+
+  // We currently only support sampled images since OpImageFetch does not work
+  // for plain images and the OpImageRead instruction needs to be materialized
+  // instead or texels need to be accessed via atomics through a texel pointer.
+  // Future work will generalize support to plain images.
+  if (auto convertedPointeeType = cast<spirv::PointerType>(
----------------
FranklandJack wrote:

Yeah I agree, I was just trying to be clever here and reduce the scope of the `convertedPointeeType` variable to the `if` block, but I agree it is at the cost of reduced readability so happy to hoist it.

https://github.com/llvm/llvm-project/pull/150978