basic solution

compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/AMDAIETileAndFuse.cpp

@@ -164,9 +164,7 @@ void AMDAIETileAndFusePass::runOnOperation() {
     options.setLoopType(scf::SCFTilingOptions::LoopType::ForallOp);
     // Here we assume there are always two levels of parallel (scf.forall)
     // loops, and the first level of tiling is always using scf.forall and
-    // mapped to blocks. Currently we are not using mapping attributes for
-    // Conv2d ops, because there could be four parallel tiling dimensions.
-    // TODO (vivian): create AIE specific mapping attributes.
+    // mapped to blocks. TODO (vivian): create AIE specific mapping attributes.
     if (!isa<linalg::ConvolutionOpInterface>(consumerOp.getOperation())) {
       if (tilingLevel == 0) {
         options.setMapping(
@@ -177,7 +175,22 @@ void AMDAIETileAndFusePass::runOnOperation() {
             {gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::DimY),
              gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::DimX)});
       }
+    } else {
+      if (tilingLevel == 1) {
+        options.setMapping({
+            gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::DimY),
+            gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::DimX),
+            gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::DimZ),
+            gpu::GPUThreadMappingAttr::get(context, gpu::MappingId::LinearDim0),
+        });
+      } else {
+        // TODO(newling) GPU block mapping attribute for convolution.
+        // AT the moment, this mapping isn't used, just gets temporally
+        // unrolled anyway.
+      }
     }
+    // TODO(vivan/newling) More sustainable way of partitioning ops...
+    // "ConvolutionOpInterface" vs "everything else" doesn't seem good!
   }
 
   IRRewriter rewriter(context);