[Dynamism][LNL] UngroupBoundedBuffersPass - support SwKernelOps with multiple outputs; and other fixes (#42)

Author: DariaMityagina

src/vpux_compiler/src/dialect/VPUIP/transforms/passes/ungroup_bounded_buffer.cpp

@@ -105,9 +105,6 @@ mlir::LogicalResult UngroupSwKernelOp::matchAndRewrite(VPUIP::SwKernelOp origOp,
         }
     }
 
-    VPUX_THROW_WHEN(origOp.getOutputBuffs().size() != 1 || origOp.getOutputs().size() != 1,
-                    "UngroupBoundedBuffers pass supports SwKernelOp with single output for now");
-
     SmallVector<mlir::Value> swKernelOutputBuffs;
     SmallVector<mlir::Value> swKernelDynamicOutputShapes;
     SmallVector<int32_t> swKernelDynamicOutputShapesMap;
@@ -134,20 +131,17 @@ mlir::LogicalResult UngroupSwKernelOp::matchAndRewrite(VPUIP::SwKernelOp origOp,
     auto args = kernelArgsRange(origOp);
     initSwKernel(swKernelOp, swKernelOperands, swKernelOutputBuffs, args, _log.nest());
 
-    for (auto result : origOp.getResults()) {
-        if (result.getType().isa<VPUIP::BoundedBufferType>()) {
-            // if result is BoundedBufferType, then it is guaranteed to have swKernelDynamicOutputShapes not empty
-            // Tracking number [E#115679]
-            // TODO: SwKernelOp's results 0 and 1 may not always be correct here, only in the case where
-            // there is only one bounded result. Similarly, using replaceOp here when there are more than
-            // one result will not behave as intended.
-            auto groupOp = rewriter.create<VPUIP::GroupBoundedBufferOp>(swKernelOp.getLoc(), swKernelOp.getResult(0),
-                                                                        swKernelOp.getResult(1));
-            rewriter.replaceOp(origOp, groupOp.getOutput());
+    SmallVector<mlir::Value> newResults;
+    for (auto i : irange(origOp.getNumResults())) {
+        if (mlir::isa<VPUIP::BoundedBufferType>(origOp.getResult(i).getType())) {
+            auto groupOp = rewriter.create<VPUIP::GroupBoundedBufferOp>(swKernelOp.getLoc(), swKernelOp.getResult(i),
+                                                                        swKernelOp.getDynamicOutputShapes()[i]);
+            newResults.push_back(groupOp.getOutput());
         } else {
-            rewriter.replaceOp(origOp, swKernelOp.getResults());
+            newResults.push_back(swKernelOp.getResult(i));
         }
     }
+    rewriter.replaceOp(origOp, newResults);
 
     return mlir::success();
 }

sw_runtime_kernels/kernels/prebuild/act_shave_bin/convert.3720xx.elf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7b9820e53c8e1b046fac50fa280a7ed21d291360cfafabd71a6772ae1c5e1d73
-size 54364
+oid sha256:493a994293f81eb38525fdec6d58e62eb88a8f4cf5df203dbe70eb0db58b4f01
+size 54332

sw_runtime_kernels/kernels/prebuild/act_shave_bin/convert.3720xx_lsu0_wo.elf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2f5173e02f7ab2d7f7f4bb4211af7e7ef9aca12337e0396bdd3c1ca45f8a7e90
-size 54444
+oid sha256:bb6697592a8bd9cb867be4dc2d55659515bbdd1b39c45a0704519c3d63701cd4
+size 54396

sw_runtime_kernels/kernels/prebuild/act_shave_bin/convert.4000xx.elf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:493226913355517e0a7b504b9caf71d058782ccec0e806cb2e1957772430480c
-size 55116
+oid sha256:1593ecaeca176873a1a64990b2bb446cff89ebcfe8484109ea12ecafc602138d
+size 56124

sw_runtime_kernels/kernels/prebuild/act_shave_bin/dynamic_reshape.4000xx.elf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f57ed9b6cdfcc3a74b4c302d4d9fa450a7b02123b46b43a367a4ef9cc32a9d4
+size 3412

sw_runtime_kernels/kernels/prebuild/act_shave_bin/reorder.4000xx.elf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fb77a3c70bb477ca5d939223a05428d65200cb70c2721ca56d30e33e8816ac68
-size 8288
+oid sha256:2858cca7dc5279decdc10bd91ce02a3a95d9cf6b3e67d8d06d05197556034983
+size 3200

sw_runtime_kernels/kernels/prebuild/act_shave_bin/shape_of.4000xx.elf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbbb9184e800d0cff27e0fe79a3cafe1f99c3f6d60c45205d3fcddc0be68d1f1
+size 820