Fix L3 Transfers

Author: Xeratec

Deeploy/Targets/PULPOpen/DMA/L3Dma.py

@@ -16,7 +16,8 @@ class L3DmaFuture(Future):
     % if comment:
     // ${comment}
     % endif
-    pi_cl_ram_req_t ${name};
+    pi_cl_ram_req_t ${name} = {0};
+
     """)
 
     _deinitTemplate = NodeTemplate("")
@@ -27,7 +28,9 @@ class L3DmaFuture(Future):
     % if comment:
     // ${comment}
     % endif
-    pi_cl_ram_copy_wait(&${name});
+    if (${name}.size != 0) {
+        pi_cl_ram_copy_wait(&${name});
+    }
     """)
 
 

Deeploy/TilingExtension/AsyncDma.py

@@ -178,9 +178,19 @@ def transfer(self,
                  future: Future,
                  comment: str = "") -> Tuple[List[CodeSnippet], List[CodeSnippet], List[CodeSnippet]]:
         callStack = []
-        callStack.extend(
-            self.dma.transfer(ctxt, externalBuffer, localBuffer, shape, strideExt, strideLoc, direction, future))
+        alloc_code, dma_code, deinit_code = self.dma.transfer(ctxt,
+                                                              externalBuffer,
+                                                              localBuffer,
+                                                              shape,
+                                                              strideExt,
+                                                              strideLoc,
+                                                              direction,
+                                                              future,
+                                                              comment = comment)
+        callStack.extend(alloc_code)
+        callStack.extend(dma_code)
         callStack.append(future.wait())
+        callStack.extend(deinit_code)
 
         return [], callStack, []
 

Deeploy/TilingExtension/CodeTransformationPasses/SingleBufferingTilingCodeGeneration.py

@@ -7,7 +7,7 @@
 from Deeploy.AbstractDataTypes import VoidType
 from Deeploy.DeeployTypes import CodeSnippet, ExecutionBlock, NetworkContext, NodeTemplate, OperatorRepresentation, \
     VariableBuffer, _ReferenceBuffer
-from Deeploy.TilingExtension.AsyncDma import AsyncDma, DmaDirection, Future
+from Deeploy.TilingExtension.AsyncDma import AsyncDma, DmaDirection, EmptyFuture, Future
 from Deeploy.TilingExtension.CodeTransformationPasses.TilingCodeGeneration import TilingCodeGeneration
 from Deeploy.TilingExtension.CodeTransformationPasses.TilingHoistingMixIn import dictOfArrays
 from Deeploy.TilingExtension.CodeTransformationPasses.TilingPrototypes import ProfilingPrototypeMixIn, \
@@ -66,6 +66,8 @@ def _generateTransferScheduleCalls(self,
                                                      override_type = VoidType)
 
             future = self.dma.getFuture(tensorName, direction)
+            _future = set([future]) - futures
+            _future = _future.pop() if len(_future) > 0 else EmptyFuture("")
             futures.add(future)
 
             callStack.extend(
@@ -76,7 +78,7 @@ def _generateTransferScheduleCalls(self,
                                                localBuffer,
                                                externalBufferRef,
                                                direction,
-                                               future,
+                                               _future,
                                                comment = comment))
 
             referenceUpdate = self._generateExternalReferenceUpdate(ctxt, tensorName, rectangles, tileIdxVar,