[AttentionTraining] Fixed PR Comments

Author: runwangdl

Deeploy/Targets/Generic/Layers.py

@@ -479,10 +479,6 @@ class SGDLayer(ONNXLayer):
     def __init__(self, maps: List[NodeMapper]):
         super().__init__(maps)
 
-    def computeOps(self):
-
-        size = self.mapper.parser.operatorRepresentation['size']
-        return size * 2
 
 
 class LinearAttentionLayer(ONNXLayer):

Deeploy/Targets/PULPOpen/Bindings.py

@@ -190,8 +190,7 @@
                 FloatAddTemplate.referenceTemplate, ForkTransformer)
 ] + [
     NodeBinding(
-        AddChecker([PointerClass(float32_t), PointerClass(float32_t),
-                    PointerClass(float32_t)], [PointerClass(float32_t)]), FloatAddTemplate.referenceTemplate,
+        AddChecker([PointerClass(float32_t), PointerClass(float32_t)], [PointerClass(float32_t)]), FloatAddTemplate.referenceTemplate,
         ForkTransformer)
 ]
 

Deeploy/Targets/PULPOpen/Templates/FloatLayernormTemplate.py

@@ -33,9 +33,9 @@
 int32_t ${nodeName}_elem_end = ${nodeName}_end * ${lastDimLength};
 int32_t ${nodeName}_elem_count = ${nodeName}_elem_end - ${nodeName}_elem_start;
  
-const float* ${nodeName}_grad_in_ptr = ${grad_in} + ${nodeName}_elem_start;
-const float* ${nodeName}_data_in_ptr = ${data_in} + ${nodeName}_elem_start;
-float* ${nodeName}_grad_out_ptr = ${grad_out} + ${nodeName}_elem_start;
+const float${grad_in_type.referencedType.typeWidth}_t* ${nodeName}_grad_in_ptr = ${grad_in} + ${nodeName}_elem_start;
+const float${data_in_type.referencedType.typeWidth}_t* ${nodeName}_data_in_ptr = ${data_in} + ${nodeName}_elem_start;
+float${grad_out_type.referencedType.typeWidth}_t* ${nodeName}_grad_out_ptr = ${grad_out} + ${nodeName}_elem_start;
  
 if (${nodeName}_elem_count > 0) {
   LayernormGrad_fp${grad_in_type.referencedType.typeWidth}_fp${grad_out_type.referencedType.typeWidth}(

Deeploy/Targets/PULPOpen/Templates/SGDTemplate.py

@@ -9,9 +9,9 @@
 // (Name: ${nodeName}, Op: ${nodeOp})
 int8_t ${nodeName}_core_id = pi_core_id();
 int8_t ${nodeName}_log2Core = log2(NUM_CORES);
-int16_t ${nodeName}_chunk = (${size} >> ${nodeName}_log2Core) + ((${size} & (NUM_CORES-1))!=0);
-int16_t ${nodeName}_chunk_start = MIN(${nodeName}_chunk*${nodeName}_core_id, ${size});
-int16_t ${nodeName}_chunk_stop = MIN(${nodeName}_chunk_start + ${nodeName}_chunk, ${size});
+int32_t ${nodeName}_chunk = (${size} >> ${nodeName}_log2Core) + ((${size} & (NUM_CORES-1))!=0);
+int32_t ${nodeName}_chunk_start = MIN(${nodeName}_chunk*${nodeName}_core_id, ${size});
+int32_t ${nodeName}_chunk_stop = MIN(${nodeName}_chunk_start + ${nodeName}_chunk, ${size});
 
 ${weight_type.typeName} ref_${weight} = ${weight};
 ${grad_type.typeName} ref_${grad} = ${grad};

Deeploy/Targets/PULPOpen/TileConstraints/GEMMTileConstraint.py

@@ -193,10 +193,12 @@ class FloatGEMMTileConstraint(TileConstraint):
     @staticmethod
     def addGeometricalConstraint(tilerModel: TilerModel, parseDict: Dict, ctxt: NetworkContext) -> TilerModel:
 
+        # Get to-be-tiled tensor's buffers
         bufferA = ctxt.lookup(name = parseDict['A'])
         bufferB = ctxt.lookup(name = parseDict['B'])
         outputBuffer = ctxt.lookup(name = parseDict['data_out'])
 
+        # Add I/O dimensions to the model as variables
         has_bias = 'C' in parseDict and parseDict['C'] is not None
         bufferC = None
         if has_bias:
@@ -222,9 +224,11 @@ def addGeometricalConstraint(tilerModel: TilerModel, parseDict: Dict, ctxt: Netw
         outputFirstDimVar = tilerModel.getTensorDimVar(tensorName = outputBuffer.name, dimIdx = dimOffsetOut)
         outputSecondDimVar = tilerModel.getTensorDimVar(tensorName = outputBuffer.name, dimIdx = dimOffsetOut + 1)
 
+        # Map output dims to inputs dims
         tilerModel.addConstraint(outputFirstDimVar == AFirstDimVar)
         tilerModel.addConstraint(outputSecondDimVar == BSecondDimVar)
 
+        # Add GEMM Geometrical constraints
         tilerModel.addConstraint(ASecondDimVar == BFirstDimVar)
 
         # Add bias constraints only if bias is present
@@ -287,7 +291,6 @@ def serializeTilingSolution(
         transB = operatorRepresentation['transB']
 
         varA = operatorRepresentation['A']
-        varB = operatorRepresentation['B']
 
         if transA == 0:
             NSize = ctxt.lookup(varA).shape[-1]