Implement remaining changes

Author: Xeratec

Deeploy/Targets/Generic/Layers.py

@@ -652,15 +652,12 @@ def computeShapes(self, inputShapes: Shape, outputShapes: Shape, operatorReprese
         weight_shape = inputShapes[1]
 
         if newOutputShapes and len(newOutputShapes[0]) >= 2:
+            # For 1D: weight_shape = [C_in, C_out // group, kW]
+            # For 2D: weight_shape = [C_in, C_out // group, kH, kW]
+            ch_out = weight_shape[1] * group
             if channels_first:
-                # For 1D: weight_shape = [C_in, C_out // group, kW]
-                # For 2D: weight_shape = [C_in, C_out // group, kH, kW]
-                ch_out = weight_shape[1] * group
                 newOutputShapes[0][1] = ch_out
             else:
-                # For 1D: weight_shape = [C_in, C_out // group, kW]
-                # For 2D: weight_shape = [C_in, C_out // group, kH, kW]
-                ch_out = weight_shape[-2] * group
                 newOutputShapes[0][-1] = ch_out
 
         return newInputShapes, newOutputShapes

Deeploy/Targets/Generic/Parsers.py

@@ -2765,7 +2765,7 @@ def parseNode(self, node: gs.Node) -> bool:
             self.operatorRepresentation['padding_y'] = int(self.operatorRepresentation['pads'][0])
             self.operatorRepresentation['stride_y'] = int(self.operatorRepresentation['strides'][0])
 
-        return wellFormed
+        return ret
 
     def parseNodeCtxt(self,
                       ctxt: NetworkContext,
@@ -2787,4 +2787,5 @@ def parseNodeCtxt(self,
                 "batchOffsetIn"] = self.operatorRepresentation["ch_im_in"] * self.operatorRepresentation["dim_im_in_y"]
             self.operatorRepresentation["batchOffsetOut"] = self.operatorRepresentation[
                 "ch_im_out"] * self.operatorRepresentation["dim_im_out_y"]
-        return ctxt, True
+            return newCtxt, True
+        return ctxt, False

Deeploy/Targets/Generic/Templates/ConvTransposeTemplate.py

@@ -77,7 +77,9 @@ def alignToContext(self, ctxt: NetworkContext,
             ${bias}, ${has_bias},
             ref_${data_out}_${data_out}, ${dim_im_out_y}
         );
-        
+
+        ref_${data_out}_${data_in} += ${batchOffsetIn};
+        ref_${data_out}_${data_out} += ${batchOffsetOut};
     }
 END_SINGLE_CORE
 """)

Deeploy/Targets/Generic/Templates/FloatConvTemplate.py

@@ -31,6 +31,10 @@
 """)
 
 reference1DTemplate = NodeTemplate("""
+<%
+batchOffsetIn = ch_im_in * dim_im_in_y
+batchOffsetOut = ch_im_out * dim_im_out_y
+%>
     // 1D FP Conv (Name: ${nodeName}, Op: ${nodeOp})
     BEGIN_SINGLE_CORE
         ${data_in_type.typeName} ref_${data_out}_${data_in} = ${data_in};
@@ -45,6 +49,9 @@
                 ref_${data_out}_${data_out},
                 ${dim_im_out_y}
             );
+
+            ref_${data_out}_${data_in} += ${batchOffsetIn};
+            ref_${data_out}_${data_out} += ${batchOffsetOut};
         }
     END_SINGLE_CORE
     """)

Deeploy/Targets/Generic/Templates/FloatMaxPoolTemplate.py

@@ -22,14 +22,22 @@
 """)
 
 reference1DTemplate = NodeTemplate("""
+<%
+batchOffsetIn = ch_im_in * dim_im_in_y
+batchOffsetOut = ch_im_out * dim_im_out_y
+%>
     // 1D Float MaxPool (Name: ${nodeName}, Op: ${nodeOp})
     BEGIN_SINGLE_CORE
+        ${data_in_type.typeName} ref_${data_out}_${data_in} = ${data_in};
+        ${data_out_type.typeName} ref_${data_out}_${data_out} = ${data_out};
         for (uint32_t n=0; n<${batch}; ++n) {
             MaxPool1d_fp32_fp32(
-                ${data_in}, ${ch_im_in}, ${dim_im_in_y},
+                ref_${data_out}_${data_in}, ${ch_im_in}, ${dim_im_in_y},
                 ${dim_kernel_y}, ${stride_y},
-                ${data_out}
+                ref_${data_out}_${data_out}
             );
+            ref_${data_out}_${data_in} += ${batchOffsetIn};
+            ref_${data_out}_${data_out} += ${batchOffsetOut};
         }
     END_SINGLE_CORE
 """)

Deeploy/Targets/Generic/Templates/FloatPadTemplate.py

@@ -65,7 +65,9 @@
 
 // 1D Float Pad (Name: ${nodeName}, Op: ${nodeOp})
 BEGIN_SINGLE_CORE
-    memset(${data_out}, 0, ${data_out_size}*sizeof(${data_out_type.referencedType.typeName}));
+    for (uint32_t i = 0; i < ${data_out_size}; i++) {
+        ${data_out}[i] = ${value};
+    }
     uint32_t xoffset_${data_out}_${data_in};
     uint32_t offset_in_${data_out}_${data_in} = 0;
 

Deeploy/Targets/MemPool/Platform.py

@@ -37,7 +37,7 @@
 
 # Fallback bindings from the generic platform
 # (they support a wider range of attribute values)
-GenericConv1D_Mapper = NodeMapper(GenericConv1DParser(), [BasicConv1DBindings])
+GenericConv1D_Mapper = NodeMapper(GenericConv1DParser(), BasicConv1DBindings)
 GenericDWConv1D_Mapper = NodeMapper(GenericDWConv1DParser(), [BasicDWConv1DBinding])
 GenericConv2D_Mapper = NodeMapper(GenericConv2DParser(), BasicConv2DBindings)
 GenericDWConv2D_Mapper = NodeMapper(GenericDWConv2DParser(), BasicDWConv2DBindings)