Merge pull request #840 from codeWorm2015/metal

update

metal/paddle-mobile/paddle-mobile/Common/MetalExtension.swift

@@ -339,6 +339,27 @@ public extension MTLTexture {
     return nil
     
   }
+  
+  // n c h w - dim
+  func toTensor(dim: (n: Int, c: Int, h: Int, w: Int)) -> [Float32] {
+    let textureArray = floatArray { (i : Float32) -> Float32 in
+      return i
+    }
+    var output: [Float32] = []
+    for s in 0..<arrayLength {
+      for c in 0..<4{
+        for h in 0..<dim.h {
+          for w in 0..<dim.w {
+            if (s * 4 + c) < dim.c {
+              let textureValue = textureArray[dim.w * dim.h * 4 * s + h * dim.w * 4 + w * 4 + c]
+              output.append(textureValue)
+            }
+          }
+        }
+      }
+    }
+    return output
+  }
 }
 
 

metal/paddle-mobile/paddle-mobile/Operators/ConvAddOp.swift

@@ -45,7 +45,10 @@ class ConvAddParam<P: PrecisionType>: OpParam {
 class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>, Runable, Creator, InferShaperable, Fusion{
   
   func delogOutput() {
-    print(" conv add: ")
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
+
+    
+//    print(" conv add: ")
     
 //    print(para.input.metalTexture)
     
@@ -53,16 +56,16 @@ class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>,
 //    let filterArray: [P] = para.filter.buffer.array()
 //    print(filterArray)
     
-    let input = para.input.metalTexture.floatArray { (p: P) -> P in
-      return p
-    }
+//    let input = para.input.metalTexture.floatArray { (p: P) -> P in
+//      return p
+//    }
 //    print(input)
     
-    let output = para.output.metalTexture.floatArray { (p: P) -> P in
-      return p
-    }
+//    let output = para.output.metalTexture.floatArray { (p: P) -> P in
+//      return p
+//    }
 //    print(para.output.metalTexture)
-    print(output)
+//    print(output)
   }
   
   

metal/paddle-mobile/paddle-mobile/Operators/ConvBNReluOp.swift

@@ -159,19 +159,23 @@ class ConvBNReluOp<P: PrecisionType>: Operator<ConvBNReluKernel<P>, ConvBNReluPa
 //      print("index: \(i) \(newBiase![i]) ")
 //    }
     
-    print(para.output.metalTexture)
-    
-    let output = para.output.metalTexture.floatArray { (p: P) -> P in
-      return p
-    }
-    print(output)
+//    print(para.output.metalTexture)
+//
+//
 //
-    writeToLibrary(fileName: "batch_norm_34.tmp_2", array: output)
-    print(" write done")
+//    let output = para.output.metalTexture.floatArray { (p: P) -> P in
+//      return p
+//    }
+//    print(output)
 //
     
+//
+//    writeToLibrary(fileName: "batch_norm_34.tmp_2", array: output)
+//    print(" write done")
+//
     
     
+    print(para.output.metalTexture.toTensor(dim: (n: para.output.tensorDim[0], c: para.output.tensorDim[1], h: para.output.tensorDim[2], w: para.output.tensorDim[3])).strideArray())
     
     //        let _: P? = para.output.metalTexture.logDesc(header: "conv add batchnorm relu output: ", stridable: true)
   }

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddBatchNormReluKernel.swift

@@ -50,7 +50,7 @@ class ConvAddBatchNormReluKernel<P: PrecisionType>: Kernel, Computable, Testable
 
     required init(device: MTLDevice, param: ConvAddBatchNormReluParam<P>) {
         
-        param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1])
+        param.output.initTexture(device: device, inTranspose: [0, 3, 1, 2])
         
         if param.filter.width == 1 && param.filter.height == 1 {
             super.init(device: device, inFunctionName: "conv_add_batch_norm_relu_1x1")

metal/paddle-mobile/paddle-mobile/Operators/Kernels/ConvAddKernel.swift

@@ -25,7 +25,7 @@ class ConvAddKernel<P: PrecisionType>: Kernel, Computable {
       super.init(device: device, inFunctionName: "conv_add_3x3")
     }
     
-    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1])
+    param.output.initTexture(device: device, inTranspose: [0, 3, 1, 2])
     
     let offsetX = param.filter.width/2 - Int(param.paddings[0])
     let offsetY = param.filter.height/2 - Int(param.paddings[1])

metal/paddle-mobile/paddle-mobile/Operators/TransposeOp.swift

@@ -49,17 +49,21 @@ class TransposeOp<P: PrecisionType>: Operator<TransposeKernel<P>, TransposeParam
     }
   }
   func delogOutput() {
-    let inputArray: [Float32] = para.input.metalTexture.floatArray { (ele: Float32) -> Float32 in
-      return ele
-    }
     
-    print(inputArray.strideArray())
+    print(para.output.metalTexture.toTensor(dim: (n: 1, c: 21, h: 19, w: 19)).strideArray())
     
-    let outputArray: [Float32] = para.output.metalTexture.floatArray { (ele: Float32) -> Float32 in
-      return ele
-    }
-  
-    print(outputArray.strideArray())
+    
+//    let inputArray: [Float32] = para.input.metalTexture.floatArray { (ele: Float32) -> Float32 in
+//      return ele
+//    }
+//
+//    print(inputArray.strideArray())
+//
+//    let outputArray: [Float32] = para.output.metalTexture.floatArray { (ele: Float32) -> Float32 in
+//      return ele
+//    }
+//
+//    print(outputArray.strideArray())
     
 //    writeToLibrary(fileName: "transpose_ouput", array: outputArray)
   }