Merge pull request #841 from codeWorm2015/metal

update

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

@@ -342,6 +342,10 @@ public extension MTLTexture {
   
   // n c h w - dim
   func toTensor(dim: (n: Int, c: Int, h: Int, w: Int)) -> [Float32] {
+    print("origin dim: \(dim)")
+    print("texture: ")
+    print(self)
+    
     let textureArray = floatArray { (i : Float32) -> Float32 in
       return i
     }
@@ -358,8 +362,42 @@ public extension MTLTexture {
         }
       }
     }
+    print(" tensor count -- \(output.count)")
     return output
   }
+  
+  func realNHWC(dim: (n: Int, h: Int, w: Int, c: Int)) -> [Float32] {
+    print("origin dim: \(dim)")
+    print("texture: ")
+    print(self)
+    
+    let textureArray = floatArray { (i : Float32) -> Float32 in
+      return i
+    }
+    var output: [Float32] = []
+
+    let numOfASlice = dim.h * dim.w * 4
+    for h in 0..<dim.h {
+      for w in 0..<dim.w {
+        for sliceIndex in 0..<arrayLength {
+          if sliceIndex * 4 + 4 > dim.c {
+            for i in 0..<((sliceIndex * 4 + 4) - dim.c) {
+              let value = textureArray[sliceIndex * numOfASlice + h * dim.w * 4 + w * 4 + i]
+              output.append(value)
+            }
+          } else {
+            for i in 0..<4 {
+              let value = textureArray[sliceIndex * numOfASlice + h * dim.w * 4 + w * 4 + i]
+              output.append(value)
+            }
+          }
+        }
+      }
+    }
+    print(" tensor count -- \(output.count)")
+    return output
+  }
+  
 }
 
 

metal/paddle-mobile/paddle-mobile/Executor.swift

@@ -14,7 +14,7 @@
 
 import Foundation
 
-let testTo = 41
+let testTo = 48
 
 public class ResultHolder<P: PrecisionType> {
   public let dim: [Int]

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

@@ -47,9 +47,9 @@ class ConvAddOp<P: PrecisionType>: Operator<ConvAddKernel<P>, ConvAddParam<P>>,
   func delogOutput() {
     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)
     
 //    print(" filter array: ")

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, 3, 1, 2])
+        param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1])
         
         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, 3, 1, 2])
+    param.output.initTexture(device: device, inTranspose: [0, 2, 3, 1])
     
     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

@@ -49,8 +49,12 @@ class TransposeOp<P: PrecisionType>: Operator<TransposeKernel<P>, TransposeParam
     }
   }
   func delogOutput() {
+    print(para.input.metalTexture.toTensor(dim: (n: para.input.tensorDim[0], c: para.input.tensorDim[1], h: para.input.tensorDim[2], w: para.input.tensorDim[3])).strideArray())
     
-    print(para.output.metalTexture.toTensor(dim: (n: 1, c: 21, h: 19, w: 19)).strideArray())
+    
+    let originDim = para.output.tensorDim
+    let outputArray = para.output.metalTexture.realNHWC(dim: (n: originDim[0], h: originDim[1], w: originDim[2], c: originDim[3]))
+    print(outputArray.strideArray())
     
     
 //    let inputArray: [Float32] = para.input.metalTexture.floatArray { (ele: Float32) -> Float32 in