commit

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

@@ -98,14 +98,14 @@ extension Array where Element: Comparable{
 }
 
 extension Array {
-  func strideArray(inCount: Int = 20) -> Array<Element> {
+  func strideArray(inCount: Int = 20) -> [(Int, Element)] {
     if count < inCount {
-      return self
+      return (0..<count).map{ ($0, self[$0]) }
     } else {
       let stride = count / inCount
-      var newArray: [Element] = []
+      var newArray: [(Int, Element)] = []
       for i in 0..<inCount {
-        newArray.append(self[i * stride])
+        newArray.append((i * stride, self[i * stride]))
       }
       return newArray
     }

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

@@ -60,7 +60,7 @@ public class Executor<P: PrecisionType> {
     queue = inQueue
     for block in inProgram.programDesc.blocks {
       //block.ops.count
-      for i in 0..<39 {
+      for i in 0..<91 {
         let op = block.ops[i]
         do {
           let op = try OpCreator<P>.shared.creat(device: inDevice, opDesc: op, scope: inProgram.scope)
@@ -128,7 +128,7 @@ public class Executor<P: PrecisionType> {
       }
 //                  return
       
-      self.ops[38].delogOutput()
+      self.ops[90].delogOutput()
 //      self.ops[91].delogOutput()
 //      self.ops[92].delogOutput()
 //      self.ops[93].delogOutput()

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

@@ -62,6 +62,7 @@ class ConcatOp<P: PrecisionType>: Operator<ConcatKernel<P>, ConcatParam<P>>, Run
     let outputArray = para.output.metalTexture.floatArray { (o: Float32) -> Float32 in
       return o
     }
+    
     print(outputArray.strideArray())
     let device: MTLDevice = MTLCreateSystemDefaultDevice()!
     

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, 2, 1])
+    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/Kernels/metal/ReshapeKernel.metal

@@ -47,20 +47,19 @@ kernel void reshape(texture2d_array<float, access::read> inTexture [[texture(0)]
     invtrans(lrp.otrans, oabcd, tabcd);
     int index = abcd2index(lrp.odim, tabcd);
     if (index < count) {
-      int c = index % 4;
-      
-      int temp0 = index % (inTexture.get_array_size() * 4);
-      int slice = temp0 / 4;
-      
-      int temp1 = index % (inTexture.get_array_size() * 4 * lrp.idim[2]);
-      int w = temp1 / (inTexture.get_array_size() * 4);
-      
-      int h = index / (inTexture.get_array_size() * 4 * lrp.idim[2]);
-      
+//      int c = index % 4;
+//
+//      int temp0 = index % (inTexture.get_array_size() * 4);
+//      int slice = temp0 / 4;
+//
+//      int temp1 = index % (inTexture.get_array_size() * 4 * lrp.idim[2]);
+//      int w = temp1 / (inTexture.get_array_size() * 4);
+//
+//      int h = index / (inTexture.get_array_size() * 4 * lrp.idim[2]);
       
-//      index2abcd(lrp.idim, index, tabcd);
-//      abcd2xyzn(iC, tabcd, ixyzn);
-      r[n] = inTexture.read(uint2(w, h), slice)[c];
+      index2abcd(lrp.idim, index, tabcd);
+      abcd2xyzn(iC, tabcd, ixyzn);
+      r[n] = inTexture.read(uint2(ixyzn[0], ixyzn[1]), ixyzn[2])[ixyzn[3]];
     } else {
       r[n] = 0;
     }

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

@@ -53,6 +53,7 @@ class ReshapeOp<P: PrecisionType>: Operator<ReshapeKernel<P>, ReshapeParam<P>>,
   func delogOutput() {
     print("reshape delog")
 //    let _: P? = para.input.metalTexture.logDesc(header: "reshape input: ", stridable: false)
-    let _: P? = para.output.metalTexture.logDesc(header: "reshape output: ", stridable: false)
+    let _: P? = para.output.metalTexture.logDesc(header: "reshape output: ", stridable: true)
+    
   }
 }