1.optimize add by channel; 2. texture reuse should consider param input and...

1.optimize add by channel; 2. texture reuse should consider param input and multi fetch; 3. should also realloc heap when max size become smaller; 4. paddle mobile test disable cache (#1769)

* optimize add by channel

* 1. texture reuse should consider param input and multi fetch; 2.  should also realloc heap when max size become smaller; 3. paddle mobile test disable cache

metal/PaddleMobileTest/PaddleMobileTest/TestViewController.swift

@@ -63,7 +63,7 @@ let device = MTLCreateSystemDefaultDevice()!
 let commandQueue = device.makeCommandQueue()!
 var timeCosts = [Double]()
 var count = 0
-var totalCount = 10
+var totalCount = 100
 
 var orderedVars: [String] = []
 var varIndex = 0
@@ -107,7 +107,9 @@ class TestViewController: UIViewController {
     }
     
     private func getTestInfo() {
-        Alamofire.request("\(hostUrlStr)/getTestInfo").validate().responseJSON { (response) in
+        var testInfoRequest = URLRequest(url: URL(string: "\(hostUrlStr)/getTestInfo")!)
+        testInfoRequest.cachePolicy = .reloadIgnoringLocalAndRemoteCacheData
+        Alamofire.request(testInfoRequest).validate().responseJSON { (response) in
             guard response.result.isSuccess else {
                 self.testLog("getTestInfo request error")
                 return
@@ -194,7 +196,9 @@ class TestViewController: UIViewController {
         testResult.model = model
         
         let modelUrlStr = "\(hostUrlStr)/getFile/\(model.name)"
-        Alamofire.request("\(modelUrlStr)/model").validate().responseData { (response) in
+        var modelRequest = URLRequest(url: URL(string: "\(modelUrlStr)/model")!)
+        modelRequest.cachePolicy = .reloadIgnoringLocalAndRemoteCacheData
+        Alamofire.request(modelRequest).validate().responseData { (response) in
             guard response.result.isSuccess, let modelData = response.result.value else {
                 let msg = "get model \(model.name) error"
                 self.testLog(msg)
@@ -205,7 +209,9 @@ class TestViewController: UIViewController {
             //let modelData2 = try! Data(contentsOf: URL(fileURLWithPath: Bundle.main.path(forResource: "yolo_model_v3_16", ofType: nil)!))
             let modelPtr = UnsafeMutablePointer<UInt8>.allocate(capacity: modelData.count)
             NSData(data: modelData).getBytes(modelPtr, length: modelData.count)
-            Alamofire.request("\(modelUrlStr)/params/\(model.paramsPrecision)").validate().responseData(completionHandler: { (response) in
+            var paramsRequest = URLRequest(url: URL(string: "\(modelUrlStr)/params/\(model.paramsPrecision)")!)
+            paramsRequest.cachePolicy = .reloadIgnoringLocalAndRemoteCacheData
+            Alamofire.request(paramsRequest).validate().responseData(completionHandler: { (response) in
                 guard response.result.isSuccess, let paramsData = response.result.value else {
                     let msg = "get params \(model.name) error"
                     self.testLog(msg)
@@ -244,7 +250,10 @@ class TestViewController: UIViewController {
                 }
                 let fetchVar = fetchVars[0]
                 net.inputDim = Dim(inDim: [dims[0], dims[2], dims[3], dims[1]])
-                Alamofire.request("\(modelUrlStr)/data/\(feedVar.name.replacingOccurrences(of: "/", with: "_"))").validate().responseData(completionHandler: { (response) in
+               
+                var feedVarRequest = URLRequest(url: URL(string: "\(modelUrlStr)/data/\(feedVar.name.replacingOccurrences(of: "/", with: "_"))")!)
+                feedVarRequest.cachePolicy = .reloadIgnoringLocalAndRemoteCacheData
+                Alamofire.request(feedVarRequest).validate().responseData(completionHandler: { (response) in
                     guard response.result.isSuccess, let inputData = response.result.value else {
                         let msg = "get var \(feedVar) error"
                         self.testLog(msg)
@@ -309,7 +318,8 @@ class TestViewController: UIViewController {
                         for i in 0..<outputSize {
                             let a = output[i]
                             let b = resultHolder.result[i]
-                            if abs(a - b) > precision && abs(a - b) / min(abs(a), abs(b)) > 0.05 {
+                            // && abs(a - b) / min(abs(a), abs(b)) > 0.05 
+                            if abs(a - b) > precision {
                                 isResultEqual = false
                                 msg = "unequal: i: \(i) target: \(output[i]) result: \(resultHolder.result[i])"
                                 self.testLog(msg)
@@ -403,7 +413,9 @@ class TestViewController: UIViewController {
                 }
             }
             if severVars.contains(varName) {
-                Alamofire.request("\(urlString)/\(varName)").validate().responseData { (response) in
+                var severVarRequest = URLRequest(url: URL(string: "\(urlString)/\(varName)")!)
+                severVarRequest.cachePolicy = .reloadIgnoringLocalAndRemoteCacheData
+                Alamofire.request(severVarRequest).validate().responseData { (response) in
                     varIndex += 1
                     guard response.result.isSuccess, let varData = response.result.value else {
                         self.compareVars(runner: runner, model: model, completion: completion)

metal/paddle-mobile-metallib/paddle-mobile-metallib/Common.metal

@@ -109,6 +109,7 @@ inline void invtrans(int32_t trans[4], int32_t ipos[4], int32_t opos[4]) {
 
 struct ElementwiseAddParam {
     int32_t fast;
+    int32_t addByChannel;
     int32_t axis;
     int32_t ylen;
     int32_t xdim[4];

metal/paddle-mobile-metallib/paddle-mobile-metallib/ConvAddReluMetal.metal

@@ -19,8 +19,10 @@ using namespace metal;
 
 half4 getBiasHalf(uint3 gid, constant ElementwiseAddParam &addParam, texture2d_array<half, access::sample> biasTexture) {
     half4 output;
-    if (addParam.fast) {
+    if (addParam.fast == 1) {
         output = biasTexture.read(gid.xy, gid.z);
+    } else if (addParam.addByChannel == 1) {
+        output = biasTexture.read(uint2(0, 0), gid.z);
     } else {
         int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
         int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
@@ -44,8 +46,10 @@ half4 getBiasHalf(uint3 gid, constant ElementwiseAddParam &addParam, texture2d_a
 
 float4 getBias(uint3 gid, constant ElementwiseAddParam &addParam, texture2d_array<float, access::sample> biasTexture) {
     float4 output;
-    if (addParam.fast) {
+    if (addParam.fast == 1) {
         output = float4(biasTexture.read(gid.xy, gid.z));
+    } else if (addParam.addByChannel == 1) {
+        output = float4(biasTexture.read(uint2(0, 0), gid.z));
     } else {
         int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
         int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];

metal/paddle-mobile-metallib/paddle-mobile-metallib/Elementwise.metal

@@ -26,12 +26,12 @@ kernel void elementwise_add(texture2d_array<float, access::read> inputX [[textur
         gid.y >= outTexture.get_height() ||
         gid.z >= outTexture.get_array_size()) return;
     float4 rx, ry;
-    
+    rx = inputX.read(gid.xy, gid.z);
     if (pm.fast == 1) {
-        rx = inputX.read(gid.xy, gid.z);
         ry = inputY.read(gid.xy, gid.z);
+    } else if (pm.addByChannel == 1) {
+        ry = inputY.read(uint2(0, 0), gid.z);
     } else {
-        rx = inputX.read(gid.xy, gid.z);
         int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
         int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
         int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};
@@ -62,12 +62,12 @@ kernel void elementwise_add_half(texture2d_array<half, access::read> inputX [[te
         gid.y >= outTexture.get_height() ||
         gid.z >= outTexture.get_array_size()) return;
     half4 rx, ry;
-    
+    rx = inputX.read(gid.xy, gid.z);
     if (pm.fast == 1) {
-        rx = inputX.read(gid.xy, gid.z);
         ry = inputY.read(gid.xy, gid.z);
+    } else if (pm.addByChannel == 1) {
+        ry = inputY.read(uint2(0, 0), gid.z);
     } else {
-        rx = inputX.read(gid.xy, gid.z);
         int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
         int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
         int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};

metal/paddle-mobile-metallib/paddle-mobile-metallib/ElementwiseAddPreluKernel.inc.metal

@@ -37,12 +37,12 @@ kernel void FUNC3_(elementwise_add, PRELU_TYPE, P)(texture2d_array<P, access::re
         gid.y >= outTexture.get_height() ||
         gid.z >= outTexture.get_array_size()) return;
     VECTOR(P, 4) rx, ry;
-    
+    rx = inputX.read(gid.xy, gid.z);
     if (pm.fast == 1) {
-        rx = inputX.read(gid.xy, gid.z);
         ry = inputY.read(gid.xy, gid.z);
+    } else if (pm.addByChannel == 1) {
+        ry = inputY.read(uint2(0, 0), gid.z);
     } else {
-        rx = inputX.read(gid.xy, gid.z);
         int32_t x_xyzn[4] = {int32_t(gid.x), int32_t(gid.y), int32_t(gid.z), 0}, x_abcd[4], t_abcd[4];
         int32_t y_abcd[4] = {0, 0, 0, 0}, y_xyzn[4];
         int32_t xtrans[4] = {pm.xtrans[0], pm.xtrans[1], pm.xtrans[2], pm.xtrans[3]};

metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvAddReluKernel.swift

@@ -160,7 +160,7 @@ class ConvAddReluKernel<P: PrecisionProtocol>: Kernel, Computable {
             try setupWithMPS(device: device, param: param)
         } else {
             if functionName == nil {
-                fatalError(" unsupport yet ")
+                throw PaddleMobileError.makeError(type: .netError, msg: "function name nil")
             }
             try super.init(device: device, inFunctionName: functionName, initContext: initContext)
             try setupWithoutMPS(device: device, param: param)
@@ -371,7 +371,7 @@ class ConvAddReluKernel<P: PrecisionProtocol>: Kernel, Computable {
     }
     
     private class func canMPSAddByElement(param: ConvAddReluParam<P>) -> Bool {
-        if let y = param.y, y.dim.dims == param.input.dim.dims {
+        if let y = param.y, y.dim.dims == param.output.dim.dims {
             return true
         }
         return false

metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ConvKernel.swift

@@ -46,7 +46,7 @@ class ConvKernel<P: PrecisionProtocol>: Kernel, Computable {
             try setupWithMPS(device: device, param: param)
         } else {
             if functionName == nil {
-                fatalError(" unsupport yet ")
+                throw PaddleMobileError.makeError(type: .netError, msg: "function name nil")
             }
             try super.init(device: device, inFunctionName: functionName, initContext: initContext)
             try setupWithoutMPS(device: device, param: param)

metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ElementwiseAddKernel.swift

@@ -16,6 +16,7 @@ import Foundation
 
 struct ElementwiseAddMetalParam {
     var fast: Int32 = 0
+    var addByChannel: Int32 = 0
     var axis: Int32 = 0
     var ylen: Int32 = 0
     var xdim: (Int32, Int32, Int32, Int32) = (0, 0, 0, 0)
@@ -91,6 +92,9 @@ class ElementwiseAddKernel<P: PrecisionProtocol>: Kernel, Computable {
             //      print("===> elementwise_add fast!!!")
             metalParam.fast = 1
         }
+        if inputY.tensorDim.cout() == 1 && (axis == 1 || (axis == -1 && inputY.tensorDim.dims[0] == inputX.padToFourDim[1])) {
+            metalParam.addByChannel = 1
+        }
         return metalParam
     }
 }

metal/paddle-mobile/paddle-mobile/Src/Operators/Kernels/ElementwiseAddPreluKernel.swift

@@ -23,27 +23,7 @@ class ElementwiseAddPreluKernel<P: PrecisionProtocol>: Kernel, Computable {
         
         try param.alpha.initBuffer(device: device, precision: GlobalConfig.shared.computePrecision)
         
-        metalParam = ElementwiseAddMetalParam.init()
-        
-        let xdim: [Int32] = (0..<4).map { Int32(param.inputX.dim[$0]) }
-        let ydim: [Int32] = (0..<4).map { Int32(param.inputY.dim[$0]) }
-        let xtrans: [Int32] = (0..<4).map { Int32(param.inputX.transpose[$0]) }
-        let ytrans: [Int32] = (0..<4).map { Int32(param.inputY.transpose[$0]) }
-        
-        metalParam.xdim = (xdim[0], xdim[1], xdim[2], xdim[3])
-        metalParam.ydim = (ydim[0], ydim[1], ydim[2], ydim[3])
-        metalParam.xtrans = (xtrans[0], xtrans[1], xtrans[2], xtrans[3])
-        metalParam.ytrans = (ytrans[0], ytrans[1], ytrans[2], ytrans[3])
-        if param.axis == -1 {
-            metalParam.axis = 4 - Int32(param.inputY.tensorDim.cout())
-        } else {
-            metalParam.axis = 4 - Int32(param.inputX.tensorDim.cout()) + Int32(param.axis)
-        }
-        metalParam.ylen = Int32(param.inputY.tensorDim.cout())
-        if (param.inputX.dim == param.inputY.dim) && (param.inputX.transpose == param.inputY.transpose) {
-            //      print("===> elementwise_add fast!!!")
-            metalParam.fast = 1
-        }
+        metalParam = ElementwiseAddKernel<P>.metalParamFrom(inputX: param.inputX, inputY: param.inputY, axis: param.axis)
         
         if GlobalConfig.shared.computePrecision == .Float32 {
             if param.mode == "channel" {

metal/paddle-mobile/paddle-mobile/Src/Program/MemoryOptimze.swift

@@ -66,6 +66,7 @@ class MemoryOptimize: MemoryManager {
         var createdNodes = [String: Node]()
         var nodesArray = [Node]()
         let scope = program.scope
+        var fetchVarNames: [String] = []
         func appendNodes(textureDic: [String: [String]], varsDic: [String: PMVarDesc]) {
             for dicPair in textureDic {
                 for varName in dicPair.value {
@@ -94,9 +95,16 @@ class MemoryOptimize: MemoryManager {
                 varsDic[varDesc.name] = varDesc
             }
             for op in block.ops {
+                if op.type == gFetchType {
+                    for names in op.inputs.values {
+                        fetchVarNames.append(contentsOf: names)
+                    }
+                }
                 appendNodes(textureDic: op.inputs, varsDic: varsDic)
+                appendNodes(textureDic: op.paraInputs, varsDic: varsDic)
                 appendNodes(textureDic: op.outputs, varsDic: varsDic)
                 appendNodes(textureDic: op.inputs, varsDic: varsDic)
+                appendNodes(textureDic: op.paraInputs, varsDic: varsDic)
             }
         }
         var nodeGroups: [[Node]] = []
@@ -106,7 +114,8 @@ class MemoryOptimize: MemoryManager {
                 node.visited = true
                 var placed = false
                 for i in 0..<nodeGroups.count {
-                    if nodeGroups[i].last?.count == 0 {
+                    let lastNode = nodeGroups[i].last
+                    if lastNode?.count == 0 && !fetchVarNames.contains(lastNode?.name ?? "") {
                         nodeGroups[i].append(node)
                         placed = true
                         break
@@ -156,7 +165,7 @@ class MemoryBucket {
     
     public func allocHeap() {
         let size = maxSizeForTextures(textures)
-        if size > (heap?.size ?? 0) {
+        if size != (heap?.size ?? 0) {
             heap?.setPurgeableState(.empty)
             heap = makeHeapForSize(size)
         }