modify ios interface

metal/paddle-mobile-demo/paddle-mobile-demo/ViewController.swift

@@ -175,19 +175,17 @@ class ViewController: UIViewController {
   override func viewDidLoad() {
     super.viewDidLoad()
     
-//    if runner.load() {
-//      print(" load success ! ")
-//    } else {
-//      print(" load error ! ")
-//    }
-//
     modelPickerView.delegate = self
     modelPickerView.dataSource = self
     threadPickerView.delegate = self
     threadPickerView.dataSource = self
+    if let image = UIImage.init(named: "test.jpg") {
+        selectImage = image
+        selectImageView.image = image
+    } else {
+        print("请添加测试图片")
+    }
     
-    selectImage = UIImage.init(named: "hand.jpg")
-    selectImageView.image = selectImage
     
 //    if platform == .CPU {
 //      inputPointer = runner.preproccess(image: selectImage!.cgImage!)

metal/paddle-mobile/paddle-mobile.xcodeproj/project.pbxproj

@@ -902,8 +902,8 @@
 			baseConfigurationReference = CDF58151D902A1CBAE56A0C2 /* Pods-paddle-mobile.debug.xcconfig */;
 			buildSettings = {
 				CLANG_ENABLE_MODULES = YES;
-				CODE_SIGN_IDENTITY = "";
-				CODE_SIGN_STYLE = Manual;
+				CODE_SIGN_IDENTITY = "iPhone Developer";
+				CODE_SIGN_STYLE = Automatic;
 				DEFINES_MODULE = YES;
 				DEVELOPMENT_TEAM = "";
 				DYLIB_COMPATIBILITY_VERSION = 1;
@@ -922,7 +922,7 @@
 					"$(inherited)",
 					"$(PROJECT_DIR)/paddle-mobile/CPU",
 				);
-				MACH_O_TYPE = staticlib;
+				MACH_O_TYPE = mh_dylib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";
@@ -939,8 +939,8 @@
 			baseConfigurationReference = E2A7957C92EDA5C3BEC0FFC2 /* Pods-paddle-mobile.release.xcconfig */;
 			buildSettings = {
 				CLANG_ENABLE_MODULES = YES;
-				CODE_SIGN_IDENTITY = "";
-				CODE_SIGN_STYLE = Manual;
+				CODE_SIGN_IDENTITY = "iPhone Developer";
+				CODE_SIGN_STYLE = Automatic;
 				DEFINES_MODULE = YES;
 				DEVELOPMENT_TEAM = "";
 				DYLIB_COMPATIBILITY_VERSION = 1;
@@ -959,7 +959,7 @@
 					"$(inherited)",
 					"$(PROJECT_DIR)/paddle-mobile/CPU",
 				);
-				MACH_O_TYPE = staticlib;
+				MACH_O_TYPE = mh_dylib;
 				MTL_LANGUAGE_REVISION = UseDeploymentTarget;
 				PRODUCT_BUNDLE_IDENTIFIER = "orange.paddle-mobile";
 				PRODUCT_NAME = "$(TARGET_NAME:c99extidentifier)";

metal/paddle-mobile/paddle-mobile/Operators/Kernels/metal/FetchKernel.metal

@@ -31,8 +31,8 @@ kernel void fetch(texture2d_array<float, access::read> inTexture [[texture(0)]],
   int output_to = 4 * input_width * input_height;
   output[gid.z * output_to + 0 * input_width * input_height + gid.y * input_width + gid.x] = input.x;
   output[gid.z * output_to + 1 * input_width * input_height + gid.y * input_width + gid.x] = input.y;
-//  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
-//  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
+  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
+  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
 }
 
 
@@ -52,8 +52,8 @@ kernel void fetch_half(texture2d_array<half, access::read> inTexture [[texture(0
   int output_to = 4 * input_width * input_height;
   output[gid.z * output_to + 0 * input_width * input_height + gid.y * input_width + gid.x] = input.x;
   output[gid.z * output_to + 1 * input_width * input_height + gid.y * input_width + gid.x] = input.y;
-//  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
-//  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
+  output[gid.z * output_to + 2 * input_width * input_height + gid.y * input_width + gid.x] = input.z;
+  output[gid.z * output_to + 3 * input_width * input_height + gid.y * input_width + gid.x] = input.w;
   
 }
 

src/io/ios_io/PaddleMobileCPU.h

@@ -27,59 +27,119 @@
 
 @end
 
+@interface  PaddleMobileCPUConfig: NSObject
+
+/**
+ @b 默认为 1, 多线程时, 建议设置为 2
+ */
+@property (assign, nonatomic) int threadNum;
+
+/**
+ @b 是否开启运行时 infershape
+ */
+@property  (assign, nonatomic) BOOL loddable;
+
+/**
+ @b 是否开启模型 op 融合优化
+ */
+@property  (assign, nonatomic) BOOL optimize;
+
+@end
+
 @interface PaddleMobileCPU : NSObject
 
-/*
-    创建对象
-*/
-- (instancetype)init;
+/**
+ @b 创建对象
 
-/*
-    load 模型, 开辟内存
-*/
-- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
+ @param config 配置
+ @return paddlemobile CPU 对象
+ */
+- (instancetype)initWithConfig:(PaddleMobileCPUConfig *)config;
 
-/*
-  加载散开形式的模型, 需传入模型的目录
-*/
+/**
+ @b 加载模型
+
+ @param modelPath 模型路径
+ @param weighsPath 权重路径
+ @return 是否加载成功
+ */
+- (BOOL)loadModel:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
+
+/**
+ @b 加载散开形式的模型, 需传入模型的目录
+
+ @param modelAndWeightPath 模型和权重的路径
+ @return 是否加载成功
+ */
 - (BOOL)load:(NSString *)modelAndWeightPath;
 
-/*
- * 从内存中加载模型
- * */
+/**
+ @b 从内存中加载模型
+
+ @param modelLen 模型大小(字节数)
+ @param modelBuf 模型在内存中的位置
+ @param combinedParamsLen 权重大小(字节数)
+ @param combinedParamsBuf 权重在内存中的位置
+ @return 是否加载成功
+ */
 - (BOOL)LoadCombinedMemory:(size_t)modelLen
                andModelBuf:(const uint8_t *)modelBuf
          andModelParamsLen:(size_t)combinedParamsLen
       andCombinedParamsBuf:(const uint8_t *)combinedParamsBuf;
 
 /*
- *  对图像进行预处理, 需要外部开辟 output 内存, 外部释放 output 内存
+ *
  * */
+
+
+/**
+ @b 对图像进行预处理, 需要外部开辟 output 内存, 外部释放 output 内存, 每一个像素经过这样的预处理 (x + means) * scale, 其中 x 为像素值
+
+ @param image 输入的图像
+ @param output 预处理后的输出
+ @param means 预处理中 means
+ @param scale 预处理中的 scale
+ @param dim 预处理后的维度
+ */
 -(void)preprocess:(CGImageRef)image
            output:(float *)output
             means:(NSArray<NSNumber *> *)means
         scale:(float)scale
         dim:(NSArray<NSNumber *> *)dim;
 
-/*
- * 预测预处理后的数据, 返回结果使用结束需要调用其 realseOutput 函数进行释放
- * */
+/**
+ 进行预测
+
+ @param input 输入
+ @param dim 输入维度
+ @return 输出结果
+ */
 - (PaddleMobileCPUResult *)predictInput:(float *)input
                                     dim:(NSArray<NSNumber *> *)dim;
 
-/*
-    进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
-*/
-- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
-
-/*
-    进行预测, 默认 means 为 0, scale 为 1.0
-*/
-- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
-
-/*
-    清理内存
-*/
+/**
+ @b 进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict, 每一个像素经过这样的预处理 (x + means) * scale, 其中 x 为像素值
+
+ @param image 输入图像
+ @param dim 输入维度
+ @param means 预处理中 means
+ @param scale 预处理中 scale
+ @return 预测结果
+ */
+- (PaddleMobileCPUResult *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
+
+/**
+ 进行预测, 预处理 means 值为 0, scale 值为 1
+
+ @param image 输入图像
+ @param dim 输入维度
+ @return 预测结果
+ */
+- (PaddleMobileCPUResult *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
+
+/**
+ @b 清理内存
+ */
 - (void)clear;
 
 @end

src/io/ios_io/PaddleMobileCPU.mm

@@ -45,21 +45,43 @@
 
 @end
 
+@implementation  PaddleMobileCPUConfig
+
+-(instancetype)init {
+  if (self = [super init]) {
+    self.threadNum = 1;
+    self.optimize = YES;
+  }
+  return self;
+}
+
+@end
 
 @interface  PaddleMobileCPU()
 {
   paddle_mobile::PaddleMobile<paddle_mobile::CPU, paddle_mobile::Precision::FP32> *pam_;
   BOOL loaded_;
 }
+
+@property (strong, nonatomic) PaddleMobileCPUConfig *config;
+
 @end
 
 @implementation PaddleMobileCPU
 
 static std::mutex shared_mutex;
 
-- (instancetype)init {
+- (instancetype)initWithConfig:(PaddleMobileCPUConfig *)config {
   if (self = [super init]) {
     pam_ = new paddle_mobile::PaddleMobile<paddle_mobile::CPU, paddle_mobile::Precision::FP32>();
+    _config = config;
+  }
+  return self;
+}
+
+-(instancetype)init {
+  if (self = [super init]) {
+    _config = [[PaddleMobileCPUConfig alloc] init];
   }
   return self;
 }
@@ -79,11 +101,11 @@ static std::mutex shared_mutex;
   return sharedManager;
 }
 
-- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath{
+- (BOOL)loadModel:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath {
   std::string model_path_str = std::string([modelPath UTF8String]);
   std::string weights_path_str = std::string([weighsPath UTF8String]);
-  pam_->SetThreadNum(2);
-  if (loaded_ = pam_->Load(model_path_str, weights_path_str, true)) {
+  pam_->SetThreadNum(self.config.threadNum);
+  if (loaded_ = pam_->Load(model_path_str, weights_path_str, self.config.optimize, false, 1, self.config.loddable)) {
     return YES;
   } else {
     return NO;
@@ -94,14 +116,14 @@ static std::mutex shared_mutex;
                andModelBuf:(const uint8_t *)modelBuf
          andModelParamsLen:(size_t)combinedParamsLen
       andCombinedParamsBuf:(const uint8_t *)combinedParamsBuf {
-  pam_->SetThreadNum(2);
+  pam_->SetThreadNum(self.config.threadNum);
   return loaded_ = pam_->LoadCombinedMemory(modelLen, modelBuf, combinedParamsLen,
-          const_cast<uint8_t*>(combinedParamsBuf));
+          const_cast<uint8_t*>(combinedParamsBuf), self.config.optimize, false, 1, self.config.loddable);
 }
 
 - (BOOL)load:(NSString *)modelAndWeightPath{
   std::string model_path_str = std::string([modelAndWeightPath UTF8String]);
-  if (loaded_ = pam_->Load(model_path_str)) {
+  if (loaded_ = pam_->Load(model_path_str, self.config.optimize, false, 1, self.config.loddable)) {
     return YES;
   } else {
     return NO;
@@ -116,6 +138,10 @@ static std::mutex shared_mutex;
         dim:(NSArray<NSNumber *> *)dim {
   std::lock_guard<std::mutex> lock(shared_mutex);
 
+  if (means == nil) {
+    means = @[@0, @0, @0];
+  }
+
   // dim to c++ vector, get numel
   std::vector<int64_t > dim_vec;
   int numel = 1;
@@ -234,7 +260,7 @@ static std::mutex shared_mutex;
   return cpuResult;
 }
 
-- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale{
+- (PaddleMobileCPUResult *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale{
 //  printf(" predict one ");
   std::lock_guard<std::mutex> lock(shared_mutex);
   if (!loaded_) {
@@ -283,28 +309,22 @@ static std::mutex shared_mutex;
   // predict
   std::vector<float> cpp_result = pam_->Predict(predict_input, dim_vec);
 
-  // result
-  long count = 0;
-  count = cpp_result.size();
-  NSMutableArray *result = [[NSMutableArray alloc] init];
-  for (int i = 0; i < count; i++) {
-    [result addObject:[NSNumber numberWithFloat:cpp_result[i]]];
-  }
-
+  float *output_pointer = new float[cpp_result.size()];
+  memcpy(output_pointer, cpp_result.data(),
+         cpp_result.size() * sizeof(float));
+  PaddleMobileCPUResult *cpuResult = [[PaddleMobileCPUResult alloc] init];
+  [cpuResult toSetOutput: output_pointer];
+  [cpuResult toSetOutputSize: cpp_result.size()];
 
   free(output);
-
-  // 待验证
-  //  if ([UIDevice currentDevice].systemVersion.doubleValue < 11.0) {
   CFRelease(cfData);
   cfData = NULL;
-  //  }
 
-  return result;
+  return cpuResult;
 }
 
-- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim {
-  [self predict:image dim:dim means:nil scale:1];
+- (PaddleMobileCPUResult *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim {
+  return [self predict:image dim:dim means:nil scale:1];
 }
 
 - (void)clear{

src/io/paddle_mobile.cpp

@@ -75,11 +75,10 @@ template <typename Dtype, Precision P>
 bool PaddleMobile<Dtype, P>::LoadCombinedMemory(size_t model_len,
                                                 const uint8_t *model_buf,
                                                 size_t combined_params_len,
-                                                uint8_t *combined_params_buf) {
-  int batch_size = 1;
-  bool optimise = true;
-  bool quantification = false;
-
+                                                uint8_t *combined_params_buf,
+                                                bool optimize,
+                                                bool quantification, int batch_size,
+                                                bool loddable) {
   if (loader_.get() == nullptr) {
     loader_ = std::make_shared<framework::Loader<Dtype, P>>();
   } else {
@@ -89,9 +88,9 @@ bool PaddleMobile<Dtype, P>::LoadCombinedMemory(size_t model_len,
   if (executor_.get() == nullptr) {
     executor_ = std::make_shared<framework::Executor<Dtype, P>>(
         loader_->LoadCombinedMemory(model_len, model_buf, combined_params_len,
-                                    combined_params_buf, optimise,
+                                    combined_params_buf, optimize,
                                     quantification),
-        batch_size, optimise);
+        batch_size, optimize, loddable);
   } else {
     LOG(kLOG_INFO) << "executor inited";
   }

src/io/paddle_mobile.h

@@ -61,7 +61,8 @@ class PaddleMobile {
 
   bool LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
                           size_t combined_params_len,
-                          uint8_t *combined_params_buf);
+                          uint8_t *combined_params_buf, bool optimize = false, bool quantification = false,
+                          int batch_size = 1, bool loddable = false);
 
   void SetThreadNum(int num);
   void Clear();