TNN MatConverter CvtColor NV21TOBGR

短命女 2022-10-31 04:29 118阅读 0赞

OpenCV 中的 [carotene][] 对于 armv7优化较好，而 armv8下则是 NEON 实现。TNN 提供了一套图像预处理接口并且进行了汇编优化。下面以 NV21TOBGR 为例进行介绍。

## [MatUtils][] ##

无成员变量，全部为静态函数。

public:
        //copy cpu <-> device, cpu<->cpu, device<->device, src and dst dims must be equal.
        static Status Copy(Mat& src, Mat& dst, void* command_queue);
    
        //src and dst device type must be same. when param scale_w or scale_h is 0, it is computed as
        // (double)dst.GetWidth() / src.GetWidth() or (double)dst.GetHeight() / src.GetHeight().
        static Status Resize(Mat& src, Mat& dst, ResizeParam param, void* command_queue);
    
        //src and dst device type must be same. when param width or height is 0, it is equal to
        //dst.GetWidth() or dst.GetHeight().
        static Status Crop(Mat& src, Mat& dst, CropParam param, void* command_queue);
    
        //src and dst device type must be same.
        static Status WarpAffine(Mat& src, Mat& dst, WarpAffineParam param, void* command_queue);
    
        //src and dst device type must be same.
        static Status CvtColor(Mat& src, Mat& dst, ColorConversionType type, void* command_queue);
    
        //src and dst device type must be same. param top, bottom, left and right must be non-negative.
        static Status CopyMakeBorder(Mat& src, Mat& dst, CopyMakeBorderParam param, void* command_queue);

## [MatUtils::CvtColor][MatUtils_CvtColor] ##

调用 [CheckSrcAndDstMat][] 输入和输出变量的设备是否相同，输入尺寸是否有效。  
构造一个转换器并调用其缩放函数。

auto ret = CheckSrcAndDstMat(src, dst, true, false, true);
        if (ret != TNN_OK) {
        
            return ret;
        }

如果输出矩阵为空可申请内存，若已有内存则不能比输入小。

if (dst.GetData() == nullptr) {
        
            // set dst size by src size and cvt type
            DimsVector dims = src.GetDims();
            dims[1] = GetCvtColorDstChannel(type);
            dst = Mat(dst.GetDeviceType(), dst.GetMatType(), dims);
        } else {
        
            if (dst.GetWidth() < src.GetWidth() || dst.GetHeight() < src.GetHeight() ||
                dst.GetChannel() < GetCvtColorDstChannel(type)) {
        
                return Status(TNNERR_PARAM_ERR, "cvt color dst size too small");
            }
        }

[MAT\_CONVERTER\_PREPARATION][MAT_CONVERTER_PREPARATION] 可在必要时为输出申请内存，并创建一个 [MatConverterAcc][] 对象。

MAT_CONVERTER_PREPARATION(src.GetDeviceType());
        return converter->CvtColor(src, dst, type, command_queue);

## [CheckSrcAndDstMat][] ##

检查输入输出的设备类型是否一致、Mat 类型是否一致，以及输入尺寸是否正常。

if (check_device_type && (src.GetDeviceType() != dst.GetDeviceType())) {
        
            return Status(TNNERR_PARAM_ERR, "src and dst DeviceType not equal");
        }
    
        if (check_mat_type && (src.GetMatType() != dst.GetMatType())) {
        
            return Status(TNNERR_PARAM_ERR, "src and dst MatType not equal");
        }
    
        if (check_src_size && (src.GetWidth() <= 0 || src.GetHeight() <= 0)) {
        
            return Status(TNNERR_INVALID_INPUT, "src size is zero or negnative");
        }
    
        return TNN_OK;

## [MAT\_CONVERTER\_PREPARATION][MAT_CONVERTER_PREPARATION] ##

[MatConverterManager::Shared][MatConverterManager_Shared] 返回一个 [MatConverterManager][] 单例。  
[MatConverterManager::CreateMatConverterAcc][MatConverterManager_CreateMatConverterAcc] 首先从字典中查找，若没有则创建一个。

#define MAT_CONVERTER_PREPARATION(device_type)                                          \
        if (dst.GetData() == nullptr) {                                                     \
            dst = Mat(dst.GetDeviceType(), dst.GetMatType(), dst.GetDims());                \
        }                                                                                   \
        auto converter = MatConverterManager::Shared()->CreateMatConverterAcc(device_type); \
        if (!converter) {                                                                   \
            return Status(TNNERR_INIT_LAYER, "image converter is nil, check device type");  \
        }

## [MatConverterManager][] ##

拥有一个 [MatConverterAccCreater][] 字典，可以实现反射。

public:
        static std::shared_ptr<MatConverterManager>& Shared();
        MatConverterManager();
        ~MatConverterManager();
        std::shared_ptr<MatConverterAcc> CreateMatConverterAcc(DeviceType device_type);
        int RegisterMatConverterAccCreater(DeviceType type, std::shared_ptr<MatConverterAccCreater> creater);
    
    private:
        std::map<DeviceType, std::shared_ptr<MatConverterAccCreater>> converter_creater_map_;

## [MatConverterManager::Shared][MatConverterManager_Shared] ##

借助 [std::once\_flag][std_once_flag] 和 [std::call\_once][std_call_once] 实现线程安全的单例模式。

static std::once_flag once;
        static std::shared_ptr<MatConverterManager> g_global_blob_converter_manager;
        std::call_once(once, []() {
         g_global_blob_converter_manager = std::make_shared<MatConverterManager>(); });
        return g_global_blob_converter_manager;

## [MatConverterManager::CreateMatConverterAcc][MatConverterManager_CreateMatConverterAcc] ##

在`converter_creater_map_`中查找设备类型，如果有相应的构造者则调用其创建函数。

auto iter = converter_creater_map_.find(device_type);
        if (iter != converter_creater_map_.end()) {
        
            return iter->second->CreateMatConverterAcc();
        }
        return nullptr;

## [MatConverterManager::RegisterMatConverterAccCreater][MatConverterManager_RegisterMatConverterAccCreater] ##

向`converter_creater_map_`字典中添加设备的构造者。

auto iter = converter_creater_map_.find(type);
        if (iter != converter_creater_map_.end()) {
        
            LOGE("Error: device_type(%d) cannot be registered twice\n", type);
            return 1;
        }
        if (!creater) {
        
            LOGE("Error: MatConverterAccCreater is nil device_type(%d)\n", type);
            return 1;
        }
        converter_creater_map_[type] = creater;
        return 0;

## [MatConverterAccCreater][] ##

为工厂方法。

public:
        virtual ~MatConverterAccCreater(){
        };
        virtual std::shared_ptr<MatConverterAcc> CreateMatConverterAcc() = 0;

## [DECLARE\_MAT\_CONVERTER\_CREATER][DECLARE_MAT_CONVERTER_CREATER] ##

定义具体设备的工厂。

#define DECLARE_MAT_CONVERTER_CREATER(device)                                                                          \
        class device##MatConverterAccCreater : public MatConverterAccCreater {                                             \
        public:                                                                                                            \
            virtual ~device##MatConverterAccCreater(){};                                                                   \
            virtual std::shared_ptr<MatConverterAcc> CreateMatConverterAcc() {                                             \
                return std::make_shared<device##MatConverterAcc>();                                                        \
            };                                                                                                             \
        }

## [REGISTER\_MAT\_CONVERTER][REGISTER_MAT_CONVERTER] ##

以定义变量的形式，向全局字典中注册特定设备的矩阵转换器。

#define REGISTER_MAT_CONVERTER(device, device_type)                                                                    \
        MatConverterAccRegister<device##MatConverterAccCreater> g_mat_converter_##device(device_type)

## [MatConverterAcc][] ##

6种操作的接口。

public:
        MatConverterAcc() {
        
            OMP_SET_THREADS_(1);
        };
        virtual ~MatConverterAcc(){
        };
        virtual Status Copy(Mat& src, Mat& dst, void* command_queue = NULL)                                      = 0;
        virtual Status Resize(Mat& src, Mat& dst, ResizeParam param, void* command_queue = NULL)                 = 0;
        virtual Status Crop(Mat& src, Mat& dst, CropParam param, void* command_queue = NULL)                     = 0;
        virtual Status WarpAffine(Mat& src, Mat& dst, WarpAffineParam param, void* command_queue = NULL)         = 0;
        virtual Status CvtColor(Mat& src, Mat& dst, ColorConversionType type, void* command_queue = NULL)        = 0;
        virtual Status CopyMakeBorder(Mat& src, Mat& dst, CopyMakeBorderParam param, void* command_queue = NULL) = 0;

## [ArmMatConverterAcc][] ##

public:
        virtual Status Copy(Mat& src, Mat& dst, void* command_queue = NULL);
        virtual Status Resize(Mat& src, Mat& dst, ResizeParam param, void* command_queue = NULL);
        virtual Status Crop(Mat& src, Mat& dst, CropParam param, void* command_queue = NULL);
        virtual Status WarpAffine(Mat& src, Mat& dst, WarpAffineParam param, void* command_queue = NULL);
        virtual Status CvtColor(Mat& src, Mat& dst, ColorConversionType type, void* command_queue = NULL);
        virtual Status CopyMakeBorder(Mat& src, Mat& dst, CopyMakeBorderParam param, void* command_queue = NULL);

## [ArmMatConverterAcc::CvtColor][ArmMatConverterAcc_CvtColor] ##

到此处时`command_queue`没有用到。  
[CheckMatConverterParams][] 检查输入输出数据是否为空，以及是否在同一设备上。  
[NV21ToBGR][]

Status ret = TNN_OK;
    
        ret = CheckMatConverterParams(src, dst, true);
        if (ret != TNN_OK)
            return ret;
    
        switch (type) {
        
            case COLOR_CONVERT_NV12TOBGR:
                NV12ToBGR((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_NV21TOBGR:
                NV21ToBGR((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_NV12TOBGRA:
                NV12ToBGRA((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_NV21TOBGRA:
                NV21ToBGRA((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_BGRTOGRAY:
                BGRToGray((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_BGRATOGRAY:
                BGRAToGray((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_RGBTOGRAY:
                RGBToGray((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            case COLOR_CONVERT_RGBATOGRAY:
                RGBAToGray((uint8_t*)src.GetData(), (uint8_t*)dst.GetData(), src.GetBatch()*src.GetHeight(), src.GetWidth());
                break;
            default:
                return Status(TNNERR_PARAM_ERR, "ArmMatConverterAcc::CvtColor, color conversion type not support yet");
        }
    
        return ret;

## [NV21ToBGR][] ##

[Float4][]  
[Float4::save][Float4_save] 保存数据到指定地址。  
[Float4::max][Float4_max] 两个对象中对应元素比较，取最大。  
[NV21ToBGR][] 每次处理8个像素。  
[NaiveYUVToBGROrBGRALoop][]

B = 1.164 ( Y − 16 ) + 2.018 ( U − 128 ) G = 1.164 ( Y − 16 ) − 0.813 ( V − 128 ) − 0.391 ( U − 128 ) R = 1.164 ( Y − 16 ) + 1.596 ( V − 128 ) \\begin\{aligned\} B &= 1.164(Y - 16) + 2.018(U - 128)\\\\ G &= 1.164(Y - 16) - 0.813(V - 128) - 0.391(U - 128)\\\\ R &= 1.164(Y - 16) + 1.596(V - 128) \\end\{aligned\} BGR=1.164(Y−16)\+2.018(U−128)=1.164(Y−16)−0.813(V−128)−0.391(U−128)=1.164(Y−16)\+1.596(V−128)

将系数左移8位运算，结果右移。  
`nn`为在每行中一次处理的像素数量。

#ifndef TNN_USE_NEON
        return NaiveYUVToBGROrBGRA(nv21, bgr, 3, h, w, false);
    #else
        const unsigned char* yptr  = nv21;
        const unsigned char* vuptr = nv21 + w * h;
    
        for (int y = 0; y < h; y += 2) {
        
            const unsigned char* yptr0 = yptr;
            const unsigned char* yptr1 = yptr + w;
            unsigned char* rgb0 = bgr;
            unsigned char* rgb1 = bgr + w * 3;
    #if __aarch64__
            int64_t nn = w >> 3;
            int remain = w - (nn << 3);
    
            int16x8_t _q1135 = vdupq_n_s16(1135);
            int8x8_t _v74    = vdup_n_s8(74);
            int8x8_t _v128   = vdup_n_s8(int8_t(128));
            int8x8_t _v102   = vdup_n_s8(102);
            int8x8_t _v52    = vdup_n_s8(52);
            int8x8_t _v25    = vdup_n_s8(25);
            // use 127 instead of 129 to prevent char overflow, add another 2 in asm
            int8x8_t _v127   = vdup_n_s8(127);
            // saturate uv to 240 to avoid b overflow
            uint8x8_t _v240  = vdup_n_u8(240);

`v0.8b-v2.8b`用于加载 YUV。

[PRFM (literal)][PRFM _literal] 为内存预取指令(prefetch memory)。  
[LD1 (vector, multiple structures)][LD1 _vector_ multiple structures] 将多个单元素结构加载到一个，两个，三个或四个寄存器。  
V0-V31 寄存器中的数据是打包的。  
[CMHI (vector, register)][CMHI _vector_ register] 比较无符号的高（向量），将目标 SIMD 和 FP 寄存器中的向量设置为1或0。  
[BSL][] 按位选择。当原始目标位为1时，该指令将目标 SIMD＆FP 寄存器中的每个位设置为来自第一源 SIMD＆FP 寄存器的对应位，否则设置为来自第二源 SIMD＆FP 寄存器的对应位。  
`ld1`读取8个字节，即4对 vu 到 V2 寄存器。  
`cmhi`和`bsl`中两个操作数互换位置，实现`v12.8b`的数值均小于240。  
[SUB (vector)][SUB _vector] 该指令从第一源 SIMD 和 FP 寄存器中的对应向量元素中减去第二源 SIMD 和 FP 寄存器中的每个向量元素，将结果放入向量中，并将该向量写入目标 SIMD 和 FP 寄存器中。  
`v2.8b`中存储  V − 128 V-128 V−128 和  U − 128 U-128 U−128。

if (nn > 0) {
        
                asm volatile(
                    "prfm  pldl1strm, [%[_vu], #128]    \n\t"
                    "ld1   {v2.8b},   [%[_vu]], #8      \n\t"
                    "cmhi  v12.8b, v2.8b, %[_v240].8b   \n\t"
                    "bsl   v12.8b, %[_v240].8b, v2.8b   \n\t"
                    "sub   v2.8b, v12.8b, %[_v128].8b   \n\t"

加载相邻行的  Y Y Y 到`v0.8b`和`v1.8b`寄存器。  
[UMULL, UMULL2 (vector)][UMULL_ UMULL2 _vector] 64位无符号数乘法指令。该指令将两个源 SIMD 和 FP 寄存器的下半部分或上半部分中的相应向量元素相乘，将结果放入向量中，然后将向量写入目标 SIMD 和 FP 寄存器中。  
`v28.8h`中为第1行的  74 Y − 1135 74Y-1135 74Y−1135  
[ORR (vector, register)][ORR _vector_ register] 在两个源 SIMD 和 FP 寄存器之间执行按位或运算，并将结果写入目标 SIMD 和 FP 寄存器。  
复制`v2.8b`中的值到`v3.8b`，得到两份  V − 128 V-128 V−128 和  U − 128 U-128 U−128。  
`v29.8h`中为第2行的  74 Y − 1135 74Y-1135 74Y−1135。  
复制`v28.16b`中的值到`v9.16b`，得到两份第1行的  74 Y − 1135 74Y-1135 74Y−1135。

"0:                                 \n\t"
                    "prfm  pldl1strm, [%[_y0], #128]    \n\t"
                    "ld1   {v0.8b},   [%[_y0]], #8      \n\t"
                    "prfm  pldl1strm, [%[_y1], #128]    \n\t"
                    "ld1   {v1.8b},   [%[_y1]], #8      \n\t"
                    "umull v28.8h, v0.8b,  %[_v74].8b   \n\t"
                    "sub   v28.8h, v28.8h, %[_q1135].8h \n\t"   // v28 -> b0
                    "orr   v3.8b,  v2.8b,  v2.8b        \n\t"
                    "umull v29.8h, v1.8b,  %[_v74].8b   \n\t"
                    "sub   v29.8h, v29.8h, %[_q1135].8h \n\t"   // v29 -> b1
                    "orr   v9.16b, v28.16b, v28.16b     \n\t"   // v9  -> g0

[TRN1 (vector)][TRN1 _vector] 和 [TRN2 (vector)][TRN2 _vector] 为转置向量 Transpose vector(primary)。[TRN1 (vector)][TRN1 _vector] 指令从两个源 SIMD 和 FP 寄存器（从零开始）读取相应的偶数个向量元素，来自第一源寄存器的矢量元素被放置到目标矢量的偶数元素中，而来自第二源寄存器的矢量元素被放置到目标矢量的奇数元素中。  
原本，`v2.8b`和`v3.8b`中的内容是相同的。转置后`v30.8b`存储重复两次的  U − 128 U-128 U−128，而`v31.8b`存储重复的  V − 128 V-128 V−128，一个元素当两个用。  
复制`v29.16b`的值到`v11.16b`，得到两份第2行的  74 Y − 1135 74Y-1135 74Y−1135。  
[SSHLL, SSHLL2 (vector)][SSHLL_ SSHLL2 _vector] 为有符号长左移位（立即），Signed Shift Left Long (immediate)。此指令从源 SIMD 和 FP 寄存器中读取每个向量元素，按指定的移位量左移每个向量元素，将结果放入向量中，然后将向量写入目标 SIMD 和 FP 寄存器。目标向量元素的长度是源向量元素的两倍。SSHLL 指令从源寄存器的下半部分提取向量元素，而 SSHLL2 指令从源寄存器的上半部分提取向量元素。  
`v27.8h`存储  2 ( V − 128 ) 2(V-128) 2(V−128)。  
[SMLSL, SMLSL2 (vector)][SMLSL_ SMLSL2 _vector] 有符号长乘减（向量）。该指令将两个源 SIMD 和 FP 寄存器的向量的下半部分或上半部分的对应有符号整数值相乘，并从目标 SIMD 和 FP 寄存器的向量元素中减去结果。目标向量元素的长度是被乘元素的两倍。SMLSL 指令从每个源寄存器的下半部分提取每个源向量，而 SMLSL2 指令从每个源寄存器的上半部分提取每个源向量。  
`v9.8h`中为第1行的  74 Y − 1135 − 52 ( U − 128 ) 74Y-1135-52(U-128) 74Y−1135−52(U−128)  
复制`v28.16b`的值到`v8.16b`，又得到一份第1行的  74 Y − 1135 74Y-1135 74Y−1135。  
复制`v29.16b`的值到`v10.16b`又得到一份第2行的  74 Y − 1135 74Y-1135 74Y−1135。  
`v11.8h`中为第2行的  74 Y − 1135 − 52 ( U − 128 ) 74Y-1135-52(U-128) 74Y−1135−52(U−128)  
[SMLAL, SMLAL2 (vector)][SMLAL_ SMLAL2 _vector] 带符号的长乘加（向量）。该指令将两个源 SIMD 和 FP 寄存器的向量的下半部分或上半部分的对应有符号整数值相乘，并将结果与目标 SIMD 和 FP 寄存器的向量元素进行累加。目标向量元素的长度是被乘元素的两倍。SMLAL 指令从每个源寄存器的下半部分提取每个源向量，而 SMLAL2 指令从每个源寄存器的上半部分提取每个源向量。  
`v8.8h`为第1行的  74 Y − 1135 + 102 ( U − 128 ) 74Y-1135 + 102(U-128) 74Y−1135\+102(U−128)  
`v28.8h`为第1行的  74 Y − 1135 + 127 ( V − 128 ) 74Y-1135 + 127(V-128) 74Y−1135\+127(V−128)  
`v10.8h`为第2行的  74 Y − 1135 + 102 ( U − 128 ) 74Y-1135 + 102(U-128) 74Y−1135\+102(U−128)

"trn1  v30.8b, v2.8b, v3.8b         \n\t"   // u
                    "trn2  v31.8b, v2.8b, v3.8b         \n\t"   // v
                    "orr   v11.16b, v29.16b, v29.16b    \n\t"   // v11 -> g1
                    "sshll v27.8h, v31.8b, #1           \n\t"
                    "smlsl v9.8h,  v30.8b, %[_v52].8b   \n\t"
                    "orr   v8.16b, v28.16b, v28.16b     \n\t"   // v8  -> r0
                    "smlsl v11.8h, v30.8b, %[_v52].8b   \n\t"
                    "orr   v10.16b, v29.16b, v29.16b    \n\t"   // v10 -> r1
                    "smlal v8.8h,  v30.8b, %[_v102].8b  \n\t"
                    "smlal v28.8h, v31.8b, %[_v127].8b  \n\t"
                    "smlal v10.8h, v30.8b, %[_v102].8b  \n\t"

[ADD (vector)][ADD _vector]  
`v28.8h`为第1行的  74 Y − 1135 + 127 ( V − 128 ) + 2 ( V − 128 ) 74Y-1135 + 127(V-128) + 2(V-128) 74Y−1135\+127(V−128)\+2(V−128)  
`v9.8h`中为第1行的  74 Y − 1135 − 52 ( U − 128 ) − 25 ( V − 128 ) 74Y-1135-52(U-128) - 25(V-128) 74Y−1135−52(U−128)−25(V−128)  
`v29.8h`中为第2行的  74 Y − 1135 + 127 ( V − 128 ) 74Y-1135 + 127(V-128) 74Y−1135\+127(V−128)  
`v11.8h`中为第2行的  74 Y − 1135 − 52 ( U − 128 ) − 25 ( V − 128 ) 74Y-1135-52(U-128)- 25(V-128) 74Y−1135−52(U−128)−25(V−128)  
`v29.8h`中为第2行的  74 Y − 1135 + 127 ( V − 128 ) + 2 ( V − 128 ) 74Y-1135 + 127(V-128) + 2(V-128) 74Y−1135\+127(V−128)\+2(V−128)

"add   v28.8h, v28.8h, v27.8h       \n\t"
                    "smlsl v9.8h,  v31.8b, %[_v25].8b   \n\t"
                    "smlal v29.8h, v31.8b, %[_v127].8b  \n\t"
                    "smlsl v11.8h, v31.8b, %[_v25].8b   \n\t"
                    "add   v29.8h, v29.8h, v27.8h       \n\t"

`v29.8h`是第2行的  74 Y − 1135 + 2 ( V − 128 ) 74Y-1135 + 2(V-128) 74Y−1135\+2(V−128)。  
[SQSHRUN, SQSHRUN2 (vector)][SQSHRUN_ SQSHRUN2 _vector] 有符号饱和右移无符号缩小（立即）。此指令读取源 SIMD 和 FP 寄存器向量中的每个有符号整数值，将每个值右移一个立即数，将结果饱和为原始宽度的一半的无符号整数值，将最终结果放入向量，并将向量写入目标SIMD和FP寄存器。  
`v26.8b`中为第1行的  R = 74 Y − 1135 + 102 ( U − 128 ) 2 6 R = \\frac\{74Y-1135 +102(U-128)\}\{2^6\} R=2674Y−1135\+102(U−128)  
`v24.8b`中为第1行的  B = 74 Y − 1135 + 129 ( V − 128 ) 2 6 B = \\frac\{74Y-1135 +129(V-128)\}\{2^6\} B=2674Y−1135\+129(V−128)  
`v6.8b`中为第2行的  R = 74 Y − 1135 − 102 ( U − 128 ) 2 6 R = \\frac\{74Y-1135 -102(U-128)\}\{2^6\} R=2674Y−1135−102(U−128)  
`v25.8b`中为第1行的  G = 74 Y − 1135 − 52 ( U − 128 ) − 25 ( V − 128 ) 2 6 G = \\frac\{74Y-1135-52(U-128)- 25(V-128)\}\{2^6\} G=2674Y−1135−52(U−128)−25(V−128)  
`v4.8b` 中为第2行的  B = 74 Y − 1135 + 129 ( V − 128 ) 2 6 B = \\frac\{74Y-1135 + 129(V-128)\}\{2^6\} B=2674Y−1135\+129(V−128)  
`v5.8b`中为第2行的  G = 74 Y − 1135 − 52 ( U − 128 ) − 25 ( V − 128 ) 2 6 G = \\frac\{74Y-1135-52(U-128)- 25(V-128)\}\{2^6\} G=2674Y−1135−52(U−128)−25(V−128)

"sqshrun v26.8b, v8.8h,  #6         \n\t"   // v24-v26: b0g0r0
                    "sqshrun v24.8b, v28.8h, #6         \n\t"
                    "sqshrun v6.8b,  v10.8h, #6         \n\t"
                    "sqshrun v25.8b, v9.8h,  #6         \n\t"   // v4-v6: b1g1r1
                    "sqshrun v4.8b,  v29.8h, #6         \n\t"
                    "sqshrun v5.8b,  v11.8h, #6         \n\t"

为何再次读取 UV 值？  
[SUBS (immediate)][SUBS _immediate] 减法（立即数），设置标志，从寄存器值中减去可选移位的立即数，然后将结果写入目标寄存器。 它根据结果更新条件标志。  
[ST3 (vector, multiple structures)][ST3 _vector_ multiple structures] 通过三个寄存器存储多个3元素结构。该指令从三个 SIMD 和 FP 寄存器将多个3元素结构交错存储到内存中。  
`v24.8b-v26.8b`为第一行的像素，`v4.8b-v6.8b`为第二行的像素。  
`beq`检查 Z 标志位，若未置位则跳转到标号`0`处循环处理。  
最后为何`%[_vu]`减8？

"prfm pldl1strm, [%[_vu], #128]     \n\t"
                    "ld1 {v2.8b},    [%[_vu]], #8       \n\t"
                    "subs %[_nn], %[_nn], #1            \n\t"
                    "prfm pstl1strm, [%[_r0]]           \n\t"
                    "st3 {v24.8b-v26.8b}, [%[_r0]], #24 \n\t"
                    "cmhi  v12.8b, v2.8b, %[_v240].8b   \n\t"
                    "bsl   v12.8b, %[_v240].8b, v2.8b   \n\t"
                    "sub   v2.8b, v12.8b, %[_v128].8b   \n\t"
                    "prfm pstl1strm, [%[_r1]]           \n\t"
                    "st3 {v4.8b-v6.8b},   [%[_r1]], #24 \n\t"
                    "bne 0b                             \n\t"
                    "sub %[_vu], %[_vu], #8             \n\t"
    
                    : [_nn]"+r"(nn),
                      [_y0]"+r"(yptr0),
                      [_y1]"+r"(yptr1),
                      [_vu]"+r"(vuptr),
                      [_r0]"+r"(rgb0),
                      [_r1]"+r"(rgb1)
                    : [_v128]"w"(_v128),
                      [_v102]"w"(_v102),
                      [_v52]"w"(_v52),
                      [_v25]"w"(_v25),
                      [_v127]"w"(_v127),
                      [_q1135]"w"(_q1135),
                      [_v74]"w"(_v74),
                      [_v240]"w"(_v240)
                    : "cc", "memory", "x0", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v8",
                      "v9", "v10", "v11", "v12", "v24", "v25", "v26","v27", "v28", "v29", "v30", "v31"
                );
            }

armv7a 的计算。

#else
            int nn         = w >> 3;
            int remain     = w - (nn << 3);
            short _s1135   = 1135;
            int8x8_t _v74  = vdup_n_s8(74);
            int8x8_t _v128 = vdup_n_s8(int8_t(128));
            // to much input w cause compile error, merge to one
            int8x8_t _vuvfilter = {
        102, 52, 25, 127, 0, 0, 0, 0};
            // saturate uv to 240 to avoid b overflow
            uint8x8_t _v240     = vdup_n_u8(240);
    
            if (nn > 0) {
        
                asm volatile(
                    "pld        [%[_vu], #128]      \n"
                    "vld1.u8    {d2}, [%[_vu]]!     \n"
                    "vcgt.u8    d27, d2, %[_v240]   \n"
                    "vbsl.u8    d27,  %[_v240], d2  \n"
                    "vsub.u8    d2, d27, %[_v128]   \n"
                    "vmov.s8    d10, %[_filt]       \n"
                    "vdup.8     d11, d10[1]         \n"   // v52
                    "vdup.8     d12, d10[2]         \n"   // v25
                    "vdup.8     d13, d10[3]         \n"   // v127
                    "vdup.16    q7,  %[_s1135]      \n"   // q1135
                    "vdup.8     d10, d10[0]         \n"   // v102
                    "0:                             \n"
                    "pld        [%[_y0], #128]      \n"
                    "vld1.u8    {d0}, [%[_y0]]!     \n"
                    "pld        [%[_y1], #128]      \n"
                    "vld1.u8    {d1}, [%[_y1]]!     \n"
                    "vmull.u8   q2, d0, %[_v74]     \n"
                    "vorr       d3, d2, d2          \n"
                    "vsub.s16   q2, q2, q7          \n"   // q2  -> b0
                    "vmull.u8   q3, d1, %[_v74]     \n"
                    "vorr       q9, q2, q2          \n"   // q9  -> g0
                    "vsub.s16   q3, q3, q7          \n"   // q3  -> b1
                    "vtrn.s8    d2, d3              \n"   // d2 -> u, d3 -> v
                    "vorr       q11, q3, q3         \n"   // q11 -> g1
                    "vshll.s8   q4, d3, #1          \n"
                    "vmlsl.s8   q9, d2, d11         \n"
                    "vorr       q8, q2, q2          \n"   // q8  -> r0
                    "vmlsl.s8   q11, d2, d11        \n"
                    "vorr       q10, q3, q3         \n"   // q10 -> r1
                    "vmlal.s8   q8, d2, d10         \n"
                    "vmlal.s8   q2, d3, d13         \n"
                    "vmlal.s8   q10, d2, d10        \n"
                    "vadd.s16   q2, q2, q4          \n"
                    "vmlsl.s8   q9, d3, d12         \n"
                    "vmlal.s8   q3, d3, d13         \n"
                    "vmlsl.s8   q11,d3, d12         \n"
                    "vadd.s16   q3, q3, q4          \n"
                    "vqshrun.s16 d26, q8, #6        \n"   // d24-d26: b0g0r0
                    "vqshrun.s16 d24, q2, #6        \n"
                    "vqshrun.s16 d4,  q3, #6        \n"
                    "vqshrun.s16 d25, q9, #6        \n"   // d4-d6: b1g1r1
                    "vqshrun.s16 d6, q10, #6        \n"
                    "vqshrun.s16 d5, q11, #6        \n"
                    "pld        [%[_vu], #128]      \n"
                    "vld1.u8    {d2}, [%[_vu]]!     \n"
                    "subs       %[_nn], #1          \n"
                    "vst3.u8    {d24-d26}, [%[_r0]]!\n"
                    "vcgt.u8    d27, d2, %[_v240]   \n"
                    "vbsl.u8    d27,  %[_v240], d2  \n"
                    "vsub.u8    d2, d27, %[_v128]   \n"
                    "vst3.u8    {d4-d6},   [%[_r1]]!\n"
                    "bne        0b                  \n"
                    "sub        %[_vu], #8          \n"
    
                    : [_nn]"+r"(nn),
                      [_y0]"+r"(yptr0),
                      [_y1]"+r"(yptr1),
                      [_vu]"+r"(vuptr),
                      [_r0]"+r"(rgb0),
                      [_r1]"+r"(rgb1)
                    : [_v128]"w"(_v128),
                      [_filt]"w"(_vuvfilter),
                      [_v74]"w"(_v74),
                      [_s1135]"r"(_s1135),
                      [_v240]"w"(_v240)
                    : "cc", "memory", "q0", "q1", "q2", "q3","q4","q5","q6","q7","q8", "q9", "q10", "q11", "q12", "q13"
                );
            }
    #endif //__aarch64__
            NaiveYUVToBGROrBGRALoop(yptr0, yptr1, vuptr, rgb0, rgb1, remain, false, 3);
            yptr  += 2*w;
            vuptr += remain;
            bgr   += 2*3*w;
        }
    #endif // TNN_USE_NEON

## [NaiveYUVToBGROrBGRA][] ##

[NaiveYUVToBGROrBGRALoop][NaiveYUVToBGROrBGRALoop 1] 每次处理两行。

const unsigned char* yptr  = yuv;
        const unsigned char* vuptr = yuv + w * h;
    
        for (int y = 0; y < h; y += 2) {
        
            const unsigned char* yptr0 = yptr;
            const unsigned char* yptr1 = yptr + w;
            unsigned char* rgb0 = bgr;
            unsigned char* rgb1 = bgr + w * channel;
    
            NaiveYUVToBGROrBGRALoop(yptr0, yptr1, vuptr, rgb0, rgb1, w, is_nv12, channel);
    
            yptr  += 2*w;
            vuptr += w;
            bgr   += 2*channel*w;
        }

## [NaiveYUVToBGROrBGRALoop][NaiveYUVToBGROrBGRALoop 1] ##

循环内每次处理两个像素。  
`u`和`v`像素值若超过240则截断。  
 \[ R G B \] = \[ 1.164 0.000 1.596 1.164 − 0.392 − 0.813 1.164 2.017 0.000 \] \[ ( Y − 16 ) ( C b − 128 ) ( C r − 128 ) \] = 1 2 6 \[ 74 0 102 74 − 25 − 52 74 129 0 \] \[ ( Y − 16 ) ( C b − 128 ) ( C r − 128 ) \] \\begin\{bmatrix\} R \\\\ G \\\\ B \\end\{bmatrix\}= \\begin\{bmatrix\} 1.164 & 0.000 & 1.596 \\\\ 1.164 & -0.392 & -0.813 \\\\ 1.164 & 2.017 & 0.000 \\end\{bmatrix\} \\begin\{bmatrix\} (Y-16) \\\\ (Cb-128)\\\\ (Cr-128) \\end\{bmatrix\}= \\frac\{1\}\{2^6\}\\begin\{bmatrix\} 74 & 0 & 102 \\\\ 74 & -25 & -52 \\\\ 74 & 129 & 0 \\end\{bmatrix\} \\begin\{bmatrix\} (Y-16) \\\\ (Cb-128)\\\\ (Cr-128) \\end\{bmatrix\} ⎣⎡RGB⎦⎤=⎣⎡1.1641.1641.1640.000−0.3922.0171.596−0.8130.000⎦⎤⎣⎡(Y−16)(Cb−128)(Cr−128)⎦⎤=261⎣⎡7474740−25129102−520⎦⎤⎣⎡(Y−16)(Cb−128)(Cr−128)⎦⎤

`SATURATE_CAST_UCHAR`用于溢出处理。

for (; remain > 0; remain -= 2) {
        
            int u, v;
            if (is_nv12) {
        
                u = (vuptr[0] > 240 ? 240 : vuptr[0]) - 128;
                v = (vuptr[1] > 240 ? 240 : vuptr[1]) - 128;
            } else {
        
                v = (vuptr[0] > 240 ? 240 : vuptr[0]) - 128;
                u = (vuptr[1] > 240 ? 240 : vuptr[1]) - 128;
            }
    
            int ruv = 102 * v;
            int guv = -52 * v + -25 * u;
            int buv = 129 * u;
    
    #define SATURATE_CAST_UCHAR(X) (unsigned char)std::min(std::max(X, 0), 255);
    
            int y00 = yptr0[0]* 74 - 1135;
            if (channel == 4)
                rgb0[3] = 255;
            rgb0[0 * channel + 2] = SATURATE_CAST_UCHAR((y00 + ruv) >> 6);
            rgb0[0 * channel + 1] = SATURATE_CAST_UCHAR((y00 + guv) >> 6);
            rgb0[0 * channel + 0] = SATURATE_CAST_UCHAR((y00 + buv) >> 6);
    
            int y01 = yptr0[1]* 74 - 1135;
            if (channel == 4)
                rgb0[7] = 255;
            rgb0[1 * channel + 2] = SATURATE_CAST_UCHAR((y01 + ruv) >> 6);
            rgb0[1 * channel + 1] = SATURATE_CAST_UCHAR((y01 + guv) >> 6);
            rgb0[1 * channel + 0] = SATURATE_CAST_UCHAR((y01 + buv) >> 6);
    
            int y10 = yptr1[0]* 74 - 1135;
            if (channel == 4)
                rgb1[3] = 255;
            rgb1[0 * channel + 2] = SATURATE_CAST_UCHAR((y10 + ruv) >> 6);
            rgb1[0 * channel + 1] = SATURATE_CAST_UCHAR((y10 + guv) >> 6);
            rgb1[0 * channel + 0] = SATURATE_CAST_UCHAR((y10 + buv) >> 6);
    
            int y11 = yptr1[1]* 74 - 1135;
            if (channel == 4)
                rgb1[7] = 255;
            rgb1[1 * channel + 2] = SATURATE_CAST_UCHAR((y11 + ruv) >> 6);
            rgb1[1 * channel + 1] = SATURATE_CAST_UCHAR((y11 + guv) >> 6);
            rgb1[1 * channel + 0] = SATURATE_CAST_UCHAR((y11 + buv) >> 6);
    
    #undef SATURATE_CAST_UCHAR
    
            yptr0 += 2;
            yptr1 += 2;
            vuptr += 2;
            rgb0  += 2*channel;
            rgb1  += 2*channel;
        }

## 参考资料： ##

*  [C++ Tutorial: Auto Registering Factory][C_ Tutorial_ Auto Registering Factory]
 *  [Automatic object factory in C++][Automatic object factory in C]
 *  [Unforgettable Factory Registration][]
 *  [Factory Method design patter with self registering derived classes][]
 *  [Factory Method in C++][Factory Method in C]
 *  [原型模式][Link 1]
 *  [Factory method pattern][]
 *  [详解设计模式 | 抽象工厂][Link 2]
 *  [4. 建造者模式][4.]
 *  [30.1　工厂方法模式VS建造者模式][30.1_VS]
 *  [建造者模式][Link 3]
 *  [建造者模式（Builder Pattern）- 最易懂的设计模式解析][Builder Pattern_-]
 *  [Chapter 4 建造者模式(Builder Pattern)][Chapter 4 _Builder Pattern]
 *  [人人都会设计模式—建造者模式–Builder][Builder]
 *  [C++11于once\_flag，call\_once分析的实现][C_11_once_flag_call_once]
 *  [C++中once\_flag、call\_once使用][C_once_flag_call_once]
 *  [Optimize RGBA->RGB arm64 assembly][Optimize RGBA-_RGB arm64 assembly]
 *  [preload-practice.zh][]
 *  [飞腾CPU体系结构（十一）][CPU]
 *  [aarch64 neon指令集拾遗][aarch64 neon]
 *  [ARM架构64位入门基础：架构分析、寄存器、调用规则、指令集、程序调试以及参考手册][ARM_64]
 *  [ARM NEON编程初探——一个简单的BGR888转YUV444实例详解][ARM NEON_BGR888_YUV444]
 *  [YUV 格式与 RGB 格式的相互转换公式及C++ 代码][YUV _ RGB _C_]
 *  [【arm】arm32位和arm64位架构、寄存器和指令差异分析总结][arm_arm32_arm64]
 *  [Armv8 指令集][Armv8]
 *  [AN12628 Optimizing Memory Copy Routines][]
 *  [ARM Neon 常用指令][ARM Neon]
 *  [ARM NEON SIMD 指令优化][ARM NEON SIMD]
 *  [Dealing with the ARM AArch64 SIMD documentation][]
 *  [Introduction to ARM64 NEON assembly][]
 *  [ARMv8常用指令][ARMv8]
 *  [AI 移动端框架常用指令·汇总（待续）][AI]
 *  [What kind of assembly instruction is this ld1 \{v0.16b\}, %\[in\]?][What kind of assembly instruction is this ld1 _v0.16b_ _in]
 *  [chromium/external/libyuv/master/./source/row\_neon64.cc][chromium_external_libyuv_master_._source_row_neon64.cc]
 *  [Arm NEON programming quick reference][]
 *  [AI移动端常用汇编指令汇总以及底层算子汇编实现（附带一点点干货）][AI 1]
 *  [ARMv8 Neon Programming][]
 *  [ARM\_NEON\_CNN编程][ARM_NEON_CNN]
 *  [6.47.2 Extended Asm - Assembler Instructions with C Expression Operands][]
 *  [How to Use Inline Assembly Language in C Code（C语言内联汇编）–continuing…][How to Use Inline Assembly Language in C Code_C_continuing]
 *  [TRN1][]
 *  [ARM64 汇编指令总结][ARM64]
 *  [浅谈移动工程师跨界机器学习之路][Link 4]
 *  [ARM指令集之乘法指令][ARM]
 *  [What kind of assembly instruction is this ld1 \{v0.16b\}, %\[in\]?][What kind of assembly instruction is this ld1 _v0.16b_ _in]
 *  [ARM Instruction Set][]
 *  [Lecture 8: ARM Arithmetic and BitweiseInstructions][Lecture 8_ ARM Arithmetic and BitweiseInstructions]
 *  [Arm A64 Instruction Set Architecture][]
 *  [arm CPU 2D卷积计算方法一览][arm CPU 2D]
 *  [ARM Assembly Programming][]
 *  [关于ARM中的tst、cmp、bne、beq指令][ARM_tst_cmp_bne_beq]
 *  [Introducing ARM assembly language][]
 *  [ARM Data Types and Registers (Part 2) | Azeria Labs][ARM Data Types and Registers _Part 2_ _ Azeria Labs]
 *  [First look at the arm64 architecture and assembly language][]
 *  [gemmlowp/internal/kernel\_neon.h][gemmlowp_internal_kernel_neon.h]
 *  [ARMv8-A A64 ISA Overview][]
 *  [What is the fastest way to index into ARMv8 registers][]
 *  [移动端arm cpu优化学习笔记第4弹–内联汇编入门][arm cpu_4]
 *  [【Arm端算法优化笔记】一，一步步优化盒子滤波算法][Arm]
 *  [ARMv8 中的 SIMD 运算][ARMv8 _ SIMD]
 *  [用NEON intrinsic实现RGB转YUV420SP(NV12)][NEON intrinsic_RGB_YUV420SP_NV12]
 *  [YUV转RGB（NV21-ARGB）的Neon优化代码][YUV_RGB_NV21-ARGB_Neon]
 *  [运用NEON指令集加速RGB与YUV相互转换][NEON_RGB_YUV]
 *  [What is arrangement specifier(.16b,.8b) in ARM assembly language instructions?][What is arrangement specifier_.16b_.8b_ in ARM assembly language instructions]
 *  [TNN新版本上线！全新特性，更加好用！][TNN]

[carotene]: https://github.com/opencv/opencv/tree/master/3rdparty/carotene
[MatUtils]: https://github.com/Tencent/TNN/blob/master/include/tnn/utils/mat_utils.h#L74
[MatUtils_CvtColor]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_utils.cc#L171
[CheckSrcAndDstMat]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_utils.cc#L38
[MAT_CONVERTER_PREPARATION]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_utils.cc#L22
[MatConverterAcc]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.h#L29
[MatConverterManager_Shared]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.cc#L23
[MatConverterManager]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.h#L49
[MatConverterManager_CreateMatConverterAcc]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.cc#L30
[MatConverterAccCreater]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.h#L43
[std_once_flag]: https://en.cppreference.com/w/cpp/thread/once_flag
[std_call_once]: https://en.cppreference.com/w/cpp/thread/call_once
[MatConverterManager_RegisterMatConverterAccCreater]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.cc#L38
[DECLARE_MAT_CONVERTER_CREATER]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.h#L73
[REGISTER_MAT_CONVERTER]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_acc.h#L82
[ArmMatConverterAcc]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/arm_mat_converter.h#L24
[ArmMatConverterAcc_CvtColor]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/arm_mat_converter.cc#L205
[CheckMatConverterParams]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/mat_converter_utils.cc#L24
[NV21ToBGR]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/arm_util.cc#L971
[Float4]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/acc/Float4.h#L33
[Float4_save]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/acc/Float4.h#L65
[Float4_max]: https://github.com/Tencent/TNN/blob/master/source/tnn/device/arm/acc/Float4.h#L212
[NaiveYUVToBGROrBGRALoop]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/naive_compute.cc#L740
[PRFM _literal]: https://developer.arm.com/documentation/dui0801/k/A64-Data-Transfer-Instructions/PRFM--literal-?lang=en
[LD1 _vector_ multiple structures]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/LD1--vector--multiple-structures-?lang=en
[CMHI _vector_ register]: https://developer.arm.com/documentation/100069/0610/A64-SIMD-Vector-Instructions/CMHI--vector--register-?lang=en
[BSL]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/BSL--vector-?lang=en
[SUB _vector]: https://developer.arm.com/documentation/100069/0610/A64-SIMD-Vector-Instructions/SUB--vector-?lang=en
[UMULL_ UMULL2 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/UMULL--UMULL2--vector-?lang=en
[ORR _vector_ register]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/ORR--vector--register-?lang=en
[TRN1 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/TRN1--vector-?lang=en
[TRN2 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/TRN2--vector-?lang=en
[SSHLL_ SSHLL2 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/SSHLL--SSHLL2--vector-?lang=en
[SMLSL_ SMLSL2 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/SMLSL--SMLSL2--vector-?lang=en
[SMLAL_ SMLAL2 _vector]: https://developer.arm.com/documentation/100069/0610/A64-SIMD-Vector-Instructions/SMLAL--SMLAL2--vector-?lang=en
[ADD _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/ADD--vector-?lang=en
[SQSHRUN_ SQSHRUN2 _vector]: https://developer.arm.com/documentation/dui0801/k/A64-SIMD-Vector-Instructions/SQSHRUN--SQSHRUN2--vector-?lang=en
[SUBS _immediate]: https://developer.arm.com/documentation/100069/0610/A64-General-Instructions/SUBS--immediate-?lang=en
[ST3 _vector_ multiple structures]: https://developer.arm.com/documentation/100069/0610/A64-SIMD-Vector-Instructions/ST3--vector--multiple-structures-?lang=en
[NaiveYUVToBGROrBGRA]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/naive_compute.cc#L725
[NaiveYUVToBGROrBGRALoop 1]: https://github.com/Tencent/TNN/blob/master/source/tnn/utils/naive_compute.cc#L669
[C_ Tutorial_ Auto Registering Factory]: http://derydoca.com/2019/03/c-tutorial-auto-registering-factory/
[Automatic object factory in C]: https://blog.noctua-software.com/object-factory-c++.html
[Unforgettable Factory Registration]: http://www.nirfriedman.com/2018/04/29/unforgettable-factory/
[Factory Method design patter with self registering derived classes]: https://rextester.com/ILKI85678
[Factory Method in C]: https://refactoring.guru/design-patterns/factory-method/cpp/example
[Link 1]: https://refactoringguru.cn/design-patterns/prototype
[Factory method pattern]: https://en.wikipedia.org/wiki/Factory_method_pattern
[Link 2]: https://juejin.im/post/6844903795235749901
[4.]: https://design-patterns.readthedocs.io/zh_CN/latest/creational_patterns/builder.html#id16
[30.1_VS]: https://www.kancloud.cn/sstd521/design/193624
[Link 3]: https://wiki.jikexueyuan.com/project/java-design-pattern/builder-pattern.html
[Builder Pattern_-]: https://blog.csdn.net/carson_ho/article/details/54910597
[Chapter 4 _Builder Pattern]: https://rongli.gitbooks.io/design-pattern/content/chapter4.html
[Builder]: https://juejin.im/post/6844903518449434638
[C_11_once_flag_call_once]: https://www.cnblogs.com/blfshiye/p/4717646.html
[C_once_flag_call_once]: https://zhuanlan.zhihu.com/p/71900518
[Optimize RGBA-_RGB arm64 assembly]: https://stackoverflow.com/questions/20662929/optimize-rgba-rgb-arm64-assembly/20663444
[preload-practice.zh]: https://github.com/Tencent/ncnn/blob/master/docs/developer-guide/preload-practice.zh.md
[CPU]: https://blog.csdn.net/lsshao/article/details/109345139
[aarch64 neon]: https://blog.csdn.net/gaussrieman123/article/details/96150214
[ARM_64]: https://soaringleefighting.github.io/2018/08/18/%E3%80%90arm%E3%80%91arm%E6%9E%B6%E6%9E%8464%E4%BD%8D%E5%85%A5%E9%97%A8%E5%9F%BA%E7%A1%80%EF%BC%9A%E6%9E%B6%E6%9E%84%E5%88%86%E6%9E%90%E3%80%81%E5%AF%84%E5%AD%98%E5%99%A8%E3%80%81%E8%B0%83%E7%94%A8%E8%A7%84%E5%88%99%E3%80%81%E6%8C%87%E4%BB%A4%E9%9B%86%E3%80%81%E7%A8%8B%E5%BA%8F%E8%B0%83%E8%AF%95%E4%BB%A5%E5%8F%8A%E5%8F%82%E8%80%83%E6%89%8B%E5%86%8C/
[ARM NEON_BGR888_YUV444]: https://segmentfault.com/a/1190000010127521
[YUV _ RGB _C_]: https://blog.csdn.net/liyuanbhu/article/details/68951683
[arm_arm32_arm64]: https://www.jianshu.com/p/a51076c33108
[Armv8]: https://blog.csdn.net/QiangLi_strong/article/details/81475227?utm_source=blogxgwz3
[AN12628 Optimizing Memory Copy Routines]: https://www.nxp.com.cn/docs/en/application-note/AN12628.pdf
[ARM Neon]: https://jzwdsb.github.io/2018/07/neon_frequently_use/
[ARM NEON SIMD]: http://images.china-pub.com/ebook4705001-4710000/4707250/ch02.pdf
[Dealing with the ARM AArch64 SIMD documentation]: http://blog.pkh.me/p/23-dealing-with-the-arm-aarch64-simd-documentation.html
[Introduction to ARM64 NEON assembly]: https://medium.com/mathieugarcia/introduction-to-arm64-neon-assembly-930c4a48bb2a
[ARMv8]: https://freshmou.github.io/2018/11/ARMv8/
[AI]: https://www.jianshu.com/p/5f75fa02c5d0
[What kind of assembly instruction is this ld1 _v0.16b_ _in]: https://stackoverflow.com/questions/32369478/what-kind-of-assembly-instruction-is-this-ld1-v0-16b-in
[chromium_external_libyuv_master_._source_row_neon64.cc]: https://chromium.googlesource.com/external/libyuv/+/master/source/row_neon64.cc
[Arm NEON programming quick reference]: https://community.arm.com/developer/tools-software/oss-platforms/b/android-blog/posts/arm-neon-programming-quick-reference
[AI 1]: https://zhuanlan.zhihu.com/p/64025085
[ARMv8 Neon Programming]: https://www.uio.no/studier/emner/matnat/ifi/INF5063/h16/pensumliste/armv8-neon-programming.pdf
[ARM_NEON_CNN]: https://github.com/Ewenwan/MVision/blob/master/CNN/HighPerformanceComputing/ARM_NEON_CNN%E7%BC%96%E7%A8%8B.md
[6.47.2 Extended Asm - Assembler Instructions with C Expression Operands]: http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
[How to Use Inline Assembly Language in C Code_C_continuing]: https://blog.csdn.net/xiaosaerjt/article/details/103893864
[TRN1]: https://developer.arm.com/documentation/ddi0602/2020-12/SIMD-FP-Instructions/TRN1--Transpose-vectors--primary--
[ARM64]: http://lazybing.github.io/blog/2019/01/24/neon-of-coding/
[Link 4]: https://zhuanlan.zhihu.com/p/89214197
[ARM]: https://blog.csdn.net/xiaoyu_750516366/article/details/110152921
[ARM Instruction Set]: https://www.csie.ntu.edu.tw/~cyy/courses/assembly/12fall/lectures/handouts/lec09_ARMisa.pdf
[Lecture 8_ ARM Arithmetic and BitweiseInstructions]: https://cseweb.ucsd.edu/classes/wi14/cse30-c/lectures/PI_WI_14_CSE30_lecture_8_post.pdf
[Arm A64 Instruction Set Architecture]: https://developer.arm.com/documentation/ddi0596/2020-12/
[arm CPU 2D]: https://zhuanlan.zhihu.com/p/81201840
[ARM Assembly Programming]: https://www.csie.ntu.edu.tw/~cyy/courses/assembly/10fall/lectures/handouts/lec10_ARMasm.pdf
[ARM_tst_cmp_bne_beq]: https://www.cnblogs.com/hangzhe/p/6916502.html
[Introducing ARM assembly language]: http://www.toves.org/books/arm/
[ARM Data Types and Registers _Part 2_ _ Azeria Labs]: https://azeria-labs.com/arm-data-types-and-registers-part-2/
[First look at the arm64 architecture and assembly language]: https://montcs.bloomu.edu/Presentations/240/200.ARMv8.pdf
[gemmlowp_internal_kernel_neon.h]: https://github.com/google/gemmlowp/blob/master/internal/kernel_neon.h
[ARMv8-A A64 ISA Overview]: https://armkeil.blob.core.windows.net/developer/Files/pdf/graphics-and-multimedia/ARMv8_InstructionSetOverview.pdf
[What is the fastest way to index into ARMv8 registers]: https://stackoverflow.com/questions/62866996/what-is-the-fastest-way-to-index-into-armv8-registers
[arm cpu_4]: https://zhuanlan.zhihu.com/p/143328317
[Arm]: https://blog.csdn.net/just_sort/article/details/107734175
[ARMv8 _ SIMD]: https://winddoing.github.io/post/13631.html
[NEON intrinsic_RGB_YUV420SP_NV12]: https://my.oschina.net/u/4383327/blog/4501638
[YUV_RGB_NV21-ARGB_Neon]: https://blog.csdn.net/jxt1234and2010/article/details/46620001
[NEON_RGB_YUV]: https://www.jianshu.com/p/e498326a55b1
[What is arrangement specifier_.16b_.8b_ in ARM assembly language instructions]: https://stackoverflow.com/questions/57294672/what-is-arrangement-specifier-16b-8b-in-arm-assembly-language-instructions
[TNN]: https://mp.weixin.qq.com/s?__biz=MzU0NjU0ODk2Mg==&mid=2247485938&idx=1&sn=b7318f9c1571b9f41c291109864eb1e9&chksm=fb5ab47fcc2d3d6968aa349105a8006a2bb310ef42ec4621db44cf27fd54fd3f34127814395c&mpshare=1&scene=1&srcid=0221RsBusKzMipOqJ25hyi9T&sharer_sharetime=1613914027319&sharer_shareid=e08f1346e20cd3a5de6b23be6624f251&exportkey=AQ7S2ta2X9STFvdCHf5Btko%3D&pass_ticket=6zrmlJYyrovpgltdH3574gVTNSuDbr4UA5meNovTYZI6ZrfQ9BBpoTRl%2FxRd7plk&wx_header=0#rd