Media data processing V2 Version (VPC) image zoom content analysis

Ascend710, Introduction to exclusive media data processing interface

1. What is? Ascend710？

Ascend710 It is a product developed by Huawei for AI Computing chip , That's what people say NPU, Through dedicated AI The operation acceleration unit can realize low-power and high-efficiency processing of a large number of AI Calculation task .

---

2. Why use a dedicated media data interface ,opencv Isn't it more general ?

opencv Very common indeed , stay CPU,GPU You can call the corresponding API Interface for some image data processing , however Ascend710 A special media data processing structure is integrated on the , It can complete the task more energy-saving and efficient than traditional media data processing , This is of great significance in some embedded scenarios , and opencv Although it can be , But considering that the traditional software decoding is mostly arranged in CPU Upper operation , It not only brings a very high amount of computation to occupy other application computing resources, but also is not efficient , and opencv You cannot call the corresponding Asend Chip capabilities . So we should have a set of special interfaces .

---

First, post links to official documents , Please check the official documents if you don't understand .
Official document link

Although media data processing v2 The version of the interface is currently in the process of experiencing , No large-scale publicity , But from the pictures on the official website, it has been well supported

flow chart

From the figure, we can see that although the interface has become the second version , But in fact, the overall thinking is still unchanged ：

System initialization ==》 Create the corresponding media processing function channel ==》 Application memory ==》 Call the processing interface ==》 Call the get result interface ==》 Memory free ==》 System logoff

Focus on several important API Interface

1. establish vpc Image processing channel interface

//  Prototype hi_s32 hi_mpi_vpc_create_chn(hi_vpc_chn chn, const hi_vpc_chn_attr *attr)//  Parameters hi_vpc_chn chn： This parameter indicates that the value range of image processing channel number is [0,256)hi_vpc_chn_attr *attrtypedef struct {    hi_s32 attr; //  Task queue depth , Value range ：[0,350], Indicates the number of tasks distributed by a channel . Memory consumption varies linearly with capacity .    hi_u32 pic_width; //  The maximum image width supported by the channel , It has no practical effect at present , The official recommendation is set to zero , Prevent compatibility problems with later versions .    hi_u32 pic_height; //  The maximum image height supported by the channel , It has no practical effect at present , The official recommendation is set to zero , Prevent compatibility problems with later versions .} hi_vpc_chn_attr;

2. Scaling interface

//  Prototype hi_s32 hi_mpi_vpc_resize (hi_vpc_chn chn, const hi_vpc_pic_info *source_pic, hi_vpc_pic_info *dest_pic, hi_double fx, hi_double fy, hi_u32 interpolation, hi_u32 *task_id, hi_s32 milli_sec)//  Parameters hi_vpc_chn chn： Image processing channel number , Value range ：[0, 256)hi_vpc_pic_info *source_pic  Input structure pointer of image information  hi_vpc_pic_info *dest_pic  Structure pointer for outputting image information typedef struct {    hi_void* picture_address; //  Storing picture data Device Address .    hi_u32 picture_buffer_size; //  Buffer size for storing picture data .( Note that this size is the width height to the following size , Not necessarily the original size of the picture )    hi_u32 picture_width; //  The picture is really wide .    hi_u32 picture_height; //  The picture is really high     hi_u32 picture_width_stride; //  Picture width stride    hi_u32 picture_height_stride; //  The picture is high stride    hi_pixel_format picture_format; //  Format of target picture } hi_vpc_pic_info;hi_double fx：  Wide scaling hi_double fy：  High zoom hi_u32 interpolation： Image zoom 0： Industry wide Bilinear Algorithm 1： Industry wide Nearest neighbor  Algorithm  hi_s32 milli_sec： How to time out  -1： Blocking mode  0： Non blocking mode  >0： Timeout mode , Configure the specific timeout . The general deviation is within a time slice of the operating system .hi_u32 *task_id： Mission ID To distinguish between different tasks 

3. Receive zoom ( Cutout ) Interface for the results of

//  Prototype hi_s32 hi_mpi_vpc_get_process_result(hi_vpc_chn chn, hi_u32 task_id, hi_s32 milli_sec) This interface can be in the same thread as the task sending interface , You can also call speed up in another thread .//  Parameters hi_vpc_chn chn：  Image processing channel number , Value range ：[0, 256)hi_u32 task_id：  Mission ID To distinguish between different tasks hi_s32 milli_sec：  How to time out  -1： Blocking mode  0： Non blocking mode  >0： Timeout mode , Configure the specific timeout . The general deviation is within a time slice of the operating system 

4. Matting task sending interface

//  Prototype hi_s32 hi_mpi_vpc_crop(hi_vpc_chn chn, const hi_vpc_pic_info *source_pic, hi_vpc_crop_region_info crop_info[], hi_u32 count, hi_u32 *task_id, hi_s32 milli_sec)//  Parameters hi_vpc_chn chn： Image processing channel number , Value range ：[0, 256)hi_vpc_pic_info *source_pic  The explanation is the same as above hi_u32 count： Number of matting areas , Value range [1,256]hi_vpc_crop_region_info crop_info[]： Matting picture information array , The array length is the same as count The parameter values are consistent typedef struct {    hi_vpc_pic_info dest_pic_info;    hi_vpc_crop_region crop_region;} hi_vpc_crop_region_info;typedef struct {    hi_u32 top_offset; //  Top left corner of matting （ Abscissa ）    hi_u32 left_offset; //  Top left corner of matting （ Ordinate ）    hi_u32 crop_width;  //  The matting area is wide     hi_u32 crop_height; //  The matting area is high } hi_vpc_crop_region;

---

Open source learning

The zoom