A low-power rasterization design method for embedded GPU

By designing rasterization acceleration units and parallel scanning algorithms in embedded GPUs, the problems of complex rasterization operations and high energy consumption are solved, efficient element rendering and performance improvements are achieved, and efficient scanning of embedded graphics processors is met. and low power consumption requirements.

CN118505887BActive Publication Date: 2024-11-08WUHAN LINGJIU MICROELECTRONICS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410959442.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-08
Estimated Expiration
2044-07-17

AI Technical Summary

Technical Problem

In the graphics rendering pipeline of embedded GPUs, rasterization operations are complex and energy-consuming. Existing delay rendering technologies are difficult to effectively reduce power consumption, and there are challenges in how multi-core computing arrays can quickly scan and unify shading clips.

Method used

Design a low-power rasterization design method for embedded GPUs, including vertex shaders and fragment shaders, primitive initialization and parallel interpolation scanning are performed through rasterization acceleration units, and data is performed using primitive initialization units and block scanning interpolation units. Storage and calculation realize efficient initialization and parallel scanning of primitives, and improve rendering efficiency.

Benefits of technology

Through parallel processing and pipeline structure, the efficiency of element rendering and the performance of the graphics processor are improved, power consumption is reduced, pixel fill rate and texture fill rate are improved, and high-performance graphics rendering needs are met.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118505887B_ABST
    Figure CN118505887B_ABST
Patent Text Reader

Abstract

The present invention provides a low-power rasterization design method for an embedded GPU, including: rasterizing the primitives after the window transformation to obtain discrete fragments, and coloring the fragments based on a fragment shader; rasterizing the primitives after the window transformation, including: receiving data sent by the window transformation, parsing the data, and storing the parsed primitive data in a general data storage method; initializing the primitives, calculating the initialization parameters of the primitives, and interpolating and scanning the primitives in a parallel scanning method based on the initialization parameters of the primitives. The present invention stores the primitive data so that the next primitive can be initialized while the previous primitive is interpolated and scanned, thereby improving the efficiency of primitive rendering, and interpolating and scanning the primitives in a parallel scanning method can accelerate the rendering of the primitives.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of graphics rendering, and more specifically, to a low-power consumption rasterization design method for an embedded GPU. Background Art

[0002] In the graphics rendering pipeline of embedded GPU (Graphics Processing Unit), rasterization is the most complex operation, and its scanning efficiency will directly affect the key performance of GPU graphics rendering. Rasterization is the process of discretizing the primitives transformed into screen space into fragments, that is, the process of converting basic primitives (points, lines, triangles) into two-dimensional images. In order to save computing power and reduce power consumption, embedded graphics processors generally use deferred rendering technology, first perform fragment tests to eliminate covered fragments, and only render visible pixels, thereby greatly reducing the amount of fragment shading tasks. In order to break through the key technology of deferred rendering, it is necessary to start from the rasterization algorithm and complete the deferred rendering design.

[0003] Rasterization operations involve a large amount of data calculations, such as initialization parameter establishment, triangle edge determination, perspective correction, and attribute interpolation. Therefore, it is particularly important to study rasterization algorithms, pipeline structures, parallel scanning methods, and low-power processing. Summary of the invention

[0004] The present invention aims at solving the technical problems existing in the prior art and provides a low-power consumption rasterization design method for an embedded GPU, comprising:

[0005] Color the vertices based on the vertex shader, assemble the colored vertices into primitives, and perform clipping, backface culling, and viewport transformation on the primitives;

[0006] Rasterize the primitives transformed by the viewport based on the rasterization acceleration unit to obtain discrete fragments, and color the fragments based on the fragment shader;

[0007] Wherein, the rasterization acceleration unit includes a primitive initialization unit and a plurality of block scanning interpolation units, and the primitive initialization unit includes a storage unit and an initialization calculation unit;

[0008] The storage unit is used to parse the data sent by the received window transformation and store the parsed image metadata in a general data storage method;

[0009] The initialization calculation unit is used to initialize the primitives and calculate the initialization parameters of the primitives;

[0010] The block scanning interpolation unit is used to perform interpolation scanning on the primitives in a parallel scanning manner based on the initialization parameters of the primitives;

[0011] The storage unit stores the image element data so that the next image element can be initialized while the previous image element is interpolated and scanned.

[0012] Based on the above technical solution, the present invention can also make the following improvements.

[0013] Optionally, the receiving the data sent by the window transformation, parsing the data, and storing the parsed metadata in a general data storage manner includes:

[0014] Receive the data sent by the window transformation, parse the data, parse out the corresponding command information and graphic element data, if the parsed data is command information, directly set the corresponding register, if the parsed data is graphic element data, store the graphic element data in the storage space of the circuit structure using a general data storage method.

[0015] Optionally, storing the image metadata in a storage space of the circuit structure in a general data storage manner includes:

[0016] Read the data sent by the window transformation, and determine whether the data is a point, line, or triangle primitive. If not, end the process;

[0017] If the data is a point primitive, read a vertex data;

[0018] If the data is a line primitive, two vertex data are read;

[0019] If the data is a triangle primitive, three vertex data are read;

[0020] The read vertex data is stored in the storage space of the circuit structure.

[0021] Optionally, the initialization parameters of the primitive include an initial value and an offset of a boundary function, an initial value and an offset of a perspective correction of an attribute, and a calculation of a triangle area and an inverse area;

[0022] The calculation process of the initialization parameters of the primitives includes addition and subtraction operations, multiplication operations and division operations, and the divider is designed by a reciprocal transcendental function, and the transcendental function is implemented by a pipeline method of quadratic polynomial approximation;

[0023] And in the calculation process of the initialization parameters of the primitives, the addition and subtraction operations, multiplication operations and division operations are sorted, and the corresponding operations are performed in sequence to complete the initialization calculation process of the primitives.

[0024] Optionally, the interpolation scanning of the primitives in a parallel scanning manner includes:

[0025] According to the different types of primitives, the corresponding interpolation scanning algorithm is used for scanning;

[0026] Wherein, when it is a point primitive, it is scanned according to the point-by-point scanning algorithm;

[0027] When it is a line element, it is scanned according to the improved Bresenham algorithm;

[0028] When it is a triangular primitive, it is scanned according to the multi-scanning block parallel scanning algorithm.

[0029] Optionally, when the image is a point primitive, scanning is performed according to a point-by-point scanning algorithm, including:

[0030] When the point size Point Size is set to 1 in the command information, the point primitive is directly output;

[0031] When the point size Point Size is set to be greater than 1 in the command information, the point primitives are output in blocks of the specified size.

[0032] Optionally, when it is a line graphic element, scanning is performed according to an improved Bresenham algorithm, including:

[0033] When the scan line width Line Width is set to 1 in the command information, it is a single line width scan;

[0034] When the scan line width Line Width set in the command information is greater than 1, it is multi-line width scanning, wherein the multi-line width scanning includes:

[0035] The current line width segment is scanned up and down with the discrete segment scanned when the scanning line width Line Width is 1 as the center, and then the next segment is calculated by Bresenham algorithm, and so on, to complete the multi-line width scanning.

[0036] Optionally, when the primitive is a triangle, scanning is performed according to a multi-scanning block parallel scanning algorithm, including:

[0037] Cutting a triangle primitive into a plurality of triangle blocks, wherein the plurality of triangle blocks are covered in different scanning blocks, and calculating an initial attribute of each triangle block;

[0038] Performing parallel line scanning on the cut multiple triangle blocks based on multiple scanning blocks;

[0039] The number of scanning blocks is determined according to the boundary function calculated by initializing the current triangle primitive, and the size of each scanning block can be adjusted dynamically.

[0040] Optionally, performing rasterization processing on the primitives after the window transformation to obtain discrete fragments and coloring the fragments includes:

[0041] The interpolation scanned fragments are output in a multi-channel parallel manner;

[0042] The output fragment is dispatched to the fragment shader through the fragment scheduler FDPR to shade the scanned fragment.

[0043] Optionally, the vertex shader and the fragment shader shade vertices and fragments respectively based on a streaming multiprocessor SMPU.

[0044] The present invention provides a low-power rasterization design method for an embedded GPU, which stores primitive data so that the next primitive can be initialized while the previous primitive is interpolated and scanned, thereby improving the efficiency of primitive rendering. The primitives are interpolated and scanned in parallel, which can accelerate the rendering of the primitives. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of a low-power rasterization design method for an embedded GPU provided by the present invention;

[0046] Figure 2 This is a schematic diagram of the primitive rendering process;

[0047] Figure 3 This is a schematic diagram of the image element scanning process;

[0048] Figure 4 A schematic diagram of storing image metadata using a general data storage method;

[0049] Figure 5 It is a schematic diagram of multi-line width scanning;

[0050] Figure 6 Scanning diagram for triangle primitives Figure 1 ;

[0051] Figure 7 Scanning diagram for triangle primitives Figure 2 . DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. This combination is not subject to the constraints of the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0053] In the face of people's growing demand for graphics processors, such as pixel fill rate, texture fill rate, computing power, power consumption, etc., embedded GPUs are also continuously improving system architecture and design methods. Since graphics processing requires a large amount of data processing such as vertex shading, fragment shading, even geometric shading and surface subdivision, it is generally completed using a multi-core computing array. The purpose of the multi-core computing array is to make data or tasks as parallel as possible, thereby accelerating processing capabilities and improving graphics rendering performance. However, how can a large number of primitives processed in parallel by the multi-core computing array be quickly scanned into fragments, and can the scanned fragments meet the requirements of unified shading for multi-core computing array utilization and chip power consumption? In addition to considering the entire hardware and software architecture, this also requires in-depth exploration of each functional module. Rasterization, as the most complex stage in the graphics rendering pipeline, can easily become a rendering bottleneck, so a high-performance rasterization acceleration unit needs to be designed.

[0054] Figure 1 A flow chart of a low-power rasterization design method for an embedded GPU provided by the present invention, such as Figure 1 As shown, the method includes:

[0055] Step 1: color the vertices based on the vertex shader, assemble the colored vertices into primitives, and perform clipping, backface culling and viewport transformation on the primitives.

[0056] It is understandable that the present invention designs multiple independent stream multiple processors SMPU (Stream Multiple Processing Unit) to complete the unified shading processing in parallel, and each SMPU is a multi-core processor array based on SIMT (Single Instruction Multiple Threads) structure.

[0057] See also Figure 2 , is the flowchart of the graphics rendering engine. First, the vertices are colored based on the vertex shader. The colored vertices are sent to the primitive assembly in the order of first-in-first-out to be assembled into primitives. Then, the primitives are clipped into a given area through three-dimensional clipping and the clipped primitives are reassembled. The invisible primitives are culled through backface culling after clipping, thereby reducing the amount of rasterization calculations. Then, the objects projected in the view volume are converted to the two-dimensional viewport plane through the window transformation. The primitives after the window transformation are rasterized, and the rasterization converts the primitives into discrete fragments. Finally, the fragment information is dispatched to the fragment shader for fragment shading through FDPR (Fragment Dispatcher). Among them, the vertex shader and fragment shader color the vertices and fragments based on SMPU (Stream Multiple Processing Unit) respectively.

[0058] Step 2: rasterize the primitives after the viewport transformation based on the rasterization acceleration unit to obtain discrete fragments, and shade the fragments based on the fragment shader.

[0059] It is understandable that the graphics primitives after the window transformation are rasterized, and in order to match the computing power of the SMPU, a rasterization acceleration unit is used to rasterize the graphics primitives.

[0060] The rasterization acceleration unit is composed of a pipelined primitive initialization unit and multiple block scanning interpolation units (parameters are adjustable). It should be noted that multiple blocks are also scanned in parallel (parameters are adjustable), but ultimately they are output to the FDPR in a 4-way parallel manner, and the FDPR sends the fragments to the SMPU for fragment shading.

[0061] Wherein, the rasterization acceleration unit includes a primitive initialization unit and a plurality of block scanning interpolation units, and the primitive initialization unit includes a storage unit and an initialization calculation unit;

[0062] The storage unit is used to parse the data sent by the received window transformation and store the parsed image metadata in a general data storage method;

[0063] The initialization calculation unit is used to initialize the primitive, calculate the initialization parameters of the primitive, and store the calculated initialization parameters of the primitive;

[0064] The block scanning interpolation unit is used to perform interpolation scanning on the primitives in a parallel scanning manner based on the initialization parameters of the primitives;

[0065] The storage unit stores the image element data so that the next image element can be initialized while the previous image element is interpolated and scanned.

[0066] Specifically, the rasterization acceleration unit receives the data sent by the window transformation and parses it to obtain the corresponding command information and metadata. If the data is command information, the corresponding register is directly set. If the data is metadata, the primitive needs to be processed according to the current configuration information. The current primitives, whether point primitives, line primitives, or triangle primitives, need to be initialized, but different primitive types have different initialization parameters and different scanning interpolation algorithms. Therefore, it is necessary to deeply explore the scanning process and design special processing into a general method to implement it. The overall structure of the rasterization acceleration unit is as follows: Figure 3 shown.

[0067] The primitive initialization process is a complex process. In addition to considering what initialization calculations different scanning algorithms need to do, it is also necessary to consider how to design it into a pipeline structure. The main purpose of the pipeline structure is to enable the next primitive to be initialized in parallel when the previous primitive is scanning interpolation calculations, thereby saving the time of primitive initialization. Since there may be a lot of primitive attribute data, the embedded GPU can support up to 32 attributes, and each attribute is a 4-component single-precision floating-point number, such as coordinate data is XYZW, color data is RGBA, etc. If the current primitive is a triangle, then the attribute data to be received is 384 single-precision floating-point numbers, which need to be stored first and then calculated. In order to meet the storage of different primitive types such as points, lines, and triangles using the same circuit structure, and only store valid attribute data, a general data storage method is designed to store primitive data. The flowchart of the general data storage method is shown below. Figure 4 shown.

[0068] First, receive the data sent by the window transformation, read the command and data, and determine whether the data is a point, line, or triangle primitive. If not, discard the data and do no subsequent processing. If the data is a point, line, or triangle primitive, read the data. If the data is a point primitive, read one vertex data; if the data is a line primitive, read two vertex data; if the data is a triangle primitive, read three vertex data. Store the read vertex data in the storage space of the circuit structure.

[0069] The present invention designs a universal circuit structure and allocates storage space, so that point primitives, line primitives, and triangle primitives can all be stored in the same circuit structure. The issued primitive data is stored in the circuit structure in a pipeline manner, so that parallel processing can be achieved later. For example, while the previous primitive is performing interpolation scanning, the initialization parameter calculation of the next primitive can be performed.

[0070] After the primitive attributes are stored, the initialization parameters need to be calculated. The initialization parameters of the primitive generally include the initial value and offset of the boundary function, the initial value and offset of the perspective correction of the attribute, the calculation of the triangle area and the inverse of the area, etc. The initialization calculation process of the primitive involves addition and subtraction, multiplication and division. In the traditional circuit structure, there are usually only addition and subtraction and multiplication operations. Since the divider generally has a large circuit area and needs to be designed as a pipeline structure, the present invention completes the divider with a transcendental function of the inverse. The transcendental function is implemented in a pipeline based on a quadratic polynomial lookup table. Among them, there are many methods for implementing transcendental functions. The present invention uses a pipeline method of quadratic polynomial approximation to implement transcendental functions. Considering that the initialization establishment of the primitive requires calculation of a large amount of data, the circuit can be reused and processed in a pipeline manner, the initialization calculation process of all primitives is sorted to meet the design requirements. The sorting here refers to sorting the addition and subtraction, multiplication and division operations in the initialization calculation process to ensure that each operation process is carried out in an orderly manner.

[0071] After the initialization is completed, different primitives are scanned and interpolated in a pipeline manner according to different interpolation algorithms. In the present invention, point primitives are scanned according to a point-by-point scanning algorithm, line primitives are scanned according to an improved Bresenham algorithm, and triangle primitives are scanned according to a multi-scanning block parallel scanning algorithm. The scanned fragment data are all selected and output in a 4-way parallel manner, and the 4 ways are completely independent and have no data correlation.

[0072] Among them, the scanning of point primitives is relatively simple. When the scanning point size Point Size set in the command information is 1, the point primitives are directly output; when the scanning point size Point Size set in the command information is greater than 1, the point primitives are output according to the blocks of the specified size.

[0073] Line primitive scanning is based on the improved Bresenham algorithm to scan and interpolate. Users can set the scan line width Line Width through command information. When the scan line width Line Width set in the command information is 1, it is a single line width scan; when the scan line width Line Width set in the command information is greater than 1, it is a multi-line width scan, where multi-line width scanning includes: scanning the current line width up and down with the discrete segment scanned when the scan line width Line Width is 1 as the center, and then calculating the next segment when Line Width is 1 through the Bresenham algorithm, and so on, to complete the multi-line width scan. The schematic diagram of multi-line width scanning can be seen in Figure 5 .

[0074] Figure 5 Line Width = 3, that is, the scanning line width is 3, and three lines of scanning are required. Figure 5 Start scanning at the vertex numbered 1 in , then calculate the vertices 2 and 3 above and below 1, so the next vertex fragment of 1 is 4, calculate the vertices 5 and 6 above and below 4, so the next vertex fragment of 4 is 7, and so on. The three lines are 1, 4, 7..., 2, 5, 8... and 3, 6, 9..., completing the multi-line width scan with Line Width=3.

[0075] The rasterization of triangle primitives is based on the TBR architecture block design, which cuts the triangle primitives into multiple triangle blocks, covers them in different scanning blocks, and calculates the initial properties of each triangle block, such as the initial position of each triangle block.

[0076] After the triangle primitives are cut, multiple triangle blocks are covered in different scanning blocks, and the multiple scanning blocks scan the triangle blocks in parallel at the same time, thereby improving the scanning speed of the triangle blocks.

[0077] The number of scanning blocks is no longer fixed but can be adjusted dynamically according to the boundary function of the current triangle initialization calculation. The size of the scanning block is also dynamically adjustable, but at least 4 rows are scanned in parallel in the block. If the current number of scanning blocks is 8 and the size of each scanning block is 4*4, then 32 fragments can be scanned in 1 clock cycle under peak conditions. Triangle rasterization does not bundle multiple scanning blocks into a group and wait until the block with the most fragment coverage is scanned before switching to the next group. It is not a serial scan of multiple blocks one by one, but multiple blocks are scanned in complete parallel and each block can be flexibly switched to the next position to be scanned after scanning. The scanning method of triangle primitives can be seen in Figure 6 and Figure 7 .

[0078] The triangle primitive is cut into multiple triangle blocks and covered into multiple scan blocks, such as Figure 6 , you can see, Figure 6 The scanning block No. 0 in is not covered by any triangle block, so the scanning block No. 0 does not need to do any work, so the scanning block No. 0 will quickly move to the next triangle block, such as Figure 7 As shown, there is no need to wait for other scanning blocks to complete scanning.

[0079] The embodiment of the present invention provides a low-power consumption rasterization design method for an embedded GPU, which has the following beneficial effects:

[0080] (1) The initialization process of different primitive types is designed to be universal and completed in a pipeline structure, which satisfies the embedded graphics processor's efficient scanning of different primitives.

[0081] (2) The rasterization unit outputs scanned fragments continuously and in parallel in a pipeline manner, greatly improving key performance of the graphics processor such as pixel fill rate. At the same time, the triangles are rasterized and scanned in a block-by-block rendering manner. Regardless of whether there is an angle between the current fragment and the original two-dimensional texture image, the texture fill rate will not drop significantly. In addition, the rasterization unit is designed with a high-reuse circuit structure and low-power processing based on clock gating to improve the performance of embedded graphics processors.

[0082] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0083] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0084] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A low-power rasterization design method for embedded GPU, It is characterized in that include: Color the vertices based on the vertex shader, assemble the colored vertices into primitives, and perform clipping, backface culling, and viewport transformation on the primitives; Rasterize the primitives transformed by the viewport based on the rasterization acceleration unit to obtain discrete fragments, and color the fragments based on the fragment shader; Wherein, the rasterization acceleration unit includes a primitive initialization unit and a plurality of block scanning interpolation units, and the primitive initialization unit includes a storage unit and an initialization calculation unit; The storage unit is used to parse the data sent by the received window transformation and store the parsed image metadata in a general data storage method; The initialization calculation unit is used to initialize the primitives and calculate the initialization parameters of the primitives; The block scanning interpolation unit is used to perform interpolation scanning on the primitives in a parallel scanning manner based on the initialization parameters of the primitives; The storage unit stores the image element data so that the next image element can be initialized while the previous image element is interpolated and scanned; The interpolation scanning of the primitives by using a parallel scanning method includes: According to the different types of primitives, the corresponding interpolation scanning algorithm is used for scanning; Wherein, when it is a point primitive, it is scanned according to the point-by-point scanning algorithm; When it is a line element, it is scanned according to the improved Bresenham algorithm; When it is a triangle primitive, it is scanned according to the multi-scanning block parallel scanning algorithm; When the primitive is a triangle, scanning is performed according to a multi-scanning block parallel scanning algorithm, including: Cutting a triangle primitive into a plurality of triangle blocks, wherein the plurality of triangle blocks are covered in different scanning blocks, and calculating an initial attribute of each triangle block; Performing parallel line scanning on the cut multiple triangle blocks based on multiple scanning blocks; The number of scanning blocks is determined according to the boundary function calculated by initializing the current triangle primitive, and the size of each scanning block can be adjusted dynamically.

2. The low-power rasterization design method for an embedded GPU according to claim 1, It is characterized in that The receiving window transform sent data, parsing the data, and storing the parsed metadata in a general data storage method include: Receive the data sent by the window transformation, parse the data, and parse out the corresponding command information and graphic data. If the parsed data is command information, directly set the corresponding register. If the parsed data is graphic data, store the graphic data in the storage space of the circuit structure using a general data storage method.

3. The low-power rasterization design method for an embedded GPU according to claim 2, It is characterized in that The step of storing the image metadata in the storage space of the circuit structure in a general data storage manner includes: Read the data sent by the window transformation, and determine whether the data is a point, line, or triangle primitive. If not, end the process; If the data is a point primitive, read a vertex data; If the data is a line primitive, two vertex data are read; If the data is a triangle primitive, three vertex data are read; The read vertex data is stored in the storage space of the circuit structure.

4. The low-power rasterization design method for an embedded GPU according to claim 1, It is characterized in that The initialization parameters of the primitives include the initial value and offset of the boundary function, the initial value and offset of the perspective correction of the attribute, and the calculation of the triangle area and the inverse of the area; The calculation process of the initialization parameters of the primitives includes addition and subtraction operations, multiplication operations and division operations, and the divider is designed by a reciprocal transcendental function, and the transcendental function is implemented by a pipeline method of quadratic polynomial approximation; And in the calculation process of the initialization parameters of the primitives, the addition and subtraction operations, multiplication operations and division operations are sorted, and the corresponding operations are performed in sequence to complete the initialization calculation process of the primitives.

5. The low-power rasterization design method for an embedded GPU according to claim 2, It is characterized in that When it is a point primitive, scanning is performed according to a point-by-point scanning algorithm, including: When the point size Point Size is set to 1 in the command information, the point primitive is directly output; When the point size Point Size is set to be greater than 1 in the command information, the point primitives are output in blocks of the specified size.

6. The low-power rasterization design method for an embedded GPU according to claim 2, It is characterized in that When it is a line element, scanning is performed according to the improved Bresenham algorithm, including: When the scan line width Line Width is set to 1 in the command information, it is a single line width scan; When the scan line width Line Width set in the command information is greater than 1, it is multi-line width scanning, wherein the multi-line width scanning includes: The current line width segment is scanned up and down with the discrete segment scanned when the scanning line width Line Width is 1 as the center, and then the next segment is calculated by Bresenham algorithm, and so on, to complete the multi-line width scanning.

7. The low-power rasterization design method for an embedded GPU according to claim 1, It is characterized in that The step of rasterizing the image primitives after the window transformation to obtain discrete fragments and coloring the fragments includes: The interpolation scanned fragments are output in a multi-channel parallel manner; The output fragment is dispatched to the fragment shader through the fragment scheduler FDPR to shade the scanned fragment.

8. The low-power rasterization design method for an embedded GPU according to claim 1, It is characterized in that The vertex shader and the fragment shader are used to shade vertices and fragments respectively based on a stream multiprocessor SMPU.