Method and system for accelerating complex scene ray tracing rendering

Abstract

The invention discloses a method for accelerating complex scene ray tracing rendering, and relates to the technical field of computer graphic analysis. The method comprises the following steps: carrying out rapid ray tracing rendering on a complex scene by utilizing a distributed rendering cluster, and improving the calculation efficiency when the distributed rendering cluster renders a frame of complex scene by matching a load balancing strategy with the distributed rendering cluster; a frame of rendering result image generated by the distributed rendering cluster and corresponding auxiliary information are input into a multi-scale denoising device for post-processing of filtering denoising and time anti-aliasing, the sense of reality of the rendering result image is improved, and the rendering result image output after processing of the multi-scale denoising device is the final high-sense-of-reality rendering result image. The invention aims at effectively carrying out vivid rendering on a complex scene, and improving the rendering efficiency and effectively denoising while keeping the rendering effect.

Description

Technical Field

[0001] The present invention relates to the technical field of computer graphics analysis, and in particular to a method and system for accelerating ray tracing rendering of complex scenes. Background Art

[0002] At present, ray tracing technology is more widely used in realistic games, film production and other fields, and it also faces higher requirements. In practical applications, realistic rendering of complex scenes is often required. The lighting conditions in complex scenes are usually more complex, ray tracing is time-consuming and difficult to guarantee rendering efficiency. Monte Carlo ray tracing is widely used for its high realism of rendering results, and its implementation basis is Monte Carlo integration. The Monte Carlo integration method requires a large number of samples to achieve reliable convergence, which also determines that Monte Carlo ray tracing needs to be rendered at a high sampling rate to obtain realistic result images; but ray tracing rendering at high sampling rates often consumes a lot of computing resources, and it takes several hours to render a frame of high-quality image.

[0003] In the prior art, in order to improve the efficiency of realistic rendering of complex scenes, a load balancing algorithm based on rendering history was proposed earlier. This algorithm uses the rendering time of the previous frame to estimate the workload of the next frame. It is simple and easy to implement. Later, a task-based dynamic load balancing algorithm was proposed. The queue mechanism was used to effectively achieve load balancing effects on multiple GPUs, which improved the computing efficiency. However, it would generate serious noise, thus affecting the realism of the rendering effect. Therefore, in order to maintain the realism of the rendering effect, the research on denoisers has been continuously developed in recent years. A denoiser based on supervised learning was proposed. A simple multi-layer perceptron (MLP) was selected for training with a small number of scenes. Subsequently, a supervised learning framework was proposed. This framework uses a deep convolutional neural network (CNN) to obtain a more complex and universal filter kernel. The complex filter kernel has a large number of filter kernel parameters, which ensures the accuracy of the denoising effect.

[0004] However, although the task-based dynamic load balancing algorithm proposed in the prior art improves the computing efficiency, the frequent communication leads to the problem that the network bandwidth is limited and the load balancing strategy has limited improvement in the ray tracing speed of complex scenes. At the same time, in order to make the rendering effect of complex scenes more realistic, the denoiser based on the supervised learning framework proposed in the prior art obtains a complex and universal filter kernel through training. Although it solves the problem of excessive blurring of the denoising effect, the large number of filter kernel parameters of the complex filter kernel slows down the filtering speed.

[0005] Therefore, how to effectively render complex scenes realistically, maintain the rendering effect while improving rendering efficiency and effectively denoising is an urgent problem to be solved. Summary of the invention

[0006] Based on this, it is necessary to provide a method and system for accelerating ray tracing rendering of complex scenes to address the above technical problems.

[0007] The present invention adopts the following technical solutions:

[0008] The present invention provides a method and system for accelerating ray tracing rendering of complex scenes, comprising:

[0009] Obtain the complex 3D scene to be rendered and the required rendering resources, and create a complex scene ray tracing task; and accelerate the rendering of the complex scene ray tracing task through a distributed rendering cluster with a load balancing strategy based on rendering history, and obtain the rendering result image and its corresponding auxiliary information;

[0010] Construct a multi-scale denoiser including an input module, a multi-scale kernel prediction module of the U_Net network shape and an output module;

[0011] The rendering result image and its corresponding auxiliary information are input into the multi-scale denoiser, the albedo of the rendering result image of the current frame is removed through the input module, and the denoising result of the previous frame is temporally reprojected to realize the accumulation of historical information in the time domain, and obtain the rendering result image containing historical information; the rendering result image containing historical information is filtered through the multi-scale kernel prediction module of the U-Net network shape to obtain the filtered rendering result image; the albedo removed by the input module is added to the filtered rendering result image through the output module to restore the detail information of the current frame to obtain the final rendering result image.

[0012] Preferably, the auxiliary information includes: geometry buffer information and pixel buffer information generated during the accelerated rendering process of the distributed rendering cluster on the ray tracing task.

[0013] Preferably, obtaining the rendering result image and its corresponding auxiliary information comprises the following steps:

[0014] The distributed rendering cluster selects squares of the same size as the division units of the screen space for subtask division. When the remaining space in the screen space is insufficient for division, the remaining part is classified into adjacent image blocks to obtain multiple subtasks, and a subtask sequence is obtained. The subtask division needs to ensure that the number of divided subtasks is greater than the number of rendering nodes in the distributed rendering cluster.

[0015] Based on the load balancing strategy of rendering history, multiple subtasks and corresponding rendering resources are distributed to each rendering node in the distributed rendering cluster, and the rendering subgraphs of each rendering node and their corresponding auxiliary information are obtained;

[0016] The control node of the distributed rendering cluster receives the rendering sub-images and their corresponding auxiliary information returned by each rendering node, and obtains the rendering result image and its corresponding auxiliary information by splicing all the rendering sub-images and their corresponding auxiliary information.

[0017] Preferably, the load balancing strategy based on rendering history distributes multiple subtasks and corresponding rendering resources to each rendering node in the distributed rendering cluster, specifically including:

[0018] Determine whether the previous frame of the current frame of the subtask sequence corresponds to a sudden scene in the subtask sequence;

[0019] If the answer is yes, the control node of the distributed rendering cluster distributes subtasks to each rendering node in turn according to the rendering node sequence number until all subtasks are distributed;

[0020] If the judgment is no, the rendering cost of each subtask of the current frame is estimated using the rendering history of the previous frame, and the control node of the distributed rendering cluster distributes the subtasks to the rendering nodes according to the greedy strategy.

[0021] Preferably, the determining whether a previous frame of the current frame of the subtask sequence corresponds to a sudden scene in the subtask sequence specifically includes:

[0022] Introduce the cost estimation accuracy and calculate the cost estimation accuracy corresponding to the previous frame of the current frame;

[0023] Set the cost estimation accuracy threshold based on experimental verification;

[0024] If the cost estimation accuracy is greater than the cost estimation accuracy threshold, the cost estimation of the previous frame is inaccurate and the previous frame corresponds to a burst scenario; otherwise, the previous frame does not correspond to a burst scenario.

[0025] Preferably, the estimating the rendering cost of each subtask of the current frame by using the rendering history of the previous frame specifically includes:

[0026] Record the total rendering cost of the previous frame of the distributed rendering system and the rendering cost consumed by the rendering node to complete and submit each subtask;

[0027] According to the inter-frame continuity principle that the rendering cost of each subtask of the previous frame is similar to the rendering cost of each subtask of the current frame, the rendering cost of each subtask of the current frame is estimated.

[0028] Preferably, the control node of the distributed rendering cluster distributes subtasks to rendering nodes according to a greedy strategy, including the following steps:

[0029] Store all subtasks in the subtask sequence into a queue in descending order of rendering cost;

[0030] According to the status report sent by the rendering node to the control node, the total amount of tasks to be calculated by each rendering node is counted, and a rendering node with the smallest total amount of tasks to be calculated is selected, and the current head element of the subtask queue is distributed to the selected rendering node; the control node repeatedly distributes the rendering nodes until the subtask queue is empty.

[0031] Preferably, obtaining the final rendering result image comprises the following steps:

[0032] Calculate the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame. The calculation formula is:

[0033] m(x i )=x i-1 -x i

[0034] Among them, m(x i ) is the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame, x i-1 With x i are the pixel values ​​of the previous frame and the current frame respectively;

[0035] According to the motion vector, the denoising result image of the previous frame is projected to the screen space of the current frame to realize the accumulation of information in the time domain and obtain the rendering result image containing historical information;

[0036] Inputting the rendering result image containing historical information into a multi-scale kernel prediction module of a U-Net network shape, the module includes a series of convolutional layers, each layer combines the retained detail information with the rendering result image containing historical information to obtain a rendering result image after multi-scale filtering; wherein the detail information is the image detail information retained after each convolutional layer in the kernel prediction module filters the rendering result image containing historical information;

[0037] The rendering result image after multi-scale filtering is input into the output module to add the albedo removed by the input module to the filtered rendering result image, restore the detail information of the current frame, and obtain the final rendering result image.

[0038] Preferably, obtaining the pixel value of the previous frame specifically includes:

[0039] Project the current frame pixels to the world coordinates of the current frame;

[0040] According to the movement of the geometric body, the world coordinates of the current frame are transformed into the world coordinates of the previous frame, and the world coordinates of the previous frame are projected into the screen space to obtain the pixel values ​​of the previous frame.

[0041] Preferably, the preliminary rendering image and feature acquisition module is used to receive the complex three-dimensional scene to be rendered and the required rendering resources, and create a complex scene ray tracing task; and accelerate the rendering of the complex scene ray tracing task through a distributed rendering cluster to obtain a rendering result image and its corresponding auxiliary information;

[0042] A denoiser building module, used to build a multi-scale denoiser including an input module, a multi-scale kernel prediction module of a U-Net network shape and an output module;

[0043] The preliminary rendering image determination module is used to input the rendering result image and its corresponding auxiliary information into the multi-scale denoiser, remove the albedo from the rendering result image of the current frame through the input module of the multi-scale denoiser, and perform time reprojection on the denoised result of the previous frame to achieve historical information accumulation in the time domain, and obtain a rendering result image containing historical information; filter the rendering result image containing historical information through the multi-scale kernel prediction module of the U-Net network shape of the multi-scale denoiser to obtain a filtered rendering result image; add the albedo removed by the input module to the filtered rendering result image through the output module of the multi-scale denoiser, restore the detail information of the current frame, and obtain the final rendering result image.

[0044] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0045] In the face of ray tracing tasks in complex scenes, the present invention proposes a load balancing strategy based on rendering history to improve rendering speed. The rendering cost of the current frame is estimated according to the rendering cost of the completed rendered frame, avoiding the rendering vacuum caused by waiting for the current frame rendering to be completed and then allocating the current frame rendering task. Then, according to the estimated rendering cost, the cluster is guided to reasonably distribute the current frame rendering task, which improves the cluster load balancing degree and thus improves rendering efficiency.

[0046] In addition, the rendering of complex scenes requires high realism of the resulting images. The present invention provides a multi-scale denoiser. By accumulating historical information in the time domain, its multi-scale kernel prediction module can effectively denoise complex scenes containing multi-scale objects, significantly improving the denoising efficiency compared to traditional technologies.

[0047] In summary, the present invention uses a distributed rendering cluster with a load balancing strategy based on historical rendering and a multi-scale denoiser to effectively improve the efficiency of accelerated rendering of complex scenes while ensuring high realism of the rendering results. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0049] Figure 1 A schematic diagram of a process flow of a method and system for accelerating ray tracing rendering of complex scenes provided by the present invention;

[0050] Figure 2 A system architecture diagram of a ray tracing acceleration method for accelerating ray tracing rendering of complex scenes provided by the present invention and a system;

[0051] Figure 3 A schematic diagram of task division when a distributed rendering cluster of a method and system for accelerating ray tracing rendering of complex scenes provided by the present invention renders one frame;

[0052] Figure 4 A schematic diagram of a distributed rendering cluster of a method and system for accelerating ray tracing rendering of complex scenes provided by the present invention distributing subtasks to a frame with an estimated rendering cost;

[0053] Figure 5 A schematic diagram of a multi-scale kernel prediction denoiser for a method and system for accelerating ray tracing rendering of complex scenes provided by the present invention;

[0054] Figure 6 A comparison chart of denoising results of a method and system for accelerating ray tracing rendering of complex scenes provided by the present invention; DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0057] A method and system for accelerating ray tracing rendering of complex scenes, the flow chart of which is as follows Figure 1 As shown in the system architecture diagram Figure 2 As shown, the method specifically comprises the following steps:

[0058] S101: receiving a complex 3D scene to be rendered and required rendering resources, creating a complex scene ray tracing task; and accelerating the rendering of the complex scene ray tracing task through a distributed rendering cluster in conjunction with a load balancing strategy based on rendering history, and obtaining a rendering result graph and its corresponding auxiliary information, specifically including:

[0059] like Figure 3 As shown, the distributed rendering cluster selects squares of the same size as the division units of the screen space for subtask division. When the remaining space in the screen space is insufficient for division, the remaining part is classified into adjacent image blocks to obtain multiple subtasks, and a subtask sequence is obtained; wherein, the subtask division needs to ensure that the number of divided subtasks is greater than the number of rendering nodes in the distributed rendering cluster;

[0060] like Figure 4 As shown, the distributed rendering cluster cooperates with the load balancing strategy based on rendering history to reasonably distribute the divided multiple subtasks and corresponding rendering resources to each rendering node in the distributed rendering cluster;

[0061] Through the distributed rendering cluster and the load balancing strategy based on historical rendering, the subtasks in the divided subtask sequence and the corresponding rendering resources are reasonably distributed to each rendering node in the distributed rendering cluster, and the rendering sub-graphs of each rendering node and their corresponding auxiliary information are obtained;

[0062] The control node of the distributed rendering cluster receives the rendering sub-images and their corresponding auxiliary information returned by each rendering node, and obtains the rendering result image and its corresponding auxiliary information by splicing all the rendering sub-images and their corresponding auxiliary information.

[0063] The load balancing strategy based on rendering history specifically includes:

[0064] Determine whether the previous frame of the current frame of the subtask sequence corresponds to a sudden scene in the subtask sequence;

[0065] If the answer is yes, the rendering history of the previous frame cannot be used to estimate the rendering cost of each subtask of the current frame. The control node of the distributed rendering cluster distributes subtasks to each rendering node in turn according to the rendering node sequence number until all subtasks are distributed;

[0066] If the judgment is no, the rendering cost of each subtask of the current frame can be estimated by using the rendering history of the previous frame, and the control node of the distributed rendering cluster distributes the subtasks to the rendering nodes according to the greedy strategy.

[0067] The method of determining whether the previous frame of the current frame of the subtask sequence corresponds to the sudden scene in the subtask sequence specifically includes: introducing a cost estimation accuracy rate, calculating the cost estimation accuracy rate corresponding to the previous frame of the current frame, and setting a cost estimation accuracy rate threshold value according to experimental verification. If the cost estimation accuracy rate is greater than the cost estimation accuracy rate threshold value, it is considered that the cost estimation of the previous frame is inaccurate and the previous frame corresponds to the sudden scene. Otherwise, it is considered that the previous frame does not correspond to the sudden scene.

[0068] The estimating the rendering cost of each subtask of the current frame by using the rendering history of the previous frame specifically includes:

[0069] Record the total rendering cost of the previous frame of the distributed rendering system and the rendering cost consumed by the rendering node to complete and submit each subtask; based on the inter-frame continuity principle that the rendering cost of each subtask in the previous frame is similar to the rendering cost of each subtask in the current frame, estimate the rendering cost of each subtask in the current frame.

[0070] The control node of the distributed rendering cluster distributes subtasks to rendering nodes according to a greedy strategy, including the following steps:

[0071] Store all subtasks in the subtask sequence into a queue in descending order of rendering cost;

[0072] According to the status report sent by the rendering node to the control node, the total amount of tasks to be calculated by each rendering node is counted, and a rendering node with the smallest total amount of tasks to be calculated is selected, and the current head element of the subtask queue is distributed to the selected rendering node;

[0073] The control node repeats the above steps of distributing subtasks until the subtask queue is empty.

[0074] S102: Construct a multi-scale denoiser including an input module, a multi-scale kernel prediction module of a U_Net network shape and an output module.

[0075] S103: inputting the rendering result image and its corresponding auxiliary information into a multi-scale denoiser, removing the albedo of the rendering result image of the current frame through the input module of the multi-scale denoiser, performing temporal reprojection on the denoised result of the previous frame, realizing historical information accumulation in the time domain, and obtaining a rendering result image containing historical information;

[0076] like Figure 5 As shown, the rendering result image containing historical information is filtered through the multi-scale kernel prediction module of the U-Net network shape of the multi-scale denoiser to obtain the filtered rendering result image; the albedo removed by the input module is added to the filtered rendering result image through the output module of the multi-scale denoiser to restore the detail information of the current frame to obtain the final rendering result image, which specifically includes:

[0077] Calculate the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame. The calculation formula is:

[0078] m(x i )=x i-1 -x i

[0079] Among them, m(x i) is the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame, x i-1 With x i are the pixel values ​​of the previous frame and the current frame respectively;

[0080] According to the motion vector, the denoising result image of the previous frame is projected to the screen space of the current frame to realize the accumulation of information in the time domain, thereby obtaining a rendering result image containing historical information;

[0081] Inputting the rendering result image containing historical information into a multi-scale kernel prediction module of a U-Net network shape, the module includes a series of convolutional layers, each layer combines the retained detail information with the rendering result image containing historical information to obtain a multi-scale filtering result, and fusing the multi-scale filtering results to obtain a filtered rendering result image; wherein the detail information is the image details retained after each convolutional layer in the kernel prediction module filters the rendering result image containing historical information;

[0082] The filtered rendering result image is input into the output module of the multi-scale denoiser, the albedo removed by the input module is added to the filtered rendering result image, the detail information of the current frame is restored, and the final rendering result image is obtained;

[0083] The acquisition of the pixel value of the previous frame specifically includes: projecting the current frame pixels to the world coordinates of the current frame, then transforming the world coordinates of the current frame into the world coordinates of the previous frame according to the geometric movement, and projecting the world coordinates of the previous frame to the screen space to obtain the pixel value of the previous frame.

[0084] In order to compare with other denoising methods, the present invention uses a sampling rate as low as 1-spp to render two complex scenes respectively, and obtains the noise result map (1-spp); the noise result map (1-spp) is input into different denoisers for denoising, and the denoising effects are compared. Figure 6 As shown in the figure, under the same denoising time, the multi-scale kernel prediction denoiser can effectively remove noise while retaining more image details; its denoising effect is better than the comparative denoiser.

[0085] In summary, through the verification of this embodiment, the present invention uses a distributed rendering cluster with a load balancing strategy based on historical rendering and combines it with a multi-scale denoiser to effectively improve the efficiency of accelerated rendering of complex scenes while ensuring high realism of the rendering results, providing scientific guidance for the practical application of ray tracing technology.

[0086] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for accelerating ray tracing rendering of complex scenes, characterized in that: The following steps are involved: Obtain the complex 3D scene to be rendered and the required rendering resources, and create a complex scene ray tracing task; and accelerate the rendering of the complex scene ray tracing task through a distributed rendering cluster with a load balancing strategy based on rendering history, and obtain the rendering result image and its corresponding auxiliary information; Construct a multi-scale denoiser including an input module, a multi-scale kernel prediction module of the U_Net network shape and an output module; The rendering result image and its corresponding auxiliary information are input into the multi-scale denoiser, the albedo of the rendering result image of the current frame is removed through the input module, and the denoising result of the previous frame is temporally reprojected to realize the accumulation of historical information in the time domain, and obtain the rendering result image containing historical information; the rendering result image containing historical information is filtered through the multi-scale kernel prediction module of the U-Net network shape to obtain the filtered rendering result image; the albedo removed by the input module is added to the filtered rendering result image through the output module to restore the detail information of the current frame to obtain the final rendering result image.

2. A method for accelerating ray tracing rendering of complex scenes as claimed in claim 1, characterized in that: The auxiliary information includes: geometry buffer information and pixel buffer information generated during the accelerated rendering process of the distributed rendering cluster on the ray tracing task.

3. A method for accelerating ray tracing rendering of complex scenes as claimed in claim 1 or 2, characterized in that: The acquisition of the rendering result image and its corresponding auxiliary information includes the following steps: The distributed rendering cluster selects squares of the same size as the division units of the screen space for subtask division. When the remaining space in the screen space is insufficient for division, the remaining part is classified into adjacent image blocks to obtain multiple subtasks, and a subtask sequence is obtained. The subtask division needs to ensure that the number of divided subtasks is greater than the number of rendering nodes in the distributed rendering cluster. Based on the load balancing strategy of rendering history, multiple subtasks and corresponding rendering resources are distributed to each rendering node in the distributed rendering cluster, and the rendering subgraphs of each rendering node and their corresponding auxiliary information are obtained; The control node of the distributed rendering cluster receives the rendering sub-images and their corresponding auxiliary information returned by each rendering node, and obtains the rendering result image and its corresponding auxiliary information by splicing all the rendering sub-images and their corresponding auxiliary information.

4. A method for accelerating ray tracing rendering of complex scenes as claimed in claim 3, characterized in that: The load balancing strategy based on rendering history distributes multiple subtasks and corresponding rendering resources to each rendering node in the distributed rendering cluster, specifically including: Determine whether the previous frame of the current frame of the subtask sequence corresponds to a sudden scene in the subtask sequence; If the answer is yes, the control node of the distributed rendering cluster distributes subtasks to each rendering node in turn according to the rendering node sequence number until all subtasks are distributed; If the judgment is no, the rendering cost of each subtask of the current frame is estimated using the rendering history of the previous frame, and the control node of the distributed rendering cluster distributes the subtasks to the rendering nodes according to the greedy strategy.

5. The method for accelerating ray tracing rendering of complex scenes as claimed in claim 4, characterized in that: The determining whether a frame before the current frame of the subtask sequence corresponds to a sudden scene in the subtask sequence specifically includes: Introduce the cost estimation accuracy and calculate the cost estimation accuracy corresponding to the previous frame of the current frame; Set the cost estimation accuracy threshold based on experimental verification; If the cost estimation accuracy is greater than the cost estimation accuracy threshold, the cost estimation of the previous frame is inaccurate and the previous frame corresponds to a burst scenario; otherwise, the previous frame does not correspond to a burst scenario.

6. The method for accelerating ray tracing rendering of complex scenes as claimed in claim 4, characterized in that: The estimating the rendering cost of each subtask of the current frame by using the rendering history of the previous frame specifically includes: Record the total rendering cost of the previous frame of the distributed rendering system and the rendering cost consumed by the rendering node to complete and submit each subtask; According to the inter-frame continuity principle that the rendering cost of each subtask of the previous frame is similar to the rendering cost of each subtask of the current frame, the rendering cost of each subtask of the current frame is estimated.

7. The method for accelerating ray tracing rendering of complex scenes as claimed in claim 4, characterized in that: The control node of the distributed rendering cluster distributes subtasks to rendering nodes according to a greedy strategy, including the following steps: Store all subtasks in the subtask sequence into a queue in descending order of rendering cost; According to the status report sent by the rendering node to the control node, the total amount of tasks to be calculated by each rendering node is counted, and a rendering node with the smallest total amount of tasks to be calculated is selected, and the current head element of the subtask queue is distributed to the selected rendering node; the control node repeatedly distributes the rendering nodes until the subtask queue is empty.

8. The method for accelerating ray tracing rendering of complex scenes as claimed in claim 1, characterized in that: The acquisition of the final rendering result image includes the following steps: Calculate the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame. The calculation formula is: m(x i )=x i-1 -x i Among them, m(x i ) is the motion vector between the scene corresponding to the previous frame and the scene corresponding to the current frame, x i-1 With x i are the pixel values ​​of the previous frame and the current frame respectively; According to the motion vector, the denoising result image of the previous frame is projected to the screen space of the current frame to realize the accumulation of information in the time domain and obtain the rendering result image containing historical information; Inputting the rendering result image containing historical information into a multi-scale kernel prediction module of a U-Net network shape, the module includes a series of convolutional layers, each layer combines the retained detail information with the rendering result image containing historical information to obtain a rendering result image after multi-scale filtering; wherein the detail information is the image detail information retained after each convolutional layer in the kernel prediction module filters the rendering result image containing historical information; The rendering result image after multi-scale filtering is input into the output module to add the albedo removed by the input module to the filtered rendering result image, restore the detail information of the current frame, and obtain the final rendering result image.

9. A method for accelerating ray tracing rendering of complex scenes as claimed in claim 8, characterized in that: The acquisition of the pixel value of the previous frame specifically includes: Project the current frame pixels to the world coordinates of the current frame; According to the movement of the geometric body, the world coordinates of the current frame are transformed into the world coordinates of the previous frame, and the world coordinates of the previous frame are projected into the screen space to obtain the pixel values ​​of the previous frame.

10. A system for accelerating ray tracing rendering of complex scenes, characterized in that: include: The preliminary rendering and feature acquisition module is used to receive the complex 3D scene to be rendered and the required rendering resources, and create a complex scene ray tracing task; And through the distributed rendering cluster, the ray tracing task of complex scenes is accelerated to obtain the rendering result image and its corresponding auxiliary information; A denoiser building module, used to build a multi-scale denoiser including an input module, a multi-scale kernel prediction module of a U-Net network shape and an output module; The preliminary rendering image determination module is used to input the rendering result image and its corresponding auxiliary information into the multi-scale denoiser, remove the albedo from the rendering result image of the current frame through the input module of the multi-scale denoiser, and perform time reprojection on the denoised result of the previous frame to achieve historical information accumulation in the time domain, and obtain a rendering result image containing historical information; filter the rendering result image containing historical information through the multi-scale kernel prediction module of the U-Net network shape of the multi-scale denoiser to obtain a filtered rendering result image; add the albedo removed by the input module to the filtered rendering result image through the output module of the multi-scale denoiser, restore the detail information of the current frame, and obtain the final rendering result image.

Images