What would it feel like to watch a fashion show in the meta universe? The friction details of various fabrics are not only clear, but also the same as real. The skirt dances with the movement of the model ~ just look, you can know whether this dress is light or not. This is the latest achievement of the graphics Daniel Wang Huamin team. A GPU based preprocessor MAS is proposed, and a GPU completes the real-time cloth collision effect.
For example, this wool coat can be simulated in real time at the speed of 37 frames per second, and the simulation accuracy can reach 5mm.
And not only clothes, but also hairy balls and soft leather toys can achieve realistic collision effects.
His thesis is included in siggraph 2022 . As the largest and highest level graphics conference in the world, the papers received by siggraph every year represent the highest level of CG in that year.
Lai Kangkang, what kind of research is this?
A GPU real-time simulation collision effect
To build a true sense of the meta universe, the first thing we need is a powerful real-time physical simulation engine.
The current common physical engine is far from reaching the ideal state in effect or efficiency.
*Clothing simulation * is the commanding point of all simulation problems.
Because the material and collision of clothing are the most complex, solving the clothing simulation is equivalent to solving most of the real-time physical simulation problems, laying the foundation for the future development of the meta universe.
Technically, real-time physical simulation can be divided into two categories: fluid simulation and shape variant simulation.
Compared with fluids, shape variants have a wider range of applications, such as elastomer simulation, cloth simulation, hair simulation, etc., and their industrial value is also clearer, such as applications in clothing and industry.
However, one of the current challenges perplexing scientists is completing higher quality real-time simulation .
The simulation of any shape variant finally belongs to a linear solution problem. The difference is nothing more than the difference in the internal value of the matrix, the degree of sparsity and the size of the matrix.
Therefore, how to quickly solve has become an important breakthrough.
Based on this background, Wang Huamin's team proposed a MAS preprocessor based on GPU.
As the name suggests, this preprocessor adopts the concept of multi-layer and domain decomposition.
Previous algorithms mainly rely on one layer of grid, so the fewer domains, the better, otherwise the error is large.
But this time, researchers are building multiple grid layers, so that using small, non overlapping domains can give full play to the parallel computing power of GPU.
Specifically, the MAS preprocessor can be divided into three stages.
In the first stage, use Morton code to quickly layer and domain .
Researchers proposed a super node segmentation method and skip method for the preliminary construction of space.
In the second stage, through the pre calculation of low-cost matrix eliminated by unidirectional Gauss Jordan , a fast inverse solution is carried out in each domain. In addition, the researchers also developed a selective update scheme to solve minor matrix modifications.
In the third stage, runtime preprocessing , so as to quickly solve the matrix vector. In order to reduce the computational cost, researchers invented a symmetric matrix vector multiplication .
Experiments show that MAS preprocessor is compatible with many linear and nonlinear solvers, which are used for shape variant simulation with dynamic contact, such as PCG, accelerated gradient descent and l-bfgs.
Taking the most common PCG solver as an example, when used together, the convergence rate on GPU is still significantly better than other preprocessing programs, including multigrid amgx on GPU.
When reaching the same convergence goal, the convergence speed of MAS preprocessor is about four times faster than that of other preprocessors.
At the same time, MAS preprocessor is also effective on CPU , and its performance is also better than other preprocessors, such as Intel's ilu0 and ilut.
In addition, it is also found that the computational overhead of MAS preprocessor is very low.
The parallel computing time is only 0.118 seconds, compared with the costs of ilu0, ichol and ilut of 2.50 seconds, 33.8 seconds and 154 seconds.
The researchers said that the cloth and shape variant simulation with the scale of 50000 fixed points to 500000 vertices has been pushed to a higher level.
In the future, we will continue to study the technology based on GPU domain decomposition, because such algorithms have a lot of space to mine.
Produced by style3d
The research team is led by Wang Huamin of style3d Research Institute.
*Wang Huamin *, former associate professor of Ohio State University, doctor of computer science of Georgia Institute of technology, master of computer science of Stanford University, and recognized as a world-class graphics scientist.
He once completed four siggraph papers as the sole author, and was a member of the sixth siggraph / siggraph Asia Technical Paper Committee. His students are all over famous factories. Because he focuses on cloth simulation, there is a saying in the industry:
If you study cloth simulation , you can't have never read Professor Wang Huamin's paper.
If you study cloth simulation , you can't have never read Professor Wang Huamin's paper.
Last August, he officially joined Lingdi technology style3d, a 3D digital company in the domestic fashion industry chain, and served as chief scientist and President of Lingdi Research Institute.
Results in less than a year, he led the team to bring the latest production, study and research results, and promoted the real-time simulation technology to a new level.
The rapid progress of research is driven by the strength of world-class technical teams and the industrialization of enterprises.
*Lingdi technology style3d *, at the beginning of its establishment, focused on the garment design and development link that most restricts the efficiency of the garment industry.
One of its core products is the 3D flexible body simulation industrial software style3d studio , which is the first to break the monopoly of foreign technology, the other is the global innovative digital fabric design software style3d fabric , the whole process collaboration platform style3d cloud , and the pioneering digital fashion content resource mall style3d market .
In the R & D process, digital presentation and collaborative cooperation can be carried out from the whole process of surface auxiliary material selection, style design, rendering simulation, collaboration and exhibition, so as to improve the R & D efficiency.
In addition, Lingdi style3d also generates a list of production materials according to the factory production standards, connecting the whole production and manufacturing links, and further promoting the digitization of the whole industry link.
It is revealed that Lingdi style3d will release mysterious products for meta universe developers at the end of June this year.
At present, its customers involve brands at home and abroad, ODM (original design and manufacturing) manufacturers, flour accessories manufacturers and pan e-commerce.
It is worth mentioning that at the same time as the siggraph collection results were released, Lingdi style3d announced its cooperation with cowboy technology company jeanologia .
The company's environmental technology is used in 35% of the jeans produced worldwide every year.
With the cooperation of the two, denim designers can also directly call jeans patterns and washing effect patterns with one click in style3d. Without making real samples of jeans, they can present the design effects of various processes such as jeans washing.
Moreover, such high simulation digital jeans can also be directly used for marketing display and sale.
In fact, in addition to promoting the efficiency of the whole garment industry chain, there are more realistic global problems that can be alleviated, namely carbon neutralization and carbon emission.
At present, textile and garment industry has become the second largest polluting industry in the world . About 12.8 million tons of clothing are buried and burned every year. The wastewater discharged from the production of clothing accounts for 20% of the global total and bears 10% of the global carbon emissions.
Even, 15% of textile products will not be used and will be burned directly, resulting in secondary pollution.
Once the whole process of the whole industry from R & D, production to the final commodity exhibition and sales is put on the digital platform, the pollution problem can be effectively solved, and sustainable fashion can really have the direction of implementation.
Industrial development of graphics
Finally, return to the research results of siggraph.
What we can see is that it doesn't look like NVIDIA and Google's annual traditional art energy - all kinds of cool and realistic special effect technologies to highlight the cutting-edge technical strength of the enterprise.
But from a more vertical industrial perspective, technological iteration is not easy to be perceived, but it effectively solves industrial problems.
On the one hand, it shows that graphics technology has come to the industrial end to create value.
Graphics is the most likely and maximum field to be implemented, not hot movies and games, but industrial fields such as clothing and industry.
On the other hand, it also shows that China's technical level in graphics is not weak and will accelerate from the frontier to the industry like AI.
From the results of the siggraph paper collection, we can see that the Chinese team has occupied more than half of the list, including representatives of colleges and universities and enterprises.
As the value of graphics industry becomes more explicit , more world-class cattle are involved. Enterprises will also attract more college talents by setting up research institutes and starting scholarships.
Taking Lingdi style3d as an example, the Lingdi graphics scholarship program was launched last year. Each winner will receive a cash reward of 30000 yuan. In the first phase, 10 masters and doctors from Tsinghua University, Zhejiang University, China University of science and technology, Nankai University and Shandong University won the scholarship.
In this way, the formation of industry university research closed loop is more conducive to the long-term development of graphics industrialization.
Now the booming meta universe finally squeezes out the foam. Like the original AI, we can only combine the underlying technology with the industry to play a longer-term value.
In this regard, Chinese enterprises have taken the lead.