I knew my life would be based around the game industry since I was 10, but I never had the opportunity to dive deeper into the industry until I first went to GDC and saw the exciting things that people are working on.

Flying back on the plane with many ideas spinning in my head, I was wondering how I could engage more with the industry and help game developers achieve better performance and improved visual quality.

A few days later, I still couldn’t sleep as ideas were spinning in my head. I imagined being in a windowed room with the sun outside travelling across the sky. Then I thought, if the room is static geometry, why do I need to render it to the shadow map every frame? The things which are moving around are everything but the room: the sun, the camera, the dynamic objects etc.  I realized that a texture could represent the whole static environment (the room in my case). Going a bit further, the alpha channel could represent how much light can enter the room. My initial assumption was that a cubemap texture would only fit a uniform  room but, as we later found out, cubemaps seem to be very good approximation of many kinds of local environment (not only square rooms but also irregular shapes such as the cave we used in the Ice Cavedemo which you can see below).

With the whole room represented by a cube texture I could access arbitrary texels of the environment from within fragment shader. With that in mind, the sun could be in any arbitrary position and the amount of light reaching a fragment calculated based on the value fetched from the cubemap.

I shared the theory with my ARM colleague, Roberto Lopez Mendez, and within a couple of hours we had a working prototype running in Unity. We were very pleased to see the prototype achieve very efficient soft shadows while maintaining a high level of visual quality and we decided to use this technique in the Ice Cave demo which was premiered at GDC 2015.

Is the video playback not good enough? Please see this link to fix the issue.

I only recently got back from staffing the ARM booth at GDC and I had a fantastic time meeting so many interesting and knowledgeable people in the game industry. I was really delighted to be able to show the Ice Cave demo and explain all the visual effects we used. The shadow technique in particular garnered a lot of interest with many developers keen to use it in their own engine.

Overview about shadow technique

In this blog I would like to give a technical overview of what you need to do in order to make the shadow technique work in your own projects. If you are not familiar with the reflections based on local cubemaps technique, I recommend reading this blog. There are also many other internet resources on the subject.

Why did I mention reflections based on local cubemaps? The main reason is that once you have implemented these reflections you are almost there with the shadow technique. The only additional thing you need to do alongside generating a reflection cube is to store alpha along with the RGB colours.

The alpha channel (transparency) represents the amount of light entering the local environment. In your scene, attach the cubemap texture to the fragment shaders that are rendering the static and dynamic objects on which you want to apply shadows. In the fragment shaders, build a vector from the current fragment to the light position in world space. As long as we are using local cubemaps we cannot use that vector directly to fetch a vectorand we need just one more calculation step which is a local correction. In order to do that we need to calculate the intersection of that vector with the bounding volume (the bounding box of the room/environment) and then we need to build another vector from the position where the cubemap has been generated to the intersection point.

This vector (Q and P) can now be used to fetch a texel from the cubemap. Once you have fetched the texel you will have information about the amount of shadow/light to be applied on the fragment which is being processed. It is as simple as that.

That was a short summary of the technique, below you will find more details and a step by step guide of what you need to do in order to make shadows look really stunning in your application.

Generating shadow cubemaps

If you are already familiar with reflections based on local cubemaps, this step is exactly the same as creating the reflection cubemap (probe).  Moreover you can reuse your reflection cubemap for this shadow technique. You just need to add an alpha channel.

Let’s assume you have a local environment within which you want to apply shadows from the light sources outside of the local environment e.g. room, cave, cage, etc. As an aside, the environment does not have to be enclosed, it can be an open space, but this is a subject for another blog. Let’s focus on the local environment, a simple room, in order to understand better how the technique works.

You need to work out the position from which you will render six faces of the cubemap – in most cases it will be the centre of the environment’s bounding volume (a box). You will require this position not only for the generation of the cubemap, but later the position needs to be passed to shaders in order to calculate a local correction vector to fetch the right texel from the cubemap.

Once you have decided where to position the centre of the cubemap you can render all faces to the cubemap texture and record the transparency (alpha) of the local environment. The more transparent an area is, the more light will come into the environment. If required, you can use the RGB channels to store the colour of the environment for coloured shadows like stained glass, reflections, refractions etc.

Rendering shadows

Applying shadows to the geometry couldn’t be simpler. All you need to do is build a vector “L” in world space from a vertex/fragment to the light(s) and fetch the cubemap shadow by using this vector.

However there is tiny step you need to do with the “L” vector before fetching each texel. You need to apply local correction to the vector. It is recommended to make the local correction in the fragment shader to obtain more precise shadows. In order to do this you need to calculate the intersection point with the bounding volume (bounding box) of the environment and use this intersection point to build another vector from the cubemap origin position to the intersection point. This gives you the final “Lp” vector which should be used to fetch the texel.

Input parameters:

• _EnviCubeMapPos – the cubemap origin position
• _BBoxMax – the bounding volume (bounding box) of the environment
• _BBoxMin – the bounding volume (bounding box) of the environment
• V – the vertex/fragment position in world space
• L – the normalized vertex-to-light vector in world space

Output value:

• Lp – the corrected vertex-to-light vector which needs to be used to fetch a texel from the shadow cubemap.

An example code snippet which you can use to correct the vector:

// Working in World Coordinate System.

vec3 intersectMaxPointPlanes = (_BBoxMax - V) / L;

vec3 intersectMinPointPlanes = (_BBoxMin - V) / L;

// Looking only for intersections in the forward direction of the ray.    

vec3 largestRayParams = max(intersectMaxPointPlanes, intersectMinPointPlanes);

// Smallest value of the ray parameters gives us the intersection.

float dist = min(min(largestRayParams.x, largestRayParams.y), largestRayParams.z);

// Find the position of the intersection point.

vec3 intersectPositionWS = V + L * dist;

// Get the local corrected vector.

Lp = intersectPositionWS - _EnviCubeMapPos;

Then use the “Lp” vector to fetch a texel from the cubemap. The texel’s alpha channel [0..1] provides information about how much light (or shadow) you need to apply for a given fragment.

float shadow = texCUBE(cubemap, Lp).a;

At this point you should have working shadows in your scene. There are then two more minor steps to improve the quality of this shadowing.

Back faces in shadow

As you may have noticed, we are not using depth information to apply shadows and that may cause some faces to be incorrectly lit when in fact they should be in shadow. The problem only occurs when a surface is facing in the opposite direction to the light. In order to fix this problem you simply need to check the angle between the normal vector and the vertex-to-light vector, L. If the angle, in degrees, is out of the range -90 to 90, the surface is in shadow. Below is a code snippet which will do this check.

if (dot(L,N) < 0)

  shadow = 0.0;

As always there is room for improvement. The above code will cause each triangle into a hard switch from light to shade, which we would like to avoid. We need a smooth transition which can be done with the following simple formula:

shadow *= max(dot(L, N), 0.0);

shadow – alpha value fetched from the shadow cubemap

L – the vertex-to-light vector in world space

N – the normal vector of the surface, also in world space

Smoothness

The last, and I would say coolest, feature of this shadow technique is its softness. If you do it right you will get realistic shadow softness in your scene.

First of all make sure you generate mipmaps for the cubemap texture and use tri-linear filtering.

Then the only thing you need to do is measure the length of a vertex-to-intersection-point vector and multiply the length by the coefficient which you will have to customize to your scene. For example, in the Ice Cave project we set the coefficient to 0.08. The coefficient is nothing but a normalizer of a maximum distance in your environment to the number of mipmap levels. If you want you can calculate it automatically against bounding volume and mipmap levels. We found it very useful to have manual control to allow us to tweak the settings to suit the environment, which helped us to improve the visual quality even further.

Let’s get to the nitty-gritty then. Here we can reuse computations which have been done in calculating local correction. We can reuse the intersection point to build the vertex-to-intersection-point vector and then we need to calculate the distance.

float texLod = length(IntersectPositionWS - V);

Then we multiply the lodDistance by the distance coefficient:

texLod *= distanceCoefficient;

To implement softness, we must fetch the correct mipmap level of the texture by using texCUBElod (Unity) or textureLod (GLSL) function and construct a vec4 where XYZ represents a direction vector and the W component represents LOD (Level of Detail).

Lp.w = texLod;

shadow = texCUBElod(cubemap, Lp).a;

Now you should see high quality smooth shadows in your scene.

Combine cubemap shadows with shadowmap

In order to get the full experience you will need to combine cubemap shadows with the traditional shadow map technique. Even if it sounds like more work it is still worth it as you will only need to render dynamic objects to the shadow map. In the Ice Cave demo, we simply added the two shadow results in the fragment shader to get our final shadow value.

Without Shadow Map	With Shadow Map

Statistics

In traditional techniques, rendering shadows can be quite expensive as it involves rendering the whole scene from the perspective of each shadow-casting light source. The technique described in this blog is mostly prebaked (for improved performance) and independent of output-resolution, producing the same visual quality for 1080p as it produces for 720p and other resolutions.

The softness filtration is calculated in hardware (via the texture pipeline) so the smoothness comes almost for free. In fact, the smoother the shadow the more efficient the technique is. This is due to the smaller mipmap levels which result in less data transferred from main memory to the GPU compared to  traditional shadow map techniques which require a large kernel to make shadows smooth enough to be more visually appealing. This obviously causes a lot more data traffic and reduces performance.

The quality you get with the shadow technique is even higher than you may expect. Apart from realistic softness, you have very stable shadows with no more shimmering on the edges. Shimmering edges can be observed when using traditional shadow map techniques due to rasterization and aliasing. None of the anti-aliasing algorithms can fix this problem entirely. Obviously enabling multi-sampling helps to improve quality, but the shimmering effect is still visible. The cubemap shadow technique is free from this problem. The edges are stable even if you use a much lower resolution than is used in the render target. You can easily use four times lower resolution than the output and see neither artefacts nor unwanted shimmering. Needless to say that having four times lower resolution saves massively on bandwidth and improves performance!

Let’s get to the point with performance data. I already covered the texture fetch above which is very efficient. Apart from fetching a texel you only need to make few calculations related to the local correction. The local correction on ARM® Mali™ GPU-based devices does not take more than three cycles per fragment. This perfectly aligns with the texture fetch pipeline and while the GPU is calculating the next vector for the cubemap, the texture pipeline is preparing the texel for the current fragment.

Conclusion

This technique can be used on any device on the market which supports shaders i.e. OpenGL® ES 2.0 and higher. But as everything else in our world it has its downsides. The technique cannot be used for everything in your scene. Dynamic objects, for instance, receive shadows from the cubemap but they cannot be prebaked to the cubemap texture. The dynamic objects should use shadow maps for generating shadows, blended with the cubemap shadow technique.

As a final word we are seeing a lot of implementations of reflections based on local cubemaps and this shadow technique is based on more or less the same paradigm. If you already know where and when to use the reflections based on local cubemaps technique, then you will be able to easily apply the shadow technique to your implementation.

More than that, you will find the shadow technique is less error prone when it comes to local correction. Our brains do not spot so many errors with shadows as we tend to spot in reflections. Please have a look at our latest demo, Ice Cave, and try to implement the technique in your own projects.

I hope you enjoy using it!

Introduction

Game developers are regularly looking for efficient methods for implementing stunning effects in their games. This is especially important when targeting mobile platforms as resources should be carefully balanced to achieve maximum performance.

When developing the Ice Cave demo, we researched in-depth the concept of local cubemaps. In a previous blog, I wrote about how implementing reflections based on local cubemaps has proven to be a very efficient technique for rendering high quality reflections. This blog is devoted to a new technique developed by the ARM demo team to render refractions also based on local cubemaps.

Refractions, what is it?

Refraction is an important effect to consider when striving for extra realism when rendering semi-transparent geometry.

Refraction is the change in direction of a wave due to a change in the transmission medium. It is essentially a surface phenomenon. The refractive index determines how much light is bent, or refracted, when entering a material. Snell’s Law establishes the relationship between the refractive indices and the sine of the incident and refracted angles, as shown in Figure 1.

Figure 1. Refraction of light as it passes through one medium to another.

Refraction implementations

Developers have tried to render refraction since the very moment they started to render reflections, since these two physical processes take place together in any semi-transparent surface. There are several well-known techniques for rendering reflections but is not the case for refractions.

Existing methods for implementing refraction at runtime (ray tracing is excluded due to its complexity) differ depending on the specific type of refraction. Nevertheless, most of the techniques render to texture, the scene behind the refractive object at runtime and apply a non-physically based distortion in a second pass to achieve the “refracted look”. This approach, which varies in the way the texture perturbation is performed, is used to render the refraction that takes place in water, heat haze and glass objects, among other effects.

Although some of these techniques can achieve credible results, texture perturbation is not physically based and the results are not always correct.  If a realistic refraction is intended by rendering to texture from the point of view of the “refraction camera”, there may be areas that are not directly visible to the camera but become visible via refraction. Nevertheless, the main limitation of runtime render-to-texture methods, beside the physical correctness and performance penalty, is the quality, as there is often pixel shimmering or pixel instability that it is easily perceived while the camera is moving.

The use of static cubemaps to implement refraction is not new. Since the very moment when cubemaps became available in 1999, developers have used the cubemaps to implement reflections as well as refractions. When using cubemaps to implement reflections in a local environment, if we don’t apply the local correction we get incorrect results. This is also true for refractions.

Refractions based on local cubemaps

We bake into a static cubemap the environment surrounding the refractive object and fetch the texel from the cubemap based on the direction of the refracted vector (after applying the local correction, see Figure 2).

Figure 2. The local correction to refraction vector.

We apply the local correction in the same way we did with reflections in a previous blog. After determining the direction of the refracted vector, we need to find where it intersects the bounding box that delimits the volume of the local scene. The next step is to build a new vector from the position where the cubemap was generated to the intersection point and use this final vector to fetch the texel from the cubemap to render what is behind the refractive object. We get a physically based refraction as the direction of the refraction vector is calculated according to Snell’s Law. Moreover, there is a built-in function we can use in our shader to find the refraction vector R strictly according to this law:

R = refract( I, N, eta);

where I is the normalized view or incident vector, N is the normalized normal vector, eta is the ratio of indices of refractions (n₁/n₂).

Shader implementation

For the simple case of a thin refractive surface, the shader implementation is straightforward, as shown in Figure 3. As for reflections, to apply the local correction in the fragment shader we need to pass the position where the cubemap was generated, as well as the minimum and maximum bounds of the bounding box (all in world coordinates).

Figure 3. Shader implementations of refraction based on local cubemap.

Once we fetch the texel corresponding to the locally-corrected refraction direction, we might want to combine the refraction colour with other lighting, for example, reflections that in most cases take place simultaneously with refraction. In this case, we just need to pass an additional view vector to the fragment shader, apply to it the local correction and use the result to fetch the reflection colour from the same cubemap. Below is a code snippet showing how reflection and refraction might be combined to produce a final output colour.

A coefficient _ReflAmount, which is passed as a uniform to the fragment shader is used to adjust the balance between reflection and refraction contributions.  You can use ReflAmount to tweak manually the visual effect for the look you are trying to achieve. You can find the implementation of the LocalCorrect function in the reflections blog. When the refractive geometry is a hollow object refractions and reflections take place in both the front and back surfaces (as shown in Figure 4.) In this case, we need to perform two rendering passes.

Figure 4. Refraction on a glass bishop based on local cubemap.  Left: First pass renders only back faces with local  refraction  and reflections.  Right: Second pass renders only front faces with local refraction and reflections and alpha  blending with the first pass.

In the first pass, we render the semi-transparent object as we would opaque geometry. Additionally, in this pass, we render the object last with front-face culling on, i.e. to avoid occluding other objects, we render only the back faces and no depth buffer writing. The colour of the back face is obtained by mixing the colours calculated from the reflection, refraction and diffuse colour of the object itself.

In the second pass, we render only the front faces (back face culling), again last in the rendering queue and with depth writing off. The front-face colour is obtained by mixing the refraction and reflection textures with the diffuse colour. In this final pass we alpha-blend the resulting colour with the previous pass. Notice the combination of environment refractions and reflections.in both pictures from Figure 4.

The refraction in the second pass will add more realism to the final rendering but we could skip this step if the refraction on the back faces is enough to highlight the effect.

Figure 5 shows the result of implementing refractions based on local cubemap on a semi-transparent phoenix in the Ice Cave demo.

Figure 5. Refractions based on local cubemaps in the Ice Cave demo.

Preparing the cubemap

Preparing the cubemap for use in the refraction implementation is a simple process. We just need to place a camera in the centre of the refractive geometry and render the surrounding static environment to a cubemap in the six directions. During the rendering process, the refractive object is hidden. This cubemap can then be used for implementing both refraction and reflection. The ARM Guide To Unity contains a simple script for rendering a cubemap.

Conclusions

Local cubemaps have proven to be an excellent technique for rendering reflections. In this blog I have shown how to use local cubemaps to implement very optimized and high quality refractions that can be combined at runtime with reflections to achieve high quality visual results (as seen in the Ice  Cave demo.) This is especially important when developing graphics for mobile devices where runtime resources must be carefully balanced. Nevertheless, this technique does have limitations due to the static nature of the cubemap. How to deal with refractions when the environment is dynamic will be the subject of another blog.

Light is sight

When we start talking about the importance of lighting in Geomerics, we often refer directly to light’s importance in setting the mood and atmosphere of a scene - but we jump way ahead of ourselves here. Step one of light in physics – without light, there is no sight. Everything we see in the real world is the result of light reflecting off surfaces and into our eyes.  If we turn off the lights in a room, close the curtains, stuff the doorframe with fabric to stop light leaking in, the objects within it are still there, our eyes are still there, but the objects remain unseen. Defining and highlighting form is the first step in lighting; it lets us see details in objects and determine how they are shaped. Smooth surfaces gradient softly between light and shadow, but sharp edges deliver distinct changes.

Going from this basic first step to effectively using lighting to set the mood, intensity and atmosphere of a scene is a long jump. There is no magic formula for getting the perfect combination of light, shadow and color to achieve the desired artistic vision for the environment, in part because it is, like any art, subjective.The combination depends so much on the ambience being created - mystery horrors, for example, will tend to use low lighting and lots of shadows, punctuated by pools of light to grab your attention and draw you in; sometimes even plunging the player into darkness with the only light source being a torch controlled by the player.

The variety of lighting

When designing a game, there are many different light sources available: for example directional, ambient, spotlight and point light. The source of a directional light is infinitely far away such that by the time they reach the viewer all light rays are parallel – a good example is sunlight; they can be stationary or movable. Ambient lighting casts soft rays equally to every part of a scene without any specific direction, and so provides light but not shadow; it has no real source.  Spotlights emit from a single source in a cone shape with an inner cone angle that provides full brightness and an outer cone angle that allow softening at the edges of where the light is falling; these are often used for torches. Point lights are much like real-world lightbulbs or candles, they emit from a single point in all directions.

Each of these light sources will provide a different type of direct lighting and their effect is computed by the rendering engine. However, to simulate the physics of lighting in the real world it is important to also calculate the indirect lighting, or global illumination, of a scene. Global illumination takes into account the way in which light travels from its source, hits an object and is then absorbed, reflected, scattered or refracted across every subsequent surface it encounters. It is perfect for producing the kind of style required by architectural visualization, interior renders, scenes with direct sunlight and photorealistic renders thanks to its calculation of indirect lighting, soft shadows, colour bleeding and more. For example, light reflecting off a red leather seat cushion will “bleed” colour onto the wall next to it – depending on the colour of the wall this could produce a reddish glow (if wall is white), or purple (if wall is blue). Alternatively, a great effect is where light can leak from one room to its neighbour, gently illuminating the new room through just an open crack in a door.

Bear in mind that an exact simulation of how light works is not necessarily required. All that’s needed is something that is good enough to fool the (admittedly, clever) human eye.

Solving the global illumination challenge

One option for achieving global illumination in a scene is an offline lightmap bake. This gives the illusion that light is being cast onto an object, but what you’re actually seeing is just the effect of the light baked onto the texture. This technique delivers high quality results, but the iteration time is slow and it has limited runtime possibilities – for example the baked light won’t have any effect on moving objects, not can it be turned on or off during play. Another technique is “bounce lighting”, where artists add light sources into the game at strategic positions in order to simulate global illumination – for example at the point where a light would be reflected, a new light source is added with the desired properties. In comparison, this has a fast iteration time, but it can take a very high number of iterations to achieve physical correctness, it is hard to achieve dynamism and the number of light sources may be limited by the engine in use.

Enlighten is a third option for achieving accurate, lightweight and dynamic global illumination. Enlighten uses real time radiosity to compute the interaction between a scene’s geometry and light. It contains a unique and highly optimised runtime library that generates lightmaps of bounce lighting in real time. The lightmap generation occurs on the CPU and is simply added to the rest of the direct lighting on the GPU. This approach can be further combined with lightmaps generated offline, so only the lights and materials that need to be updated at run time incur any cost. In this way, Enlighten offers a highly scalable solution suitable for all gaming platforms, from PC and console right the way down to mobile, and all lighting requirements, from fully baked to totally dynamic. Because the scene’s lighting and materials are also able to be updated dynamically at runtime in the editor (as well as in the game), rapid iteration is possible. By taking into account the indirect light, surface properties, and specularity in a scene it generates an extremely high quality and realistic output. For example, by enabling the bounce lighting to pick up the colour properties of the surfaces in the scene, Enlighten naturally ties together the geometry and lighting in an environment. In addition its ability to update materials’ properties at runtime create a host of new gameplay opportunities, as demonstrated with the Subway demo where destruction was achieved by making walls transparent.

More information on Enlighten is available at www.geomerics.com.

About me

Hi, I am Hans-Kristian Arntzen! This is my first post here. I work in the Mali use cases team where we explore the latest mobile APIs to find efficient ways of implementing modern graphics techniques on the ARM Mali architecture.

Sometimes, we create small tech demos which result in Mali SDK samples, smaller code examples which you can take inspiration from when developing your own applications.

Since August, I've been writing quite a lot of code for OpenGL ES 3.1 and I will summarize what we have done with OpenGL ES 3.1 the last months.

About OpenGL ES 3.1

OpenGL ES 3.1 is an update to OpenGL ES 3.0 which recognizes the fact that OpenGL ES 3.0 capable hardware is already capable of much more, for example compute. OpenGL ES 3.1 now brings GPU compute support directly to OpenGL ES, so there is no longer any need to interface with external APIs to expose the compute capabilities of the hardware. The interface for compute is very clean, powerful and easy to use.

Compute support in graphics APIs means there are many more opportunities now for applications to offload parallel work to the GPU than before and being able to do this on mobile hardware is very exciting.

See Here comes OpenGL® ES 3.1! for more details.

Mali driver support for OpenGL ES 3.1

We released the r5p0 driver in December with support for OpenGL ES 3.1. The driver for Linux and Android platforms can be found here: Drivers - Mali Developer Center Mali Developer Center.

Update to the Mali OpenGL ES SDK

The latest Linux and Android OpenGL ES SDK has new sample code for compute shaders.

Mali OpenGL ES SDK for Linux - Mali Developer Center Mali Developer Center

Mali OpenGL ES SDK for Android - Mali Developer Center Mali Developer Center

The samples can be built for Linux development platforms with fbdev.

There is also OpenGL ES emulator support included (OpenGL ES Emulator - Mali Developer Center Mali Developer Center) so you can run the Linux fbdev samples on your desktop on Linux and Windows.

If your desktop implementation supports X11/EGL in Linux, you should be able to run the samples without emulator by leveraging the GL_ARB_ES3_1_compatibility extension which went  into core in OpenGL 4.5.

Introduction to Compute Shaders

Introduction to compute shaders - Mali Developer Center Mali Developer Center

Compute is a new subject for many graphics programmers. This document tries to explain the different mind set you need to effectively use GPU compute and the new APIs found in OpenGL ES 3.1.

It goes through the major features of compute, and in-depth into some more difficult subjects like synchronization, memory ordering and execution barriers.

It is recommended that you read this before studying the examples below unless you're already familiar with compute shaders.

Particle Flow Simulation with Compute Shaders

Particle Flow Simulation with Compute Shaders - Mali Developer Center Mali Developer Center

This sample implements a modern particle system. It uses compute shaders to sort particles back-to-front which is critical to obtain correct alpha blending.

Since we can sort now on the GPU, we can offload the entire particle system to the GPU.

It also implements cool things like 4-layer opacity shadow map for some sweet volumetric shadow effects and simplex noise to add turbulence to the particles.

Combining all these techniques together allow you to create a very nice particle system.

Occlusion Culling with Hierarchical-Z

Occlusion Culling with Hierarchical-Z - Mali Developer Center Mali Developer Center

Culling is important in complex scenes to keep vertex work down as mentioned in this blog post: Mali Performance 5: An Application's Performance Responsibilities

For game objects, there are many sophisticated CPU-based solutions which often rely on baking data structures based on how the scene is put together.

For example, in indoor scenes with separate rooms, it makes sense to only consider rendering the room you're in and objects from rooms with are visible from the room you're standing in. Doing this computation on-the-fly could get expensive, but once the information is baked, it is fairly simple.

However, when we add a large amount of "chaotic" elements to a more dynamic scene, it becomes more difficult to bake anything and we need to compute this on the fly. We have to look for some more general solutions for these scenarios.

The sample shows how you can use a low-resolution depth map and bounding spheres to efficiently cull entire instances in parallel before they are even rendered. It can also be combined with level-of-detail sorting to reduce geometry load even further.

Finally, the result is drawn with indirect draws, a new feature of OpenGL ES 3.1.

Using these kinds of techniques allow you to offload big "particle-like" systems to the GPU efficiently.

Game Developers Conference 2015

At GDC2015 we presented updated best practices for GLES 3.1 on Mali along with a newly developed tech demo. I manned our tech booth at the expo floor most of the time where I got to show my demo to other people, which was quite exciting.

Best Practices for OpenGL ES 3.1 on Mali

There are certain things you should think about when developing for Mali. During our work with OpenGL ES 3.1, we have found some general performance tips you should take into account.

Compute exposes more low level details about the architecture, and to get optimal performance for a particular architecture, you might need some specific optimizations.

If you are experienced with compute on desktop, you might find that many general truths about performance on desktop don't necessarily apply to mobile! Sometimes, performance tips are opposite of what you'd want to do on desktop.

If you have used OpenCL on Mali before, best practices for OpenCL also apply for compute shaders.

Presentations

I presented at the GDC2015 along with Tom Olson (Chair of Khronos OpenGL ES and Vulkan working groups, Director of Graphics Research at ARM) and Dan Galpin (Developer Advocate, Google).

The presentation goes through OpenGL ES 3.1 (focus on compute), some of the techniques I mentioned in this post, best practices for OpenGL ES 3.1 and AEP on Mali and a small sneak peak on early Vulkan experiments on Mali.

Unleash the Benefits of OpenGL ES 3.1 and Android Extension Pack:

http://malideveloper.arm.com/downloads/GDC15/Unleash%20the%20Benefits%20of%20OpenGL%20ES%203.1and%20the%20Android%20Exte…

At the ARM Lecture Theater at GDC2015, I also did a short presentation with focus exclusively on compute. It goes a bit more in detail on compute shader basics compared to the full length GDC talk:

http://malideveloper.arm.com/downloads/GDC15/Lecture%20Theatre%20-%20Wednesday/Unleash%20the%20Benefits%20of%20OpenGL%20…

Caveats with r5p0 release

Unfortunately, there are some performance bugs with some features in r5p0 release. You might stumble into them when developing for OpenGL ES 3.1.

• Indirect draws can slow down a lot compared to regular draws.
• Compute shaders with smaller work groups (e.g. 4 or 8 threads) are much slower (3-4x) than compute shaders using 64 or 128 threads.

If you run into these issues, they have been addressed and should be fixed in future driver releases.

Occlusion Culling with Compute Shaders demo

I am very excited about compute shaders and culling, so much that I wanted to create a demo for it at GDC. We do have the Occlusion Culling sample code in the SDK, but it is far too bare-bones to show at an event.

I attended GDC 2015, where I manned our tech booth most of the time and I got to show this demo to many people passing by.

Instead of dull green spheres I went for some procedurally generated asteroids. All the asteroids look slightly different even if they are instanced due to the use of a 4-component RGBA8 heightmap. All asteroids have different random weighting factors which make them look a bit different. They have independent radii, rotation axes and rotation speeds as well which makes the scene look fairly complex. Diffuse textures and normal textures are shared for all asteroids. They are also generated procedurally with perlin noise and compressed with ASTC LDR.

There are over 27000 asteroids in the scene here spread out across a big sphere around the camera.

At highest quality, each individual asteroid has over 2500 triangles. If we were to just naively draw this without any kind of optimization, we would get a triangle count in the ballpark of 50+ million which is extremely overkill.

We need some culling. The first and obvious optimization is frustum culling, which can remove most of the asteroids outright. We can do this on the GPU very efficiently and parallel since it's just a couple of dot products per instance after all.

All the asteroids in the scene are represented as a flat linear array of per-instance data such as position, base radius, rotation axis, rotation speed and heightmap weighting factors. We combine frustum culling with the physics update (rotating the asteroids and creating a final rotation quaternion per asteroid). Since we need to update every asteroids anyways, might as well do frustum culling while they are in cache!

Now we're looking at ~2000 asteroids being rendered, but just frustum culling is not enough! We also need LOD sorting to get the vertex count low enough.

The idea behind LOD sorting is that objects far away don't need high detail. We can add this technique to plain frustum culling and reduce the vertex count a lot. After these optimizations, we're looking at 500-600k triangles per frame, a 100x reduction from before. We can also use cheaper vertex shaders for objects far away, which reduces the vertex load even more. We can also do this efficiently in compute shaders, it's just a question of pushing per-instance data to one of many instance buffers if it passes the frustum test.

Here we see close objects in white and it gets darker as the LOD factor increases.

We can also use different shading for close objects. Here, close asteroids have full bling with normal mapping and specular highlights from the skydome, objects farther away are only diffuse with spherical harmonics for diffuse lighting.

This kept fragment shading load down quite a bit. Screenshot shows the debugged normals. The normals without normal mapping look a bit funky, but that's because the normals are computed directly in the vertex shader by sampling the heightmap multiple times. With shading applied it looks fine however

But we can do even better. You might have noticed the transparent "glasslike" wall in front of the asteroids? It is supposed to be opaque. We wanted this to be a space station interior or something cool, but unfortunately we didn't have time for GDC

The main point here is that there is a lot of stuff going on behind the occluder in the scene. There is no reason why we should waste precious bandwidth and cycles on vertex shading asteroids which are never seen.

Enter Occlusion Culling!

We can go from this:

to this:

After this optimization we cull over half the asteroids in the scene on average, and we are looking at a very manageable 200-300k triangles.

My hope for the future is that we'll be able to easily do all kinds of scene management directly on the GPU. It's not feasible to do everything on the GPU quite yet, the CPU is still very capable of doing things like these, but we can definitely accelerate massively instanced cases like these

Skydome

The skydome is procedurally generated with FBM noise. It is HDR and is used for all the lighting in the scene. I compressed it with ASTC HDR instead of RGB9E5, a 32-bit shared exponent format which is pretty much the only reasonable alternative if I didn't have ASTC HDR.

Performance

I squeezed out 60 fps at 1080p/4xMSAA on a Samsung Galaxy Note 10.1 (Mali-T628 MP6) and Samsung Galaxy Note 4 (Korean version, Mali-T760 MP6) when all culling was applied, which I'm quite happy with .

I used DS-5 (ARM DS-5 Streamline - Mali Developer Center Mali Developer Center) to find bottlenecks when tuning along with (Mali Offline Compiler - Mali Developer Center Mali Developer Center) to fine-tune the shaders (mediump varyings can make a lot of difference!).

OpenGL ES is the standard API for 2D and 3D graphics on embedded systems, which includes mobile phones, tablets, smart TVs, consoles, and other appliances and vehicles. The API is a well-defined subset of desktop OpenGL.

In our ARM Mali DeveloperCenter, we have two OpenGL ES Software Development Kits, one for Android OS and one for Linux environments. These SDKs are aimed mainly from beginners to intermediate users, with a guide on how to get your system properly configured and set up, as well as how to build the OpenGL ES 2.0 and the latest OpenGL ES 3.x sample code and run them. 

 

The SDKs contain tutorials and sample code. The tutorials start from the basic concepts, such as introduction to shaders, and setting up the graphics pipeline to render to the display, followed by examples to render basic geometry and add textures and light onto them. The more advanced tutorials talk about our latest ARM Mali features like implementing deferred shading using the tile buffer available through the Shader Pixel Local Storage OpenGL ES extension, as well as samples illustrating the use of AdaptiveScalable Texture Compression (ASTC), and showing the use of the very latest OpenGL ES 3.1 APIs, released in the latest Android Lollipop.

The most important OpenGL ES 3.1 API feature is Compute Shaders, which allow the GPU to be used for general-purpose computing and therefore, the SDK includes a dedicated tutorial to it called introduction to compute shaders.  The tutorials refer to sample code, and a detailed summary of the Compute Shader sample code available in the SDK is given in Hans-Kristian’s blog.

ARM Mali-T6xx and ARM Mali-T7xx series support OpenGL ES 3.1, and the latest Android L OS is capable of running OpenGL ES 3.1 applications. At MWC, Samsung launched their Galaxy S6, based on Android L and supporting OpenGL ES 3.1. Other existing devices in the market with support for the latest API are the Nexus10 tablet (It might need upgrading the OS to Android L from here) and the Galaxy Note 4 smartphone, upgrading the OS to Android L, too.

Most people use Mali Graphics Debugger (MGD) to help debug OpenGL® ES applications on Linux or Android. However graphics is not the only API that is supported by MGD, in-fact MGD supports applications that use the OpenCL™ API as well. This means that if you run MGD with an application that uses OpenCL you will get the same function level tracing that you would get with OpenGL ES. You will also get access to the kernel source code in much the same way you would get access to your OpenGL ES shader source code.

With the release 2.1 of Mali Graphics Debugger the OpenCL feature set has been improved by the inclusion of GPUVerify support. GPUVerify is a tool for formal analysis of kernels written in OpenCL and was partly funded by the EU FP7 CARP project http://carpproject.eu/. The tool can prove that kernels don’t suffer from the following three issues:

• Intra-group data races: This is when there is a data race between work items in the same work group.
• Inter-group data races: This is when there is a data race between work items in different work groups.
• Barrier divergence: This is when a kernel breaks the rules for barrier synchronization in conditional code defined in the OpenCL documentation.

The tool was created by Imperial College London and more information about the tool and the issues it can diagnose can be found by visiting http://multicore.doc.ic.ac.uk/GPUVerify.

In essence if you provide GPUVerify with the source code of a OpenCL kernel along with the local and global work group sizes it can check your kernel for the issues highlighted above. One of the main reasons for including support for this tool in MGD is that the information required to run GPUVerify is already known from tracing your application. The steps needed to use this tool with MGD are outlined below:

Step 1: Download Mali Graphics Debugger v2.1 and GPUVerify

Mali Graphics Debugger can be downloaded from http://malideveloper.arm.com/develop-for-mali/tools/software-tools/mali-graphics-debugger/. If you are using an older version of MGD you will need to upgrade. Version 2.1 has much more included than just GPUVerify support, a few items are listed below:

• Support for ARMv8 (Android 64-bit)
• Support for tracing Android Extension Pack
• Many improvements to the Frame Overrides feature.

GPUVerify can be downloaded in binary form for both Linux and Windows from the following location: http://multicore.doc.ic.ac.uk/tools/GPUVerify/download.php. GPUVerify does support Mac OS X as well but you will have to build it from source.

Step 2: Make sure GPUVerify is working stand-alone

Using GPUVerify successfully inside MGD is often easier once it is established that GPUVerify works in a stand-alone capacity. Once it has been downloaded the process of making it work is easy thanks to the documentation provided on the GPUVerify's website. Along with a section that provides common troubleshooting advice. A good way to check that it is working correctly is to try GPUVerify with some of the examples that are provided with the Mali OpenCL SDK.

The following image is of the output from GPUVerify running the HelloWorld example from the Mali OpenCL SDK:

Step 3: Making MGD Aware of the Location of GPUVerify

As GPUVerify is not shipped with MGD you must specify the location of where you installed it to MGD. To do this in MGD do the following:

• Click Edit -> Preferences

• Click on the Text box next to "Path to GPUVerify" and provide the location of the GPUVerify binary.

Step 4: Run a normal OpenCL trace in MGD

As mentioned previously MGD will try to fill in the prerequisite information for GPUVerify from the trace. It does this by looking for several functions; the most important are clCreateProgramWithSource, clCreateKernel and clEnqueueMapBuffer.

If you don't use clCreateKernel to create your kernels, MGD can also obtain the information from clCreateKernelsInProgram. As shown in the image above MGD also captures the build options used in clBuildProgram to pass to GPUVerify. The more details that can be passed to GPUVerify the more accurate it can analyze your kernels.

Step 5: Running GPUVerify from MGD

To run the tool you need to select Debug -> Launch GPU Verify. MGD will then present a new dialog box which summarizes all of the information MGD managed to pull out of the trace. You are free to fill this information in or even change the information. One of the reasons you may want to do this is to try new work group sizes or global work sizes, to see if there are any unforeseen issues with different kernel parameters.

Stage 6: Analyzing the Results

The results should be placed in the console view as part of MGD. Here are the results of the HelloWorld example from above running through MGD:

Summary

Mali Graphics Debugger can be used to do much more than debug and trace graphics applications. It can be used to debug and trace OpenCL applications as well. With the inclusion of GPUVerify support it is now possible to debug possible data race conditions in your kernels as well as barrier divergence issues. MGD will send as much data as it can to GPUVerify by analyzing the trace of your OpenCL application.

Since our successful demonstration at SIGGRAPH, ARM and Collabora have continued to work together on providing the best possible platform for media playback and presentation. Numerous applications such as digital signage, IVI, tablets, remote monitoring, and more, all require accurate, high-quality and low-power video presentation, with as low a thermal envelope as possible.

Collabora has made significant contributions to, and maintenance of, both the standard open-source GStreamer media framework, and the next-generation Wayland window system. Combining the two has allowed us to bring out the full extent of the capabilities of GStreamer, for years used in broadcast television with its exacting standards, and Wayland's lightweight and flexible design, which above all else emphasises accuracy and perfect end results.

The result is a system providing perfectly synchronised network video presentation. Three displays, all powered by separate ODROID-XU3 systems using the Samsung EXYNOS 5422 SoC with an ARM MALI-T620 GPU. Each system displays one segment of the video, with one acting as the co-ordinator to keep timing consistent across all three segments. From the user's point of view, the video appears as one consistent whole.

Network synchronisation with GStreamer

GStreamer is the reference open-source media framework, used in everything from audio playback on embedded systems, to huge farms powering broadcast TV. GStreamer's pipeline concept provides a flexible and lightweight transport to suit almost any usecase. In this particular instance, we are using GStreamer to load H.264 content from disk, pass it to a hardware H.264 decoder, feed the resulting frames to Wayland, and then feed the timing information from Wayland back to the master device.

Core to GStreamer's flexibility and applicability has been its excellent support for clock control, being able to synchronise multiple disparate sources. Its clock control supports both hardware and software sources and sinks and allows the most precise matching possible between input audio and video clocks, and the output device's actual capabilties.

GStreamer's measurement was then supplemented by an open-source distributed media control system called Aurena, which uses these measurement reports and targets from GStreamer to synchronise playback across all three devices.

The work we did with GStreamer to enhance its Wayland and H.264 hardware decoding support is both already merged to the upstream open-source project, and fully hardware-independent.

Accurate display with Wayland

The next-generation Wayland window system allows us to make the most efficient possible use of the hardware IP blocks, not only maximising throughput (thus increasing the highest achievable resolution, or number of streams, without sacrificing quality), but also providing predictable presentation.

Wayland's design goal of 'every frame is perfect' means that the content shown to the user must always be complete, coherent, and well-timed. The frame-based model employed is a significant stride over legacy X11 and DirectFB systems, and the consideration given to timing concerns allows us to make sure that the media is always delivered as close to on time as possible, without unsightly visual artifacts such as tearing.

Building on this solid and well-tested core, Collabora developed multiple extensions to Wayland. The first ensures that no copies of the video data are made in the compositing process, preserving precious memory bandwidth, latency, and overall system responsiveness. This extension uses the latest Khronos Group EGL extensions, as supported by ARM's MALI GPU.

However, even without this copy stage, as video usage continues to push at the margins of hardware performance – one recent customer project involved 4K output of nine 1080p H.264 streams on an embedded system – we realise that it might not be physically possible to obtain full frame-rate at all times. To compensate for this, Collabora developed an additional Wayland extension, allowing not only real-time feedback of actual hardware presentation time, but ahead-of-time frame queueing.

This feedback mechanism allows GStreamer to dynamically adjust its clock to obtain perfect synchronisation both across devices and between audio/video, whilst the ahead-of-time queueing gives the hardware the best possible chance to make frame deadlines, as well as preserving power by allowing the hardware to enter sleep states for longer.

This work is all either included with current Wayland releases, or actively being discussed and developed as part of the upstream open-source community.

Hardware enablement

Neither GStreamer nor Wayland required any hardware-specific development or tweaking. However, in order to make this work possible, Collabora has worked extensively on the kernel drivers for the Exynos 5422 SoC found inside the ODROID-XU3. Bringing the Exynos hardware support up to speed with the latest developments in the Kernel Modesetting and Video4Linux 2 subsystems, as well as fixing bugs found in our automated stress-testing laboratory, allowed this work to proceed without a hitch.

Far from being throwaway, this work is being merged into the upstream Linux kernel and U-Boot projects, as part of our ongoing commitment to working closely with the open source community to raise the bar for quality and functionality. As this and other platforms rapidly adopt these improvements and become able to run this work, device manufacturers are able to select from the greatest possible choice of vendors. Our work with our partners, including ARM, the wider open source community, and membership of the Khronos Group, continues to deliver benefits for the entire ecosystem, not just one platform or device.

This open standards-based approach allows platform selection to be driven by the true capabilities of the hardware and cost/logistics concerns, rather than having to fret about software capability and vendor lock-in.

Further development

Of course, the power of GStreamer, Wayland, and a standards-based Linux system does not just stop there: its extensibility includes support for OpenGL ES and EGL, as well as arbitrary client-defined content. Not only can applications like web browsers take advantage of this synchronisation in order to embed seamless media content in HTML5 displays, but the source data can be anything from a single OpenGL ES application, a web browser, or anything else imaginable. Whether providing for immersive gaming experiences or large-scale digital signage, the underlying technology is flexible and capable enough to deal with any needs.

ARM is a registered trademark of ARM Limited (or its subsidiaries) in the EU and/or elsewhere. All rights reserved.

It's now been a year since we announced the Shader Pixel Local Storage extension. Here I'll recap what we've done since the time of the release.

What is Pixel Local Storage?

I recommend reading Jan-Harald Fredriksen's blog post Pixel Local Storage on ARM® Mali™ GPUs for background information about what Pixel Local Storage is and the advantages of exposing it.

Order Indepdendent Transparency

At SIGGRAPH 2014 we presented "Efficient Rendering with Tile Local Storage", with detailed use-cases mixing advanced techniques such as deferred shading and order independent transparency. The problem with transparency is that blending operations tend to be non-commutative, meaning that the end result is highly sensitive to the shading order of the blended fragments. Using Pixel Local Storage we implemented both a full depth-peeling approach and also compared that against approximate approaches such as Multi-Layer Alpha Blending and Adaptive Range blending. Not only that; we implemented all of this very efficiently on-top of a fully deferred shading renderer. Please see Efficient Rendering with Tile Local Storagefor more details.

Collaboration with Epic

We integrated Pixel Local Storage into Epic's Unreal Engine 4. This enabled more efficient HDR rendering as well as features such as bloom and soft particles.

Sample code

We've also release a couple of samples showing how to use Pixel Local Storage in your own code.

Shader Pixel Local Storage

This sample implements deferred shading using pixel local storage.

http://malideveloper.arm.com/develop-for-mali/sample-code/shader-pixel-local-storage-sample/

Translucency

This sample uses Pixel Local Storage to render translucent geometry.

http://malideveloper.arm.com/downloads/deved/tutorial/SDK/android/2.0/translucency.html

References

1. Pixel Local Storage on ARM® Mali™ GPUs
2. Supporting the development of mobile games at GDC 2015
3. Efficient Rendering with Tile Local Storage

At GDC 2015, ARM and PlayCanvas unveiled the Seemore WebGL demo. If you haven’t seen it yet, it takes WebGL graphics to a whole new level.

CLICK HERE
TO LAUNCH SEEMORE

So why did we build this demo? We had two key goals:

Put amazing demo content in the hands of you, the developer

Seemore WebGL is the first conference demo that has been developed to run specifically in the web browser. This is great, because you can run it for yourself and do so on any device. Nothing to download and install - hit a link, and you’re immediately dropped into a stunning 3D experience. And better yet, you can learn from the demo and use that knowledge in your own creations.

Demonstrate console quality graphics on mobile

ARM Mali GPUs pack a serious graphical punch and Seemore is designed to fully demonstrate this power. We have taken advanced graphical features seen in the latest generation of console titles and optimized them to be completely mobile friendly. And best of all, all of this technology is open sourced on GitHub.

It's not practical to examine all of the engine updates we made to bring Seemore to life. So instead, let’s examine three of the more interesting engine features that were developed for the project.

Prefiltered Cubemaps

This is the generation and usage of prefiltered cubemaps. Each mipmap level stores environment reflection at different level of surface roughness - from mirror-like to diffuse.

How did we do it?
First, we added a cubemap filtering utility to the engine (GPU-based importance sampling). The next step was to expose this functionality in the PlayCanvas Editor. This technique uses Phong lobes of different sizes to pre-blur each mip level. Runtime shaders use either the EXT_shader_texture_lod extension (where supported) or reference mip levels stored as individual textures that are interpolated manually.

Show me the code!

https://github.com/playcanvas/engine/pull/202

Further reading:

http://http.developer.nvidia.com/GPUGems3/gpugems3_ch20.html
https://seblagarde.wordpress.com/2012/06/10/amd-cubemapgen-for-physically-based-rendering/

Box-projected cubemaps

This feature makes cubemaps work as if projected onto the insides of a box, instead of being infinitely far away (as with a regular skybox cubemap). This technique is widely used in games for interior reflection and refraction.

How did we do it?

This effect is implemented using a world-space AABB projection. Refraction uses the same code as reflection but with a different ray direction, so the projection automatically applies to it as well.

Show me the code!

https://github.com/playcanvas/engine/pull/183

Further reading:

http://www.gamedev.net/topic/568829-box-projected-cubemap-environment-mapping/

Custom shader chunks

Standard material shaders in PlayCanvas are assembled from multiple code 'chunks'. Often, you don't want to replace the whole shader, but you'd like to only change some parts of it, like adding some procedural ambient occlusion or changing the way a surface reflects light.

This feature was required in Seemore to achieve the following:

• Dual baked ambient occlusion. The main plant uses 2 AO maps for open and closed mouth states which are interpolated dynamically.

• Fake foliage translucency. This attenuates emission to make it appear as though light is scattered on the back-faces of leaves in a hemispherically lit room. The plant’s head uses a more complex version of the effect, calculating per-vertex procedural light occlusion.

• Plant/tentacle animation. Procedural code that drives vertex positions/normals/tangents.

How did we do it?

Shader chunks are stored in the engine sourcebase as .vert and .frag files that contain snippets of GLSL. You can find all of these files here. Here’s an example chunk that applies exponential squared fog to a fragment:

uniform vec3 fog_color;
uniform float fog_density;

vec3 addFog(inout psInternalData data, vec3 color)
{
    float depth = gl_FragCoord.z / gl_FragCoord.w;    float fogFactor = exp(-depth * depth * fog_density * fog_density);    fogFactor = clamp(fogFactor, 0.0, 1.0);    return mix(fog_color, color, fogFactor);
}

Each chunk file’s name becomes its name at runtime, with PS or VS appended, depending on whether the chunk forms part of a vertex or pixel shader. In the case above, the filename is fogExp2.frag. It’s a simple matter to replace this fragment on a material. Simply do:

  material.chunks.fogExp2PS = myCustomShaderString;

Show me the code!

https://github.com/playcanvas/engine/pull/172

So there you have it. A brief insight into some of the latest technology in the PlayCanvas Engine. Want to find out more? Head over to GitHub, watch, star and fork the codebase - get involved today!

The Embedded Vision Summit 2015 is nearly upon us.  This annual gathering of experts interested in this highly dynamic area is a day of fascinating presentations and demonstrates of leading-edge developments in vision-enabled products.  The Summit is on 12 May.

This year ARM® is hosting a special half-day seminar connected with the Summit on the day before.  Titled “Enabling Computer Vision on ARM” the event will see a number of industry-leading developers presenting their experiences in computer vision working over a variety of ARM platforms and use cases.

Growth in Computer Vision

Computer vision is seeing phenomenal growth in adoption and deployment.  The increased power, efficiency and variety of processors is enabling many new use cases while revolutionizing existing vision applications previously confined to the desktop.  These are now increasingly possible on energy efficient mobile devices and across market segments from automotive, retail, medical and industry.

In this seminar, a selection of computer vision experts and leaders in their fields will present their experiences working with ARM-based systems across a variety of real use cases.  Attendees will learn how to:

• Resolve common issues encountered when implementing complex vision algorithms on embedded processors in areas such as object recognition and augmented reality
• Balance workloads across processors and processor types
• Debug heterogeneous vision applications on ARM-based systems to remove design bottlenecks and improve performance and efficiency

Seminar program:

• Jeff Bier, President, BDTI
  Title: Benchmarking Metrics and Processor Selection for Computer Vision
  Abstract: This presentation looks at the long term trends in computer vision applications and processors and the challenges these pose to benchmarking vision applications.  As the complexity of applications, processors and the heterogeneous design of systems increases, so do the challenges in measuring their performance in meaningful ways.  Being able to assess the relative performance of processors and processor types under various combinations and configurations is a vital factor in matching systems to particular use cases.  For example, mobile use cases are becoming the a key focus for software development and these systems increasingly rely on heterogeneous configurations to increase processing efficiency.  To use these processors efficiently, developers must determine the optical mapping of their applications onto the SoC’s heterogeneous processing cores.

• Dr. Masaki Satoh, Morpho Inc
  Title: Development of Image and Vision Processing Software and Optimizations for ARM
  Abstract:This presentation will give technical insight on the benefit of NEON Acceleration, including detailing actual performance improvement.  From the developer perspective, the presentation will examine specific algorithms and how they are optimized with NEON™.  The future of imaging will also be examined, looking at the potential of GPU compute and other heterogeneous combinations, deep learning image recognition engines accelerated through NEON and OpenCL, and research and development into automotive products.
  
  
• Dr. Piotr Stec, Project Manager in the Imaging Field, FotoNation
  Title: Video Image Stabilization for Mobile Devices
  Abstract:The presentation will show the processing steps needed to perform video stabilization on mobile devices. We will show the algorithm flow indicating the steps that need to be taken to perform the algorithm and the data flow between various components of the algorithm. Some parts of the algorithm proven to be particularly challenging in terms of achieving suitable performance. Not always the most complex parts turned out to be the bottleneck. We will show how those difficulties were overcome on the device using ARM chipset and what gains were achieved in terms of processing time.  The last part of the presentation will be a live demonstration of the working algorithm.
  
  
• Gian Marco Iodice, Compute Engineer, ARM
  Title: Real-time Dense Passive Stereo Vision: A Case Study in Optimizing Computer Vision Applications Using OpenCL on ARM
  Abstract: Passive stereo vision is a powerful visual sensing technique aimed at inferring depth without using any structured light. Nowadays, as it offers low cost and reliability solutions, it finds application in many real use cases, such as natural user interfaces, industrial automation, autonomous vehicles, and many more. Since stereo vision algorithms are extremely computationally expensive, resulting in very high CPU load, the aim of this presentation is to demonstrate the feasibility of this task on a low power mobile ARM Mali GPU. In particular, the presentation will focus on a local stereo vision method based on a novel extension of census transform, which exploits the highlyparallel execution feature of mobile Graphic Processing Units with OpenCL.  The presentation will show also the approaches and the strategies used to optimize the OpenCL code in order to reach significant performance benefits on the GPU.
  
  
• Martin Lechner, CTO and Markus Eder, Head of Computer Vision, Wikitude
  Title: Utilizing NEON for Accelerated Computer Vision Processing in Augmented Reality Scenarios
  Abstract:In the core of the Wikitude SDK runs an engine that heavily relies on different computer vision algorithms to get information about the current environment of the user. As those algorithms can be very computationally intensive, a major part in our work is to optimize and specifically design the algorithms for the architecture on mobile devices.  As most of the current mobile phones have either armv7 or armv8 architectures the ARM NEON SIMD-instruction set provides a huge possibility for improving the performance of computer vision algorithms. This presentation will focus on how the NEON instruction set can be used to improve the performance of general image processing functions. It will also include a discussion of our experience with the NEON instruction set, specifically the process on how to find the hotspots in the code and how those functions can be tested and debugged as well as our experience with porting the NEON functionality from armv7 to armv8.
  
  
• Tim Hartley, Technical Marketing Manager, ARM
  Title:Measuring the Whole System: Holistic Profiling of CPU and GPU for Optimal Vision Applications on ARM Platforms
  Abstract:Developers of sophisticated vision applications need all the processing power they can lay their hands on, and using OpenCL on a GPU can be a vital additional compute resource.  But spreading the workload amongst processors and processor types brings its own problems and difficulties, and traditional application optimization techniques are not always effective in this brave new heterogeneous world.  The key to achieving performance is twofold: getting access to hardware counters for all the processors in your system, and then understanding what those numbers are telling you.  In this talk, I will examine the tools and techniques available to profile these sorts of applications and will use real case studies from vision applications. Using tools like DS5 Streamline I will show how to extract meaningful performance numbers and how to interpret them.
  
  
• Ken Lee, Founder and CEO, Van Gogh Imaging
  Title:Using ARM Processors to Implement Real-Time 3D Object Recognition on Mobile Devices
  Abstract:Diverse applications such as 3D printing, augmented reality, medical, parts inspections, and ecommerce can benefit significantly from the ability of 3D computer vision to separate a scene into discrete objects and then recognize and analyze them reliably. The 3D approach is much more robust and accurate than the traditional 2D approach and is now possible with embedded 3D sensors and powerful processors in mobile devices.  This discussion will focus on how real-time 3D computer vision can now be implemented in the ARM CPU.  Further, we will discuss how these algorithms can be further accelerated using ARM Mali GPU with OPENCL implementation.

The seminar will include an industry panel discussion. Experts from the worlds of vision IP, ADAS (Automatic Driver Assistance Systems), and image sensors and recognition software will discuss future trends in technology for ARM-based systems.

To register for free for the ARM Seminar:
http://www.eventbrite.com/e/seminar-enabling-computer-vision-on-arm-tickets-10231417445?aff=EVS

There’s more about the Embedded Vision Summit here:
http://www.embedded-vision.com/summit

For a long time, I've not really been interested in 3D programming.

(Well, I've done some minor OpenGL programming many years ago, but I must admit that I'm more of a dev-tools programmer)

After watching the video Alban linked to in ARM Processor - Sowing the Seeds of Success - Computerphile, I decided to watch some more from Computerphile.

The videos by John Chapman are spoken in a very clear and very easy to understand english, and they explain advanced technology in a way so it's easy to follow.

If you're new to 3D programming, this might be a good starting point.

A Universe of Triangles

True Power of the Matrix

Triangles to Pixels

Visibility Problem

Lights and Shadows in Graphics

Originally posted on ASTC Evaluation Codec now on GitHub - Mali Developer Center

The evaluation codec for Adaptive Scalable Texture Compression (ASTC) technology is now available on GitHub, making it easier for you to access, browse and contribute feedback. Previously the ASTC codec was released as a ZIP archive on malideveloper.arm.com, including the source code and example binaries, however we wanted to use GitHub to bring ASTC to the heart of the developer open-source community.

What is ASTC technology?

ASTC technology developed by ARM and AMD is an official extension to both the OpenGL® and OpenGL® ES graphics APIs. ASTC is a major step forward in terms of image quality, reducing memory bandwidth and thus energy use. It is the outcome of many years of research, development and engineering, and it is now implemented in a number of GPUs across the whole industry.

ASTC is widely supported by all major hardware vendors and it is free to use. Google’s Android Extension Pack (GL_ANDROID_extension_pack_es31a) also requires support for ASTC. If you are a game developer, Unreal Engine 4 and Unity 4.3 already support ASTC and for those of you building your own game engine, you can clone the GitHub repository to start using ASTC yourself.

A number of cutting-edge developer tools also support ASTC such as:

• ARM® Mali GPU Texture Compression Tool– can compress and decompress textures to ASTC, show the preview and let you experiment with different parameters and block sizes.
• OpenGL ES Emulator– an OpenGL ES library for desktop PCs lets you experiment with ASTC textures even where they are not supported in hardware.
• Mali Graphics Debugger– can trace, capture and visualize ASTC textures used by your application.

For the technical details on ASTC, please refer to the blog posts by Tom Olson and Sean Ellis.

Example of ASTC used to encode normal maps

Why is this now on GitHub?

With the source code available on GitHub, it is even easier for every developer to access, clone, browse and contribute feedback and improvements. Since ASTC is a standard, the whole community benefits from it, and there is now a straightforward way to share fixes and new features.

Availability and support

As always, tools provided by ARM are supported in the ARM Connected Community. You can ask a question in the Mali Developer Forum, follow us on Twitter, Sina Weibo, or watch our ARM YouTube and ARM YouKu channels.

My last blog looked at some of the critical areas which an application has to implement efficiently to get best performance out of 3D content, such as broad-brush culling of large sections of a scene which are guaranteed not to be visible so they are not sent to the GPU at all. In one of the follow on comments to this blog seanlumly01 asked "Is there a performance penalty for an application modifying textures between draw calls?". It is a really good question, but the answer is non-trivial, so I deferred to this blog post to answer it fully.

Pipelined Rendering

The most important thing to remember when it comes to resource management is the fact that OpenGL ES implementations are nearly all heavily pipelined.  This is discussed in more detail in this earlier blog, but in summary ...

When you call glDraw...() to draw something the draw does not happen instantly, instead the command which tells the GPU how to perform that draw is added to a queue of operations to be performed at some point in future. Similarly, eglSwapBuffers() does not actually swap the front and back buffer of the screen, but really just tells the graphics stack that the application has finished composing a frame of rendering and queues that frame for rendering. In both cases the logical specification of the behaviour  - the API calls - and the actual processing of the work on the GPU are decoupled by a buffering process which can be tens of milliseconds in length.

Resource Dependencies

For the most part, OpenGL ES defines a synchronous programming model. Apart from a few explicit exceptions, when you make a draw call rendering must appear to have happened at the point that the draw call was made, with pixels on screen correctly reflecting the state of any command flags, textures, or buffers at that point in time (based either on API function calls or previously specified GPU commands). This appearance of synchronous rendering is an elaborate illusion maintained by the driver stack underneath the API, which works well but does place some constraints on the application behavior if you want to achieve the best performance and lowest CPU overheads.

Due to the pipelining process outlined earlier, enforcing this illusion of synchronicity means that a pending draw call which reads a texture or buffer effectively places a modification lock on that resource until that draw operation has actually completed rendering on the GPU.

For example, if we had a code sequence:

glBindTexture(1)       // Bind texture 1, version 1
glDrawElements(...)    // Draw reading texture 1, version 1
glTexSubImage2D(...)   // Modify texture 1, so it becomes version 2
glDrawElements(...)    // Draw reading the texture 1, version 2

... then we cannot allow the glTexSubImage2D() to modify the texture memory until the first draw call has actually been processed by the GPU, otherwise the rendering of the first draw call will not correctly reflect the state of the GL at the point the API call was made (we need it to render the draw using the contents of the physical memory which reflect texture version 1, not version 2). A lot of what OpenGL ES drivers spend their time doing is tracking resource dependencies such as this one to make sure that the synchronous programming "illusion" is maintained, ensuring that operations do not happen too early (before the resources are available) or too late (after a later resource modification has been made).

Breaking Resource Dependencies

In scenarios where a resource dependency conflict occurs - for example a buffer write is requested when that buffer still has a pending read lock - the Mali drivers cannot apply the resource modification immediately without some special handling; here are multiple possible routes open to the drivers to resolve the conflict automatically.

Pipeline Finish

We could drain the rendering pipeline to the point where all pending reads and writes from the GPU for the conflicted resource are resolved. After the finish has completed we can process the modification of the resource as normal. If this happens part way through the drawing of a framebuffer you will incur incremental rendering costs where we are forced to flush the intermediate render state to main memory; see this blog for more details.

Draining the pipeline completely means that the GPU will then go idle waiting for the CPU to build the next workload, which is a poor use of hardware cycles, so this tends to be a poor solution in practice.

Resource Ghosting

We can maintain both the illusion of the synchronous programming model and process the application update immediately, if we are willing to spend a bit more memory. Rather than modifying the physical contents of the current resource memory, we can simply create a new version of the logical texture resource, assembling the new version from both the application update and any of the data from the original buffer (if the modification is only a partial buffer or texture replacement). The latest version of the resource is used for any operations at the API level, older versions are only needed until their pending rendering operations are resolved, at which point their memory can be freed. This approach is known as resource ghosting, or copy-on-write.

This is the most common approach taken by drivers as it leaves the pipeline intact and ensures that the GPU hardware stays busy. The downsides of this approach are additional memory footprint while the ghost resources are alive, and some additional processing load to allocate and assemble the new resource versions in memory.

It should also be noted that resource ghosting isn't always possible; in particular when resources are imported from external sources using a memory sharing API such as UMP, Gralloc, dma_buf, etc. In these cases other drivers, such as cameras, video decoders, and image processors may be writing into these buffers and the Mali drivers have no way to know whether this is happening or not. In these cases we generally cannot apply copy-on-write mechanisms, so the driver tends to block and wait for pending dependencies to resolve. For most applications you don't have to worry about this, but if you are working with buffers sourced from other media accelerators this is one to watch out for.

Application Overrides

Given that resource dependencies are a problem on all hardware rendering systems due to pipeline depth, it should come as no surprise that more recent versions of OpenGL ES come with some features which allow application developers to override the purely synchronous rendering illusion to get more fine control if it is needed.

The function glMapBufferRange() function in OpenGL ES 3.0  allows application developers to map a buffer into the application's CPU address space. Mapping buffers allows the application to specify an access flag of GL_MAP_UNSYNCHRONIZED_BIT, which loosely translates as the "don't worry about resource dependencies, I know what I am doing" bit. When a buffer mapping is unsynchronized the driver does not attempt to enforce the synchronous rendering illusion, and the application can modify areas of the buffer which are still referenced by pending rendering operations and therefore cause incorrect rendering for those operations if the buffer updates are made erroneously.

Working With Resource Dependencies

In addition to the direct use of features such as GL_MAP_UNSYCHRONIZED_BIT, many applications work with the knowledge that the resource usage is pipelined to create flexible rendering without causing excessive ghosting overheads.

Separate Out Volatile Resources

Ghosting can be made less expensive by ensuring that volatile resources are separated out from the static resources, making the memory regions which need to be allocated and copied as small as possible. For example, ensuring that animated glyphs which are updated using glTexSubImage2D() are not sharing a texture atlas with static images which are never changed, or ensuring that models which are animated in software on the CPU (either via attribute or index update) are not in the same buffer as static models.

Batch Updates

The overheads related to buffer updates can be reduced, and the number of ghosted copies minimized, by performing most of the resource updates in a single block (either one large update or multiple  sequential sub-buffer/texture updates), ideally before any rendering to a FBO has occurred. Avoid interleaving resource updates with draw calls like this ...

glBufferSubData
glDraw
glBufferSubData
glDraw
glBufferSubData
glDraw

... unless you are able to use GL_MAP_UNSYNCHORNIZED_BIT. It is usually much more efficient to make the same set of updates like this:

glBufferSubData
glBufferSubData
glBufferSubData
glDraw
glDraw
glDraw

Application Pipelined Resources

If the application wants to make performance more predictable and avoid the overheads of ghosting reallocating memory in the driver, one technique it can apply is to explicitly create multiple copies of each volatile resource in the application, one for each frame of latency present in the rendering pipeline (typically 3 for a system such as Android). The resources are used in a round-robin sequence, so when the next modification of a resource occurs the pending rendering using that resource should have completed. This means that the application modifications can be committed directly to physical memory without needing special handling in the driver.

There is no easy way to determine the pipeline length of an application, but it can be empirically tested on a device by inserting a fence object by calling glFenceSync() after a draw call using a texture, and then polling that fence object by calling glClientWaitSync() with a timeout of zero just before making the modifications N frames later. If this wait returns GL_TIMEOUT_EXPIRED then the rendering is still pending and you need to add an additional resource version to the resource pool you are using.

Thanks to Sean for the good question, and I hope this answers it!

Cheers,
Pete

Pete Harris is the lead performance engineer for the Mali OpenGL ES driver team at ARM. He enjoys spending his time working on a whiteboard and determining how to get the best out of combined hardware and software compute sub-systems. He spends his working days thinking about how to make the ARM Mali GPUs even better.

Introduction

Have you ever heard about the Taoyuan effect in reflections? Probably not, so I invite you to read this blog to learn about this effect and the application of local cubemaps in graphics.

Several blogs published in this community cover different uses of local cubemaps. The first was about reflections  followed by two more blogs that describe novel uses of local cubemaps to render shadows  and refractions  with great quality and performance.

These new techniques, completely developed by the ecosystem demo team in ARM’s Media Processing Group, are especially relevant to developing for mobile devices where runtime resources must be carefully balanced. They not offer only great performance but also high-quality rendering, which makes these techniques appealing to desktop computers and consoles’ developers as well. Fetching a texture from a static cubemap guarantees very precise and stable shadows, reflections and refractions, compared to the pixel instabilities we would get when rendering these effects at runtime with a moving camera using conventional techniques.

Thanks to these different uses of the local cubemap techniques, we can achieve very high quality graphics with existing hardware.

Unity Unite events in Asia

Recently Sylwester Bala and I attended Unity Unite events in Seoul, Beijing and Taipei where we presented the talk “Enhancing Your Unity Mobile Games”. I delivered the talk at Seoul, Nathan Li at Beijing and Sylwester at Taipei. In Seoul and Beijing there were joint talks with Carl Callewaert from Unity (see here).

In the talk we introduced the concept of local cubemap and how it can be used to render reflections, and we expanded this technique for rendering shadows in an innovative way. During the talk we used several short videos to illustrate the concept and the advantages of these techniques, which helped to deliver a clear and understandable message to the audience. A couple of videos were particularly useful to show how our new shadows technique can render dynamic and soft shadows.

In the final part of the talk Carl gave a live demo about some of the most important improvements in Unity 5: Global Illumination, Physically-Based Shading and Reflection Probes. Unity’s implementation of Reflection Probes is based on the local cubemap rendering technique. Now Unity developers have access to a simpler but powerful technique to add reflections in their games in a further optimized way, which is particularly relevant for mobile platforms. Carl also mentioned that Unity might consider implementing the new shadows technique in the engine as it is tightly close to the reflection probe feature.

In all cities, the talks were received well with great expectation and had a high number of attendees, reflecting the interest in these techniques. Carl showed a live Unity demo with a great example on how reflections based on local cubemaps work. His example clearly demonstrated the advantages of this technique and how easy it can be used in the Unity engine.

Top left: Roberto Lopez Mendez at Seoul. Top right: Nathan Li at Beijing. Bottom Left: Carl Callewaert at Beijing. Bottom right: Sylwester Bala at Taipei.

Reflections and shadows in games

During a long night walk in Seoul, Sylwester Bala and I talked about how reflections and shadows can contribute to improving the visual experience of games. We walked along Yeongdong Avenue and there were high buildings full of neon and glass. Reflections and shadows were practically in every surface. Modern cities are full of polished, high-reflective glass and metal surfaces, virtually everywhere.

Therefore, when rendering this kind of environment in games, we need to consider the fact that the local cubemap technique might offer a very efficient way of rendering reflections and shadows in combination with other rendering techniques.

Figure 1. Reflections on the facade of the Coex building at Seoul.

It is well known that reflections and shadows are key topics in any game. Without reflections and shadows any virtual world would look plain and unrealistic. Let’s take for example reflections. Reflections change every time the camera updates its position and orientation. We could prebake in a texture the reflections from the static environment but the final effect will be disappointing. An effective way of rendering high quality and optimized reflections is using local cubemaps. Using this technique we prebake the environment into static cubemaps and later at runtime we render the reflections by retrieving the texture from the cubemaps, using the local corrected view vector and then combining the contribution of local cubemaps for a certain position.

And what do you do with dynamic geometry? We obviously can’t prebake the reflections from dynamic geometry. In this case, we can render at runtime reflections/shadows from dynamic geometry using traditional techniques and combine them with reflections/shadows rendered using the local cubemap technique.

Unfortunately, the use of local cubemaps in reflections is not yet widely implemented, despite the fact that it has been available for 15 years. Now with the implementation of the Reflection Probe in Unity 5 the technique of reflections based on local cubemap is becoming available to more than the half of developers all over the world, which is pretty good. To provide support for our shadows technique based on local cubemaps in the Unity engine, it would be as simple as rendering the transparency of the environment into the alpha channel of the same cubemap used for reflections.

The Taoyuan effect

At the very end of our journey when heading to the flight gate at the Taiwan Taoyuan International Airport in Taipei, Sylwester and I were walking through a long corridor with a very polished and reflective floor. On one side the wall was projecting some message with Chinese symbols, which were perfectly reflected on the floor. The picture drew my attention and I pointed out to Sylwester that was a clear use case of reflections based on local cubemaps.

Nevertheless, Sylwester, who always pays extra attention to details, pointed out that further away from the wall reflections were more blurred. We looked at each other because it was clear that this effect can be implemented in reflections in the same way we implemented soft shadows.

Figure 2. Reflections on the corridor at the Taiwan Taoyuan International Airport.

In the shadows technique, we implemented the blurred effect by fetching the texture from an interpolated cubemap mipmap level. The magnitude passed to the fetching function is proportional to the distance from the fragment to the intersection point of the fragment-to-light vector with the scene-bounding box (a more detailed explanation of the shadows technique is in this blog).

As it happens with reflections based on local cubemap, our technique offers very clear advantages in terms of quality and resource saving when rendering shadows. Regardless of its limitations -mainly derived from its static nature- I would like to encourage developers to try the technique and explore it beyond the use case we presented in our talk. You never know what a technique can do for you until you start to use it and push it to new extremes. If you are skeptic about that then please continue reading.

The effect was caused by the fact that the Chinese symbols were carved in to the wall and the light source was behind in such a way that the light scattered in the borders of the symbols provoking a soft pattern of light beams.

I decided to implement this effect in the same demo we used to show reflections and shadows: the chess room demo. The results are below. Fig. 3 shows the reflections on the chessboard using the standard technique of local cubemap. Fig. 4 shows a blurred reflection based on the distance from the fragment to the intersection point of the reflection vector with the scene-bounding box. The further the reflection is from the real object, the more blurred it is rendered.

Figure 3. Standard reflections on the chess board based on local cubemap.

Figure 4. Reflections on the chessboard based on local cubemap using an interpolated mipmap level based on the distance from the fragment to the intersection point of the fragment-to-light vector with the scene-bounding box.

This example shows how powerful and flexible the technique of local cubemaps is. It allows rendering of not only soft shadows but also the reflections coming from soft light beams.

Conclusion

What other new applications of local cubemaps will come? As local cubemaps are becoming more popular, developers will find new applications for this technique - it is just a matter of time. But one thing is certain; this simple technique has proved to be a very effective and powerful way to render reflections, shadows and refractions.

Hey everybody,

ARM have arrived and the talks have started! We're eagerly awaiting the Game Jam kicking off at 8pm CEST. See all the details on the Mali Developer Center

We have with us a load of Samsung GearVR kits for developers to work on and optimise for mobile as well as our help and advice available all weekend! The GearVR kits work with the Samsung Galaxy Note 4 which has the latest Mali-T760 GPU inside. Full spec below:

Samsung Galaxy Note 4

•Samsung Exynos 7 Octa

•ARM Mali-T760 GPU (MP6)

•ARM big.LITTLE™ processing technology

•ARM Cortex®-A53 CPU (MP4)

  ARM Cortex-A57 CPU (MP4)

We're excited to see all the new innovations in the VR space and what the developers can put together in just a weekend.

You can see the full agenda for the event at the Shayla Games website

Machine Vision is perhaps one of the few remaining areas in technology that can still lead you to say “I didn’t know computers could do that”.  The recent pace of development has been relentless.  On the one hand you have the industry giants competing to out-do each other on the ImageNet challenge, surpassing human vision recognition capabilities along the way, and on the other there is significant, relentless progress in bringing this technology to smart, mobile devices.  May 11 and 12 saw the annual Santa Clara California gathering of industry experts and leaders to discuss latest developments at the Embedded Vision Alliance Summit.  Additionally, this year ARM was proud to host a special seminar linked to the main event to discuss developments in computer vision on ARM processor technologies.  In this blog I’m going to provide my perspective of some of the highlights from both events.

The Santa Clara Convention Centre, California.  Host to both the ARM workshop and the EVA Summit

Computer Vision on ARM Seminar, 11 May, Santa Clara, CA

It was my great pleasure to host this event and for those of you who were there I hope you enjoyed the afternoon’s presentations and panel discussion.  The proceedings from the seven partner presentations can all be downloaded from here.  The idea of this event – the first of its kind ARM has held on computer vision – was intended to bring together leaders and experts in computer vision from across the ARM ecosystem.  The brief was to explore the subjects of processor selection, optimisation, balancing workloads across processors, debugging and more, all in the context of developing computer vision applications.  This covered both CPU and NEON™ optimisations, as well as working with Mali™ GPUs.

With a certain degree of cross-over, the seminar program was divided into three broad themes:

Optimising Computer Vision for NEON

• Dr. Masaki Satoh, a research engineer from Morpho, talked about the benefits and main technical aspects of NEON acceleration, including a focus on specific algorithmic optimisations using NEON SIMD (Single Instruction, Multiple Data) instructions.
• Wikitude are a leader in Augmented Reality applications on mobile devices and in his talk CTO Martin Lechner highlighted the use of NEON to accelerate this particular computer vision use case.

Real Computer Vision Use Cases on ARM

• Ken Lee, founder and CEO of Van Gogh Imaging, showcased their work developing real-time 3D object recognition applications using 3D stereoscopic sensors, including optimisation via NEON and their early exploration of further acceleration via Mali GPUs.
• Gian Marco Iodice, Compute Engineer at ARM, discussed his work on accelerating a real-time dense passive stereo vision algorithm using OpenCL™ on ARM Mali GPUs.
• Real-time image stabilization running entirely in software was the subject of the presentation by Dr. Piotr Stec, Project Manager at FotoNation.  His analysis covered the complete processing pipeline for this challenging use case and discussed where optimisations were most effective.

Processor selection, Benchmarking and Optimising

• Jeff Bier, president of BDTIand founder of the Embedded Vision Alliance discussed the important area of processor selection and making intelligent choices when selecting benchmarking metrics for computer vision applications.
• Tim Hartley (that’s me!) discussed the importance of whole-system measurements when developing computer vision applications and demonstrated profiling techniques that can be applied across heterogeneous CPU and GPU processor combinations.

Jeff Bier from BDTI gets things going with his presentation about processor selection and benchmark criteria

Panel Discussion

In addition to the above presentationsRoberto Mijathosted a panel discussion looking at current and future trends in computer vision on mobile and embedded platforms.  The panel included the following industry experts:

It was a fascinating and wide-ranging discussion with some great audience questions. Roberto asked the panelists what had stood out for them with computer vision developments to date.  Lazlo talked about the increasing importance of intelligence embedded in small chips within cameras themselves.  Michael Tusch echoed this, highlighting the problem of high quality video in IP cameras causing saturation over networks.  Having analysis embedded within the cameras and then only uploading selective portions, or even metadata describing the scene, would mitigate this significantly.  Tim Droz stressed the importance of the industry moving away from the pixel count race and concentrating instead on sensor quality.

Roberto then asked about the panelist’s view on the most compelling future trends in the industry.  Michael Tusch discussed the importance in the smart home and businesses of the future of being able to distinguish and identify multiple people within a scene, in different poses and sizes, and being able to determine trajectories of objects.  This will need flexible vision processing abstractions with the the aim of understanding the target you are trying to identify: you cannot assume one size or algorithm will fit all cases.  Michael forsees, just as GPUs do for graphics, the advent of engines cable of enabling this flexible level of abstraction for computer vision applications.

Lazlo Kishonti talked about future health care automation including sensor integration in hospitals and the home, how artificial intelligence in computer vision for security is going to become more important and how vision is going to enable the future of autonomous driving.  Lazlo also described the need for what he sees as the 3rd generation of computer vision algorithms.  These will require levels of sophistication that will reach, for example, the ability of differentiating between a small child walking safely hand-in-hand with an adult with one where there might be a risk of running out into the road.  This kind of complex mix of recognition and semantic scene analysis was, said Lazlo, vital before fully autonomous vehicles can be realized.  It brought home to me both the importance of ongoing research in this area and perhaps how much further computer vision technology has to develop as a technology.

Tim Droz talked about the development of new vector processors flexible enough for a variety of inputs, HDR – high dynamic range, combining multiple images from different exposures – becoming ubiquitous, along with low-level OpenCL implementations in RTL.  He also talked about Plenoptic, light-field cameras that allow re-focusing after an image is taken, becoming much smaller and more efficient in the future.

The panel ended with a lively set of questions from the audience, wrapping up a fascinating discussion.

Gian Marco Iodice talks about accelerating a real-time dense passive stereo vision algorithm

Overall it was a real pleasure to see so many attendees so engaged with the afternoon and we are grateful to all of you who joined us on the day.  Thanks also to all our partners and panelists whose efforts led to a fascinating set of presentations and discussions.

The presentations from the seminar can be downloaded here: EVA Summit 2015 and ARM’s Computer Vision Seminar - Mali Developer Center

Embedded Vision Summit, 12 May, Santa Clara, CA

The annual Embedded Vision Summit is the industry event hosted by the Embedded Vision Alliance, a collection of around 50 companies working in the computer vision field.  Compared to the 2014 event, this year saw the Summit grow by over 70%, a real reflection of the growing momentum and importance in embedded vision across all industries.  Over 700 attendees had access to 26 presentations on a wide range of computer vision subjects arranged into 6 conference tracks.  The exhibition area show-cased the latest work from 34 companies.

See below for links to more information about the proceedings and for downloading the presentations.

Dr. Ren Wu, Distinguished Scientist from Baidu delivered the first of two keynotes, exploring what is probably the hottest topic of the hour: visual intelligence through deep learning.  Dr. Wu has pioneered work in this area, from training supercomputers through to deployment on mobile and Internet of Things devices.  And for robot vacuum cleaner fans – and that’s all of you surely – the afternoon keynote was from Dr. Mike Aldred from Dyson who talked about the development of their 360° vision (and ARM!) enabled device which had earlier entertained everyone as it trundled around the exhibition area, clearing crumbs thrown at it by grown men and women during lunch.

ARM showcased two new partner demos at the Summit exhibition: SLAMBench acceleration on Mali GPU by the PAMELA consortium and video image stabilization in software with Mali acceleration by FotoNation

The six conference tracks covered a wide range of subject areas.  Following on from Ren Wu’s keynote, Deep Learning and CNNs (Convolutional Neural Networks) made a notable mark with its own track this year.  And there were tracks covering vision libraries, vision algorithm development, 3D vision, business and markets, and processor selection.  In this final track, Roberto Mijat followed on from ARM’s previous day’s seminar with an examination of the role of GPUs in accelerating vision applications.

Roberto Mijat discusses the role of the integrated GPU in mobile computer vision applications

A list of all the speakers at this year's Summit can be found here: 2015 Embedded Vision Summit Speakers

All the papers from the event can be downloaded here (registration required): 2015 Embedded Vision Summit Replay

Chinese version中文版：百视通/ARM HTML5技术论坛顺利举行（附资料下载）

Dear All, good day. HTML5 is really one of the hottest word now in Internet technology area. I believe most of you heard it before. These years HTML5 commercial promotion have a number of ups and downs as a result that it is said that every year is the Year One of HTML5. But with the HTML5 specification fixed, 2015 would be the hottest year of HTML5. According to statistics, there are over 50 HTML5 related meetings in 2015 in China. News keep coming about that HTML5 companies launched IPO and HTML5 startups  won billion level valuations.

ARM as the Global lead IP technology company, we will never be absent from the HTML5 boom. Yesterday, BesTV/ARM HTML5 technical forum was held in BesTV New Media Institute (Shanghai) besTV This forum is run by besTV and ARM, inviting upstream and downstream guests from SoC ,ODM/OEM, ISP, solution vendors, channel and developers to discuss the HTML5 technology status and future.

Dr. Wen Li , senior VP of besTV and vice dean of BesTV New Media Institute gave opening speech and technical speech HTML5 will be main technical choice of smart TV and home entertainment area. He analyzed the advantage of HTML5 and introduced their excellent work in China government-sponsered core technology project and China SARFT leading TVOS project.

Matt Spencer the UI and browser marketing manager from ARM UK HQ, gave the speech HTML5 newest trends and the latest progress of global HTML5 technologies and specification information, as well as the positive role of ARM and our strategy.

The Top 2 HTML5 game engines in China came on the stage one after the other, Jianyun Bao , senior engineer of developer service department of Cocos2d-x Community , the giant mobile game company in China gave the speech Cocos HTML5 solution and introduced the total solution support to HTML5 games in open source engine, tool chain, runtime, channel access, test/analysis and payment system. And next to him, the technical evangelist Xinlei Zhang , who is from the fast rising HTML5 game engine – Egret, presented the audience their total solution support to HTML5 game in open source engine , tool chain, running platform , channel access, test/analysis and payment system too. He also introduced the new launched mobile application framework Lark.

Oversea HTML5 game engine also came to our forum. Joel Liang, senior ecosystem engineer from ARM gave the speech PlayCanvas: 3D game engine base on WebGL. He showed the advanced GPU algorithms based on WebGL in PlayCanvas engine. It can give users great visual experience of HTML5 game like a Hollywood movie. The most of the technology development are done under the cooperation between PlayCanvas and ARM.

Finally, Xiao Shen the development director from Thunder Software Technology Co., Ltd the leading solution vendor in China, gave the speech Thundersoft experience in HTML5 development . They have many maturing products, reference designs and technical storage of HTML5 solution for SoC, ODM/OEM and IPS to do quick prototype/production development.

In this forum  the attendees raised many questions andthe discussion was very hot. We got fully success in this forum. Since many other people were not able to appear in this event but keen to the contents, please let us share the slides in the enclosed files. Welcome to share your comments.

Thanks a lot.

At a recent press event in China, Huawei announced their latest flagship smartphone, the Honor 7. This device is special. It marks a major milestone in the use of innovative heterogeneous computing technologies.

A special device…

Initial sales figures indicate the device is doing pretty well: a record 200,000 units sold in two minutes in the first flash sale (that is 1,667 handsets a second!), followed by 9 million units pre-ordered in the first week. There are also plans to make this available in Europe, I look forward to get my hands on one.

Powering the  Honor 7 is the in-house Kirin 935 SoC, featuring an octa-core 64-bit system with big.LITTLE technology (two quad-core ARM Cortex-A53 processor clusters) and a Mali-T628 MP4 GPU.

Here is the official launch video:

… with a special camera

Among the many leading features, the device includes a 20-megapixel camera with f/2.0 aperture and phase detection auto-focus, which allows the camera to focus in just 0.1 seconds, and an 8-megapixel front camera with fixed focus and f/2.4 aperture. What is really special about this phone, is every time you take a photo, the Mali GPU processes it to improve its appearance.

The OpenCL standard API has been used by Huawei to offload key image processing steps onto the GPU. Inside the camera stack the processing is optimally balanced between the CPU and the GPU. Using the processor that is most suited to each step ensures increased efficiency.

This is a break away from the common approach of using dedicated h/w IP, and has enabled a market leading OEM such as Huawei to work on advancing the algorithm with new techniques and optimizations all the way to the device launch. And of course further improvements can still be rolled out with over the air updates whilst the device are in the field, whereas hardware updates are not possible.

ARM and Huawei engineers have collaborated very closely in this project. Key algorithms such as de-noising have been ported to the Mali architecture using OpenCL and optimized at micro and macro level to operate more efficiently. Now every photo the user takes with this device is processed through the Mali GPU, in real time.

Is it that simple?

De-noise may sound diminutive, however it can indeed be a very complex nut to crack, in particular in challenging lighting conditions, which is where most noise occurs, and where most photos are taken.

Any implementation of a de-noise pipeline normally includes a mix of the common steps, for example: Haar features detection, Gaussian blurring, Sobel operators, biliteral filtering, down and up-scaling applied to various channels, all interleaved in more or less complex pipelines, depending on the implementation. These type of filter chains can easily exceeded 20 stages, and need to operate on high resolution images and critically, in real time and to a sensible power overhead.

On paper each individual block would seem suitable for GPU acceleration. A lot of work went into getting this pipeline optimized, as well as fine tuning interoperation between CPU and GPU, and integrating this new functionality in the existing device camera framework.

But, wait, why not use the ISP?

Although using an ISP, or any dedicated hardware for the matter, can often provide advantages in power, performance and area for the specific use case it is designed for, it also often has and important shortcoming: limited flexibility. Fixed logic cannot be changed once is it committed to silicon. The algorithm in question was modified, improved, developed aggressively right to the wire. The choice of using OpenCL on the Mali GPU has essentially enabled the OEM, Huawei in this instance, to give the end user a better camera experience, using existing hardware. In addition to superior image quality, in their public launch Huawei claimed a performance improvement of 2x through the use of the GPU, so faster pictures too!

It’s a journey, and it’s getting exciting.

The Honor 7 device marks a major adoption milestone for ARM’s vision of heterogeneous computing and GPU compute. But it is just the latest milestone of a fantastic journey in the adoption of this technology. Here is a refresher of what took us here.

GPU Compute in mobile and embedded systems was introduced in shipping devices in late 2012, when the Google Nexus 10 tablet was launched with RenderScript support on the Mali-T604 GPU.

Since then a large number of partners have endorsed this technology to improve a variety of end-user applications.

In early 2013 many partners (I have personally being involved in excess of 25 independent projects) have gradually enabled their technologies to use OpenCL on Mali. We have since seen:

• RenderScript accelerated image filters by MulticoreWare and advanced camera and looks improvement by ThunderSoft
• An entire interleaved HDR ISP pipelines using OpenCL by the sensor vendor Aptina (now On Semi)
• Ittiam Systems showcasing at various events incremental improvements to their HEVC and then VP9 decoders using OpenCL. This included demonstration of how GPU Compute can save power in a real silicon implementation.
• Google adding OpenCL support to VP9 and open-sourcing this in the WebM project for all to use
• Gesture middleware and face tracking applications using OpenCL from our partner eyeSight Technologies of Israel.

This snowball was certainly rolling! And much continued to happen early this year:

At CES 2015 Omnivision announced the availability of an advanced imaging library aimed at complementing their camera module ISP using the Mali GPU. The library includes advanced imaging features such as 3D noise suppression, chroma noise reduction, de-fringe and de-haze and is targeted at smartphones and tablets.

At the Mobile World Congress in Barcelona earlier this year ArcSoft demonstrated their latest middleware camera products running on MediaTek MT6752 based chip. This included pre-ISP image stabilization, real-time dynamic video HDR and other camera middleware with the objective of bringing to the mass market features that are typically available at the high end, this made possible by the use of GPU Compute.

ARM has recently taken part in the Embedded Vision Alliance Summit in Santa Clara, where we hosted a workshop on computer vision on ARM based system. Fotonation, Morpho, ArcSoft, Wikitude and many others have discussed how they have been using NEON and Mali OpenCL to improve their latest products. Our own Gian Marco Iodice and Tim Hartley have also detailed how Mali can help with problems such as stereo processing and efficiently use heterogeneous computing. You can access the proceeding of this event here.

OEMs have an important challenge in their hands: how to deliver the best user experiences whilst respecting the lowest possible energy budgets. Huawei’s approach using GPU Compute in their camera, illustrates the merits of this technology. This is the latest exciting step in the adoption of this technology.

A bright future for GPU Compute

Beyond smartphone cameras, GPU compute enables a new field of innovation and new customer experiences driven by real-time visual computing. Target applications include computational photography, computer vision, deep learning, and the enablement of new emerging multimedia codecs and algorithms. So far we have only scratched the surface of the potential that GPU Compute can deliver. ARM is always collaborating with new partners in exciting use-cases areas and I personally look forward to continue this journey and see GPU Compute proliferate and delivering more and more benefits to users.

ARM organised an HTML5 technology workshop last week hosted by Liberty Global/Virgin Media at the Digital TV Group’s offices in central London,next to the MI6 building! This was a very well attended event with more than 45 delegates plus speakers from ARM, LibertyGlobal, Linaro, Opera Software, PlayCanvas and YouView.  The agenda covered HTML engine trends, UI implementation case studies, WebGL, developer support, Encrypted Media Extensions, GPU and browser integration, Chromium Embedded Framework and the Chrome process.

My thanks go out to Richard Stamvik, Gemma Paris and Ryan Booth for helping with the logistics for the event, and to our co-hosts from Virgin Media for sponsoring the event.

Trends in HTML5 Engines [slides]

I opened the conference by talking about the current trends in HTML5 engines, with a focus on activities that are going to make HTML5 application development more relevant for embedded application developers.  The presentation introduces modern HTML5 techniques such as Web Components and Service Workers plus infrastructure changes such as HTTP/2.  All links are contained in the slides, so check them out for more information.

Implementing a Cutting Edge HTML5 UI [slides]

Nico Verhout, VP applications development and Anne Bakker, director web development at LibertyGlobal (LGI), biggest worldwide cable provider with 27M customers in Europe & Latin America, spoke about their cutting edge HTML5 UI. The Horizon UI targets all screens from in house hardware to 3^rd party STB's and second screen devices.  It uses HTML5, WebGL, CSS3 animations and pixi.js animation physics. Besides openness, multiple vendor, large ecosystem and portability, they talk about how HTML5 offers stability, performance, productivity and security.

Creating Performant HTML5 Apps for TV's [slides]

Thomas Kurowski, Opera, talked about performant HTML5 applications for TV and the importance of simple design and effective implementation to provide good performance. In some instances, this can be achieved by avoiding animations and using Blink-based Opera’s optimisations.

Developer Support and Challenges [slides]

Ian Renyard and Andrea Fassina presented their view on the HTML landscape from the viewpoint of supporting third party developers for the YouView platform.  They discusses d the transition from Adobe Air for TV in 2012 to HTML5 in 2014 with the challenges faced in that transition.  Currently only the BBC apps, iPlayer and red button are in HTML5, but plans are in motion to move more apps over.

The main theme of this presentation was to understand your production environment, make sure you test and validate on target and understanding the footprint and characteristics of the libraries that you use in your apps.  They presented some great recommendations to developers and are willing to share these findings with the community.

Real World WebGL - The making of SeeMore WebGL [slides]

Will Eastcott of PlayCanvas presented their Open Source WebGL engine and how they implemented the SeeMore demo providing solid jank free performance on a midrange ARM-based tablet. The very impressive demo showcases physically based rendering supporting materials textures, light mapping and shadows through WebGL, and various optimisations, e g halos implemented using transparent camera-aligned sprites.

The Challenge of Integrating Media Playback on Embedded Devices [slides]

Alex Ashley, YouView, talked about challenges with HbbTV 2.0 supporting multiple playing/paused videos; combining live video and advertisements; HTML5  trick play; testing live stream seek; and lack of data sharing, e g cookies, between browser and media player.

Check out his slides for some sound recommendations for media playback.

Chromium Embedded Framework Integration [slides]

Zoltan Kuscsuk, Linaro, talked about HTML5 in embedded Linux; how to embed a browser in a native application; hardware accelerated video playback; and media encryption extensions.

He presented GPU-accelerated rendering in Chromium running in an ARM and Linux-based STB environment.

Encrypted Media Extensions on OP TEE [slides]

A second session by Zoltan covered Trusted Execution Environment (TEE)-protected DRM, and Encrypted Media Extensions (EME)  in the open source OPTEE which uses HTML’s MediaElement (there was some but not much TrustZone and TEE experience in the audience). The draft EME specification is also supported by Chromium.

Integrating the Mali GPU with Browsers [slides]

Wasim Abbas, technical lead for middleware graphics in ARM’s Mali ecosystem team, presented ARM’s GPU and browser integration work, focusing on embedded devices’ low power budget.  This presentation covered low level aspects of embedded GPU’s, and some great tips for developers targeting integrating against an embedded system.

The Chrome Process [slides]

I close the conference out with a look at how Chrome development actually happens, it looks into the intent process and explores how all development is discussed with open forums in the blink-dev community.  This presentation is a call to arms for those who have a vested interest in the web platform to get involved in the platforms development.

The presentation then goes on to look at some of the cool new platform enhancements that are in flight, and a look into future work that will have a positive impact on our lives as embedded HTML5 application engineers.

Conclusion

This conference pulled in attendees from across the value chain for the embedded world including chipset vendors, STB OEM's middleware providers and operators.  Feedback from the event was very positive with great networking opportunities.  Bring like minded people into one room to work on common problems across the segment.

We hope to replicate the format again, so watch this space and we hope to see you all at a future event soon.

http://blog.siggraph.org/2015/07/shadertoy-competition-at-siggraph-2015.html/