1 2 3 Previous Next

ARM Mali Graphics

253 posts

Heterogeneous applications – those running code on multiple processors like a CPU and a GPU at the same time – are inherently difficult to optimize.  Not only do you need to consider how optimally the different parts of code that run on the different processors are performing, but you also need to take into account how well they are interacting with each other.  Is either processor waiting around unnecessarily for the other?  Are you copying large amounts of memory unnecessarily?  What level of utilisation are you making of the GPU?  Where are the bottlenecks?  The complexities of understanding all these are not for the squeamish.


Performance analysis tools are, of course, the answer, at least in part.  DS-5 Streamline performance analyzer is one of these tools and recently saw the addition of some interesting new features targeting OpenCL.  Streamline is one of the components of ARM DS-5 Development Studio, the end-to-end suite of tools for software development on any ARM processor.


So, armed with DS-5 Streamline and a complex, heterogeneous application how should you go about optimization?  In this blog I aim to give you a starting point, introducing the DS-5 tool and a few concepts about optimization along the way.


DS-5 Streamline Overview


DS-5 Streamline allows you to attach to a live device and retrieve hardware counters in real time.  The counters you choose are displayed in a timeline, and this can include values from both the CPU and GPU in the same trace.  The image above, for example, shows a timeline with a number of traces.  From the top there’s the dual-core CPU activity in green, the GPU’s graphics activity in light blue and the GPU’s compute activity in red.  Following that are various hardware counter and other system traces.


As well as the timeline, on the CPU side you can drill down to the process you want to analyse and then profile performance within the various parts of the application, right down to system calls.  With Mali GPUs you can specify performance counters and graph them right alongside the CPU.  This allows you to profile both graphics and OpenCL compute jobs, allowing for highly detailed analysis of the processing being done in the cores and their components. A recently added feature, the OpenCL timeline, takes this a step further making it possible to analyse individual kernels amongst a chain of kernels.


Optimization Workflow

So with the basics described, what is the typical optimization process for complex heterogeneous applications?


When the intention is to create a combined CPU and GPU solution for a piece of software you might typically start with a CPU-only implementation.  This gets the gremlins out of the algorithms you need to implement and then acts both as a golden reference for the accuracy of computations being performed, and as a performance reference so you know the level of benefit the move to multiple processor types is giving you.


Often the next step is then to create a “naïve” port.  This is where the transition of code from CPU to GPU is functional but relatively crude. You wouldn’t necessarily expect a big – or indeed any – uplift in performance at this stage, but it’s important to establish a working heterogeneous model if nothing else.


At this point you would typically start thinking about optimization.  Profiling the naïve port is probably a good next step as this can often highlight the level of utilisation within your application and from there you can deduce where to concentrate most of your efforts.  Often what you’re looking for at this stage is a hint as to the best way to implement the parallel components of your algorithm.


Of course to get the very best out of the hardware you’re using it is vital to have a basic understanding at least of the architecture you are targeting.  So let’s start with a bit of architectural background for the Mali GPU.


The OpenCL Execution Model on Mali GPUs

Firstly, here’s how the OpenCL execution model maps onto Mali GPUs.




Work items are simply threads on the shader pipeline, each one with its own registers, program counter, stack pointer and private stack. Up to 256 of these can run on a core at a time, each capable of natively processing vector data.


OpenCL work groups – collections of work items – also work on an individual core. Workgroups can have barriers, local atomics and cached local memory.


The ND Range, the entire work load for an OpenCL job, splits the workgroups up and assigns them around the available Mali GPU cores. Global atomics are supported, and we have cached global memory.


As we’ll see, relative to some other GPU architectures, Mali GPU cores are relatively sophisticated devices capable of handling hundreds of threads in flight at any one time.


The Mali GPU Core


Let’s take a closer look inside one of these cores:




Here we see the dual ALU, the load/store and the texture pipelines. Threads come in at the top and enter one of these pipes, circle round back up to the top for the next instruction until the thread completes, at which point it exits at the bottom.  We would typically have a great many threads running this way spinning around the pipelines instruction by instruction.




So let’s imagine the first instruction is a load.  It enters and is executed in the load/store pipe.  If the data is available, the thread can loop round on the next cycle for the next instruction.  If the data hasn’t yet arrived from main memory, the instruction will have to wait in the pipe until it’s available.




Imagine then the next instruction is arithmetic.  The thread now enters one of the arithmetic pipes.  ALU instructions support SIMD – single instruction, multiple data – allowing operations on several components at a time.  The instruction format itself is VLIW – very long instruction word – supporting several operations per instruction.  This could include, for example, a vector add, a vector multiply and various scalar operations all in one instruction.  This can give the effect of certain operations appearing “as free” because the arithmetic units within the ALU can perform many of these in parallel within a single cycle.  Finally there is a built in function library – the “BIFL” – which has hardware acceleration for many mathematical and other operational functions.


So this is a complex and capable core, designed to keep many threads in flight at a time, and thereby hide latency.  Latency hiding is what this is ultimately all about. We don’t care if an individual thread has to wait around for some data to arrive as long as the pipelines can get on with processing other threads.


Each of these pipelines is independent from the other and likewise the threads are entirely independent from other threads.  The total time for a program to be executed is then defined by the pipeline that needs the most cycles to let every thread execute all the instructions in its program. If we have predominantly load/store operations for example, the load/store pipe will become the limiting factor.  So in order to optimize a program we need to find which pipeline this is allowing us to target optimization efforts effectively.


Hardware Counters


To help determine this we need to access the GPU’s hardware counters. These will identify which parts of the cores are being exercised by a particular job.  In turn this helps target our efforts towards tackling bottlenecks in performance.


There are a large number of these hardware counters available. For example there are counters for each core and counters for individual components within a core, allowing you to peek inside and understand what is going on with the pipelines themselves.  And we have counters for the GPU as a whole, including things like the number of active cycles.


Accessing these counters is where we come back to DS-5 Streamline.  Let’s look at some screenshots of Streamline at work.



The first thing to stress is that what we see here is a whole-system view.  The vertical green bars in the top line shows the CPU, the blue bars below that show the graphics part of the application running on the GPU, and the red bars show the compute-specific parts of the application on the GPU.




There are all sorts of ways to customise this – I’m not going to go into huge amounts of detail here, but you can select from a wide variety of counter information for your system depending on what it is you need to measure. Streamline allows you to isolate counters against specific applications for both CPU and GPU, allowing you to focus in on what you need to see.



Looking down the screen you can see an L2 cache measurement - the blue wavy trace in the middle there -  and further down we’ve got a counter showing activity in the Mali GPU’s arithmetic pipelines.  We could scroll down to find more and indeed zoom in to get a more detailed view at any point.


DS-5 Streamline can often show you very quickly where the problem lies in a particular application.  The next image was taken from a computer vision application running on the CPU and using OpenCL on the GPU.  It would run fine for a number of seconds, and then seemingly randomly would suddenly slow down significantly, with the processing framerate dropping in half.



You can see the trace has captured the moment this slowdown happened. To the left of the timeline marker we can see the CPU and GPU working reasonably efficiently.  Then this suddenly lengthens out, we see a much bigger gap between the pockets of GPU work, and the CPU activity has grown significantly.  The red bars in amongst the green bars at the top represent increased system activity on the platform.  This trace and others like it were invaluable in showing that the initial problem with this application lay with how it was streaming and processing video.


One of the benefits of having the whole system on view is that we get a holistic picture of the performance of the application across multiple processors and processor types, and this was particularly useful in this example.



Here we’ve scrolled down the available counters in the timeline to show some others – in particular the various activities within the Mali GPU’s cores.  You can see counter lines for a number of things, but in particular the arithmetic, load-store and texture pipes – along with cache hits, misses etc.  Hovering over any of these graphs at any point in the timeline will show actual counter numbers.




Here for example we can see the load/store pipe instruction issues at the top, and actual instructions on the bottom.  The difference in this case is a measure of the load/store re-issues necessary at this point in the timeline – in itself a measure of efficiency of memory accesses.  What we are seeing at this point represents a reasonably healthy position in this regard.


The next trace is from the same application we were looking at a little earlier, but this time with a more complex OpenCL filter chain enabled.




If we look a little closer we can see how efficiently the application is running.  We’ve expanded the CPU trace – the green bars at the top – to show both the cores we had on this platform.  Remember the graphics elements are the blue bars, with the image processing filters represented by the red.




Looking at the cycle the application is going through for each frame:


  1. Firstly there is CPU activity leading up to the compute job.
  2. Whilst the compute job then runs, the CPU is more or less idle.
  3. With the completion of the compute filters, the CPU does a small amount of processing, setting up the graphics render.
  4. The graphics job then runs, rendering the frame before the sequence starts again.


So in a snapshot we have this holistic and heterogeneous overview of the application and how it is running.  Clearly we could aim for much better performance here by pipelining the workload to avoid the idle gaps we see.  There is no reason why the CPU and GPU couldn’t be made to run more efficiently in parallel, and this trace shows that clearly.


OpenCL Timeline


There are many features of DS-5 Streamline, and I’m not going to attempt to go into them all.  But there’s one in particular I’d like to show you that links the latest Mali GPU driver release to the latest version of DS-5 (v5.20), and that’s the OpenCL Timeline.




In this image we’ve just enabled the feature – it’s the horizontal area at the bottom.  This shows the running of individual OpenCL kernels, the time they take to run, any overhead of sync-points between CPU and GPU etc.




Here we have the name of each kernel being run along with the supporting host-side setup processes   If we hover over any part of this timeline…



… we can see details about the individual time taken for that kernel or operation.  In terms of knowing how then to target optimizations, this is invaluable.


Here’s another view of the same feature.




We can click the “Show all dependencies” button and Streamline will show us visually how the kernels are interrelated.  Again, this is all within the timeline, fitting right in with this holistic view of the system.  Being able to do this – particularly for complex, multi-kernel OpenCL applications is becoming a highly valuable tool for developers in helping to understand and improve the performance of ever-more demanding applications.


Optimizing Memory Accesses


So once you have these hardware counters, what sort of use should you make of them?


Generally speaking, the first thing to focus on is the use of memories. The SoC only has one programmer controlled memory in the system – in other words, there is no local memory, it’s all just global.  The CPU and GPU have the same visibility of this memory and often they’ll have a shared memory bus.  Any overlap with memory accesses therefore might cause problems.


If we want to shift back and forth between CPU and GPU, we don’t need to copy memory (as you might do on a desktop architecture).  Instead, we only need to do cache flushes.  These can also take time and needs minimising. So we can take an overview with Streamline of the program allowing us to see when the CPU was running and when the GPU was running, in a similar way to some of the timelines we saw earlier.  We may want to optimize our synchronisation points so that the GPU or CPU are not waiting any longer than they need to. Streamline is very good at visualising this.


Optimizing GPU ALU Load


With memory accesses optimized, the next stage is to look more closely at the execution of your kernels.  As we’ve seen, using Streamline we can zoom into the execution of a kernel and determine what the individual pipelines are doing, and in particular determine which pipeline is the limiting factor.  The Holy Grail here – a measure of peak optimization – is for the limiting pipe to be issuing instructions every cycle.


I mentioned earlier that we have a latency-tolerant architecture because we expect to have a great many threads in the system at any one time. Pressure on register usage, however, will limit the number of threads that can be active at a time.  And this can introduce latency issues once the number of threads falls sufficiently.  This is because if there are too many registers per thread, there are not enough registers for as many threads in total.  This manifests itself in there being too few instructions being issued in the limiting pipe.  And if we’re using too many registers there will be spilling of values back to main memory, so we’ll see additional load/store operations as a result.  The compiler manages all this, but there can be performance implications of doing so.


An excessive register usage will also result in a reduction in the maximum local workgroup size we can use.


The solution is to use fewer registers.  We can use smaller types – if possible.  So switching from 32 bit to 16 bit if that is feasible.  Or we can split the kernel into multiple kernels, each with a reduced number of registers.  We have seen very large kernels which have performed poorly, but when split into 2 or more have then overall performed much better because each individual kernel needs a smaller number of registers.  This allows more threads at the same time, and consequently more tolerance to latency.


Optimizing Cache Usage


Finally, we look at cache usage.  If this is working badly we would see many L/S instructions spinning around the L/S pipe waiting for the data they have requested to be returned. This involves re-issuing instructions until the data is available.  There are GPU hardware counters that show just what we need, and DS-5 can expose them for us.


This has only been a brief look at the world of compute optimization with Mali GPUs.  There’s a lot more out there.  To get you going I’ve included some links below to malideveloper.arm.com for all sorts of useful guides, developer videos, papers and more.


Download DS-5 Streamline: ARM DS-5 Streamline - Mali Developer Center Mali Developer Center

Mali-T600 Series GPU OpenCL Developer Guide: Mali-T600 Series GPU OpenCL Developer Guide - Mali Developer Center Mali Developer Center

GPU Compute, OpenCL and RenderScript Tutorials: http://malideveloper.arm.com/develop-for-mali/opencl-renderscript-tutorials/

Epic Giveaway.png

samsung note 4.jpgThe Samsung Galaxy Note 4 has quickly made itself popular among the ARM Mali graphics team. And it’s not its 515PPI Quad HD Super AMOLED display,

Exynos 7 Octa.png

vivid colors and intuitive UI that has earned it a place in our hearts – it is, as you would expect from ARM engineers, its stunning processor that has caught our eyes.

The Samsung Exynos 7 Octa is the latest mobile application processor to come out of the Samsung LSI team and it boasts considerable improvements over the previous generation – including up to 74% enhanced graphics performance with the ARM Mali-T760 MP6 GPU.  With this boost the Samsung Galaxy Note 4 can deliver superior, more life-like 3D gaming experiences on super HD screens, as well as a smoother, more responsive user interface and performance intensive, up-and-coming applications such as instant image stabilization, video editing or facial recognition. The Mali-T760 incorporates many of ARM’s popular energy efficient technologies, such as ARM Frame Buffer Compression, Adaptive Scalable Texture Compression, Smart Composition and Transaction Elimination – these together with the micro-architectural improvements to the Mali-T760, in particular to the L2 cache interconnect, result in an Exynos SoC that delivers a fantastic graphics experience without overexerting its intrinsic thermal and

samsung exynos 7.png

power budget.

When combined with a 1.9GHz ARM Cortex-A57 MP4, a 1.3GHz Cortex-A53 MP4 processor in big.LITTLE™ configuration with Samsung's HMP (Heterogeneous Multi-Processing) solution, every process can intelligently use the high processing power in such a way that no matter what the multitasking needs are, or what application is being run, there will be no lags and ultimately no unnerving power consumption. In all, the HMP technology, when used with the Cortex-A57 cores and Cortex-A53 cores, provides a 57% CPU performance increase from the previous generation Exynos 5 Octa.


The sum of all this is a device that has not only impressed across a range of benchmarks but also delighted critics and the public at large. The Samsung Galaxy Note 4 is an extremely desirable device that delivers the very latest advances in mobile technology – and it can be yours! ARM is giving away a Samsung Galaxy Note 4 as part of the 2014 Epic Giveaway in partnership with HEXUS. To find out more and to enter the EPIC Giveaway for your chance to win, click here and go to HEXUS’ Facebook page.


The 2014 Epic Giveaway is underway today. In partnership with HEXUS, ARM is giving you the chance to win amazing new prizes this holiday season! Every day for the next few weeks, we'll be giving away a brand-new ARM-based device. We'll have an array of prizes from ARM and our partners, including Atmel, Nvidia and Samsung, plus many, many more! Each prize draw will be open for seven days, so visit the dedicated competition page to keep tabs on what's up for grabs and what's coming soon.

For GDC 2014 I wrote a presentation intended to capture people’s interest and get them to the ARM lecture theatre using a lot of the buzz surrounding the current uptake in 3D tech. Alas, we never got to see the talk in the lecture theatre as so many partners offered to speak on our stand that we didn’t have time for it, so instead we recorded it in the office to present in ARMflix format.


When I first started at ARM on the demo team, the first demo I saw was True//Force, running in stereoscopic 3D on a 36 inch screen. Connected to that screen was a single core ARM® Mali™-400 MP GPU based development board.


That demo was adapted to many other platforms through the years, and later we reinvented the 3D camera system for a gesture input UI, again designed for large stereoscopic scenes. Throughout this development we made several interesting observations about pitfalls faced in generating stereoscopic content, which this year we were able to share over the ARMflix channel.



Admittedly the video above may be a little overlong in its description of 3D screen technology, most of this is already very common knowledge, although the advice  about splitting the signal top/bottom instead of left/right still throws some developers. Since screens with alternating polarised lines already have a reduction in vertical resolution, the reduction in vertical resolution from a top/bottom split screen isn’t seen. A left/right split screen will reduce the horizontal resolution, forcing the screen to interpolate between pixels, and the vertical detail gets lost in the interlacing of lines anyway.


The basics of stereoscopy can be understood by thinking about lines of sight coming out from your eyes and intersecting at an object. If the object gets closer those lines converge nearer, if it moves into the distance those lines converge further away. When an object is drawn on a stereoscopic display, each eye sees the object in different places, and where those eye lines cross is where it appears to be in space, which could be in front of or even behind (or inside) the screen.




This projection into different views in order to emulate this difference in what each eye sees is the basis of stereoscopy, but stereoscopy is only one factor used by your brain to discern the shape of the world. If you’re going to implement stereoscopy, it has to remain congruent with other mental models of how the world works.


One such model is occlusion. For example, you would never expect to see something through a window which was further away than the object you’re seeing through it. Anything which comes out of the screen in a stereoscopic setup is basically this exact scenario. It’s okay so long as the object is in the middle of the screen, because then it could conceivably be in front of it, but if it clips the edge of the screen your brain becomes aware that it is seeing through the screen like a window, and yet the object being occluded by it is closer. This creates a visual dissonance which breaks the spell, so to speak.



As it approaches the edge what you will observe is a kind of impossible monocular defect, where the eyes disagree to a degree that it becomes hard to discern spatial information. Monocular defects occur in real life too, imagine peering round the corner of a wall such that only one eye sees an object in the distance, or looking at an object which has a highly angle sensitive iridescence or noisy high gloss texture such that it seems a different colour or pattern in each eye. The difference in these cases is that they arise from genuine circumstance and we can rely on other factors to provide additional information, from small lateral movements of the head to using optical focus as a secondary source of depth information.


In a simulation, the focus of the eyes is always on the screen, as there is currently no way to defocus the image on screen to converge at a different focal point for your eyes, and without sophisticated head tracking hardware, your TV can’t correct perspective for you leaning to the left or coming closer.


More importantly, as the generation of stereoscopic viewpoints is being faked on a 2D screen, it’s possible to do odd things like occlude something with an object that appears to be behind it. This is a mistake frequently made by overlaying the menu on the screen for each eye but not thinking about how far in or outside the screen it will appear.


You could also occlude protrusions from the screen by the screen itself. The imagined reality of watching something stereoscopic on a screen is that the screen is a window you’re looking through into a fully 3D world. Occasionally things may bulge out of this window, and that’s fine, but if they then slide sideways out of view, they’re being occluded by an object behind them: the frame of this virtual window.


Best to try to keep the space in front of the screen a little bit special, by keeping the collision plane of the camera and the convergence plane of the projections in the same place. With that in mind, let’s talk about convergence planes.

This is a good point to remind you that this is advice for stereoscopy on screens, as virtual headsets use a very different method of projection.


So you’ve got two cameras pointing forwards, some distance apart. This will give you a kind of 3D, but unfortunately you’re not quite done yet. As mentioned before, the distance at which your eyes converge on the screen is the place in 3D space where objects should appear in the same place on the two eye images. But if you draw two visibility frustums pointing parallel, you’ll see that actually there is no point where an object appears in the same place in both. It will tend towards convergence at infinity, when the distance between the frustums is negligible compared to the width of the view plane. That means that the screen is essentially the furthest away anything can be in this setup, whereas what you probably want is a sense of depth into the screen. To achieve this you need to cross the streams, make the eyes look inward a little.


It’s not as simple as just rotating the frustums however. Unlike the spherical optics of an actual eye (admittedly rather exaggerated in the video), the flat projection of a frustum will intersect but never truly converge. For convergence of the view plane you must skew the frustum.


The equation for this is far simpler than it looks, a multiplication by a matrix that looks like this:

skew matrix.png


g is the gradient of the skew, defined as s/d (where s is half the separation between the two cameras, and d is the perpendicular distance between the desired convergence plane and the cameras).


So now you need to decide what the eye distance and the convergence plane are. The best advice I can give is to attempt to replicate as best you can the relationship between the physical world and the virtual. Basically, you would expect someone playing a game to sit at least a meter from the screen and have eyes which are between 6 and 7cm apart. However, more important than an exact emulation is the ability to parameterise it for the far distance. The draw frustum has an angular factor in the field of view.


Typically the field of view is around 90 degrees, which means the screen size at the convergence plane is sin(45)d or roughly 0.7d so in the scenario above where we assume the screen is about 1 meter away from the cameras, the screen width is 70cm. I’m currently sat in front of a 24 inch monitor and I measure the width as approximately 50cm. If the eyes were placed 6cm apart in this scenario, the far plane would tend towards the virtual 7cm eye separation required for the eye vectors to be parallel. Scaled down onto my 50cm screen however this would be a 5cm eye separation. My eye lines would not be parallel and so the far plane would actually have a tangible optical distance*.


Ideally you want the distance of the far plane separation on screen to be the same as your user’s actual physical eye separation. This of course requires parameterisation for user eye separation and screen size. Sticking with 7cm eye separation and a 50cm screen for a moment, let’s look at our options.


1. You can move the cameras further apart (by a ratio of 7/5 to a virtual 9.8cm in this case) and adjust the skew gradient to suit.

2. You can make the field of vision (FOV) narrower (to sin-1(0.5) or 30 degrees in this case) so that the convergence plane is the same size as the screen.


In case it isn’t obvious which of these is the better option (a 30 degree FOV, really you think that’s okay?), consider the math if the screen is 10cm across. Obviously you wouldn’t see much through a 5.7 degree FOV, so that option is right out.


Option 1 has one major drawback however, which is that the whole world’s scale will scale with the eye distance.  In this case the whole world would be 5/7ths its original size. Again, do the math with a 10cm screen and you’ll see that the world would be 1/7th the size.


There is a temptation to move the cameras closer to the convergence plane, but to keep the screen the same size and place that plane closer, the FOV would have to widen, creating a mismatch between the perspective in the real and virtual space. What this mismatch does is move things towards the far plane more rapidly, as the sense of depth is an asymptotic power curve linked to perspective. The wider the FOV goes, the higher the power and the tighter the curve. If this curve is different to what you’re used to, the depth will feel very odd and less real.  However, an incorrect convergence or even a lack of convergence is far preferable to the worst sin of stereoscopy: divergence.



Since the entire basis of stereoscopy is your eyes pointing to two different places on the screen and the imaginary sight lines crossing at the depth being simulated, we know that these lines being parallel means the object is infinitely far away. But what if the images for each eye move even further apart? If the gap is wider than the gap between your eyes the imaginary lines of sight are no longer parallel but actually diverge and this is one of the biggest head ache inducers in stereoscopy.


It is a scenario which you can literally never replicate in real life without optical equipment. If the images diverge by a reasonable amount it’s not too bad because you just can’t focus on them, which ruins the picture but not your eyes. If they diverge by a tiny amount your eyes will have a go at achieving binocular focus and you’ll probably end up with the mother of all migraines in very little time. This scenario is the most likely to arise if you don’t allow users to fine tune their own stereoscopy settings.


There are other simple mistakes and pitfalls covered in the video at the top of this blog, such as forgetting to make menus stereoscopic and being careful with the perceived depth of overlays. It makes sense for a pop-up menu to have its text at screen depth on the convergence plane so that if the stereoscopy needs adjusting, the menu is still easy to read… but the menu has to be drawn over everything else, so how to avoid objects behind the menu overlapping the convergence plane at these times? Again, this comes down to keeping space in front of the convergence plane special.


Which just leaves me time to add one side ramble which isn’t in the video. By far the most popular stereoscopic display currently in use is that on the 3DS hand held game console from Nintendo, and when it first came out people talked about the weirdness of the little slider on the side which seemed to scale the 3D effect on the screen. If it was at the bottom it went 2D, and sliding it up slowly took it to full depth simulation. At the time people wondered why you would want a game to be in half 3D, where it was still stereoscopic, but appeared semi-flattened.


The answer is simple: The slider was actually adjusting the eye-separation, so that full depth stereoscopy could be tuned for use by children of all ages. If the anatomical eye gap of the user is wider than the highest setting of the device, it doesn’t matter too much, it just means that the far plane won’t look as far. But if the eye separation was fixed width and someone with a smaller eye gap tried to use it, it would cause line of sight divergence, and all the associated headaches.


So make sure you’ve got that slider in your software.


* I know you’ll be knocking yourselves out over this so I’ll do the math for you. The far plane would appear to be 2.5 meters behind the screen.

ARM has released new major upgrades for the graphics debugger (Mali Graphics Debugger v2.0) and the OpenGL® ES emulator (Mali OpenGL ES Emulator v2.0), and an update for the offline shader compiler (Mali Offline Compiler v4.5), adding support for OpenGL ES 3.1 and many other features. This is the last update for 2014, and concludes a year full of many releases and amazing new features, like overdraw and shader map modes, support for Android KitKat and Lollipop, ASTC textures, frame replay and support for all the ARM® Mali™ GPU cores in the offline compiler. Additionally to those, today we present you three new releases, including key features like live shader editing, support for unrooted Android devices, compute shaders and indirect draw calls.

Content from Major Upgrade for Mali Graphics Development Tools Mali Developer Center

Mali Graphics Debugger v2.0

Mali Graphics Debugger allows developers to trace OpenGL ES and OpenCL™ API calls in their application and understand frame-by-frame the effect on the application to help identify possible issues. We support Android and Linux devices with ARM Cortex® CPUs and Mali GPUs & Linux, Windows and Mac OS X hosts.

Key New Features

  • OpenGL ES 3.1 support
    This means that all OpenGL ES 3.1 function calls will be present in a trace.
  • Live editing
    • Support added for changing both the fragment and vertex shader of a program and then replaying the frame to view the results.
    • Support added to override textures in an application and replace them with a new texture that will aid in diagnosing any issues with incorrect texture coordinates.
    • Support added for overriding the precision of all elements in a shader and then replaying the frame to view the results (force highp/mediump/lowp modes)
    • New mid-level hierarchy in the outline view added so now draw calls are separated per framebuffer as well as per frame allowing the user to better visualize render passes.
  • New Android application provided to support unrooted devices
  • New view for compute shaders.
  • User can now filter by frame feature to highlight interesting frames in larger traces.
  • Support for making notes alongside function calls has been added. This allows important functions in a trace to be located easily.
  • Support for most OpenGL ES extensions.

To learn more about Mali Graphics Debugger and performance analysis of graphics applications, you can watch the tutorials and presentations on YouTube.

Get Mali Graphics Debugger

Mali Graphics Debugger Daemon Application for Android

With the objective of making the installation of the graphics debugger on Android targets easier, we have developed an Android application that runs the required daemon. This eliminates the need to manually install executables on the Android device. The application (APK) works on rooted and unrooted devices.




Mali OpenGL ES Emulator v2.0

The OpenGL ES Emulator is a library that maps OpenGL ES 3.1 API calls to the OpenGL API. By running on a standard PC, the emulator helps software development and testing of next generation OpenGL ES 3.1 applications since no embedded platform is required. We support Linux and Windows PCs.Key new features

  • OpenGL ES 3.1 support

We have put a lot of effort to make the emulator as close as possible to the Khronos standard OpenGL ES 3.1 API, executing the conformance suite and getting good conformance score on different platforms. Some tests cannot easily pass on desktop PCs, due to the nature of the underlying OpenGL driver; nevertheless we passed more than 94% of the conformance tests on our test platform.

Get OpenGL ES Emulator



Mali Offline Compiler v4.5

Mali Offline Compiler is a command line tool that translates vertex shaders and fragment shaders written in the OpenGL ES Shading Language (ESSL) into binary vertex shaders and binary fragment shaders for execution on the Mali GPUs. It generates statistics that are useful to optimize shader code.Key new features

  • OpenGL ES 3.1 shader language support, including compute shaders
  • Support for Mali ‘Midgard‘ driver version r5p0 (for Mali-T600 and Mali-T700 Series)
  • Support for Mali ‘Utgard‘ driver version r5p0 (for Mali-400 MP and Mali-450 MP)


Get Mali Offline Compiler




Support and Other Activities

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


Lorenzo Dal Col is the Product Manager of DS-5 Streamline and Mali GPU Tools. He first used ARM technology when, in 2007, he created a voice-controlled robot at university. He has experience in machine learning, image processing and computer vision. He joined ARM in 2011 to work on 3D graphics, developing performance analysis and debug tools.

I have previously shared how to install OpenCL on the Samsung XE303C12 Chromebook powered by the ARM® Mali-T604 GPU. I have found that things are slightly different on the newer Samsung XE503C12 Chromebook ("Chromebook 2") powered by the ARM® Mali-T628 GPU, so decided to provide an update. As before, please bear in mind that this is not ARM's "official guide" (which can be found here). However, it's a useful alternative to the official guide if, for example, you don't have a Linux PC or just want to use Chrome OS day in and day out.


You will need:


How fast you will complete the installation will depend on how fast you can copy-and-paste instructions from this guide (Ctrl-C) into the shell (Shift-Ctrl-C), how fast your Internet connection is and how fast your memory card is. (I will give an approximate time for each step measured when using a rather slow 30 MB/s card). The basic OpenCL installation should take up to half an hour; PyOpenCL and NumPy about an hour; further SciPy libraries about 4 hours. Most of the time, however, you will be able to leave the Chromebook unattended, beavering away while compiling packages from source.


Finally, the instructions are provided "as is", you use them at your own risk, and so on, and so forth... (The official guide also contains an important disclaimer.)


Installing OpenCL

Enabling Developer Mode

NB: Enabling Developer Mode erases all user data - do a back up first.


Enter Recovery Mode by holding the ESC and REFRESH (↻ or F3) buttons, and pressing the POWER button. In Recovery Mode, press Ctrl+D and ENTER to confirm and enable Developer Mode.


Entering developer shell (1 min)

Open the Chrome browser and press Ctrl-Alt-T.

Welcome to crosh, the Chrome OS developer shell.

If you got here by mistake, don't panic!  Just close this tab and carry on.

Type 'help' for a list of commands.

Don't panic, keep the tab opened and carry on to enter the shell:

crosh> shell
chronos@localhost / $ uname -a
Linux localhost 3.8.11 #1 SMP Wed Dec 10 14:41:54 PST 2014 armv7l SAMSUNG EXYNOS5 (Flattened Device Tree) GNU/Linux


Preparing a Micro SD card (5 min)

Insert a blank Micro SD card (denoted as /dev/mmcblk1 in what follows):

chronos@localhost / $ mount | grep "SD Card"
/dev/mmcblk1p1 on /media/removable/SD Card type vfat (rw,nosuid,nodev,noexec,relatime,dirsync,uid=1000,gid=1000,fmask=0022,dmask=0022,codepage=437,iocharset=iso8859-1,shortname=mixed,utf8,flush,errors=remount-ro)

Unmount the card and run fdisk:

chronos@localhost / $ sudo umount /dev/mmcblk1p1
chronos@localhost / $ sudo /sbin/fdisk /dev/mmcblk1

Welcome to fdisk (util-linux 2.24).
Changes will remain in memory only, until you decide to write them. Be careful before using the write command. Command (m for help):

Enter 't' to change a partition type, then '83' to change the partition type to 'Linux', and finally 'w' to apply the change:

Command (m for help): t
Selected partition 1
Hex code (type L to list all codes): 83
If you have created or modified any DOS 6.x partitions, please see the fdisk documentation for additional information.
Changed type of partition 'W95 FAT32 (LBA)' to 'Linux'.

Command (m for help): w
The partition table has been altered.
Calling ioctl() to re-read partition table.
Syncing disks.
chronos@localhost / $ 

Format the card e.g. using ext3:

chronos@localhost / $ sudo /sbin/mkfs.ext3 /dev/mmcblk1p1

NB: If you use a card that is less than 8 GB, you may need to reserve enough inodes when you format the card e.g.:

chronos@localhost / $ sudo /sbin/mkfs.ext3 /dev/mmcblk1p1 -j -T small

Mount the card and check that it's ready:

chronos@localhost / $ sudo mkdir -p ~/gentoo
chronos@localhost / $ sudo mount -o rw,exec -t ext3 /dev/mmcblk1p1 ~/gentoo
chronos@localhost / $ df -h ~/gentoo
Filesystem      Size  Used Avail Use% Mounted on
/dev/mmcblk1p1   15G   38M   14G   1% /home/chronos/user/gentoo
chronos@localhost / $ df -hi ~/gentoo
Filesystem     Inodes IUsed IFree IUse% Mounted on
/dev/mmcblk1p1   951K    11  951K    1% /home/chronos/user/gentoo

Installing Gentoo Linux (15 min)

chronos@localhost / $ cd ~/gentoo
chronos@localhost ~/gentoo $ ls -la
total 32
drwxr-xr-x  3 root    root            4096 Dec  9 21:31 .
drwx--x--- 30 chronos chronos-access 12288 Dec  9 21:38 ..
drwx------  2 root    root           16384 Dec  9 21:31 lost+found

Download the latest stage 3 archive for armv7a_hardfp:

chronos@localhost ~/gentoo $ sudo wget http://distfiles.gentoo.org/releases/arm/autobuilds/latest-stage3-armv7a_hardfp.txt
chronos@localhost ~/gentoo $ sudo wget http://distfiles.gentoo.org/releases/arm/autobuilds/`cat latest-stage3-armv7a_hardfp.txt | grep stage3-armv7a_hardfp`

Extract the downloaded archive right onto the card e.g.:

chronos@localhost ~/gentoo $ sudo tar xjpf stage3-armv7a_hardfp-20141023.tar.bz2

Clean up:

chronos@localhost ~/gentoo $ sudo rm latest-stage3-armv7a_hardfp.txt
chronos@localhost ~/gentoo $ sudo rm stage3-armv7a_hardfp-20141023.tar.bz2


Downloading OpenCL drivers (4 min)

Go to the page listing Mali-T6xx Linux drivers and download release r4p0-02rel0 for Mali-T62x fbdev (mali-t62x_r4p0-02rel0_linux_1+fbdev.tar.gz). Make sure you carefully read and accept the associated licence terms.

chronos@localhost ~/gentoo $ sudo tar xvzf ~/Downloads/mali-t62x_r4p0-02rel0_linux_1+fbdev.tar.gz 

This will create ~/gentoo/fbdev which we will use later.


Entering Gentoo Linux (2 min)

Similar to crouton, we will use chroot to enter our Linux environment.


Create two scripts and make them executable:

chronos@localhost ~/gentoo $ sudo vim ~/gentoo/setup.sh
mount -t proc /proc $GENTOO_DIR/proc
mount --rbind /sys  $GENTOO_DIR/sys
mount --rbind /dev  $GENTOO_DIR/dev
cp /etc/resolv.conf $GENTOO_DIR/etc
chronos@localhost ~/gentoo $ sudo vim ~/gentoo/enter.sh
LC_ALL=C chroot $GENTOO_DIR /bin/bash
chronos@localhost ~/gentoo $ sudo chmod u+x ~/gentoo/setup.sh ~/gentoo/enter.sh

Execute the scripts:

chronos@localhost ~/gentoo $ sudo ~/gentoo/setup.sh
chronos@localhost ~/gentoo $ sudo ~/gentoo/enter.sh
localhost / # 

Note that the ~/gentoo directory will become the root (/) directory once we enter our new Linux environment. For example, ~/gentoo/fbdev will become /fbdev inside the Linux environment.


Installing OpenCL header files (2 min)

Download OpenCL header files from the Khronos OpenCL registry:

localhost / # mkdir /usr/include/CL && cd /usr/include/CL
localhost / # wget http://www.khronos.org/registry/cl/api/1.1/opencl.h
localhost / # wget http://www.khronos.org/registry/cl/api/1.1/cl_platform.h
localhost / # wget http://www.khronos.org/registry/cl/api/1.1/cl.h
localhost / # wget http://www.khronos.org/registry/cl/api/1.1/cl_gl.h
localhost / # wget http://www.khronos.org/registry/cl/api/1.1/cl_ext.h


Installing OpenCL driver (2 min)

Change properties on the downloaded OpenCL driver files and copy them to /usr/lib:

localhost / # chown root /fbdev/*
localhost / # chgrp root /fbdev/*
localhost / # chmod 755 /fbdev/*
localhost / # mv /fbdev/* /usr/lib
localhost / # rmdir /fbdev



By now you should have a mint Linux installation complete with the OpenCL drivers and headers, so you can start playing with OpenCL!

When you reboot, you just need to mount the card and execute the setup script again:

chronos@localhost / $ sudo mount -o rw,exec -t ext3 /dev/mmcblk1p1 ~/gentoo
chronos@localhost / $ sudo ~/gentoo/setup.sh

Then you can pop in and out of the Linux environment with:

chronos@localhost / $ sudo ~/gentoo/enter.sh
localhost / # exit
chronos@localhost / $

But the fun just begins here! Follow the instructions below to install PyOpenCL and SciPy libraries for scientific computing.


Installing PyOpenCL

Configuring Portage (15 min)

Portage is Gentoo's package management system.

localhost / # echo "MAKEOPTS=\"-j4\"" >> /etc/portage/make.conf
localhost / # echo "ACCEPT_KEYWORDS=\"~arm\"" >> /etc/portage/make.conf
localhost / # mkdir /etc/portage/profile
localhost / # mkdir /etc/portage/package.use
localhost / # mkdir /etc/portage/package.unmask
localhost / # mkdir /etc/portage/package.accept_keywords
localhost / # mkdir /etc/portage/package.keywords
localhost / # touch /etc/portage/package.keywords/dependences

Perform an update:

localhost / # emerge --sync --quiet
localhost / # emerge --oneshot portage
localhost / # eselect news read

NB: If any emerge command below fails, rerun it with the --autounmask-write flag; then run etc-update and answer '-3' followed by 'y'. Running emerge again should now get the build started e.g.:

localhost / # emerge --autounmask-write dev-python/pandas
localhost / # etc-update
Scanning Configuration files...
The following is the list of files which need updating, each
configuration file is followed by a list of possible replacement files.
1) /etc/portage/package.keywords/dependences (1)
Please select a file to edit by entering the corresponding number.
              (don't use -3, -5, -7 or -9 if you're unsure what to do)
              (-1 to exit) (-3 to auto merge all files)
                           (-5 to auto-merge AND not use 'mv -i')
                           (-7 to discard all updates)
                           (-9 to discard all updates AND not use 'rm -i'): -3
Replacing /etc/portage/package.keywords/dependences with /etc/portage/package.keywords/._cfg0000_dependences
mv: overwrite '/etc/portage/package.keywords/dependences'? y
Exiting: Nothing left to do; exiting.
localhost / # emerge dev-python/pandas


Setting up Python (3 min)

localhost / # eselect python set python2.7
localhost / # emerge dev-python/setuptools


Installing NumPy (40 min)

Install NumPy with LAPACK as follows.

localhost / # echo "dev-python/numpy lapack" >> /etc/portage/package.use/numpy
localhost / # echo "dev-python/numpy -lapack" >> /etc/portage/profile/package.use.mask
localhost / # emerge dev-python/numpy
localhost / # python -c "import numpy; print numpy.__version__"


Installing PyOpenCL (7 min)

Install PyOpenCL.

localhost / # cd /tmp
localhost tmp # wget https://pypi.python.org/packages/source/p/pyopencl/pyopencl-2014.1.tar.gz
localhost tmp # tar xzf pyopencl-2014.1.tar.gz
localhost tmp # cd pyopencl-2014.1
localhost pyopencl-2014.1 # python configure.py
localhost pyopencl-2014.1 # make install
localhost pyopencl-2014.1 # cd examples
localhost examples # python demo.py
Choose device(s):
[0] <pyopencl.Device 'Mali-T628' on 'ARM Platform' at 0x-49b96370>
[1] <pyopencl.Device 'Mali-T628' on 'ARM Platform' at 0x-49b96270>
Choice, comma-separated [0]:0
Set the environment variable PYOPENCL_CTX='0' to avoid being asked again.
(0.0, 241.52145)
localhost examples # python -c "import pyopencl; print pyopencl.VERSION_TEXT"

(That's right! The Exynos 5420 chip effectively has two Mali-T62x GPUs: GPU 0 has 4 cores; GPU 1 has 2 cores. How cool is that?)


Installing scientific libraries

If you would like to follow my posts on benchmarking (e.g. see the intro), I recommend you install packages from the SciPy family.


Installing IPython (45 min)

localhost / # emerge dev-python/ipython
localhost / # ipython --version


Installing IPython Notebook (5 min)

Install IPython Notebook to enjoy a fun blend of Chrome OS and IPython experience.


localhost / # emerge dev-python/jinja dev-python/pyzmq www-servers/tornado
localhost / # ipython notebook
2014-05-08 06:49:08.424 [NotebookApp] Using existing profile dir: u'/root/.ipython/profile_default'
2014-05-08 06:49:08.440 [NotebookApp] Using MathJax from CDN: http://cdn.mathjax.org/mathjax/latest/MathJax.js
2014-05-08 06:49:08.485 [NotebookApp] Serving notebooks from local directory: /
2014-05-08 06:49:08.485 [NotebookApp] The IPython Notebook is running at:
2014-05-08 06:49:08.486 [NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
2014-05-08 06:49:08.486 [NotebookApp] WARNING | No web browser found: could not locate runnable browser.

Open in a new Chrome tab to start creating your own IPython Notebooks!


Installing Matplotlib (50 min)

localhost / # emerge dev-python/matplotlib
localhost / # python -c "import matplotlib; print matplotlib.__version__"


Installing SciPy (60 min)

localhost / # emerge sci-libs/scipy
localhost / # python -c "import scipy; print scipy.__version__"


Installing Pandas (80 min)

localhost / # emerge dev-python/pandas
localhost / # python -c "import pandas; print pandas.__version__"

TyGL is now open source!

Posted by mattspencer Dec 18, 2014

You might remember that back in August we released details of a program ARM had been working on in conjunction with Szeged University and Samsung Research UK.


TyGL is a new backend for WebKit which demonstrates a huge acceleration in mobile web rendering. While it was developed and tested on an ARM Mali-T628 GPU based Chromebook, it will work on any GPU conforming to OpenGL ES 2.0 and higher and has been shown to achieve a performance uplift of up to eleven times. Full details about the process through which TyGL manages this huge boost are available in one of our previous blogs, TyGL: Hardware Accelerated Web Rendering.


tygl svg tiger.png


Key Features of TyGL include:

  • GPU involvement in web rendering pipeline

Clever batching of draw calls delivers better results on GPUs (see previous blogs on batching for more information).  While the Graphics Context API can result in frequent state changes if implemented in the wrong way, TyGL is designed to catch this problem and group commands together to reduce draw calls – and in this way improve performance.


  • Automatic shader generation

TyGL enables efficient batching by generating complex shaders from multiple shader fragments – and ensures the batches fit into the shader cache of the GPU.


  • Trapezoid based path rendering

This section of the engine is under continuous improvement and is planning to take advantage of new GPU capabilities such as the Pixel Local Storage extension for OpenGL ES in order to get that extra bit of performance.


  • No software fallback

This is a complete GPU-based hardware accelerated solution with no dependency on legacy software and no need to sacrifice optimizations for compatibility.

This month we were delighted to open source the TyGL port and it is now available for the world to see on GitHub and we will truly value community involvement in implementing additional features and improving the code. Take a look, let us know what you think and feel free to contact us if you want more information about the project!


Further Reading

TyGL: Hardware Accelerated Web Rendering

TyGL on GitHub

Blog by our partners, Szeged University

How to build TyGL

This year at GDC I gave a presentation on our exhibition booth about using ASTC with different types of textures to get the best visual results. It’s interesting that in the past whenever I spoke about ASTC it was always about how it works, rather than how to use it, which is bizarre because that’s not really what developer education is about.


It would be like a driving instructor turning up and lecturing you for the full hour on the science behind the internal combustion engine.


I did go on to write a fairly long guide to understanding the various settings and options you get when compressing in ASTC for GPU Pro, and the release of that roughly coincided with my booth talk at GDC. Those present on the day may have noticed the presentation wasn’t up to my usual standard. I can only apologise, I was very ill and dragged myself out of the hotel to give the talk before immediately returning at the end.


I’d like to use this as an opportunity to reiterate some of that content in the form of a blog, to clarify some parts I missed or stumbled over on the day. For those who weren’t there or just want to relive the presentation, I’ve attached a recording of it here.



The first topic I covered is a really basic introduction to texture compression in general, including a few notes on why textures should be compressed in the first place. With the ubiquity of gif, jpg and png image compression formats, surprisingly few people stop to think about the size of raw pixel data. Whether you have an alpha channel or not, cache alignment means you’re essentially always packing one in raw image data, making a 32 bits per pixel (bpp) cost. With even modest textures weighing in around a million pixels each, you can see how this might get quite heavy.


It’s not the size of the texture that causes the real problem, it’s the fact that you have to constantly look that data up, as the GPU taps into the main memory to pull that data into its cache whilst shading fragments, all of which compounds the bandwidth usage of the application. The solution to this is not compressed image files unpacked into GPU memory, it’s compressed textures in the GPU memory that the GPU can unpack as needed. This places interesting constraints on the compression formats used.


Firstly the pixels need to be accessed randomly. PNG is all well and good for compressing a whole image but to unpack a single pixel you have to unpack the entire line it’s on. Maybe it’d be worthwhile if you were reading in order along that axis, but if you’re sampling across the lines, you end up unpacking far more data than you need. Compression relies on grouping data to compressed bundles, so optimally these bundles need to be blocks of pixels, not lines, allowing the block to be decompressed into the cache and sampled randomly in any direction.


As this implies, the blocks have to be completely standalone. Other than the compression format, there must be no external information such as a dictionary or symbol table to decode the block. Finally, blocks have to line up in memory in a regular formation, or your decompressor won’t know where to look in the data to find a specific block. Scanning through to find it is not an option.


This is why texture compression has its own specialist formats for the task. In the older ARM® Mali™ GPUs, we only supported the ETC and later ETC2 formats because those are Khronos standards. There’s a pretty good reason for sticking to standards because  the capability and availability of different compression formats is rather sparsely populated. Your choice of format might not just lock your application into a certain bitrate or channel specification; a proprietary format could also lock it to specific hardware.


ASTC is a texture compression format designed to solve this problem from the ground up, allowing different bit rates, different pixel formats, even different combinations of spatial dimensions to be picked for any given texture. So maybe you want a 2D high bit rate normal map with just X and Y data, or maybe you want a low bit rate 3D HDR RGBA texture? ASTC can do both, and more.


If you want to know how that even works, I already wrote about that at length here. If you want to know how to get the best results from it, you’re in the right place.


The quality of a compressed texture is controlled with three main factors: the bit rate, the limits and the error factors. I’ll tackle these from the easiest to understand to the hardest.


Bit Rates and Block Size


ASTC, as you may know, can encode in different block modes. The dimensions of a single block are called its footprint. Whereas other texture formats have fixed footprints, ASTC has various block footprints from 4x4 to 12x12 (and from 3x3 to 6x6x6). What stays the same in ASTC is the data size used to encode it, at exactly 128 bits. So if those 128 bits encode a 4x4 block (16 pixels), that’s 8bpp, whereas the 12x12 block (144 pixels) is a staggering 0.98bpp. If you think that’s impressive, a 6x6x6 block is 216 pixels, making it 0.59 bits per pixel. Ordinarily at this point there would just be a reminder that higher bit rate leads to higher quality and move on, but you’ve spent the time to read this far so I’ll explode that myth for you.  A 128 bit block can represent 2128 different binary combinations, each of which will map to a specific layout of pixels. The smallest block size, 4x4, contains 16 pixels, which at 32bpp (RGBA) can represent 2512 different combinations of pixel data. For those not used to thinking in binary, that means you have less than one in a googol’s chance of getting an exact match (a googol is one with a hundred zeros). That may seem very small, but the whole point is that you don’t need an exact match for every outcome, and the best texture compression formats are geared towards the 2128 pixel layouts most likely to make sense as part of a larger image.


The point is, if you’re using 12x12 blocks, there are 24608 combinations. The probability of getting an exact match on a block that size is less than one in one with a thousand zeroes, which we don’t even have a proper name for; it also means there’s a much lower chance of even getting a passable match for it. The compressor will have to pick the best configuration it can, and hope you don’t notice.


Limits and Leeway


Which leads us neatly onto limits, or how hard should the compressor try to find a good match? The whole point of texture compression algorithms is that they have a fast deterministic decompression function, but after a few intelligent choices, the best the compressor can do is try out different combinations and see how close they are. This means the more it checks, the more likely it is to find a good one. You don’t necessarily want to check them all; that would take a very long time. This is why you have to set limits. The limits can be things like “only try so many combinations, then give up and pick the best we found” or “if you find one that’s suitably close to the original, use that and stop looking” or even “if you try a few patterns with two partitions and it’s no better than those using a single partition, don’t bother trying three or four partitions” (the concept of partitions is explained in this blog post)


It’s fair to say most people wouldn’t know where to begin setting signal to noise decibel ratings for these kinds of decisions so, handily, the compressor has a few in-built presets from very fast to exhaustive. There’s a chance that it will find the best combination in the very fast presets, but it’s a very low chance. The probability is much higher if you’re willing to wait. The best advice therefore is to iterate your assets with fast or very fast compression, then ship with thorough or exhaustive compression. Curiously there’s very little difference between the result from thorough and exhaustive but exhaustive will take a lot longer, this again is down to the relative probabilities involved.

The one question remaining, therefore, is if it’s trying all these different blocks of pixels to see how close they are to the same block in the raw image, how is it comparing them?


Priority and Perception


In order to tell which one out of a hundred or a thousand or even a hundred thousand proposed blocks is the best, you need to be able to compare any pair and say that one is objectively better than the other, then repeat with the best and the next attempt.  The standard way to compare two images is called PSNR or percentage signal to noise ratio, so you take your original image, subtract all of the colour values from your resulting image, convert all the negative numbers to positive (the absolute difference) and then sum them. The ratio part comes from a sort of imaginary maximum error, which would be if an all white image came out all black or vice versa.

But there are different things you might want to preserve.


When the numbers are added together they can have weightings applied to them. Little known fact, the human optic system is more sensitive to high frequency detail in green light than red or blue. Using this knowledge you can add a pre-multiplier to different channels. If you gave a weight of two to the green error, and there were two tiles which differed by roughly the same amount, one mostly in the red channel, one mostly in the green channel, the error in the green channel would be doubled, meaning the one with the red error would be considered a better match.  Alternatively, you could be more concerned about angular error. This is particularly relevant in normal maps where the pixels represent not a colour to be displayed on screen but a field of vectors. In this scenario the ratio of the channels is far more important than simple per channel or overall magnitude differences, and this can be reflected by giving a weight to the angular component.


One interesting thing that arises as the result of block based comparisons is that errors near the edge of a tile may have positive error within the defined limits, and the errors on the adjacent tile may be negative error within the defined limits, making the step change between two blocks, which should match up of course, larger than the desired error bounds. Block weighting reduces that error by applying additional error weight to boundary mismatches.  If you really want to get under the hood, there are a few settings that tinker with the way individual pixel errors are combined into a full block error. These work by applying weights and pre-multipliers to the mean average error and the standard deviation of the error in a certain radius. I could talk at length about how this may be weighted to favour a tile with a few big errors over a tile with lots of little errors, or certain settings can favour a noisy looking tile over one which smoothes minor details out, and I haven’t even researched all the possibilities yet. Either way it’s a huge topic and one that although I touch upon in the presentation, I’m going to leave alone for now and go into much greater detail at a later date.


Getting Started with ASTC


If you want to try out ASTC you’ve got quite a few options. There are commercial devices available right now with the appropriate hardware to decode ASTC on the GPU even though it’s still a very new technology.


If you’d like to see how it looks without the hardware advantages of memory bandwidth reduction, the OpenGL ES 3.0 emulator can handle ASTC textures (although its underlying technique is to decode them to raw images, the compression artefacts are left intact) so you can try them out in your shaders. To generate ASTC images you have two options: the command line evaluation codec or the texture compression toolBoth of these tools have a lot of preset modes and switches for different use cases.


Things already mentioned like block or channel weighting can be set easily in either tool to clean up specific error cases. Also there are preset modes for normal maps, which map the angular weighted X and Y of the normal to Luminance and Alpha for better compression, and data masking, which tells the encoder to treat errors in different channels separately so that they can encode unrelated non colour data.


Both tools are also capable of encoding volumetric 3D textures. Either of them will accept an uncompressed 3D image file, and the command line tool has commands for accepting an array of 2D slices to build the 3D volume.


In my main auditorium talk at GDC I gave a few more tips on working with compressed textures, and I’ll share those in another blog real soon. For now, download a compressor and have fun playing around with the future of texture compression.



Epic Giveaway.png

Another year has almost passed for the Mali team and, crikey, has it been a busy one. With the mobile market booming - particularly in the mainstream segment - demand for ARM® Mali GPUs has never been stronger and continues to mount. Our engineers are taking it in their stride, releasing new products across the board to drive the spread of stunning media experiences across the world. Our product roadmap is at its most flexible ever, offering a wide range of GPU, video and display solutions that scale from a single core Mali-400 for smartwatches to the recently announced Mali-T860 to power the next generation of premium smartphones. Meanwhile, in the current generation of premium smartphones, the Mali-T600 Series is shipping in volume and delivering the stunning graphics capabilities of the performance efficient Mali GPU roadmap to consumers who demand only the best media experiences.

If you’re looking this Holiday Season for a mobile device that delivers high quality graphics experiences without draining the battery, why not check out some of the following – all of which are new to the market this year and sport an ARM Mali GPU. You will also have the chance to win some of these fantastic devices as part of the Epic Giveaway, continue reading to find out more.

1.     Samsung Galaxy Note 4

Featuring an Exynos 7 Octa processor with a Mali-T760 MP6 GPU, the Samsung Galaxy Note 4 was released in September 2014 and will be part of the Epic Giveaway in January. With its pixel dense display, vibrant colours and powerful 16MP back facing camera the Samsung Galaxy Note 4 is a feature-rich phone that also provides astonishing performance for compute intensive applications such as high resolution, high end games.

2.     Huawei Honor 6

The Honor 6’s HiSilicon Kirin 920 processor features four ARM® Cortex®-A15 cores and four Cortex-A7 cores in big.LITTLE configuration alongside a Mali-T628 MP4 GPU to offer a good HD display and high performance within an affordable price bracket.

3.     HTC Desire 820s

With the 64-bit, Mali-T760 GPU-based Mediatek MT6752 SoC at its heart the HTC Desire 820s was announced on Sina Weibo, the Chinese social network site, in November.

4.     Pipo Pad-P1

Rockchip were the first to market with a Mali-T760 GPU-based SoC, our most advanced GPU shipping to date, releasing the first devices half a year after the cores were first announced.epic giveaway 2.jpg  Their fast time to market is setting the pace for the industry and the implication of the RK3288 is that, in the future, the latest graphics processing will be arriving to the hands of consumers quicker than ever before.

5.     Omate TrueSmart


With a single core Mali-400 GPU at the heart of the Mediatek MT6572, this little smartwatch delivers a smooth UI along with the ability to make calls, navigate and use Android apps independently of your smartphone.

6.     Samsung Galaxy S5

A final device of mention is the Samsung Galaxy S5 which has just entered ARM’s 2014 Epic Giveaway! Released at the start of the year in Barcelona, it sports a Samsung Exynos 5 Octa chipset with octa-core big.LITTLE technology, splitting CPU work across four Cortex-A7 and four Cortex-A15 processors, with an implementation of the Mali-T628 GPU IP delivering the multimedia experience. Enter the Epic Giveaway competition for your chance to get your hands on a Samsung Galaxy S5 this Holiday Season.



The 2014 Epic Giveaway gets underway today. In partnership with HEXUS, ARM is giving you the chance to win amazing new prizes this holiday season! Every day for the next few weeks, we'll be giving away a brand-new ARM-based device. We'll have an array of prizes from ARM and our partners, including Atmel, Nvidia and Samsung, plus many, many more! Each prize draw will be open for seven days, so visit the dedicated competition page to keep tabs on what's up for grabs and what's coming soon.

Click here to find out more and to enter the EPIC Giveaway for your chance to win.

The global smartphone market has witnessed extraordinary growth in recent years with shipments rising by 23% in 2014 to exceed the 1.2bn unit threshold. Market research firm CCS Insight forecasts smartphone shipments of 1.89bn units by 2018. This growth predominantly comes from the lower spectrum of the market with consumers upgrading from feature phones to the mainstream smartphones. ARM® Cortex® CPUs are in the heart of over 95% of the mobile and smartphones in the world ARM® Mali™ is the #1 licensable GPU with a mature and growing ecosystem of partners.


ARM and its partners have been at the forefront of innovation, delivering energy-efficient and affordable devices to millions of consumers who would like to be able to enjoy features of modern technology like mobile email, web browsing, video streaming and mobile gaming without paying a high price.


The ARM® Mali™-450 MP GPU has been enjoying popularity since its launch and is found in millions of entry to mid-range level smartphones, tablets and set-up-box. The Mali-450 MP has been designed for the volume market and optimized with a focus on energy and bandwidth savings. It is a perfect energy-efficient, cost and area optimized solution for a market that requires OpenGL ES 2.0 implementation.


Recently we announced a 64-bit driver for Mali-450MP enabling developers to take full advantage of the latest technology available on the market.

Previous blog in the series: Mali Performance 3: Is EGL_BUFFER_PRESERVED a good thing?

5 Principles.png


In my previous blogs in this series I have looked at the bare essentials for using a tile-based rendering architecture as efficiently as possible, in particular showing how to structure an application's use of framebuffers most efficiently to minimize unnecessary memory accesses. With those basics out of the way I can now start looking in more detail about how to drive the OpenGL ES API most efficiently to get the best out of a platform using Mali, but before we do that, I would like to introduce my five principles of performance optimization.


Principle 1: Know Your Goals


When starting an optimization activity have a clear set of goals for where you want to end up. There are many possible objectives to optimization: faster performance, lower power consumption, lower memory bandwidth, or lower CPU overhead to name the most common ones. The kinds of problems you look for when reviewing an application will vary depending on what type of improvement you are trying to make, so getting this right at the start is of critical importance.


It is also very easy to spend increasingly large amounts of time for smaller and smaller gains, and many optimizations will increase complexity of your application and make longer term maintenance problematic. Review your progress regularly during the work to determine when to say "we've done enough", and stop when you reach this point.


Principle 2: Don't Just Try to Make Things Fast


I am often asked by developers working with Mali how they can make a specific piece of content run faster. This type of question is then often quickly followed up by more detailed questions on how to squeeze a little more performance out of a specific piece of shader code, or how to tune a specific geometry mesh to best fit the Mali architecture. These are all valid questions, but in my opinion often unduly narrow the scope of the optimization activity far too early in the process, and leave many of the most promising avenues of attack unexplored.


Both of the questions above try to optimize a fixed workload, and both make the implicit assumption that the workload is necessary at all. In reality graphics scenes often contain a huge amount of redundancy - objects which are off screen, objects which are overdrawn by other objects, objects where half the triangles are facing away from the user, etc - which contribute nothing to the final render. Optimization activities therefore need to attempt to answer two fundamental questions:


  1. How do I remove as much redundant work from the scene as possible, as efficiently as possible?
  2. How do I fine tune the performance of what is left?


In short - don't just try to make something faster, try to avoid doing it at all whenever possible! Some of this "work avoidance" must be handled entirely in the application, but in many cases OpenGL ES and Mali provides tools which can help provided you use them correctly. More on this in a future blog.


Principle 3: Graphics is Art not Science


If you are optimizing a traditional algorithm on a CPU there is normally a right answer, and failure to produce that answer will result in a system which does not work. For graphics workloads we are simply trying to create a nice looking picture as fast as possible; if an optimized version is not bit-exact it is unlikely anyone will actually notice, so don't be afraid to play with the algorithms a little if it helps streamline performance.


Optimization activities for graphics should look at the algorithms used, and if their expense does not justify the visual benefits they bring then do not be afraid to remove them and replace them with something totally different. Real-time rendering is an art form, and optimization and performance is part of that art. In many cases smooth framerate and fast performance is more important than a little more detail packed into a single frame.


Principle 4: Data Matters


GPUs are data-plane processors, and graphics rendering performance is often dominated by data-plane problems. Many developers spend a long time looking at OpenGL ES API function call sequences to determine problems, without really looking at the data they are passing into those functions. This is nearly always a serious oversight.


OpenGL ES API call sequences are of course important, and many logic issues can be spotted by looking at these during optimization work, but remember that the format, size, and packing of data assets is of critical importance and must not be forgotten when looking for opportunities to make things faster.


Principle 5: Measure Early, Measure Often


The impact of a single draw call on scene rendering performance is often impossible to tell from the API level, and in many cases seemingly innocuous draw calls often have some of the largest performance overheads. I have seen many performance teams sink days or even weeks of time into optimization something, only to belatedly realise that the shader they have been tuning only contributes 1% of the overall cost of the scene, so while they have done a fantastic job and made it 2x faster that only improves overall performance by 0.5%.


I always recommend measuring early and often, using tools such as DS-5 Streamline to get an accurate view of the GPU hardware performance via the integrated hardware performance counters, and Mali Graphics Debugger to work out which draw calls are contributing to that rendering workload. Use the performance counters not only to identify hot spots to optimize, but also to sanity check what your application is doing against what you expect it to be doing. For example, manually estimate the number of pixels, texels, or memory accesses, per frame and compare this estimate against the counters from the hardware. If you see twice as many pixels as expected being rendered then there are possibly some structural issues to investigate first which could give much larger wins than simple shader tuning.


Next Time


The best optimizations in graphics are best tackled when made a structural part of how an application presents data to the OpenGL ES API, so in my next blog I will be looking at some of the things an application might want to consider when trying very hard to not do any work at all.




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 drivers even better.



This last blog in the Movie Vision App series, following on from The Movie Vision App: Part 1 and The Movie Vision App: Part 2, will discuss two final movie effect filters.



Movie Vision Filters: “Follow The White Rabbit…”


This is the most intriguing and complex filter in the Movie Vision demonstration. The camera preview image is replaced by a grid of small characters (primary Japanese Kana). The characters are coloured varying shades of green reminiscent of old computer displays. Additionally, the brightness is also manipulated to create the appearance of some characters ‘falling’ down the image. The overall impression is that the image is entirely composed of green, computer-code like characters.


      //Run the WhiteRabbitScript with the RGB camera input allocation.
      mWhiteRabbitScript.forEach_root(mWhiteRabbitInAlloc, mWhiteRabbitOutAlloc);
      //Make the heads move, dependant on the speed.
      for(int hp = 0; hp < mScreenWidth / mCharacterSize; hp++) {
          //If the character string has reached the bottom of the screen, wrap it back around.
          if(mHeadPos[hp] > mScreenHeight + 150) {
              mHeadPos[hp] = 0;
              mStrChar[hp] = mGen.nextInt(8)+1;
              mStrLen[hp] = mGen.nextInt(100)+50;
              mUpdate = true;
      //If a character string has reached the bottom, update the allocations with new random values.
      if(mUpdate) {
          mUpdate = false;

“Follow the White Rabbit” excerpt from processing of each camera frame


The Java component of this image filter does the standard RenderScript set up, but also populates several arrays to use in mapping the image to characters. The number of columns and rows of characters is calculated and a random index set for each column. A set of header positions and string lengths is also randomly generated for each column. These correspond to areas that will be drawn brighter than the rest of the image, to give the impression of falling strings of characters. On the reception of each camera preview frame, the standard YUV to RGB conversion is performed. Then, the RenderScript image effect script’s references to the character, position and length arrays are updated. The script kernel is executed. Afterwards, the header positions are adjusted so that the vertical brighter strings appear to fall down the image (and wrap back to the top).


static const int character1[mWhiteRabbitArraySize] = {0, 0, 1, 0, 0, 0,
                                                  0, 0, 1, 0, 0, 0,
                                                  1, 1, 1, 1, 1, 1,
                                                  0, 0, 1, 0, 0, 1,
                                                  0, 1, 0, 0, 0, 1,
                                                  1, 0, 0, 0, 1, 1};
static const int character2[mWhiteRabbitArraySize] = {0, 0, 1, 1, 1, 0,
                                                  1, 1, 1, 1, 1, 1,
                                                  0, 0, 0, 1, 0, 0,
                                                  0, 0, 0, 1, 0, 0,
                                                  0, 0, 1, 0, 0, 0,
                                                  0, 1, 0, 0, 0, 0};

“Follow the White Rabbit” RenderScript character setup


This is by far the most complicated RenderScript kernel in the Movie Vision app. The script file starts with eight statically defined characters from the Japanese Kana alphabet. These are defined as 6x6 arrays. The first line of the script execution is a conditional statement – the script only executes on every eighth pixel in both the x and y direction. So, the script executes ‘per character’ rather than ‘per pixel’. As we use 6x6 characters, this gives a one pixel border to each character. The output colour for the current position is set to a default green value, based on the input colour. The character index, header position and length values are retrieved from the arrays managed by the Java class. Next, we determine if the character corresponding to the current pixel is in our bright ‘falling’ string, and adjust the green value appropriately: brightest at the head, gradually fading behind and capped at a lower maximum value elsewhere. If the current character position isn’t at the front of the falling string, we also pseudo randomly change the character to add a dynamic effect to the image. Next, some basic skin tone detection is used to further brighten the output if skin is indeed detected. Finally, the output values for all pixels in the current character position are set.


      //Sets the initial green colour, which is later modified depending on the in pixel.
      refCol.r = 0;
      refCol.g = in->g;
      refCol.b = in->g & 30;
//If the Y position of this pixel is the same as the head position in this column.
        if(y == currHeadPos)
            refCol.g = 0xff; //Set it to solid green.
        //If the character is within the bounds of the falling character string for that column, make it darker the further away
        //from the head it is.
        else if((y < currHeadPos && y >= (currHeadPos - currStringLength)) || (y < currHeadPos && (currHeadPos - currStringLength) < 0))
            refCol.g = 230 - ((currHeadPos - y));
        else if(refCol.g > 150) //Cap the green at 150.
            refCol.g -= 100;
            refCol.g += refCol.g | 200; //For every other character, make it brighter.
      //If the current character isn't the head, randomly change it.
      if(y != currHeadPos)
            theChar += *(int*)rsGetElementAt(stringChars, (y/mWhiteRabbitCharSize));
      //Basic skin detection to highlight people.
      if(in->r > in->g && in->r > in->b) {
            if(  in->r > 100 && in->g > 40
              && in->b > 20 && (in->r - in->g) > 15)
                refCol.g += refCol.g & 255;
      //Loop through the binary array of the current character.
      for(int py = 0; py < mWhiteRabbitCharSize; py++){
          for(int px = 0; px < mWhiteRabbitCharSize; px++){
                out[(py*mWidth)+px].r = 0;
                out[(py*mWidth)+px].g = 0;
                out[(py*mWidth)+px].b = 0;
                if(theChar == 1) {
                    if(character1[(py*(mWhiteRabbitCharSize))+px] == 1)
                      out[(py*mWidth)+px] = refCol;
                }else if(theChar == 2) {
                    if(character2[(py*(mWhiteRabbitCharSize))+px] == 1)
                      out[(py*mWidth)+px] = refCol;

Excerpts of “Follow the White Rabbit” Renderscript Kernel root function



Movie Vision Filters: “Why So Serious?”



This filter mimics a sonar vision effect. Part of this is a simple colour mapping to a blue toned image. In addition, areas of the image are brightened relative to the amplitude of sound samples from the microphone.


    mRecorder = new MediaRecorder();
    try {
        mRecording = true;
    } catch (IOException ioe){
        mRecording = false;

“Why so serious?” setting up the microphone


The Java side of this filter does the standard configuration for a RenderScript kernel. It also sets up the Android MediaRecorder to constantly record sound, but dumps the output to /dev/null. A set of look-up tables, similar to the ‘Get to the chopper’ filter, are used to do the colour mapping. References to these are passed to the script. For each camera preview frame, the maximum sampled amplitude since the last frame and a random x and y position are passed to the RenderScript kernel. The image is converted to RGB and then the image effect kernel is executed.


    //If the current pixel is within the radius of the circle, apply for 'pulse' effect colour.
    if (((x1*x1)+(y1*y1)) < (scaledRadius*scaledRadius)){
        dist = sqrt((x1*x1)+(y1*y1));
        if (dist < scaledRadius){
            effectFactor = (dist/scaledRadius) * 2;
            lightLevel *= effectFactor;
            blue -= lightLevel;
    //Lookup the RGB values based on the external lookup tables.
    uchar R = *(uchar*)rsGetElementAt(redLUT, blue);
    uchar G = *(uchar*)rsGetElementAt(greenLUT, blue);
    uchar B = *(uchar*)rsGetElementAt(blueLUT, blue);
    //Clamp the values between 0-255
    R > 255? R = 255 : R < 0? R = 0 : R;
    G > 255? G = 255 : G < 0? G = 0 : G;
    B > 255? B = 255 : B < 0? B = 32 : B;
    //Set the final output RGB values.
    out->r = R;
    out->g = G;
    out->b = B;
    out->a = 0xff;

“Why So Serious?” RenderScript Kernel root function


The RenderScript kernel calculates a brightness, radius and offset for a ‘pulse’ effect based on the amplitude and position passed to it. If the current pixel is within the pulse circle, it is brightened considerably. The output colour channels for the pixel are then set based on the lookup tables defined in the Java file.



Movie Vision: Conclusions


Can you guess which movies inspired “Follow the White Rabbit” and “Why So Serious?” ?


At the beginning of this blog series we stated that the Movie Vision app was conceived as a demonstration to highlight heterogeneous computing capabilities in mobile devices. Specifically, we used RenderScript on Android to show the GPU Compute capabilities of ARM® Mali™ GPU technology. As a proof of concept and a way to explore one of the emerging GPU computing programming frameworks, Movie Vision has been very successful: RenderScript has proven to be an easy to use API. It is worth noting that it is highly portable, leveraging both ARM CPU and GPU technology. The Movie Vision App explored a fun and entertaining use-case, but it is only one example of the potential of heterogeneous approaches like GPU Compute.


We hope you have enjoyed this blog series, and that this inspires you to create your own applications that explore the capabilities of ARM technology.




Creative Commons License

This work by ARM is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. However, in respect of the code snippets included in the work, ARM further grants to you a non-exclusive, non-transferable, limited license under ARM’s copyrights to Share and Adapt the code snippets for any lawful purpose (including use in projects with a commercial purpose), subject in each case also to the general terms of use on this site. No patent or trademark rights are granted in respect of the work (including the code snippets).

We are pleased to release a new version of the Mali OpenGL ES Emulator, v2.0-BETA*, which adds support for OpenGL ES 3.1.

From Khronos website:

“OpenGL ES 3.1 provides the most desired features of desktop OpenGL 4.4 in a form suitable for mobile devices,” said Tom Olson, chair of the OpenGL ES working group and Director of Graphics Research at ARM. “It provides developers with the ability to use cutting-edge graphics techniques on devices that are shipping today.”

The OpenGL® ES Emulator is a library that maps OpenGL ES 3.1 API calls to the OpenGL API. By running on a standard PC, the emulator helps software development and testing of next generation OpenGL ES 3.1 applications since no embedded platform is required.

Get the Mali OpenGL ES Emulator




Following on from The Movie Vision App: Part 1, in Part 2, we’ll immediately move on and discuss two more image filters implemented for the project.



Movie Visions Filters: “An Old Man Dies…”



This filter is only slightly more complex than the “I’ll be back” effect described in the previous blog. The camera preview image is filtered to a black and white, grainy ‘comic book’ style, but any objects detected as red retain their colour.

The Java portion of the filter does the standard RenderScript initialisation. A Script Intrinsic is used to convert the YUV camera preview data to an RGB array, and a second custom script applies the actual visual effect.


Function: root
param uchar4 *in        The current RGB pixel of the inout allocation.
param uchar4 *out      The current RGB pixel of the output allocation.
param uint32_t x        The X position of the current pixel.
param uint32_t          The Y position of the current pixel.
void root(const uchar4 *in, uchar4 *out, uint32_t x, uint32_t y){
    //The black and white output char.
    uchar4 bw;
    //Range between -120 and 120, 120 being the highest contrast.
    //We're applying this to make a high-contrast image.
    int contrast = 120;
    float factor = (255 * (contrast + 255)) / (255 * (255 - contrast));
    int c = trunc(factor * (in->r - 128) + 128)-50;
    if(c >= 0 && c <= 255)
        bw.r = c;
        bw.r = in->r;
    //Now determine if we apply a 'grain' effect to this pixel - every 4th pixel
    //If the current pixel is divisible by 4, apple a 'grain' effect.
    if(x % 4 == 0 && y % 4 == 0)
        bw.r &= in->g;
    //Finally determine if this pixel is 'red' enough to be left as red...
    //Red colour threshhold.
    if (in->r > in->g+55 && in->r > in->b+60) {
        //Only show the red channel.
        bw.g = 0;
        bw.b = 0;
    } else {
        //Make all colour channels the same (Black & White).
        bw.g = bw.r;
        bw.b = bw.r;
    //Set the output pixel to the new black and white one.
    *out = bw;

“An Old Man Dies” RenderScript Kernel root function


First, we apply a formula to calculate a colour value for the pixel that will result in a high contrast black & white image. The value for every fourth pixel is further modified to stand out, resulting in a slight grain effect. Finally, if the pixel’s red colour value exceeds a certain threshold, only the red channel for that pixel is shown. Otherwise, the blue and green channels are set to the same value as the red to achieve the black & white look.

Once again, can you guess the movie that inspired this filter?



Movie Vision Filters: “Get To The Chopper…”


This filter creates a ‘thermal camera’ effect, and also applies a Heads Up Display (HUD) type decoration. The colour filtering utilises RenderScript, whilst the HUD leverages Android’s built in face detection. A set of look-up tables map specific colour ranges to output colours. Thermal cameras generally map infrared to a narrow set of output colours. This image filter mimics this by mapping input image colours to a similar set of output colours.


* Creates the lookup table use for the 'heat map' splitting the image int
* 16 different colours.
private void createLUT() {
    final int SPLIT = 8;
    for (int ct = 0; ct < mMaxColour/SPLIT; ct++){
        for (int i = 0; i < SPLIT; i++){
            switch (ct) {
                * The following cases define a set of colours.
                case (7):
                    mRed[(ct*SPLIT) +i] = 0;
                    mGreen[(ct*SPLIT) +i] = 255;
                    mBlue[(ct*SPLIT) +i] = 0;
                case (6):
                    mRed[(ct*SPLIT) +i] = 128;
                    mGreen[(ct*SPLIT) +i] = 255;
                    mBlue[(ct*SPLIT) +i] = 0;

“Get to the Chopper” creating look-up tables


On the Java side, along with setting up the typical set up of Allocation objects to pass input and receive output from RenderScript, three look-up tables are defined: one each for the red, green and blue colour channels. Each look-up table is essentially an array of 255 values, giving the output value for each of the possible input values of the colour channel. Each frame is again first converted to RGB before being passed to the image effect RenderScript kernel. After the RenderScript filtering, the decoration drawing callback is used to draw a red, triangular ‘targeting’ reticule on any faces that were detected by the Android face detection API.


    //Basic skin detection.
    //These values specifically filter out skin colours.
    if(in->r > in->g+10 && in->r > in->b+5 && in->g < 120) {
        //If skin has been detected, apply the 'hotter' colours.
        out->r = in->r & 40;
        out->g = in->g & 40;
        out->b = 24;
        out->a = 0xff;
    //Use the external lookup allocations to dertermine the colour.
    out->r = *(uchar*)rsGetElementAt(redLUT, in->r);
    out->g = *(uchar*)rsGetElementAt(greenLUT, in->g);
    out->b = *(uchar*)rsGetElementAt(blueLUT, in->b);

“Get to the Chopper” RenderScript Kernel root function


The RenderScript script for this effect is very simple. For each pixel, it first checks if the RGB values fall within a range considered a ‘skin tone’. If so, the output is forced to the ‘hot’ output colours. Otherwise, the output values for the pixel are set directly from the pre-configured look-up tables for each channel.


Which movie inspired “Get to the Chopper”? As a hint, it features the same actor as “I’ll be back”.


That concludes this second Movie Vision App blog. Read on for the most complex image effects of the Movie Vision App and some concluding comments in The Movie Vision App: Part 3!





Creative Commons License

This work by ARM is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. However, in respect of the code snippets included in the work, ARM further grants to you a non-exclusive, non-transferable, limited license under ARM’s copyrights to Share and Adapt the code snippets for any lawful purpose (including use in projects with a commercial purpose), subject in each case also to the general terms of use on this site. No patent or trademark rights are granted in respect of the work (including the code snippets).



The Applied Systems Engineering (ASE) team at ARM is responsible for both implementing and presenting technical demonstrations that show ARM® technology in context. These demonstrations make their way to trade shows and other events and are also often discussed in online features such as this series of blogs. Here we introduce our Movie Vision App which explores the use of GPU Compute.



The Movie Vision Concept


The Movie Vision app was conceived as a demonstration to highlight emerging heterogeneous computing capabilities in mobile devices. Specifically, it makes use of Android’s RenderScript computation framework. There is a great deal of discussion about GPU Compute’s capabilities and a variety of exciting new use-cases have been highlighted. On the ARM Connected Community, you can find discussions on the advantages and benefits of GPU Compute, the methods available and details on the compute architecture provided by ARM® Mali™ GPUs. The objective of the Movie Vision application is to provide a visually compelling demonstration that leverages the capabilities of ARM technology and GPU Compute, to explore the RenderScript API and to provide an example application for the Mali Developer community.


Movie Vision takes the preview feed from an Android device’s camera, applies a visual effect to that feed and displays it on the device screen. A number of visual effect filters have been implemented, each modeled on various special effects seen in popular movies over the years. Frame-by-frame, this breaks down to applying one or more mathematical operations to a large array of data – a task well suited to the kind of massive parallelism that GPU Compute provides.


In this series of blogs, we’ll go through each of the visual effect filters we implemented and use these to explore and understand the capabilities of the RenderScript API.





RenderScript is well described on the Android Developers website, an excellent place to start for details and instructions on its use. To summarize, from a developer’s standpoint, in using RenderScript you will be writing your high-performance ‘kernel’ in C99 syntax and then utilizing a Java API to manage the use of this by your application, and to manage the data going into and out of it. These kernels are parallel functions executed on every element of the data you pass into the script. Under the hood, RenderScript code is first compiled to intermediate byte-code, and then further compiled at runtime to machine code by a sequence of Low Level Virtual Machine (LLVM) compilers. This is not conceptually dissimilar to the standard Android or Java VM model, but obviously more specialized. The final machine code is generated by an LLVM on the device, and optimized for that device. On the Google Nexus 10 used for development of the Movie Vision application, RenderScript would thus make use of either the dual core ARM Cortex®-A15 CPU or the GPU Compute capabilities of the Mali-T604 GPU.



Movie Vision Application Structure


The Movie Vision app has the following structure:




- Main Android activity class

- UI layout/functionality

- Setup of camera preview

- Setup of face detection

ImageFilterOverlayView- Allows rendering of icons & text decorations on top of filtered camera preview image
Image Filters (Java)

- Set-up for sharing data with RenderScript kernels

- Callback for each camera preview frame

- Callback for rendering decorations

Image Filters (RenderScript)- Application of image filter operations to image data



The functionality of the app is fairly simple. The Android Camera API provides a hook to receive a camera preview callback, each such call delivering a single frame from the camera. The main Movie Vision Activity receives this and passes the frame data to the currently selected image filter. After the frame has been processed, the resulting image is rendered to the screen. A further call back to the selected filter allows decorations such as text or icons to be rendered on top of the image. The Android AsyncTask class is used to decouple image processing from the main UI thread.


Each Image Filter shares some relatively common functionality. All of them perform a conversion of the camera preview data from YUV to RGB. The data from the camera is in YUV, but the filter algorithms and output for Movie Vision require RGB values. The Android 4.2 releases included updates to RenderScript which added an “Intrinsic” script to perform this operation. A RenderScript Intrinsic is an efficient implementation of a common operation. These include Blends, Blurs, Convolutions and other operations – including this YUV to RGB conversion. More information can be found on the Android Developer Website. Each Image Filter Java class also configures its ‘effect’ script. Configuration generally consists of allocating some shared data arrays (using the Allocation object) for input and output and allocating or setting any other values required by the script.


* RenderScript Setup
mTermScript = new ScriptC_IllBeBack(mRS, res, R.raw.IllBeBack);

Type.Builder tb = new Type.Builder(mRS, Element.RGBA_8888(mRS));

mScriptOutAlloc = Allocation.createTyped(rs, tb.create());
mScriptInAlloc = Allocation.createSized(rs, Element.U8(mRS), (mHeight * mWidth) + ((mHeight / 2) * (mWidth / 2) * 2));

Initial set up of a RenderScript kernel



Movie Vision Filters: “I’ll Be Back”



This filter applies a famous movie effect with a red tint and an active Heads-Up Display that highlights faces. It is probably the simplest Movie Vision effect in terms of the RenderScript component. However, a desired additional feature of this filter highlights some challenges.


* Processes the current frame. The filter first converts the YUV data
* to RGB via the conversion script. Then the Renderscript kernel is run, which applies a
* red hue to the image. Finally the filter looks for faces and objects, and on finding one
* draws a bounding box around it.
* @param data The raw YUV data.
* @param bmp Where the result of the ImageFilter is stored.
* @param lastMovedTimestamp Last time the device moved.
public void processFrame(byte[] data, Bitmap bmp, long lastMovedTimestamp){
    mConv.forEach_root(mScriptInAlloc, mScriptOutAlloc);
    mTermScript.forEach_root(mTermOutAlloc, mTermOutAlloc);

“I’ll Be Back” processing each camera frame


The Java component of this filter is relatively straight forward. The YUV/RGB conversion and image effect RenderScript kernels are configured. For each camera preview frame, we convert to RGB and pass the image to the effect kernel. After that, in a second pass on the frame, we render our HUD images and text if any faces have been detected. This draws a box around the face and prints out some text to give the impression that the face is being analyzed.


Function: root
param uchar4 *in        The current RGB pixel of the inout allocation.
param uchar4 *out      The current RGB pixel of the output allocation.
void root(const uchar4 *in, uchar4 *out){
  uchar4 p = *in;
  //Extracting the red channel, ignoring the green and blue channels. Creates the red 'HUE' effect.
  out->r = p.r & 0xff;
  out->b = p.g & 0x00;
  out->g = p.b & 0x00;
  out->a = 0xff;

“I’ll Be Back” RenderScript Kernel root function


The RenderScript component is very simple. The output green and blue channels are zeroed, so that just the red channel is visible in the final image.


Initially, an attempt was made to add pseudo object detection to this effect, such that ‘objects’ as well as faces would be highlighted by the HUD. A prototype using the OpenCV library was implemented, using an included library implementation of an algorithm for Contour Detection. It is worth noting that this approach would not utilise GPU Compute and run only on the CPU. Contour Detection is a relatively complex multi-stage computer vision algorithm. First, a Sobel Edge Detection filter is applied to bring out the edges of the image. Then, a set of steps to identify joined edges (contours) is applied. The prototype object detection then interpreted this to find contours in the image that were likely to be objects. Generally, large, rectangular regions were chosen. One of these would be selected and highlighted with the HUD decorations as an ‘object of interest’.


The issue with this object detection prototype was that it required several passes of algorithmic steps, with intermediate data sets. Porting this to RenderScript to take advantage of GPU Compute would have resulted in several chained together RenderScript scripts.  At the time of initial development this resulted in some inherent inefficiencies, although the addition of ‘Script Groups’ in Android 4.2 will have helped to address this. For now, porting the Contour Detection algorithm to RenderScript remains an outstanding project.


That concludes the first blog in this series on the Movie Vision App. Carry on reading with The Movie Vision App: Part 2 for examples of increasingly complex visual effect filters. Can you guess the movie that inspired the “I’ll be back” filter?




Creative Commons License

This work by ARM is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. However, in respect of the code snippets included in the work, ARM further grants to you a non-exclusive, non-transferable, limited license under ARM’s copyrights to Share and Adapt the code snippets for any lawful purpose (including use in projects with a commercial purpose), subject in each case also to the general terms of use on this site. No patent or trademark rights are granted in respect of the work (including the code snippets).

Mali Graphics Debugger v1.3.2 now supports the most recent version of Android, 5.0 “Lollipop”.

Mali Graphics Debugger is an advanced API tracer tool for OpenGL ES, EGL and OpenCL. It allows developers trace their graphics and compute applications to debug issues and analyze the performance.

It also supports Linux targets and runs on Windows, Linux and Mac.

Get Mali Graphics Debugger

Mali Graphics Debugger v1.3.2

Filter Blog

By date:
By tag: