GPU optimization

Introduction

The demand for new graphics features and progress almost guarantees that you will encounter graphics bottlenecks. Some of these can be on the CPU side, for instance in calculations inside the Redot engine to prepare objects for rendering. Bottlenecks can also occur on the CPU in the graphics driver, which sorts instructions to pass to the GPU, and in the transfer of these instructions. And finally, bottlenecks also occur on the GPU itself.

Where bottlenecks occur in rendering is highly hardware-specific. Mobile GPUs in particular may struggle with scenes that run easily on desktop.

Understanding and investigating GPU bottlenecks is slightly different to the situation on the CPU. This is because, often, you can only change performance indirectly by changing the instructions you give to the GPU. Also, it may be more difficult to take measurements. In many cases, the only way of measuring performance is by examining changes in the time spent rendering each frame.

Draw calls, state changes, and APIs

Note

The following section is not relevant to end-users, but is useful to provide background information that is relevant in later sections.

Redot sends instructions to the GPU via a graphics API (OpenGL, OpenGL ES or Vulkan). The communication and driver activity involved can be quite costly, especially in OpenGL and OpenGL ES. If we can provide these instructions in a way that is preferred by the driver and GPU, we can greatly increase performance.

Nearly every API command in OpenGL requires a certain amount of validation to make sure the GPU is in the correct state. Even seemingly simple commands can lead to a flurry of behind-the-scenes housekeeping. Therefore, the goal is to reduce these instructions to a bare minimum and group together similar objects as much as possible so they can be rendered together, or with the minimum number of these expensive state changes.

2D batching

In 2D, the costs of treating each item individually can be prohibitively high - there can easily be thousands of them on the screen. This is why 2D batching is used. Multiple similar items are grouped together and rendered in a batch, via a single draw call, rather than making a separate draw call for each item. In addition, this means state changes, material and texture changes can be kept to a minimum.

For more information on 2D batching, see Optimization using batching.

3D batching

In 3D, we still aim to minimize draw calls and state changes. However, it can be more difficult to batch together several objects into a single draw call. 3D meshes tend to comprise hundreds or thousands of triangles, and combining large meshes in real-time is prohibitively expensive. The costs of joining them quickly exceeds any benefits as the number of triangles grows per mesh. A much better alternative is to join meshes ahead of time (static meshes in relation to each other). This can either be done by artists, or programmatically within Redot.

There is also a cost to batching together objects in 3D. Several objects rendered as one cannot be individually culled. An entire city that is off-screen will still be rendered if it is joined to a single blade of grass that is on screen. Thus, you should always take objects' location and culling into account when attempting to batch 3D objects together. Despite this, the benefits of joining static objects often outweigh other considerations, especially for large numbers of distant or low-poly objects.

For more information on 3D specific optimizations, see Optimizing 3D performance.

Reuse Shaders and Materials

The Redot renderer is a little different to what is out there. It's designed to minimize GPU state changes as much as possible. SpatialMaterial does a good job at reusing materials that need similar shaders. If custom shaders are used, make sure to reuse them as much as possible. Redot's priorities are:

• Reusing Materials: The fewer different materials in the scene, the faster the rendering will be. If a scene has a huge amount of objects (in the hundreds or thousands), try reusing the materials. In the worst case, use atlases to decrease the amount of texture changes.

• Reusing Shaders: If materials can't be reused, at least try to reuse shaders. Note: shaders are automatically reused between SpatialMaterials that share the same configuration (features that are enabled or disabled with a check box) even if they have different parameters.

If a scene has, for example, 20,000 objects with 20,000 different materials each, rendering will be slow. If the same scene has 20,000 objects, but only uses 100 materials, rendering will be much faster.

Pixel cost versus vertex cost

You may have heard that the lower the number of polygons in a model, the faster it will be rendered. This is really relative and depends on many factors.

On a modern PC and console, vertex cost is low. GPUs originally only rendered triangles. This meant that every frame:

1. All vertices had to be transformed by the CPU (including clipping).

2. All vertices had to be sent to the GPU memory from the main RAM.

Nowadays, all this is handled inside the GPU, greatly increasing performance. 3D artists usually have the wrong feeling about polycount performance because 3D DCCs (such as Blender, Max, etc.) need to keep geometry in CPU memory for it to be edited, reducing actual performance. Game engines rely on the GPU more, so they can render many triangles much more efficiently.

On mobile devices, the story is different. PC and console GPUs are brute-force monsters that can pull as much electricity as they need from the power grid. Mobile GPUs are limited to a tiny battery, so they need to be a lot more power efficient.

To be more efficient, mobile GPUs attempt to avoid overdraw. Overdraw occurs when the same pixel on the screen is being rendered more than once. Imagine a town with several buildings. GPUs don't know what is visible and what is hidden until they draw it. For example, a house might be drawn and then another house in front of it (which means rendering happened twice for the same pixel). PC GPUs normally don't care much about this and just throw more pixel processors to the hardware to increase performance (which also increases power consumption).

Using more power is not an option on mobile so mobile devices use a technique called tile-based rendering which divides the screen into a grid. Each cell keeps the list of triangles drawn to it and sorts them by depth to minimize overdraw. This technique improves performance and reduces power consumption, but takes a toll on vertex performance. As a result, fewer vertices and triangles can be processed for drawing.

Additionally, tile-based rendering struggles when there are small objects with a lot of geometry within a small portion of the screen. This forces mobile GPUs to put a lot of strain on a single screen tile, which considerably decreases performance as all the other cells must wait for it to complete before displaying the frame.

To summarize, don't worry about vertex count on mobile, but avoid concentration of vertices in small parts of the screen. If a character, NPC, vehicle, etc. is far away (which means it looks tiny), use a smaller level of detail (LOD) model. Even on desktop GPUs, it's preferable to avoid having triangles smaller than the size of a pixel on screen.

Pay attention to the additional vertex processing required when using:

• Skinning (skeletal animation)

• Morphs (shape keys)

• Vertex-lit objects (common on mobile)

Pixel/fragment shaders and fill rate

In contrast to vertex processing, the costs of fragment (per-pixel) shading have increased dramatically over the years. Screen resolutions have increased (the area of a 4K screen is 8,294,400 pixels, versus 307,200 for an old 640×480 VGA screen, that is 27x the area), but also the complexity of fragment shaders has exploded. Physically-based rendering requires complex calculations for each fragment.

You can test whether a project is fill rate-limited quite easily. Turn off V-Sync to prevent capping the frames per second, then compare the frames per second when running with a large window, to running with a very small window. You may also benefit from similarly reducing your shadow map size if using shadows. Usually, you will find the FPS increases quite a bit using a small window, which indicates you are to some extent fill rate-limited. On the other hand, if there is little to no increase in FPS, then your bottleneck lies elsewhere.

You can increase performance in a fill rate-limited project by reducing the amount of work the GPU has to do. You can do this by simplifying the shader (perhaps turn off expensive options if you are using a SpatialMaterial), or reducing the number and size of textures used.

When targeting mobile devices, consider using the simplest possible shaders you can reasonably afford to use.

Reading textures

The other factor in fragment shaders is the cost of reading textures. Reading textures is an expensive operation, especially when reading from several textures in a single fragment shader. Also, consider that filtering may slow it down further (trilinear filtering between mipmaps, and averaging). Reading textures is also expensive in terms of power usage, which is a big issue on mobiles.

If you use third-party shaders or write your own shaders, try to use algorithms that require as few texture reads as possible.

Texture compression

By default, Redot compresses textures of 3D models when imported using video RAM (VRAM) compression. Video RAM compression isn't as efficient in size as PNG or JPG when stored, but increases performance enormously when drawing large enough textures.

This is because the main goal of texture compression is bandwidth reduction between memory and the GPU.

In 3D, the shapes of objects depend more on the geometry than the texture, so compression is generally not noticeable. In 2D, compression depends more on shapes inside the textures, so the artifacts resulting from 2D compression are more noticeable.

As a warning, most Android devices do not support texture compression of textures with transparency (only opaque), so keep this in mind.

Note

Even in 3D, "pixel art" textures should have VRAM compression disabled as it will negatively affect their appearance, without improving performance significantly due to their low resolution.

Post-processing and shadows

Post-processing effects and shadows can also be expensive in terms of fragment shading activity. Always test the impact of these on different hardware.

Reducing the size of shadowmaps can increase performance, both in terms of writing and reading the shadowmaps. On top of that, the best way to improve performance of shadows is to turn shadows off for as many lights and objects as possible. Smaller or distant OmniLights/SpotLights can often have their shadows disabled with only a small visual impact.

Transparency and blending

Transparent objects present particular problems for rendering efficiency. Opaque objects (especially in 3D) can be essentially rendered in any order and the Z-buffer will ensure that only the front most objects get shaded. Transparent or blended objects are different. In most cases, they cannot rely on the Z-buffer and must be rendered in "painter's order" (i.e. from back to front) to look correct.

Transparent objects are also particularly bad for fill rate, because every item has to be drawn even if other transparent objects will be drawn on top later on.

Opaque objects don't have to do this. They can usually take advantage of the Z-buffer by writing to the Z-buffer only first, then only performing the fragment shader on the "winning" fragment, the object that is at the front at a particular pixel.

Transparency is particularly expensive where multiple transparent objects overlap. It is usually better to use transparent areas as small as possible to minimize these fill rate requirements, especially on mobile, where fill rate is very expensive. Indeed, in many situations, rendering more complex opaque geometry can end up being faster than using transparency to "cheat".

Multi-platform advice

If you are aiming to release on multiple platforms, test early and test often on all your platforms, especially mobile. Developing a game on desktop but attempting to port it to mobile at the last minute is a recipe for disaster.

In general, you should design your game for the lowest common denominator, then add optional enhancements for more powerful platforms. For example, you may want to use the GLES2 backend for both desktop and mobile platforms where you target both.

Mobile/tiled renderers

As described above, GPUs on mobile devices work in dramatically different ways from GPUs on desktop. Most mobile devices use tile renderers. Tile renderers split up the screen into regular-sized tiles that fit into super fast cache memory, which reduces the number of read/write operations to the main memory.

There are some downsides though. Tiled rendering can make certain techniques much more complicated and expensive to perform. Tiles that rely on the results of rendering in different tiles or on the results of earlier operations being preserved can be very slow. Be very careful to test the performance of shaders, viewport textures and post processing.