Everybody's heard stories about Mobile GPUs being different from desktop GPUs in that they do "deferred" rendering instead of "immediate" rendering.

Do we actually know what this means? What are the implications for the performance of our mobile gfx code? What should we change?

Gathering raw documentation from the source

There doesn't seem to exist a good source of information on "deferred" GPUs in general. Worse, "deferred" means different things on different GPUs with different performance implications.

So for lack of something better, let's start with the only information available: documentation from GPU vendors, that's going to be biased for their own GPUs. Once we have that, we try below to aggregate that information into more generally useful knowledge.

ARM (Mali)

Mali Developer Center

Mali GPU Application Optimization Guide (2011)

Qualcomm (Adreno)

Adreno 200 Performance Optimization (2010) Adreno Tiling

NVIDIA (Tegra)

Tegras are the only major mobile GPUs that are immediate, like desktop GPUs --- and not deferred like other mobile GPUs.

NVIDIA White Papers

NVIDIA Tegra 4 Family GPU Architecture

Imagination Technologies (PowerVR)

POWERVR Series5 Graphics

PowerVR -- A Master Class in Graphics Technology and Optimization (2012)

More documentations seems to exist at their "PowerVR insider" site but there appears to be a paywall there.

Intel

Intel currently just licenses PowerVR. They seem to be preparing different hardware for the future, but for now we only need to care about PowerVR-based Intel.

Other GPU vendors without known public documentation

Broadcom VideoCore

Vivante

What does "deferred" actually mean in various GPUs?

The best document that I could find on this is POWERVR Series5 Graphics, section 3. However, we won't use exactly the terminology of this document because it reserves the word "deferred" solely for PowerVR's version of "deferred".

In our terminology, there are 3 types of GPUs: immediate, tile-based deferred rasterization, and tile-based deferred HSR, where HSR stands for Hidden Surface Removal.

We will abbreviate "tile-based deferred rasterization" as tbd-rast and "Tile-based deferred HSR" as tbd-hsr.

Here's a table summarizing how this maps to various GPU vendors' terminology, what GPUs fall into which category, and what each term actually means.

Our terminology	Immediate	Deferred
		Tile-based deferred rasterization, abbreviated as tbd-rast	Tile-based deferred HSR, abbreviated as tbd-hsr
ImgTec terminology	Immediate rendering	Tile-based rendering (TBR)	Tile-based deferred rendering (TBDR)
ARM terminology	Immediate rendering	Interchangeably "tile-based rendering" or "tile-based deferred rendering"
Hardware	NVIDIA Tegra, desktops	ARM Mali, Qualcomm Adreno	ImgTec PowerVR
Meaning	Submitted geometry is immediately rendered; no tiling is used.	Submitted geometry is immediately transformed and stored in per-tile lists. Rasterization is then done separately for each tile.	Submitted geometry is immediately transformed and stored in per-tile lists. HSR is then done for each tile, yielding a list of visible fragments.
Performance implications	Good old desktop GPU optimization	Optimizations discussed below for deferred GPUs	Optimizations discussed below for deferred GPUs; the only difference is that there is no need for front-to-back sorting, as HSR is efficiently handled by hardware.

In a tbd-rast GPU, upon submitting geometry, vertex shaders are run and resulting triangles are clipped, but instead of proceeding further down the pipeline as an immediate renderer would, the resulting triangles are only recorded in tile-specific triangle lists. The actual rasterization of the triangles in each tile is delayed until the frame needs to be resolved, whence the name: tile-based deferred rasterization. Deferring rasterization until all the triangles in a given tile are known, allows tbd-rast GPUs to achieve higher efficiency over immediate GPUs, if only through higher cache coherency of framebuffer accesses --- in practice, the tile size is small enough that the framebuffer tile will fit in cache memory, considerably limiting framebuffer memory bandwidth. There probably are more gains too, although they will depend on GPU specifics. For example, deferred rendering may allow GPUs to sort primitives by textures, achieving higher texture cache coherency.

All of that also applies to tbd-hsr GPUs such as PowerVR's, which are similar to tbd-rast GPUs except for an additional optimization that they automatically perform: when a tbd-hsr GPU is about to start rasterizing the triangles in a given tile, it first identifies for each fragment which primitives may be visible at that fragment: see Section 4.4 in this PowerVR document. What this means in practice is that a tbd-hsr GPU will be equally efficient regardless of the ordering of opaque primitives, whereas other types of GPUs will perform better if opaque geometry is submitted in front-to-back order.

Notes about tile-based GPUs

Traditional OpenGL pipeline requires a lot of memory bandwidth which is very bad for power consumption. Mobile GPUs try to alleviate that by moving the frame buffer out of main memory, into high-speed on-chip memory (GMEM). This memory is very fast and power-efficient, but also very small. So the GPU will break up the framebuffer into smaller tiles and render to them using the high-speed memory one after the other. The size of tile buffers varries accross hardware, but it can be as small as 16x16 pixels tiles. To avoid overdraw, the GPU collects all the geometry, computes culling based on the output of the vertex shader and store the result in a spatial data structure for later use. When a tile is rendered, the data structre is consulted to see which triangles are relevent.

Performance implications of "deferred"

Draw-calls are expensive

They are always expensive regardless of the type of GPU, but talking with ARM people it sounded like there was something about deferred that made it worse. The ARM document explains: "The processing required by draw calls include allocating memory, copying data, and processing data. The overhead is the same whether you draw a single triangle or thousands of triangles in a draw call."

Replacing a texture image can cause a pipeline stall

Since deferred GPUs accept new GL calls while they are still resolving the previous frame, if a texImage2D call replaces an image in a texture that is still being sampled, this will stall the pipeline.

Anything that can force immediate framebuffer resolving is expensive

Forcing immediate framebuffer resolving negates the benefits of deferred renderers and turns them into liabilities.

This can be caused by changing the framebuffer binding, or by anything that will depend on the framebuffer's pixel values.

Framebuffer bindings are expensive

Changing the framebuffer binding forces immediately resolving the rendering of the current framebuffer. Therefore it is important to sort rendering to minimize framebuffer bindings. The Adreno 200 document, Section 3.2.4, has a useful explanation.

glCopyTexImage2D is expensive

Traditional GPU optimization lore says that glReadPixels is expensive. The updated version of this story to account for deferred GPUs says that glCopyTexImage2D is quite expensive too (although not as terrible as glReadPixels), because it forces resolving the framebuffer's rendering.

Always call glClear immediately after glBindFramebuffer

See the Adreno 200 document, section 3.2.1: "it is imperative to (a) use clears when switching Frame Buffer Objects (FBOs) to avoid having the driver tries to save/restore GMEM contents, and (b) always clear the depth-buffer at the start of a frame."

That makes sense, so we should always do it. Concretely, this means that we should do a glClear after every glBindFramebuffer call, ideally right after it, or at least before any draw-call.

Overdraw is still expensive on tbd-rast GPUs

While tbd-hsr GPUs optimize away overdraw regardless of the order of submission of geometry, tbd-rast GPUs don't. Therefore it remains important to sort opaque geometry in front-to-back order.

Incremental frame updates

While only rerendering the part of the framebuffer that has changed since the last swap look like an intersting performance trick, it turns out to cause bad performance problems on tile-based GPUs. This is because if the previous tile buffer is not cleared, the GPU must restore its state which requires a bandwidth cost that can be higher than just redrawing the entire frame. Some vendors provide with extensions (like QCOM_tiled_rendering) to address this use-case. Such extensions make it possible to minimize the amount of tiles that need to be restored.

Per pixel hidden surface removal

Some GPUs like PowerVR family do per pixel hidden surface removal to shad as few pixels as possible. Blending, sample masking and fragment shaders that may output transparent pixels or use the discard keyword will disable this optimization.

Misc mobile GPU performance topics

Add here anything that's not directly related to deferred rendering.

Use only triangle strips

Reason: Adreno 200's have a vertex cache size of 2.

Avoid using a stencil and depth buffers if you don't need them

Reason: Adreno 200's have 256Kb of GMEM. All of the render targets need to fit in GMEM and so only using a color buffer means we can have a larger tile size. The Freedreno driver doesn't currently (Apr 5 2013) optimize for this and always has all three buffers. It would be good to confirm with qualcomm that there driver does optimize this.

Make sure that our framebuffer's bit depth matches the hardware framebuffer's

Or else we'll get inefficient conversions between 16bpp and 32bpp. The pitfall here is typically we may be requesting FBConfigs with at least 16bpp and blindly taking the first returned one, which may actually be 32bpp.

Avoid MakeCurrent's

The Adreno 200 document lists it as very expensive. On B2G at least we should be able to land bug 749678.

Always clear depth and stencil framebuffers at the same time

The underlying format may be a DEPTH24_STENCIL8 buffer and failing to clear the two components at the same time may force the GPU to go down a slow path.

Memory Bandwidth

Unagi can do memcpy of 320*480*4 in 0.7ms or 834MB/s It can do memmove at 200MB/s