Reversing .stbsp

[snake-biscuits]

.stbsp files contain helper data for texture streaming

Our biggest official resource on their creation & use is this GDC talk:

• GDC on YouTube: Efficient Texture Streaming in Titanfall 2

bsp_tool.extensions.stbsp.StreamBsp maps the format

[snake-biscuits]

How Respawn Generates .stbsp

As part of their internal commit & build systems at respawn, they used Jenkins to generate .stbsp

This work was dispatched to their "Precomputation Cruncher", pictured below:

From what I understand of the GDC talk, they would do a histogram of texture scale relative to a "probe"
Essentially a virtual player that looks at the map & decides what texture resolution looks best

.stbsp files seem to store the results of these histograms in some way, indexing specific material names
Mostly targetting .rpak materials, though some .vmt support exists

[snake-biscuits]

Columns

.stbsp contains columns (not unlike CM_GRID)
Each column is linked to some binary data we haven't reversed yet
@r-ex actually looked at the files in ida & figured out that what I thought was a histogram colour is instead coverage_scale
idk what that does specifically, so I'm visualising it as a colour

Exporting

>>> from bsp_tool.extensions.stbsp import StreamBsp
>>> StreamBsp.from_file(".../Titanfall2/r2/maps/mp_complex3.stbsp").columns_as_obj("complex3.obj")

Colourising

In Blender python terminal

>>> for m in D.materials:
...     if len(m.name) == 8:
...         rgba = int(m.name, 16)
...         r = rgba >> 24
...         g = (rgba >> 16) & 0xFF
...         b = (rgba >> 8) & 0xFF
...         a = rgba & 0xFF
...         m.diffuse_color = [255 / x if x != 0 else 0 for x in (r, g, b, a)]

NOTE: might need to delete / rename the default material before this works

[snake-biscuits]

at ~8mins in the GDC Youtube video, they define "coverage" as how much of a 256x256 screen a 4096x4096 texture fills

[snake-biscuits]

Column Data

Empty Columns

Since we use columns, there's loads of empty space
Empty columns have some bytes of column data, but all bytes are 0x00
If a column is empty: coverage_scale is 1.1754943508222875e-38 (00 00 80 00)
2177/3355 columns are empty in mp_complex3

[snake-biscuits]

rexx & royalblue found code references in ida
it's an array of pairs of shorts
1st short is bin (mip) & MaterialInfo index
2nd short is coverage (multiply by column.coverage_scale)

The arrays are preceded by a short count of pairs (+1 for some reason) & 0xFFFF
not always at the start, sometime preceeded by nulls
remember, empty columns start with nulls too