SnapshotUnpremultiply.h File Reference

The snapshot-unpremultiply compute program, expressed in the donner::gpu::shader IR. More...

#include "donner/gpu/shader/IrModule.h"
#include "donner/gpu/shader/programs/SnapshotUnpremultiplyBindings.h"

Include dependency graph for SnapshotUnpremultiply.h:

Namespaces
namespace	donner
	Top-level Donner namespace, which is split into different sub-namespaces such as donner::svg and donner::css.

Functions
ShaderResult< IrModule >	donner::gpu::shader::programs::BuildSnapshotUnpremultiplyModule ()
	Builds the snapshot-unpremultiply compute program: one @compute @workgroup_size(8, 8, 1) entry point named cs_main that reads one texel of the premultiplied render target and writes the straight-alpha texel to a write-only storage texture.

Detailed Description

The snapshot-unpremultiply compute program, expressed in the donner::gpu::shader IR.

Function Documentation

BuildSnapshotUnpremultiplyModule()

ShaderResult< IrModule > donner::gpu::shader::programs::BuildSnapshotUnpremultiplyModule

(

)

Builds the snapshot-unpremultiply compute program: one @compute @workgroup_size(8, 8, 1) entry point named cs_main that reads one texel of the premultiplied render target and writes the straight-alpha texel to a write-only storage texture.

The arithmetic reproduces the CPU reference loop byte for byte:

a == 0  ->  (0, 0, 0, 0)
a != 0  ->  min(255, (premul * 255 + a / 2) / a)     (round half up)

It is integer arithmetic on the recovered 8-bit channels rather than float division, because the readback's contract is that the GPU and CPU paths produce identical bytes, and float division would agree only to within a rounding mode.

Recovering those channels multiplies the normalized texel by 255 and rounds. The three backends spell rounding slightly differently - half to even in WGSL and SPIR-V, half away from zero in MSL - and they agree here because a unorm8 texel times 255 lands exactly on an integer in f32, so no value is ever at a half-way tie.

The destination extent is the smaller of the two textures. Clamping to both matters because the staging texture is pooled: writing only the region the input covers would leave the rest of a larger pooled texture holding a previous snapshot's pixels instead of raising a validation error. Invocations outside that extent return without writing, so a dispatch rounded up to whole workgroups is safe.