[Feat] (Subgroups): use native subgroups or emulate them on both backends

This commit is contained in:
2026-10-13 06:03:57 -04:00
parent cfbaffd757
commit d6b0eba8fe
50 changed files with 909 additions and 360 deletions
@@ -184,6 +184,20 @@ namespace MobileGL::MG_Util::SelfTest {
// ProbeGeometryStageSsboWriteAfterEmitDropped(), evaluated at most once per process.
Bool GeometryStageSsboWriteAfterEmitDropped(const MG_External::GLESFunctionsTable& gl);
// Every invocation of a 32-wide compute workgroup stores 1 into its shared slot, barrier(),
// reads slot 0, barrier(), stores 7, barrier(), reads slot 0 again, and writes both reads
// to its own result word: the second read must see 7. The Adreno ES compiler treats ordinary shared loads as pure - it hands back the
// first read (and, with other addresses, moves a read past a later store) - while an
// atomicAdd(x, 0u) read of the same words is kept in place. That atomic twin is the
// control: the verdict is "stale" only when the plain program goes wrong while the atomic
// one stays right. Deterministic (a compiler transform), so latched; false when compute
// shaders or an entry point are missing or anything fails to set up. Restores the GL state
// it touches.
Bool ProbeSharedMemoryLoadsGoStale(const MG_External::GLESFunctionsTable& gl);
// ProbeSharedMemoryLoadsGoStale(), evaluated at most once per process.
Bool SharedMemoryLoadsGoStale(const MG_External::GLESFunctionsTable& gl);
// Samples one R32F GL_TEXTURE_2D_MULTISAMPLE texel through a swizzled alpha channel, twice,
// with a separately linked program each time. Returns true only when the swizzled read goes
// wrong while every control read stays right.