UPSTREAM PR #1336: feat: add generic DiT support to spectrum cache

[loci-dev]

Note

Source pull request: leejet/stable-diffusion.cpp#1336

Implementing the paper literally is heavier because it depends on architecture-specific internal feature hooks and final-block forecasting inside each DiT family, which would require invasive model-side changes and separate validation per model. This patch keeps Spectrum generic at the sampler level

Comparaison with easycache with 20 steps:

	Baseline	EasyCache	Spectrum
Image
Speedup	—	x1.82	x1.82
Config	—	0.2, 0.15, 0.95	default

[loci-review]

Overview

Analysis of 49,806 functions following commit 2cbf294 ("add dit") shows negligible net performance impact with mixed compiler-driven optimizations. Modified: 56 functions (0.11%), New: 0, Removed: 0.

Binaries analyzed:

• build.bin.sd-cli: -0.016% power consumption (-76.48 nJ)
• build.bin.sd-server: +0.146% power consumption (+768.09 nJ)

Function Analysis

All performance changes occur in C++ standard library functions (not application code) due to compiler optimization variations between builds:

Improvements:

• std::vector::begin() (sd-cli): -68% response time (-181ns) — entry block consolidation
• std::vector::empty() (sd-server): -41% response time (-190ns) — eliminated intermediate jump
• __gnu_cxx::__normal_iterator::operator+ (sd-server): -40% response time (-66ns) — merged stack canary operations
• std::__shared_ptr::operator= (sd-server): -8% response time (-80ns) — entry block optimization

Regressions:

• std::_Sp_counted_ptr_inplace::_M_destroy (sd-server): +61% response time (+189ns) — unnecessary branch insertion
• std::vector::back() (sd-server): +26% response time (+184ns) — added entry indirection
• std::vector::erase() (sd-server): +6% response time (+81ns) — entry block split

Other analyzed functions showed similar compiler-driven variations with offsetting improvements and regressions.

Assessment: The ±0.15% power consumption variance indicates improvements and regressions largely cancel out. No source code changes to these standard library functions were detected. Changes stem from compiler code generation differences, not intentional optimizations.

🔎 Full breakdown: Loci Inspector
💬 Questions? Tag @loci-dev