0.3.3

• Pixel-major CSF weighting + fast log10 approximation, ~20% faster

Note: scores will be slightly different, but their meaning isn't deeply affected. Here's the updated MOS correlation:

Metric	0.3.2	0.3.3	Diff
KRCC	0.6319	0.6333	+0.0014
SRCC	0.8274	0.8287	+0.0013
PCC	0.8221	0.8229	+0.0008

This is technically an improvement, but it isn't big enough to register.

Full Changelog: 0.3.1...0.3.2

0.3.2

• Runtime SIMD dispatch for x86-64 systems, which falls back to scalar C if the host doesn't have FMA or AVX2.

Full Changelog: 0.3.1...0.3.2

0.3.1

• Force zig build to use the LLVM backend at all times, as it fails with the native backend
• Use mathematically correct sRGB transfer curve instead of approximation; encoder slows slightly (~3-4%), but it is worth it for correctness.
• Add new mcos display model

Full Changelog: 0.3.0...0.3.1

0.3.0

• NEON on ARM via cvvdp_neon.h
• AVX2 on x86-64 via cvvdp_avx2.h
• Scalar fallbacks in cvvdp_c.h
• cvvdp_create now sets *out_c = NULL and uses cvvdp_destroy(c) on early failures so the thread pool is released along with any partial allocations
• cvvdp_temporal_filter_init now nulls freed kernel pointers on partial allocation failure so destructor-based cleanup is safe
• remove duplicated CVVDP_GAUSSIAN_SIZE

SIMD changes are visible via #14, resulting in up to 51% better performance on x86-64 and up to 33% on ARM. User+sys time improves by up to 58% on both.

Note: fcvvdp is configured to use AVX2 on every x86-64 machine, meaning some ≤3% of x86-64 users may see compilation errors on hardware from before Haswell (2013). Other architectures fall back to C (hence the README benchmark remaining the same). If you are affected, please report it.

Full Changelog: 0.2.4...0.3.0

0.2.4

• Utilize allyourcodebase/libspng for PNG decoding instead of relying on system
• Add many new display modes, documented in doc/cvvdp.md
• Update copyright

Full Changelog: 0.2.3...0.2.4

0.2.3

• Fix help menu to print all options (prev wasn't printing -t/--threads)

Full Changelog: 0.2.2...0.2.3

0.2.2

• Update to Zig 0.16
• More accurately requantize 10-bit inputs to 8-bit

Full Changelog: 0.2.1...0.2.2

0.2.1

• Default to using system zlib instead of zlib-rs, with flag -Dlibz-rs=true to toggle zlib-rs. Users complained that zlib-rs was relatively inaccessible for most people, and building with system zlib reduces a lot of friction.

Full Changelog: 0.2.0...0.2.1

0.2.0

• Task threading for video inputs (#8)
• Memory usage optimizations (#9)
• FFmpeg patch (applies on FFmpeg n8.0)

Testing on 8-bit 360p Y4M inputs, fcvvdp runs a bit over 50fps on a Core i7-13700k test machine. At 1080p, this is reduced to ~7fps.

Full Changelog: 0.1.1...0.2.0

Binaries:

• fcvvdp.zst: x86-64-v3 Linux
• fcvvdp.xz: ARM macOS

0.1.1

• fix: temporal filtering bug (thanks @Line-fr for the heads-up!)
• fix: add newline before help print
• style: colors in help print for format string (similar to fssimu2)

Full Changelog: 0.1.0...0.1.1