You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Implemented high-throughput 256-bit AVX2 vector kernel in cpu_simd_me.c unrolled by 4 rows with dual vector accumulators.
Fully exploits AMD Zen 2 native 256-bit AVX2 execution units, halving instruction count and latency for CPU-side motion estimation.
Automatic runtime CPUID detection with seamless SSE2 and scalar fallbacks.
2. HEVC (H.265) Dynamic Governor Parity
Brought full 4-tier dynamic hybrid load balancing to the H.265 encoder (encoder_h265.c).
When GPU contention occurs during HEVC streaming, Motion Estimation (~70% of compute work) is seamlessly offloaded to idle Zen 2 CPU cores, preserving stream frame delivery.
Governor Tier 3 Latch Bug: Fixed a stall where emergency failover (P_Skip) could lock the driver in Tier 3 indefinitely by adding dynamic_governor_notify_failover_handled().
Asymmetric vkMapMemory Leak: Resolved host GPU address space leak by structuring nested mapping checks with guaranteed vkUnmapMemory execution.
1080p Bottom Row ME Cutoff: Corrected boundary check in cpu_simd_me.c to use macroblock-padded buffer dimensions, restoring motion search for row 67 (y = 1072..1087).
Fast Spatial Motion Predictor: Evaluates left neighbor MB vector prior to diamond search, allowing ~75% of non-static blocks to converge in <= 2 SAD checks (<0.5 ms CPU time).
馃摝 Bundled Artifacts
bc250-driver-linux-x86_64.tar.gz: 64-bit VA-API driver with precompiled SPIR-V shaders, setup scripts, and audio clock fix.
bc250-driver-linux-i386.tar.gz: 32-bit VA-API driver for 32-bit gaming runtimes (Steam Link).