ExynosTools 1.5.0 introduces a low-level architectural redesign focused on improving stability, latency, and efficiency in BCn texture decompression for Samsung devices equipped with Xclipse GPUs (RDNA 2, 3, and 3.5 architectures).
Key Changes
- Staging Buffer Pool (Ring Buffer)
The previous static temporary buffer system has been fully replaced with a multi-threaded preallocated ring buffer (16MB).
- Enables continuous, non-blocking transfers to VRAM
- Supports parallel operations without pipeline stalls
- Dynamically scales using temporary buffers under heavy workloads
Impact: Reduced memory bottlenecks and improved data throughput.
- Multi-Mipmap Batching
A batching system has been implemented for small mipmap levels (<32×32).
- Combines multiple sublevels into a single dispatch call
- Significantly reduces CPU/GPU overhead
Impact: Up to 7× reduction in overhead per processed texture.
- Adaptive Shaders (LDS / RDNA)
Compute shaders (BC5, BC7, mipmaps) have been rewritten using Specialization Constants.
- Dynamically scales workgroup size based on detected architecture:
- Exynos 2500 (RDNA 3.5): up to 128 threads
- Exynos 2200/2400 (RDNA 2/3): 64 threads
- Improved utilization of Local Data Share (LDS)
Impact: Better parallel efficiency and improved hardware utilization.
- Vulkan Layer (Format Interception)
A custom Vulkan layer intercepts and overrides GPU-reported format properties:
- Hooks into "vkGetPhysicalDeviceFormatProperties"
- Reports BCn support as native hardware capability
Impact: Improved compatibility with translation layers (e.g., DXVK) and more direct hardware feature usage.
Source Code
The complete project source code is included in the archive:
- "codigo_abierto.zip" — contains the full implementation of ExynosTools, including all core modules, shaders, and Vulkan layer components.
Conclusion
Version 1.5.0 delivers a refined architecture that:
- Minimizes texture processing overhead
- Optimizes CPU → GPU data flow
- Enhances stability in complex execution environments
This release establishes a solid and efficient foundation for RDNA-based Xclipse GPU integrations.