Libgomp.texi: Mention ROCR_VISIBLE_DEVICES in context of USM
Documentation for libgomp now clarifies the use of `ROCR_VISIBLE_DEVICES` for unified shared memory with AMD GPUs.
This commit updates the libgomp.texi documentation to better explain the use of unified shared memory (USM) with AMD GPUs, particularly on mixed systems. It re-adds a crucial half-sentence that was previously lost, clarifying that ROCR_VISIBLE_DEVICES serves as a workaround when not all AMD GPUs support USM, especially in configurations with integrated and discrete GPUs. The wording also now correctly differentiates between USM/SVM support and XNACK capabilities.
In Details
The libgomp.texi file provides documentation for libgomp, GCC's OpenMP runtime library, specifically detailing the GCN (Graphics Core Next) offloading capabilities for AMD GPUs. This commit refines the explanation of Unified Shared Memory (USM) support, which is critical for efficient data transfer between the host and GPU. On heterogeneous systems where not all AMD GPUs support USM, the ROCR_VISIBLE_DEVICES environment variable becomes a necessary control mechanism to ensure host fallback doesn't occur unnecessarily. This documentation update also corrects a subtle confusion between USM/…
For Context
Modern computer systems often combine a main processor (CPU) with specialized processors like Graphics Processing Units (GPUs) for faster computation. To make these work together efficiently, data needs to be easily accessible by both. Unified Shared Memory (USM) is a technology that allows the CPU and GPU to share the same memory space, simplifying programming and improving performance. However, some older or entry-level GPUs might not fully support USM, especially in systems with both an integrated GPU and a separate, more powerful one. This update to the GCC documentation for OpenMP (a programming standard for parallel computing) clarifies that if you encounter issues with USM on AMD GPUs in such 'mixed' systems, you can use the ROCR_VISIBLE_DEVICES setting. This setting lets you tell the system which specific GPUs should be used, effectively working around the limitation and ensuring your programs run correctly with USM.