Pipeline Barriers in Vulkan 1.3 Dynamic Rendering: Fine‑Grained vs Coarse‑Grained Decision

Answer

Use fine‑grained barriers when the image is accessed by different pipeline stages with distinct access types (e.g., graphics fragment writes followed by compute shader reads) or when sub‑passes need different masks; otherwise a single coarse barrier with VK_ACCESS_2_MEMORY_READ_BIT|WRITE_BIT and VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT is sufficient and reduces command‑buffer complexity.

Confirmed facts

• Vulkan 1.3 requires a memory barrier (or image memory barrier) whenever an image changes layout or is accessed with different access types between commands.
• When transferring ownership between queues, the barrier must include srcQueueFamilyIndex and dstQueueFamilyIndex, and its access mask must cover all accesses performed on the source queue and all accesses that will be performed on the destination queue.
• Fine‑grained barriers let you specify exact srcStageMask, dstStageMask, srcAccessMask, dstAccessMask; coarse barriers use the ALL_COMMANDS bits and broad access masks.

Likely explanation / decision guidance

1. Identify the set of pipeline stages that will touch the image before and after the barrier (e.g., VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT vs VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT).
2. If the before‑and‑after stage sets overlap and the access types are identical (e.g., both are VK_ACCESS_SHADER_READ_BIT), a coarse barrier is likely sufficient.
3. If the stages differ or the access types differ (write vs read, or different shader stages), prefer fine‑grained barriers to avoid false dependencies.
4. For queue ownership transfer, add a barrier with the appropriate queue family indices; you can still apply fine‑grained masks inside that barrier to limit stalls.
5. Measure the overhead of the chosen barrier on your target hardware (GPU timestamp delta) and compare it to your latency budget; if the coarse barrier stays below the threshold, you may accept its simplicity.

Missing diagnostic detail

Please provide the measured GPU‑side overhead (e.g., timestamp query delta) for a coarse barrier on your target GPU so we can confirm whether it meets your performance threshold.