RADV Driver Enables Host Image Copy By Default For RDNA2 & Newer
Introduced back in 2023 with Vulkan 1.3.258 was VK_EXT_host_image_copy to copy data between host memory and images on the host processor without needing to stage the data through a CPU-accessible buffer. This direct CPU-to-GPU image data transfer path can reduce memory usage during asset loads and all around more efficiency and performance. Finally now the RADV open-source Radeon driver is enabling support by default.
Last year RADV landed experimental support for https://www.phoronix.com/news/RADV-Experimental-HIC but was disabled by default due to possible performance issues. Finally now though with updated AMD ADDRLIB library, the performance is good enough for enabling by default on RDNA 2 (GFX10.3) and newer GPUs.
Samuel Pitoiset of Valve's Linux graphics team explained in a now-merged Mesa patch:
The AVX2 optimizations in ADDRLIB pay off big time as noted in the merge request:
RDNA 1 (GFX10) can still force it on via RADV_DEBUG=hic if desired while moving forward RDNA 2 and newer will have this important performance and memory efficiency feature enabled by default.
Last year RADV landed experimental support for https://www.phoronix.com/news/RADV-Experimental-HIC but was disabled by default due to possible performance issues. Finally now though with updated AMD ADDRLIB library, the performance is good enough for enabling by default on RDNA 2 (GFX10.3) and newer GPUs.
Samuel Pitoiset of Valve's Linux graphics team explained in a now-merged Mesa patch:
"Latest addrlib supports SIMD (AVX2) and it's definitely fast enough to be used in production now.
GFX10 is still not enabled by default due to some regressions from the addrlib bump, also still missing AVX for some formats."
The AVX2 optimizations in ADDRLIB pay off big time as noted in the merge request:
"This new addrlib bump contains AVX optimizations for upload (mem->img) and it's way faster, the order of magnitude is like x10, it's roughly 20GiB/s on NAVI48."
RDNA 1 (GFX10) can still force it on via RADV_DEBUG=hic if desired while moving forward RDNA 2 and newer will have this important performance and memory efficiency feature enabled by default.