Nowadays GPU drivers can dynamically allocate DMA buffers to be used in addition to VRAM. It's mainly done by translation-table maps (ttm) module. Some GPU modules may still use GEM subsystem to manage buffers though.
It seems to me that htop counts TTM memory as free on my system. Looks like TTM module usage can determined by reading file /sys/class/drm/ttm/memory_accounting/pool/pool_allocation_size.
Should we add GPU memory class and calculate TTM + GEM allocated memory?
Nowadays GPU drivers can dynamically allocate DMA buffers to be used in addition to VRAM. It's mainly done by translation-table maps (
ttm) module. Some GPU modules may still use GEM subsystem to manage buffers though.It seems to me that htop counts TTM memory as free on my system. Looks like TTM module usage can determined by reading file
/sys/class/drm/ttm/memory_accounting/pool/pool_allocation_size.Should we add
GPUmemory class and calculate TTM + GEM allocated memory?