Describe the proposed feature and its relevance to inferencing
Currently, we have three backend, Metal, Vulkan and VMVX (or cpu). The issue of VMVX is that it is slow from my experience (could be outdated information since the last time I test it is a year ago). Moreover, about 85% of android supports Vulkan, based on the android developer website. I believe it would be wise to just let android run on Vulkan.
This doesn't means that we dump everything about VMVX. We could wait for the technology for VMVX to be fast enough to be run on devices. For now, we could introduce CMake build option to force a VMVX backend for people who are opted to try out this widely supported but not too fast backend.
Context or use case
No response
Proposed solution
Let android's backend to be Vulkan. Introduce CMake build option for force VMVX backend and implement preprocessor directive for it.
fire
commented
1 week ago
Describe the proposed feature and its relevance to inferencing
Currently, we have three backend, Metal, Vulkan and VMVX (or cpu). The issue of VMVX is that it is slow from my experience (could be outdated information since the last time I test it is a year ago). Moreover, about 85% of android supports Vulkan, based on the android developer website. I believe it would be wise to just let android run on Vulkan.
This doesn't means that we dump everything about VMVX. We could wait for the technology for VMVX to be fast enough to be run on devices. For now, we could introduce CMake build option to force a VMVX backend for people who are opted to try out this widely supported but not too fast backend.
Context or use case
No response
Proposed solution
Let android's backend to be Vulkan. Introduce CMake build option for force VMVX backend and implement preprocessor directive for it.
Additional information
No response