arsenm wrote:

> I don't understand why we are now adding the concept of triples to SPIR-V and 
> GPU runtimes. And I really really really don't want that we imply in any way 
> that applications would need to provide their OpenCL runtime specific SPIR-V 
> binaries. Please make no assumptions about how FMA is implemented, no 
> assumptions about denorm behavior or anything, because we'll have to support 
> the same spirv across varies hardware that have different implementation 
> details.

I'm increasingly confused by how you're using libclc. Do you have a description 
of your system written somewhere? End user code and implementation provider 
code are fundamentally different uses. You're using it as the compiler runtime 
library, but also don't want it to be target specific? These are contradictory. 
If libclc is providing the ultimate implementation, it's necessarily hardware 
specific? This is my point about decomposing the libraries. The target specific 
codegen bits ought to be separate from the user facing language entry points 

https://github.com/llvm/llvm-project/pull/199618
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to