[libclc] Add an option to build SPIR-V targets with the LLVM backend #151347
Conversation
This removes the dependency on an external tool to build the SPIR-V files. It may be of interest to projects such as Mesa. Note that the option is off by default as using the SPIR-V backend, at least on my machine, uses a *lot* of memory and the process is often killed in a parallelized build. It does complete, however. Fixes llvm#135327.
I see the same behavior, goes up to something like 30GB. The input is not particularly large (just 300k lines of IR). From a quick run through massif, it looks like MachineInstrs in SPIRVGlobalRegistry are the memory hog? |
| ) | ||
| if ( LIBCLC_USE_SPIRV_BACKEND ) | ||
| add_custom_command( OUTPUT ${libclc_builtins_lib} | ||
| COMMAND ${clang_exe} --target=${ARG_TRIPLE} -x ir -o ${libclc_builtins_lib} ${builtins_link_lib} |
There was a problem hiding this comment.
should -c be added to clang command line? It avoids spirv-link step.
There was a problem hiding this comment.
Good question! I'm not sure, to be honest. I tried using -c and it generates a slightly larger file (6900 vs 6792). It seems to take a few fewer seconds to do so, which is an upside. I haven't looked into what exactly it's doing.
There was a problem hiding this comment.
perhaps spirv-link does optimization, but obviously there is only one file to link. LGTM as the final size is smaller, so this is not meant to be a blocker.
|
Can this be backported to 21? |
|
I don't think it makes sense to backport this. At least at the time it landed it only worked in the sense of "does not crash during compilation". I'm not sure if the big compile-time and memory usage issue has been resolved since, or anyone has checked whether the produced SPIRV actually works. |
|
Yeah, it looks like at least 638383c is needed on 21 to make it work Just that infact |
This removes the dependency on an external tool to build the SPIR-V files. It may be of interest to projects such as Mesa.
Note that the option is off by default as using the SPIR-V backend, at least on my machine, uses a lot of memory and the process is often killed in a parallelized build. It does complete, however.
Fixes #135327.