• Warnings in GPU to NVVM pipeline Hi, I’m currently in the process of trying to understand the conversion from the GPU dialect to LLVM via the NVVM dialect and GPU code-generation in general. • When I look at the tablegen file for the GPU to NVVM pipeline, I see that there are warnings on the index bitwidth and bare pointer call convention options that state “warning this should be 64 until the GPU layering is fixed” and “warning this should be false until the GPU layering is fixed” respectively. • I have two questions related to this: What is meant by “GPU layering is fixed”? • I have set the index bitwidth to 32 (which seems like a good idea from a performance perspective) and used the bare pointer call convention and the generated kernels compile just fine (though their performance is worse than hand written CUDA kernels, but that could be for a lot of other reasons). • What is meant by “GPU layering is fixed”? • I have set the index bitwidth to 32 (which seems like a good idea from a performance perspective) and used the bare pointer call convention and the generated kernels compile just fine (though their performance is worse than hand written CUDA kernels, but that could be for a lot of other reasons).
Article Summaries:
- Hi, I’m currently in the process of trying to understand the conversion from the GPU dialect to LLVM via the NVVM dialect and GPU code-generation in general. When I look at the tablegen file for the GPU to NVVM pipeline, I see that there are warnings on the index bitwidth and bare pointer call convention options that state “warning this should be 64 until the GPU layering is fixed” and “warning this should be false until the GPU layering is fixed” respectively. I have two questions related to this: - What is meant by “GPU layering is fixed”? I have set the index bitwidth to 32 (which seems lik
Sources: