[OpenMP][WIP] `spir64` device runtime library template

[spoutn1k]

openmp/libomptarget/DeviceRTL/src/State.cpp

69–86

This is only defined for AMDGCN with a TODO, in a way that seems to overwrite the default malloc. Should I do the same for this implementation ?

openmp/libomptarget/DeviceRTL/src/Utils.cpp

87–91

Found shuffle below in the spirv interface but not pack. Should I start using bitwise operation the way it is done for AMD ?

[tianshilei1992]

Can you update the patch with context?

[jdoerfert]

openmp/libomptarget/DeviceRTL/src/State.cpp

69–86

On AMDGPU there is no "default malloc". On NVPTX there is, on Intel we need to check. You can probably go with the AMD route for now.

openmp/libomptarget/DeviceRTL/src/Utils.cpp

87–91

There is a default implementation for Pack and Unpack, you can just not overwrite them. I think that is what we do for AMD.

[spoutn1k]

Added context

[spoutn1k]

Add pack/unpack and malloc/free for spir64 target runtime implementation.

[tianshilei1992]

llvm/lib/Frontend/OpenMP/OMPContext.cpp

55

SPIR-V can be for host as well right? OpenCL CPU runtime can also take SPIR-V.

openmp/libomptarget/DeviceRTL/src/LibC.cpp

37

I'm thinking potentially we can have a generic implementation of omp_vprintf that just returns -1 and having the NVPTX version calls vprintf. In this way we don't need to do it for every target.

[spoutn1k]

llvm/lib/Frontend/OpenMP/OMPContext.cpp

55

It can ! The objective of this work is to use it with intel gpus. I put it there in order to have it understand the variants in DeviceRTL. Is there a way to signify both host and device ?

openmp/libomptarget/DeviceRTL/src/LibC.cpp

37

How do I go about doing this ? Declare omp_vprintf as weak ?

[spoutn1k]

Added contents of DeviceRTL/src/Mapping.cpp; ignored activemask for the time being.

[tianshilei1992]

openmp/libomptarget/DeviceRTL/src/LibC.cpp

37

just something like:

int32_t omp_vprintf(const char *Format, void *Arguments, uint32_t) {
  return -1;
}
#pragma omp begin declare variant match(device = {arch(nvptx64)})
int32_t omp_vprintf(const char *Format, void *Arguments, uint32_t) {
  return vprintf(...);
}
#pragma omp end declare variant

In this way, when the arch is NVPTX64, that variant will replace the default one.

openmp/libomptarget/DeviceRTL/src/Mapping.cpp

162

Do we have these builtins in LLVM right now?

[spoutn1k]

openmp/libomptarget/DeviceRTL/src/LibC.cpp

37

Oh I see. I was under the impression this would create a symbol conflict. Will do !

openmp/libomptarget/DeviceRTL/src/Mapping.cpp

162

That was my question but Johannes told me to go ahead and use them.

[spoutn1k]

Defined spirv methods as extern and defined spirv enums.