[OpenMP][libomptarget] Add support for declare target to clause under unified memory

[gtbercea]

Summary

This patch adds support for handling variables under the:

#pragma omp declare target to()

clause when the

#pragma omp requires unified_shared_memory

is used.

The address of the host variable is copied into the device pointer just like for the declare target link case.

[gtbercea]

ping

[grokos]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

Maybe this OR is redundant; it always evaluates to true. Since the symbol we are processing has size!=0 it is a variable and variables are always either "to" or "link", there are no other possibilities (at least now, maybe in the future more attributes will be added).

[gtbercea]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

So I could remove the attribute check and just have:

if (DeviceInfo.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY) {..}

Would that work?

The reason I included this explicit check (even if it evaluates to true always) is that if in the future this is extended with other attributes then this check will not automatically work for that. I think this disambiguates the cases that are supported today and guard against this path being taken unintentionally.

[grokos]

LGTM

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

Yes, that's what I thought as well. Let's leave the check there, it makes the implementation more future-proof.

[gtbercea]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

Awesome. I'll leave it as is for now then.

[jcownie]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

The code which is checking the flags looks a bit weird. One test is done with & (suggesting these are bit flags which could be or-ed together), the other with == suggesting this is a field full of mutually exclusive enumerations.

[grokos]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

This is because actually "TO" is not a flag, it's the complete absence of flags (0x00). If something has been "declared target" and is not a ctor/dtor or a linked variable then it is "TO".

I don't like this fact either - if we want to be 100% correct we should introduce a dedicated flag for "TO" but that involves changes in the compiler/library interface and is definitely not part of this patch. We can pursue such a patch anyway if there are strong arguments that warrant the effort.

[gtbercea]

• Fix condition.

[gtbercea]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

I've move the checks into a comment and simplified the condition.

[grokos]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

This is what I had in mind when I posted the first comment. It looks better now!

[gtbercea]

libomptarget/plugins/cuda/src/rtl.cpp

452–453 ↗

(On Diff #203934)

Awesome I'll commit this version!

openmp/trunk/libomptarget/plugins/cuda/src/rtl.cpp

if (e->size) {
         CUDA_ERR_STRING(err);
         return NULL;
       }
 
       DP("Entry point " DPxMOD " maps to global %s (" DPxMOD ")\n",
           DPxPTR(e - HostBegin), e->name, DPxPTR(cuptr));
       entry.addr = (void *)cuptr;
 
-      if (DeviceInfo.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
-          e->flags & OMP_DECLARE_TARGET_LINK) {
-        // If unified memory is present any target link variables
+      // Note: In the current implementation declare target variables
+      // can either be link or to. This means that once unified
+      // memory is activated via the requires directive, the variable
+      // can be used directly from the host in both cases.
+      // TODO: when variables types other than to or link are added,
+      // the below condition should be changed to explicitely
+      // check for to and link variables types:
+      //  (DeviceInfo.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY &&
+      //   (e->flags & OMP_DECLARE_TARGET_LINK ||
+      //    e->flags == OMP_DECLARE_TARGET_TO))
+      if (DeviceInfo.RequiresFlags & OMP_REQ_UNIFIED_SHARED_MEMORY) {
+        // If unified memory is present any target link or to variables
         // can access host addresses directly. There is no longer a
         // need for device copies.
         cuMemcpyHtoD(cuptr, e->addr, sizeof(void *));
         DP("Copy linked variable host address (" DPxMOD ")"
            "to device address (" DPxMOD ")\n",
           DPxPTR(*((void**)e->addr)), DPxPTR(cuptr));
       }