[Libomptarget] Add checks for CUDA subarchitecture using new info

[jhuber6]

Summary

This patch extends the is_valid_binary routine to also check if the
binary's architecture string matches the one parsed from the runtime.
This should allow us to only use the binary whose compute capability
matches, allowing us to support basic multi-architecture binaries for
CUDA.

Depends on D127432

[tianshilei1992]

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

1549

I think here we should check if all devices are compatible instead of one.

  if (ArchStr != image->Info.Arch)
    return true;
}

return true.

[tianshilei1992]

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

1548–1551

[jhuber6]

Should I add a test for this? Something like

clang --offload-arch=sm_35 --offload-arch=sm_37 --offload-arch=sm_50 --offload-arch=sm_52 --offload-arch=sm_53 --offload-arch=sm_60 --offload-arch=sm_61 --offload-arch=sm_62 --offload-arch=sm_70 --offload-arch=sm_72 --offload-arch=sm_75 --offload-arch=sm_80 --offload-arch=sm_86

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

1549

Yeah I wasn't sure how to handle this in the case of multiple devices, but it's probably more reasonable to expect each one to match the architecture for it to be compatible. I'll change it.

[tianshilei1992]

In D127505#3573696, @jhuber6 wrote:

Should I add a test for this? Something like

clang --offload-arch=sm_35 --offload-arch=sm_37 --offload-arch=sm_50 --offload-arch=sm_52 --offload-arch=sm_53 --offload-arch=sm_60 --offload-arch=sm_61 --offload-arch=sm_62 --offload-arch=sm_70 --offload-arch=sm_72 --offload-arch=sm_75 --offload-arch=sm_80 --offload-arch=sm_86

I don't think our current runtime libraries support the case that there are multiple images for one plugin.

[jhuber6]

In D127505#3573729, @tianshilei1992 wrote:

In D127505#3573696, @jhuber6 wrote:

Should I add a test for this? Something like

clang --offload-arch=sm_35 --offload-arch=sm_37 --offload-arch=sm_50 --offload-arch=sm_52 --offload-arch=sm_53 --offload-arch=sm_60 --offload-arch=sm_61 --offload-arch=sm_62 --offload-arch=sm_70 --offload-arch=sm_72 --offload-arch=sm_75 --offload-arch=sm_80 --offload-arch=sm_86

I don't think our current runtime libraries support the case that there are multiple images for one plugin.

That's the whole point of this patch set. Now we will embed all of these images into an offloading binary which contains the architecture string. Inside the runtime we can then check if the architecture string matches and use that if it matches.

[tianshilei1992]

In D127505#3573731, @jhuber6 wrote:

In D127505#3573729, @tianshilei1992 wrote:

In D127505#3573696, @jhuber6 wrote:

Should I add a test for this? Something like

clang --offload-arch=sm_35 --offload-arch=sm_37 --offload-arch=sm_50 --offload-arch=sm_52 --offload-arch=sm_53 --offload-arch=sm_60 --offload-arch=sm_61 --offload-arch=sm_62 --offload-arch=sm_70 --offload-arch=sm_72 --offload-arch=sm_75 --offload-arch=sm_80 --offload-arch=sm_86

I don't think our current runtime libraries support the case that there are multiple images for one plugin.

That's the whole point of this patch set. Now we will embed all of these images into an offloading binary which contains the architecture string. Inside the runtime we can then check if the architecture string matches and use that if it matches.

Oh, sorry didnt see its parent patch.

[jhuber6]

Rebase on the previous changes.

[jhuber6]

Rebase to use new function.

[jdoerfert]

LG

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

 #ifdef __cplusplus
 extern "C" {
 #endif
 
 int32_t __tgt_rtl_is_valid_binary(__tgt_device_image *Image) {
   return elf_check_machine(Image, /* EM_CUDA */ 190);
 }
 
+int32_t __tgt_rtl_is_valid_binary_info(__tgt_device_image *image,
+                                       __tgt_image_info *info) {
+  if (!__tgt_rtl_is_valid_binary(image))
+    return false;
+
+  // A subarchitecture was not specified. Assume it is compatible.
+  if (!info->Arch)
+    return true;
+
+  int32_t NumberOfDevices = 0;
+  if (cuDeviceGetCount(&NumberOfDevices) != CUDA_SUCCESS)
+    return false;
+
+  for (int32_t DeviceId = 0; DeviceId < NumberOfDevices; ++DeviceId) {
+    CUdevice Device;
+    if (cuDeviceGet(&Device, DeviceId) != CUDA_SUCCESS)
+      return false;
+
+    int32_t Major, Minor;
+    if (cuDeviceGetAttribute(&Major,
+                             CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR,
+                             Device) != CUDA_SUCCESS)
+      return false;
+    if (cuDeviceGetAttribute(&Minor,
+                             CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR,
+                             Device) != CUDA_SUCCESS)
+      return false;
+
+    std::string ArchStr = "sm_" + std::to_string(Major) + std::to_string(Minor);
+    if (ArchStr != info->Arch)
+      return false;
+  }
+
+  DP("Image has compatible compute capability: %s\n", info->Arch);
+  return true;
+}
+
 int32_t __tgt_rtl_number_of_devices() { return DeviceRTL.getNumOfDevices(); }
 
 int64_t __tgt_rtl_init_requires(int64_t RequiresFlags) {
   DP("Init requires flags to %" PRId64 "\n", RequiresFlags);
   DeviceRTL.setRequiresFlag(RequiresFlags);
   return RequiresFlags;
 }
 

openmp/libomptarget/plugins/exports

 VERS1.0 {
   global:
     __tgt_rtl_is_valid_binary;
+    __tgt_rtl_is_valid_binary_info;
     __tgt_rtl_is_data_exchangable;
     __tgt_rtl_number_of_devices;
     __tgt_rtl_init_requires;
     __tgt_rtl_init_device;
     __tgt_rtl_deinit_device;
     __tgt_rtl_load_binary;
     __tgt_rtl_data_alloc;
     __tgt_rtl_data_submit;