[SYCL][E2E] Fix some tests in multi-device mode (#13510)

The atomic tests can be run by multiple SYCL devices, in which case the
compiler needs to be told which device the `--cuda-gpu-arch` parameter
applies to. This is done in the same vein as other occurrences of this
parameter.

The bfloat16 tests fail on some CPU devices due to a known issue with
opaque pointers.

Author: frasercrmck

sycl/test-e2e/BFloat16/bfloat16_builtins.cpp

@@ -14,6 +14,8 @@
 
 // Currently the feature isn't supported on FPGA.
 // UNSUPPORTED: accelerator
+// FIXME: enable opaque pointers support on CPU.
+// XFAIL: cpu
 #include "bfloat16_builtins.hpp"
 
 int main() {

sycl/test-e2e/BFloat16/bfloat16_builtins_cuda_generic.cpp

@@ -12,6 +12,8 @@
 
 // Currently the feature isn't supported on FPGA.
 // UNSUPPORTED: accelerator
+// FIXME: enable opaque pointers support on CPU.
+// XFAIL: cpu
 #include "bfloat16_builtins.hpp"
 
 int main() {

sycl/test-e2e/BFloat16/bfloat16_type.cpp

@@ -7,6 +7,7 @@
 // UNSUPPORTED: accelerator
 
 // FIXME: enable opaque pointers support on CPU.
+// XFAIL: cpu
 
 //==----------- bfloat16_type.cpp - SYCL bfloat16 type test ----------------==//
 //

sycl/test-e2e/syclcompat/atomic/atomic_class.cpp

@@ -32,7 +32,7 @@
 
 // UNSUPPORTED: hip || (windows && level_zero)
 
-// RUN: %clangxx -std=c++20 -fsycl -fsycl-targets=%{sycl_triple} %if any-device-is-cuda %{ -Xsycl-target-backend --cuda-gpu-arch=sm_70 %} %s -o %t.out
+// RUN: %clangxx -std=c++20 -fsycl -fsycl-targets=%{sycl_triple} %if any-device-is-cuda %{ -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_70 %} %s -o %t.out
 // RUN: %{run} %t.out
 
 #include <sycl/detail/core.hpp>

sycl/test-e2e/syclcompat/atomic/atomic_memory_acq_rel.cpp

@@ -32,7 +32,7 @@
 
 // UNSUPPORTED: hip
 
-// RUN: %clangxx -fsycl -fsycl-targets=%{sycl_triple} %if any-device-is-cuda %{ -Xsycl-target-backend --cuda-gpu-arch=sm_70 %} %s -o %t.out
+// RUN: %clangxx -fsycl -fsycl-targets=%{sycl_triple} %if any-device-is-cuda %{ -Xsycl-target-backend=nvptx64-nvidia-cuda --cuda-gpu-arch=sm_70 %} %s -o %t.out
 // RUN: %{run} %t.out
 
 #include <iostream>