[SYCL][HIP] sssp results mismatch

[zjin-lcf]

To reproduce issue for the example https://github.com/zjin-lcf/HeCBench/tree/master/sssp-sycl

The example built with the DPC++ with CUDA support finishes without result mismatch.

cp ../sssp-cuda/data.tar.gz .
tar -zxf data.tar.gz
make HIP=yes run
./main -g 120 -t 1 -w 10 -r 1
Number of nodes = 264346        Number of edges = 733846        Initialization Time (ms): 349.456024
Allocation Time (ms): 0.116000
Copy To Device Time (ms): 9.258000
Kernel Time (ms): 2.901000
Copy Back and Merge Time (ms): 1.888000
FAIL: Computed node 2 cost (-2147483647 != 50999) does not match the expected value

[npmiller]

I had a quick look and it looks like this sample uses local memory accessors, unfortunately I believe this isn't supported by the HIP backend at the moment.

The NVPTX backend has a special IR pass to handle these which doesn't exist yet for the AMDGCN backend so any local accessor argument will not be setup correctly. I suspect using global buffer for the two arguments may work as a workaround, however we do need to add local argument support for HIP.

[zjin-lcf]

There are quite a few SYCL programs that have local accessors for shared local memory, but their results match. Is that right ? Thanks.