Using SYCL to create hash table

cmlee — Thu, 21 Mar 2024 11:17:10 GMT

hi,

I am working to use SYCL to create hash table, similar to the code in following page.

https://developer.nvidia.com/blog/maximizing-performance-with-massively-parallel-hash-maps-on-gpus/

auto [old_k, old_v] = buckets[i].load(memory_order_relaxed);
// if the bucket is empty we can attempt to insert the pair
if (old_k == empty_sentinel) { // try to atomically replace the current content of the bucket with the input pair
bool success = buckets[i].compare_exchange_strong(
{old_k, old_v}, {k,v}, memory_order_relaxed);
if (success) { // store was successful
return true;
}

To create hash table parallel, will need to compare and exchange key and value pair in one atomic operation.

SYCL supports atomic compare_exchange_strong but only limited to scalar data type. It doesn't support some pair of key and value.

May I know any way to enable parallel hash table if atomic operation only support scalar data type?

Re: Using SYCL to create hash table

cmlee — Thu, 04 Apr 2024 08:41:00 GMT

sycl support atomic operation using a memory pointer. So, manage to create hash table by storing all key/value into some memory buffer. When insert the hash to the hash table, will use compare_exchange_strong with the pointer of the memory location that contains the key/value.

topic Using SYCL to create hash table in Intel® oneAPI DPC++/C++ Compiler

Using SYCL to create hash table

Re: Using SYCL to create hash table