From: Beignet [mailto:[email protected]] On Behalf Of Dehuan Xin Sent: Sunday, November 15, 2015 5:50 PM To: [email protected] Subject: [Beignet] Question on some implementation details: async copy and vector functions
Hi, I have two questions on the implementation detail of Beignet: 1. I see in this file that async copy is implemented as a for-loop. Does this mean async copy is currently a processor copy and is not offloaded to hardware like DMA and it's not truly `async`, either? http://cgit.freedesktop.org/beignet/tree/backend/src/libocl/src/ocl_async.cl async_copy is done through GPU I/O read/write messages. AFAIK, this is not DMA like transfer. 2. In this source file, I see vector load is scalarized to scalar load, is this behavior disabled when `__attribute__((vec_type_hint(<typen>)))` is used? http://cgit.freedesktop.org/beignet/tree/backend/src/libocl/src/ocl_vload.cl generally vload is not recommended. The problem is take vload4(uchar *) as an example, the pointer passed in is only guaranteed to be 1 aligned. But a uchar4* read access will ensure the pointer is 4 aligned. That is a big difference to compiler. The hardware performance is quite different. So, the guide is try to use uchar4* pointer which cl programmer need to make sure the pointer is 4 aligned. If you cannot guarantee the pointer is 4 aligned, you can choose to use vload which is easy to program and . We will re-merge the scalarized load in compiler backend. Don’t worry. See llvm_loadstore_merge.cpp. There is no performance difference between vload4(int/float*, ) and int/float4* read. Because they are both 4 aligned We haven’t done anything special to `__attribute__((vec_type_hint(<typen>)))`. Thanks! Ruiling Regards Dehuan
_______________________________________________ Beignet mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/beignet
