[PATCH] D15603: [OpenCL] Pipe type support

Pekka Jääskeläinen via cfe-commits cfe-commits at lists.llvm.org
Mon Dec 28 05:28:38 PST 2015


pekka.jaaskelainen added inline comments.

================
Comment at: lib/CodeGen/CGOpenCLRuntime.cpp:108
@@ +107,3 @@
+    PipeTy = llvm::PointerType::get(llvm::StructType::create(
+      CGM.getLLVMContext(), "opencl.pipe_t"), PipeAddrSpc);
+  }
----------------
I'm not sure if touching the built-in fingerprints for this is a good idea. It might lead to problems and confusion. Cannot one pass pipes as arguments to user functions too? Are the fingerprints of those functions then modified accordingly? It might become messy.

Because the packet size is known at host side, the pipes can be implemented as structs which holds the packet size as one of the member variables. The problem with this approach is how to exploit wider reads/writes instead of a scalar read/write loop + unpack/pack in case of vector datatypes. 

If the size is known only at runtime, one cannot easily generate vector reads/writes even if the element is a vector datatype and it would be efficient to keep the packet in a vector register as long as possible. For helping this I'd add a metadata which can be utilized at compile time to make reading/writing from the pipe faster.  But in a way that is already an optimization, not a requirement, to make pipes work.

The reading itself is platform dependent as the pipe can be even a hardware FIFO accessed using special instructions.


http://reviews.llvm.org/D15603





More information about the cfe-commits mailing list