[Polly] [PPCGCodeGeneration] Add managed memory support to GPU code
authorSiddharth Bhat <siddu.druid@gmail.com>
Fri, 28 Apr 2017 11:16:30 +0000 (11:16 +0000)
committerSiddharth Bhat <siddu.druid@gmail.com>
Fri, 28 Apr 2017 11:16:30 +0000 (11:16 +0000)
commitabed49699b8157b4981c6d9e69080c73364d0d87
tree08c74449c16dabb91b2b76df00d085563eec86f0
parent04928fd021e48bfff735d3c3c4554d240f1dbc56
[Polly] [PPCGCodeGeneration] Add managed memory support to GPU code
generation.

This needs changes to GPURuntime to expose synchronization between host
and device.

1. Needs better function naming, I want a better name than
"getOrCreateManagedDeviceArray"

2. DeviceAllocations is used by both the managed memory and the
non-managed memory path. This exploits the fact that the two code paths
are never run together. I'm not sure if this is the best design decision

Reviewed by: PhilippSchaad

Tags: #polly

Differential Revision: https://reviews.llvm.org/D32215

llvm-svn: 301640
polly/lib/CodeGen/PPCGCodeGeneration.cpp
polly/test/GPGPU/cuda-managed-memory-simple.ll [new file with mode: 0644]
polly/tools/GPURuntime/GPUJIT.c
polly/tools/GPURuntime/GPUJIT.h