AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make host memory allocations coherent by default.

Browse Source 
Associated change is to optimize event recording so it uses
agent-scope releaes (since it was only using system-scope release
to support non-coherent host mem).

Flags and environment variables exist to obtain previous behavior
if desired.  Options are documented in new performance guide.


[ROCm/clr commit: 6576201ec2]
This commit is contained in:
Ben Sander
2017-07-26 18:40:37 -05:00
parent 7fab0b5ecc
commit 235b60165f
7 changed files with 96 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_hcc.cpp
+3 -3
View File
@@ -83,8 +83,8 @@ int HIP_HIDDEN_FREE_MEM = 256;
int HIP_FORCE_SYNC_COPY = 0;
// TODO - set these to 0 and 1
int HIP_EVENT_SYS_RELEASE=1;
int HIP_COHERENT_HOST_ALLOC = 0;
int HIP_EVENT_SYS_RELEASE=0;
int HIP_HOST_COHERENT = 1;
// TODO - set to 0 once we resolve stability.
// USE_ HIP_SYNC_HOST_ALLOC
@@ -1226,7 +1226,7 @@ void HipReadEnv()
READ_ENV_I(release, HIP_SYNC_NULL_STREAM, 0, "Synchronize on host for null stream submissions");
READ_ENV_I(release, HIP_COHERENT_HOST_ALLOC, 0, "If set, all host memory will be allocated as fine-grained system memory. This allows threadfence_system to work but prevents host memory from being cached on GPU which may have performance impact.");
READ_ENV_I(release, HIP_HOST_COHERENT, 0, "If set, all host memory will be allocated as fine-grained system memory. This allows threadfence_system to work but prevents host memory from being cached on GPU which may have performance impact.");
READ_ENV_I(release, HCC_OPT_FLUSH, 0, "When set, use agent-scope fence operations rather than system-scope fence operationsflush when possible. This flag controls both HIP and HCC behavior.");