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

SWDEV-193956

Browse Source 
[hipclang-vdi-rocm][perf]~45% to 50% of Performance drop on
rocBLAS_int8 test

- Enable AMD_OPT_FLUSH optimization by default to match HCC
- Disable CPU writes to GPU memory on boards with large bar,
because it requires HDP flush tracking.
- Enable L2 cache on kernel arguments, because L2 will be
invalidated on memory reuse .

Change-Id: I124cf250bdd4d19c523ce542c163813828f8fbdc
This commit is contained in:
German Andryeyev
2020-02-18 13:56:17 -05:00
parent afa685234c
commit 374f612b7c
3 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/utils/flags.hpp
+1 -1
View File
@@ -226,7 +226,7 @@ release(bool, PAL_ALWAYS_RESIDENT, false, \
release(uint, HIP_HOST_COHERENT, 0, \
"Coherent memory in hipHostMalloc, 0x1 = memory is coherent with host"\
"0x0 = memory is not coherent between host and GPU") \
release(uint, AMD_OPT_FLUSH, 0, \
release(uint, AMD_OPT_FLUSH, 1, \
"Kernel flush option , 0x0 = Use system-scope fence operations." \
"0x1 = Use device-scope fence operations when possible.") \
release(uint, HIP_HIDDEN_FREE_MEM, 0, \