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

P2p checkpoint.

Browse Source 
- set USE_PEER_TO_PEER=3 (requires HCC "am_memtracker_update_peers")
- when enabling peer, turn it on for previously allocated memory.
- hipDeviceCanAccessPeer is no longer self-ware (self does not qualify
  as a peer)
- device peerlist always includes self, so when we call allow_access
  we never remove self access.
- hipDeviceReset() removes old peer mappings.
This commit is contained in:
Ben Sander
2016-04-11 12:52:18 -05:00
parent c2d37b4bbd
commit 83f0de7806
9 changed files with 202 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/hip_memory.cpp
+3 -2
View File
@@ -132,7 +132,7 @@ hipError_t hipMalloc(void** ptr, size_t sizeBytes)
hc::am_memtracker_update(*ptr, device->_device_index, 0);
{
LockedAccessor_DeviceCrit_t crit(device->criticalData());
if (crit->peerCnt()) {
if (crit->peerCnt() > 1) { // peerCnt includes self so only call allow_access if other peers involved:
hsa_status_t hsa_status = hsa_amd_agents_allow_access(crit->peerCnt(), crit->peerAgents(), NULL, *ptr);
if (hsa_status != HSA_STATUS_SUCCESS) {
hip_status = hipErrorMemoryAllocation;
@@ -173,8 +173,9 @@ hipError_t hipHostMalloc(void** ptr, size_t sizeBytes, unsigned int flags)
}else{
hc::am_memtracker_update(*ptr, device->_device_index, flags);
{
// TODO - allow_access only works for device memory, need to change am_alloc to allocate host directly.
LockedAccessor_DeviceCrit_t crit(device->criticalData());
if (crit->peerCnt()) {
if (crit->peerCnt() > 1) { // peerCnt includes self so only call allow_access if other peers involved:
hsa_status_t hsa_status = hsa_amd_agents_allow_access(crit->peerCnt(), crit->peerAgents(), NULL, *ptr);
if (hsa_status != HSA_STATUS_SUCCESS) {
hip_status = hipErrorMemoryAllocation;