diff --git a/projects/hip/src/hip_hcc.cpp b/projects/hip/src/hip_hcc.cpp index fe69810615..5218f4e2eb 100644 --- a/projects/hip/src/hip_hcc.cpp +++ b/projects/hip/src/hip_hcc.cpp @@ -76,6 +76,7 @@ int HIP_NUM_KERNELS_INFLIGHT = 128; int HIP_WAIT_MODE = 0; int HIP_FORCE_P2P_HOST = 0; +int HIP_DENY_PEER_ACCESS = 0; // Force async copies to actually use the synchronous copy interface. int HIP_FORCE_SYNC_COPY = 0; @@ -1374,7 +1375,7 @@ void ihipInit() READ_ENV_I(release, HIP_WAIT_MODE, 0, "Force synchronization mode. 1= force yield, 2=force spin, 0=defaults specified in application"); - READ_ENV_I(release, HIP_FORCE_P2P_HOST, 0, "Force use of host/staging copy for peer-to-peer copiecopies"); + READ_ENV_I(release, HIP_FORCE_P2P_HOST, 0, "Force use of host/staging copy for peer-to-peer copies.1=always use copies, 2=always return false for hipDeviceCanAccessPeer"); READ_ENV_I(release, HIP_FORCE_SYNC_COPY, 0, "Force all copies (even hipMemcpyAsync) to use sync copies"); READ_ENV_I(release, HIP_NUM_KERNELS_INFLIGHT, 128, "Max number of inflight kernels per stream before active synchronization is forced."); @@ -1856,7 +1857,7 @@ void ihipStream_t::resolveHcMemcpyDirection(unsigned hipMemKind, *forceUnpinnedCopy = false; if (canSeeMemory(*copyDevice, dstPtrInfo, srcPtrInfo)) { - if (HIP_FORCE_P2P_HOST ) { + if (HIP_FORCE_P2P_HOST & 0x1) { *forceUnpinnedCopy = true; tprintf (DB_COPY, "P2P. Copy engine (dev:%d) can see src and dst but HIP_FORCE_P2P_HOST=0, forcing copy through staging buffers.\n", (*copyDevice)->getDeviceNum()); } else { diff --git a/projects/hip/src/hip_hcc.h b/projects/hip/src/hip_hcc.h index 66502cb7a4..0040263194 100644 --- a/projects/hip/src/hip_hcc.h +++ b/projects/hip/src/hip_hcc.h @@ -54,6 +54,7 @@ extern int HIP_DB; extern int HIP_STAGING_SIZE; /* size of staging buffers, in KB */ extern int HIP_STREAM_SIGNALS; /* number of signals to allocate at stream creation */ extern int HIP_VISIBLE_DEVICES; /* Contains a comma-separated sequence of GPU identifiers */ +extern int HIP_FORCE_P2P_HOST; //--- diff --git a/projects/hip/src/hip_peer.cpp b/projects/hip/src/hip_peer.cpp index b0e4eeef52..b7dca06e5f 100644 --- a/projects/hip/src/hip_peer.cpp +++ b/projects/hip/src/hip_peer.cpp @@ -43,13 +43,18 @@ hipError_t ihipDeviceCanAccessPeer (int* canAccessPeer, hipCtx_t thisCtx, hipCtx if ((thisCtx != NULL) && (peerCtx != NULL)) { + if (thisCtx == peerCtx) { *canAccessPeer = 0; - tprintf(DB_COPY, "Can't be peer to self. (this=%s, peer=%s)\n", + tprintf(DB_MEM, "Can't be peer to self. (this=%s, peer=%s)\n", thisCtx->toString().c_str(), peerCtx->toString().c_str()); + } else if (HIP_FORCE_P2P_HOST & 0x2) { + *canAccessPeer = false; + tprintf(DB_MEM, "HIP_FORCE_P2P_HOST denies peer access this=%s peer=%s canAccessPeer=%d\n", + thisCtx->toString().c_str(), peerCtx->toString().c_str(), *canAccessPeer); } else { *canAccessPeer = peerCtx->getDevice()->_acc.get_is_peer(thisCtx->getDevice()->_acc); - tprintf(DB_COPY, "deviceCanAccessPeer this=%s peer=%s canAccessPeer=%d\n", + tprintf(DB_MEM, "deviceCanAccessPeer this=%s peer=%s canAccessPeer=%d\n", thisCtx->toString().c_str(), peerCtx->toString().c_str(), *canAccessPeer); } @@ -58,6 +63,7 @@ hipError_t ihipDeviceCanAccessPeer (int* canAccessPeer, hipCtx_t thisCtx, hipCtx err = hipErrorInvalidDevice; } + return err; } @@ -93,7 +99,7 @@ hipError_t ihipDisablePeerAccess (hipCtx_t peerCtx) LockedAccessor_CtxCrit_t peerCrit(peerCtx->criticalData()); bool changed = peerCrit->removePeerWatcher(peerCtx, thisCtx); if (changed) { - tprintf(DB_COPY, "device %s disable access to memory allocated on peer:%s\n", + tprintf(DB_MEM, "device %s disable access to memory allocated on peer:%s\n", thisCtx->toString().c_str(), peerCtx->toString().c_str()); // Update the peers for all memory already saved in the tracker: am_memtracker_update_peers(peerCtx->getDevice()->_acc, peerCrit->peerCnt(), peerCrit->peerAgents()); @@ -127,7 +133,7 @@ hipError_t ihipEnablePeerAccess (hipCtx_t peerCtx, unsigned int flags) // Add thisCtx to peerCtx's access list so that new allocations on peer will be made visible to this device: bool isNewPeer = peerCrit->addPeerWatcher(peerCtx, thisCtx); if (isNewPeer) { - tprintf(DB_COPY, "device=%s can now see all memory allocated on peer=%s\n", + tprintf(DB_MEM, "device=%s can now see all memory allocated on peer=%s\n", thisCtx->toString().c_str(), peerCtx->toString().c_str()); am_memtracker_update_peers(peerCtx->getDevice()->_acc, peerCrit->peerCnt(), peerCrit->peerAgents()); } else { diff --git a/projects/hip/tests/src/hipPeerToPeer_simple.cpp b/projects/hip/tests/src/hipPeerToPeer_simple.cpp index a0bf6abac1..1dfbdafdfc 100644 --- a/projects/hip/tests/src/hipPeerToPeer_simple.cpp +++ b/projects/hip/tests/src/hipPeerToPeer_simple.cpp @@ -45,7 +45,7 @@ void help(char *argv[]) { printf ("usage: %s [OPTIONS]\n", argv[0]); printf (" --memcpyWithPeer : Perform memcpy with peer.\n"); - printf (" --mirrorPeersi : Mirror memory onto both default device and peerdevice. If 0, memory is mapped only on the default device.\n"); + printf (" --mirrorPeers : Mirror memory onto both default device and peerdevice. If 0, memory is mapped only on the default device.\n"); printf (" --peerDevice N : Set peer device.\n"); }; @@ -175,6 +175,9 @@ void enablePeerFirst(bool useAsyncCopy) HIPCHECK (myHipMemcpy(A_d1, A_d0, Nbytes, hipMemcpyDefault, 0/*stream*/, useAsyncCopy)); // This is P2P copy. // Copy data back to host: + // Have to wait for previous operation to finish, since we are switching to another one: + HIPCHECK(hipDeviceSynchronize()); + HIPCHECK (hipSetDevice(g_peerDevice)); HIPCHECK (myHipMemcpy(A_h, A_d1, Nbytes, hipMemcpyDeviceToHost, 0/*stream*/, useAsyncCopy)); HIPCHECK(hipDeviceSynchronize()); @@ -241,12 +244,13 @@ void allocMemoryFirst(bool useAsyncCopy) HIPCHECK (hipSetDevice(p_memcpyWithPeer ? g_peerDevice : g_currentDevice)); HIPCHECK (myHipMemcpy(A_d1, A_d0, Nbytes, hipMemcpyDefault, 0/*stream*/, useAsyncCopy)); + syncBothDevices(); // TODO - remove me, should handle this in implementation. + // Copy data back to host: HIPCHECK (hipSetDevice(g_peerDevice)); HIPCHECK (myHipMemcpy(A_h, A_d1, Nbytes, hipMemcpyDeviceToHost, 0/*stream*/, useAsyncCopy)); - HIPCHECK(hipDeviceSynchronize()); - HIPCHECK (hipSetDevice(g_currentDevice)); + syncBothDevices(); // TODO - remove me, should handle this in implementation. //--- @@ -287,15 +291,15 @@ void testPeerHostToDevice(bool useAsyncCopy) size_t Nbytes = N*sizeof(char); - char *A_d0, *A_d1; + char *A_host_d0, *A_d1; char *A_h; A_h = (char*)malloc(Nbytes); // allocate and initialize memory on device0 HIPCHECK (hipSetDevice(g_currentDevice)); - HIPCHECK (hipHostMalloc(&A_d0, Nbytes) ); - HIPCHECK (hipMemset(A_d0, memsetval, Nbytes) ); + HIPCHECK (hipHostMalloc(&A_host_d0, Nbytes) ); + HIPCHECK (hipMemset(A_host_d0, memsetval, Nbytes) ); // allocate and initialize memory on peer device HIPCHECK (hipSetDevice(g_peerDevice)); @@ -314,15 +318,15 @@ void testPeerHostToDevice(bool useAsyncCopy) if (p_memcpyWithPeer) { // p_memcpyWithPeer=1 case is HostToDevice. // if p_mirrorPeers = 1, this is accelerated copy over PCIe. - // if p_mirrorPeers = 0, this should fall back to host (because peer can't see A_d0) + // if p_mirrorPeers = 0, this should fall back to host (because peer can't see A_host_d0) HIPCHECK (hipSetDevice(g_peerDevice)); - HIPCHECK (myHipMemcpy(A_d1, A_d0, Nbytes, hipMemcpyHostToDevice, 0/*stream*/, firstAsyncCopy)); // This is P2P copy. + HIPCHECK (myHipMemcpy(A_d1, A_host_d0, Nbytes, hipMemcpyHostToDevice, 0/*stream*/, firstAsyncCopy)); // This is P2P copy. } else { // p_memcpyWithPeer=0 case is HostToDevice. // if p_mirrorPeers = 1, this is accelerated copy over PCIe. // if p_mirrorPeers = 0, this should fall back to host (because device0 can't see A_d1) HIPCHECK (hipSetDevice(g_currentDevice)); - HIPCHECK (myHipMemcpy(A_d1, A_d0, Nbytes, hipMemcpyHostToDevice, 0/*stream*/, firstAsyncCopy)); // This is P2P copy. + HIPCHECK (myHipMemcpy(A_d1, A_host_d0, Nbytes, hipMemcpyHostToDevice, 0/*stream*/, firstAsyncCopy)); // This is P2P copy. } syncBothDevices();