add IPC event support (#1996)

include/hip/hcc_detail/hip_runtime_api.h

@@ -97,8 +97,6 @@ typedef int hipDevice_t;
 
 typedef struct ihipStream_t* hipStream_t;
 
-// TODO: IPC implementation
-
 #define hipIpcMemLazyEnablePeerAccess 0
 
 #define HIP_IPC_HANDLE_SIZE 64
@@ -107,12 +105,9 @@ typedef struct hipIpcMemHandle_st {
     char reserved[HIP_IPC_HANDLE_SIZE];
 } hipIpcMemHandle_t;
 
-// TODO: IPC event handle currently unsupported
-struct ihipIpcEventHandle_t;
-typedef struct ihipIpcEventHandle_t* hipIpcEventHandle_t;
-
-
-// END TODO
+typedef struct hipIpcEventHandle_st {
+    char reserved[HIP_IPC_HANDLE_SIZE];
+} hipIpcEventHandle_t;
 
 typedef struct ihipModule_t* hipModule_t;
 
@@ -3154,10 +3149,8 @@ hipError_t hipIpcOpenMemHandle(void** devPtr, hipIpcMemHandle_t handle, unsigned
 hipError_t hipIpcCloseMemHandle(void* devPtr);
 
 
-// hipError_t hipIpcGetMemHandle(hipIpcMemHandle_t* handle, void* devPtr);
-// hipError_t hipIpcCloseMemHandle(void *devPtr);
-// // hipError_t hipIpcOpenEventHandle(hipEvent_t* event, hipIpcEventHandle_t handle);
-// hipError_t hipIpcOpenMemHandle(void** devPtr, hipIpcMemHandle_t handle, unsigned int flags);
+hipError_t hipIpcGetEventHandle(hipIpcEventHandle_t* handle, hipEvent_t event);
+hipError_t hipIpcOpenEventHandle(hipEvent_t* event, hipIpcEventHandle_t handle);
 
 
 /**

src/hip_event.cpp

@@ -24,6 +24,74 @@ THE SOFTWARE.
 #include "hip_hcc_internal.h"
 #include "trace_helper.h"
 
+#include <errno.h> // errno, ENOENT
+#include <fcntl.h> // O_RDWR, O_CREATE
+#include <sys/mman.h> // shm_open, shm_unlink, mmap, munmap, PROT_READ, PROT_WRITE, MAP_SHARED, MAP_FAILED
+#include <unistd.h> // ftruncate, close
+
+namespace {
+
+    inline
+    const char* hsa_to_string(hsa_status_t err) noexcept
+    {
+        const char* r{};
+
+        if (hsa_status_string(err, &r) == HSA_STATUS_SUCCESS) return r;
+
+        return "Unknown.";
+    }
+
+    template<std::size_t m, std::size_t n>
+    inline
+    void throwing_result_check(hsa_status_t res, const char (&file)[m],
+                               const char (&function)[n], int line) {
+        if (res == HSA_STATUS_SUCCESS) return;
+
+        throw std::runtime_error{"Failed in file " + (file +
+                                 (", in function \"" + (function +
+                                 ("\", on line " + std::to_string(line))))) +
+                                 ", with error: " + hsa_to_string(res)};
+    }
+
+    template<std::size_t m, std::size_t n>
+    inline
+    void throwing_retval_check(int good, int retval, const char (&file)[m],
+                                const char (&function)[n], int line) {
+        if (retval == good) return;
+
+        throw std::runtime_error{"Failed in file " + (file +
+                                 (", in function \"" + (function +
+                                 ("\", on line " + std::to_string(line))))) +
+                                 ", with error: " + strerror(retval)};
+    }
+
+    template<std::size_t m, std::size_t n, std::size_t o>
+    inline
+    void throwing_msg_check(bool bad, const char (&msg)[o],
+                                const char (&file)[m],
+                                const char (&function)[n], int line) {
+        if (!bad) return;
+
+        throw std::runtime_error{"Failed in file " + (file +
+                                 (", in function \"" + (function +
+                                 ("\", on line " + std::to_string(line))))) +
+                                 ", with error: " + msg};
+    }
+
+    template<std::size_t m, std::size_t n>
+    inline
+    void throwing_errno_check(bool bad, const char (&file)[m],
+                              const char (&function)[n], int line) {
+        if (!bad) return;
+
+        throw std::runtime_error{"Failed in file " + (file +
+                                 (", in function \"" + (function +
+                                 ("\", on line " + std::to_string(line))))) +
+                                 ", with error: " + strerror(errno)};
+    }
+
+} // Unnamed namespace.
+
 //-------------------------------------------------------------------------------------------------
 //-------------------------------------------------------------------------------------------------
 // Events
@@ -49,6 +117,43 @@ void ihipEvent_t::attachToCompletionFuture(const hc::completion_future* cf, hipS
 }
 
 
+static void createIpcEventShmemIfNeeded(ihipEventData_t &ecd) {
+    if (!ecd._ipc_name.empty()) return;
+
+    // create random shmem name
+    char name_template[] = "/tmp/eventXXXXXX";
+    int temp_fd = mkstemp(name_template);
+    throwing_errno_check(-1 == temp_fd, __FILE__, __func__, __LINE__);
+
+    // copy shmem name into event data, reformat to use a single slash
+    ecd._ipc_name = name_template;
+    ecd._ipc_name.replace(0, 5, "/hip_");
+
+    // open shmem
+    ecd._ipc_fd = shm_open(ecd._ipc_name.c_str(), O_RDWR | O_CREAT, 0777);
+    throwing_errno_check(ecd._ipc_fd < 0, __FILE__, __func__, __LINE__);
+
+    // size it
+    throwing_retval_check(0, ftruncate(ecd._ipc_fd, sizeof(ihipIpcEventShmem_t)), __FILE__, __func__, __LINE__);
+
+    // mmap it
+    ecd._ipc_shmem = (ihipIpcEventShmem_t*)mmap(0, sizeof(ihipIpcEventShmem_t), PROT_READ | PROT_WRITE, MAP_SHARED, ecd._ipc_fd, 0);
+    throwing_errno_check(NULL == ecd._ipc_shmem, __FILE__, __func__, __LINE__);
+
+    // initialize shared state
+    ecd._ipc_shmem->owners = 1;
+    ecd._ipc_shmem->read_index = -1;
+    ecd._ipc_shmem->write_index = 0;
+    for (int i=0; i < IPC_SIGNALS_PER_EVENT; i++) {
+        ecd._ipc_shmem->signal[i] = 0;
+    }
+
+    // remove temp file
+    throwing_errno_check(-1 == close(temp_fd), __FILE__, __func__, __LINE__);
+    throwing_errno_check(-1 == unlink(name_template), __FILE__, __func__, __LINE__);
+}
+
+
 static std::pair<hipEventStatus_t, uint64_t> refreshEventStatus(ihipEventData_t &ecd) {
     if (ecd._state == hipEventStatusRecording && ecd.marker().is_ready()) {
         if ((ecd._type == hipEventTypeIndependent) ||
@@ -75,9 +180,9 @@ static std::pair<hipEventStatus_t, uint64_t> refreshEventStatus(ihipEventData_t
 hipError_t ihipEventCreate(hipEvent_t* event, unsigned flags) {
     hipError_t e = hipSuccess;
 
-    // TODO-IPC - support hipEventInterprocess.
     unsigned supportedFlags = hipEventDefault | hipEventBlockingSync | hipEventDisableTiming |
-                              hipEventReleaseToDevice | hipEventReleaseToSystem;
+                              hipEventReleaseToDevice | hipEventReleaseToSystem |
+                              hipEventInterprocess;
     const unsigned releaseFlags = (hipEventReleaseToDevice | hipEventReleaseToSystem);
 
     const bool illegalFlags =
@@ -105,29 +210,66 @@ hipError_t hipEventCreate(hipEvent_t* event) {
     return ihipLogStatus(ihipEventCreate(event, 0));
 }
 
-
 hipError_t hipEventRecord(hipEvent_t event, hipStream_t stream) {
     HIP_INIT_SPECIAL_API(hipEventRecord, TRACE_SYNC, event, stream);
     if (!event) return ihipLogStatus(hipErrorInvalidHandle);
     stream = ihipSyncAndResolveStream(stream);
     LockedAccessor_EventCrit_t eCrit(event->criticalData());
-    if (eCrit->_eventData._state == hipEventStatusUnitialized) return ihipLogStatus(hipErrorInvalidHandle);
+    auto &ecd{eCrit->_eventData};
+    if (ecd._state == hipEventStatusUnitialized) return ihipLogStatus(hipErrorInvalidHandle);
     if (HIP_SYNC_NULL_STREAM && stream->isDefaultStream()) {
         // TODO-HIP_SYNC_NULL_STREAM : can remove this code when HIP_SYNC_NULL_STREAM = 0
         // If default stream , then wait on all queues.
         ihipCtx_t* ctx = ihipGetTlsDefaultCtx();
         ctx->locked_syncDefaultStream(true, true);
-        eCrit->_eventData.marker(hc::completion_future());  // reset event
-        eCrit->_eventData._stream = stream;
-        eCrit->_eventData._timestamp = hc::get_system_ticks();
-        eCrit->_eventData._state = hipEventStatusComplete;
+        ecd.marker(hc::completion_future());  // reset event
+        ecd._stream = stream;
+        ecd._timestamp = hc::get_system_ticks();
+        ecd._state = hipEventStatusComplete;
+        // TODO handle IPC case?
     }
     else {
         // Record the event in the stream:
-        eCrit->_eventData.marker(stream->locked_recordEvent(event));
-        eCrit->_eventData._stream = stream;
-        eCrit->_eventData._timestamp = 0;
-        eCrit->_eventData._state = hipEventStatusRecording;
+        ecd.marker(stream->locked_recordEvent(event));
+        ecd._stream = stream;
+        ecd._timestamp = 0;
+        ecd._state = hipEventStatusRecording;
+        if (event->_flags & hipEventInterprocess) {
+            createIpcEventShmemIfNeeded(ecd);
+            int write_index = ecd._ipc_shmem->write_index++; // fetch add
+            int offset = write_index % IPC_SIGNALS_PER_EVENT;
+            // While event still valid and still locked, spin.
+            while (ecd._ipc_shmem->signal[offset] != 0) {
+                // TODO backoff
+            }
+            // Lock signal.
+            ecd._ipc_shmem->signal[offset] = 1;
+            // forward signal state from local signal to IPC signal via host callback
+            // create callback that can be passed to hsa_amd_signal_async_handler
+            // this function decrements the IPC signal by 1 to indicate completion
+            std::atomic<int> *signal = &ecd._ipc_shmem->signal[offset];
+            auto t{new std::function<void()>{[=]() {
+                signal->store(0);
+            }}};
+            // register above callback with HSA runtime to be called when local signal
+            // is decremented from 1 to 0 by CP
+            auto local_signal = *reinterpret_cast<hsa_signal_t*>(eCrit->_eventData.marker().get_native_handle());
+            hsa_amd_signal_async_handler(local_signal, HSA_SIGNAL_CONDITION_LT, 1,
+                [](hsa_signal_value_t x, void* p) {
+                    (*static_cast<decltype(t)>(p))();
+                    delete static_cast<decltype(t)>(p);
+                    return false;
+                }, t);
+            // Update read index to indicate new signal.
+            int expected = write_index-1;
+            while (!ecd._ipc_shmem->read_index.compare_exchange_weak(expected, write_index)) {
+                throwing_msg_check(
+                    expected >= write_index,
+                    "IPC event record update read index failure",
+                    __FILE__, __func__, __LINE__);
+                expected = write_index-1;
+            }
+        }
     }
     return ihipLogStatus(hipSuccess);
 }
@@ -137,8 +279,18 @@ hipError_t hipEventDestroy(hipEvent_t event) {
     HIP_INIT_API(hipEventDestroy, event);
 
     if (event) {
+        {
+            LockedAccessor_EventCrit_t crit(event->criticalData());
+            auto &ecd{crit->_eventData};
+            if (ecd._ipc_shmem) {
+                int owners = --ecd._ipc_shmem->owners;
+                throwing_errno_check(-1 == munmap(ecd._ipc_shmem, sizeof(ihipIpcEventShmem_t)), __FILE__, __func__, __LINE__);
+                throwing_errno_check(-1 == close(ecd._ipc_fd), __FILE__, __func__, __LINE__);
+                if (0 == owners)
+                    throwing_errno_check(-1 == shm_unlink(ecd._ipc_name.c_str()), __FILE__, __func__, __LINE__);
+            }
+        }
         delete event;
-
         return ihipLogStatus(hipSuccess);
     } else {
         return ihipLogStatus(hipErrorInvalidHandle);
@@ -148,31 +300,44 @@ hipError_t hipEventDestroy(hipEvent_t event) {
 hipError_t hipEventSynchronize(hipEvent_t event) {
     HIP_INIT_SPECIAL_API(hipEventSynchronize, TRACE_SYNC, event);
 
-    if (event){
-        if (!(event->_flags & hipEventReleaseToSystem)) {
-            tprintf(DB_WARN,
-                "hipEventSynchronize on event without system-scope fence ; consider creating with "
-                "hipEventReleaseToSystem\n");
-        }
-        auto ecd = event->locked_copyCrit();
+    if (!event) return ihipLogStatus(hipErrorInvalidHandle);
 
-        if (ecd._state == hipEventStatusUnitialized) {
-            return ihipLogStatus(hipErrorInvalidHandle);
-        } else if (ecd._state == hipEventStatusCreated) {
-            // Created but not actually recorded on any device:
-            return ihipLogStatus(hipSuccess);
-        } else if (HIP_SYNC_NULL_STREAM && (ecd._stream->isDefaultStream())) {
-            auto* ctx = ihipGetTlsDefaultCtx();
-            // TODO-HIP_SYNC_NULL_STREAM - can remove this code
-            ctx->locked_syncDefaultStream(true, true);
-            return ihipLogStatus(hipSuccess);
-        } else {
-            ecd.marker().wait((event->_flags & hipEventBlockingSync) ? hc::hcWaitModeBlocked
-                                                                     : hc::hcWaitModeActive);
-            return ihipLogStatus(hipSuccess);
+    if (!(event->_flags & hipEventReleaseToSystem)) {
+        tprintf(DB_WARN,
+            "hipEventSynchronize on event without system-scope fence ; consider creating with "
+            "hipEventReleaseToSystem\n");
+    }
+
+    auto ecd = event->locked_copyCrit();
+
+    if (event->_flags & hipEventInterprocess) {
+        // this is an IPC event
+        int previous_read_index = ecd._ipc_shmem->read_index;
+        if (previous_read_index >= 0) {
+            // we have at least one recorded event, so proceed
+            int offset = previous_read_index % IPC_SIGNALS_PER_EVENT;
+            // While event still valid and still locked, spin.
+            while (ecd._ipc_shmem->read_index < previous_read_index+IPC_SIGNALS_PER_EVENT && ecd._ipc_shmem->signal[offset] != 0) {
+                // TODO backoff
+            }
         }
-    } else {
+        return ihipLogStatus(hipSuccess);
+    }
+
+    if (ecd._state == hipEventStatusUnitialized) {
         return ihipLogStatus(hipErrorInvalidHandle);
+    } else if (ecd._state == hipEventStatusCreated) {
+        // Created but not actually recorded on any device:
+        return ihipLogStatus(hipSuccess);
+    } else if (HIP_SYNC_NULL_STREAM && (ecd._stream->isDefaultStream())) {
+        auto* ctx = ihipGetTlsDefaultCtx();
+        // TODO-HIP_SYNC_NULL_STREAM - can remove this code
+        ctx->locked_syncDefaultStream(true, true);
+        return ihipLogStatus(hipSuccess);
+    } else {
+        ecd.marker().wait((event->_flags & hipEventBlockingSync) ? hc::hcWaitModeBlocked
+                                                                 : hc::hcWaitModeActive);
+        return ihipLogStatus(hipSuccess);
     }
 }
 
@@ -239,9 +404,80 @@ hipError_t hipEventQuery(hipEvent_t event) {
 
     auto ecd = event->locked_copyCrit();
 
-    if (ecd._state == hipEventStatusRecording && !ecd.marker().is_ready()) {
-        return ihipLogStatus(hipErrorNotReady);
+    // this event is either from an ipc handle, or the owner of a local ipc event
+    if (event->_flags & hipEventInterprocess) {
+        if (ecd._ipc_shmem) {
+            int previous_read_index = ecd._ipc_shmem->read_index;
+            int offset = previous_read_index % IPC_SIGNALS_PER_EVENT;
+            if (ecd._ipc_shmem->read_index < previous_read_index+IPC_SIGNALS_PER_EVENT && ecd._ipc_shmem->signal[offset] != 0) {
+                return ihipLogStatus(hipErrorNotReady);
+            }
+            else {
+                return ihipLogStatus(hipSuccess);
+            }
+        }
+    }
+    // normal event
+    else {
+        if (ecd._state == hipEventStatusRecording && !ecd.marker().is_ready()) {
+            return ihipLogStatus(hipErrorNotReady);
+        }
     }
 
     return ihipLogStatus(hipSuccess);
 }
+
+hipError_t hipIpcGetEventHandle(hipIpcEventHandle_t* handle, hipEvent_t event)
+{
+    HIP_INIT_API(hipIpcGetEventHandle, handle, event);
+
+#if USE_IPC && ATOMIC_INT_LOCK_FREE == 2
+    if (!handle) return ihipLogStatus(hipErrorInvalidHandle);
+    if (!event) return ihipLogStatus(hipErrorInvalidHandle);
+    if (!(event->_flags & hipEventInterprocess)) return ihipLogStatus(hipErrorInvalidHandle);
+    if (!(event->_flags & hipEventDisableTiming)) return ihipLogStatus(hipErrorInvalidHandle);
+
+    LockedAccessor_EventCrit_t crit(event->criticalData());
+
+    auto &ecd{crit->_eventData};
+    createIpcEventShmemIfNeeded(ecd);
+    // copy name into handle
+    ihipIpcEventHandle_t* iHandle = (ihipIpcEventHandle_t*)handle;
+    memset(iHandle->shmem_name, 0, HIP_IPC_HANDLE_SIZE);
+    ecd._ipc_name.copy(iHandle->shmem_name, std::string::npos);
+
+    return ihipLogStatus(hipSuccess);
+#else
+    return ihipLogStatus(hipErrorNotSupported);
+#endif
+}
+
+hipError_t hipIpcOpenEventHandle(hipEvent_t* event, hipIpcEventHandle_t handle)
+{
+    HIP_INIT_API(hipIpcOpenEventHandle, event, &handle);
+
+#if USE_IPC && ATOMIC_INT_LOCK_FREE == 2
+    if (!event) return ihipLogStatus(hipErrorInvalidHandle);
+
+    // create a new event with timing disabled, per spec
+    auto hip_status = ihipEventCreate(event, hipEventDisableTiming | hipEventInterprocess);
+    if (hip_status != hipSuccess) return ihipLogStatus(hip_status);
+
+    LockedAccessor_EventCrit_t crit((*event)->criticalData());
+    auto &ecd{crit->_eventData};
+    ihipIpcEventHandle_t* iHandle = (ihipIpcEventHandle_t*)&handle;
+    ecd._ipc_name = iHandle->shmem_name;
+    // open shmem
+    ecd._ipc_fd = shm_open(ecd._ipc_name.c_str(), O_RDWR, 0777);
+    throwing_errno_check(ecd._ipc_fd < 0, __FILE__, __func__, __LINE__);
+    // mmap it
+    ecd._ipc_shmem = (ihipIpcEventShmem_t*)mmap(0, sizeof(ihipIpcEventShmem_t), PROT_READ | PROT_WRITE, MAP_SHARED, ecd._ipc_fd, 0);
+    throwing_errno_check(NULL == ecd._ipc_shmem, __FILE__, __func__, __LINE__);
+    // update shared state
+    ecd._ipc_shmem->owners += 1;
+
+    return ihipLogStatus(hipSuccess);
+#else
+    return ihipLogStatus(hipErrorNotSupported);
+#endif
+}

src/hip_hcc.cpp

@@ -332,12 +332,55 @@ void ihipStream_t::locked_wait() {
     locked_wait(waited);
 };
 
+typedef struct {
+    int previous_read_index;
+    ihipIpcEventShmem_t *shmem;
+    hsa_signal_t signal;
+} callback_data_t;
+
+static void WaitThenDecrementSignal(callback_data_t *data) {
+    int offset = data->previous_read_index % IPC_SIGNALS_PER_EVENT;
+    // While event valid and locked, spin.
+    while (data->shmem->read_index < data->previous_read_index+IPC_SIGNALS_PER_EVENT && data->shmem->signal[offset] != 0) {
+    }
+    hsa_signal_store_relaxed(data->signal, 0);
+    delete data;
+}
+
 // Causes current stream to wait for specified event to complete:
 // Note this does not provide any kind of host serialization.
 void ihipStream_t::locked_streamWaitEvent(ihipEventData_t& ecd) {
     LockedAccessor_StreamCrit_t crit(_criticalData);
 
-    crit->_av.create_blocking_marker(ecd.marker(), hc::accelerator_scope);
+    // if event is an IPC event, it doesn't have a marker
+    // we use a host callback to block stream with a signal wait
+    if (ecd._ipc_shmem) {
+        // create first marker
+        auto cf = crit->_av.create_marker(hc::no_scope);
+        // get its signal
+        auto signal = *reinterpret_cast<hsa_signal_t*>(cf.get_native_handle());
+        // increment its signal value
+        hsa_signal_add_relaxed(signal, 1);
+
+        // create callback that can be passed to hsa_amd_signal_async_handler
+        // this function will host wait on IPC signal, then sets first packet's signal to 0 to indicate completion
+        auto t{new callback_data_t{ecd._ipc_shmem->read_index, ecd._ipc_shmem, signal}};
+
+        // register above callback with HSA runtime to be called when first packet's signal
+        // is decremented from 2 to 1 by CP (or it is already at 1)
+        // the HSA async handler is single threaded, so we can't block, therefore use a detached thread
+        hsa_amd_signal_async_handler(signal, HSA_SIGNAL_CONDITION_EQ, 1,
+            [](hsa_signal_value_t x, void* p) {
+                std::thread(WaitThenDecrementSignal, static_cast<decltype(t)>(p)).detach();
+                return false;
+            }, t);
+
+        // create additional marker that blocks on the first one
+        crit->_av.create_blocking_marker(cf, hc::accelerator_scope);
+    }
+    else {
+        crit->_av.create_blocking_marker(ecd.marker(), hc::accelerator_scope);
+    }
 }
 
 

src/hip_hcc_internal.h

@@ -389,18 +389,28 @@ const hipStream_t hipStreamNull = 0x0;
 
 
 /**
- * HIP IPC Handle Size
+ * HIP IPC Mem Handle Size
  */
-#define HIP_IPC_RESERVED_SIZE 24
+#define HIP_IPC_MEM_RESERVED_SIZE 24
 class ihipIpcMemHandle_t {
    public:
 #if USE_IPC
     hsa_amd_ipc_memory_t ipc_handle;  ///< ipc memory handle on ROCr
 #endif
     size_t psize;
-    char reserved[HIP_IPC_RESERVED_SIZE];
+    char reserved[HIP_IPC_MEM_RESERVED_SIZE];
 };
 
+/**
+ * HIP IPC Event Handle Size
+ */
+#define HIP_IPC_EVENT_RESERVED_SIZE 32
+class ihipIpcEventHandle_t {
+   public:
+#if USE_IPC
+    char shmem_name[HIP_IPC_HANDLE_SIZE];
+#endif
+};
 
 struct ihipModule_t {
     std::string fileName;
@@ -686,6 +696,14 @@ enum ihipEventType_t {
     hipEventTypeStopCommand,
 };
 
+#define IPC_SIGNALS_PER_EVENT 32
+typedef struct ihipIpcEventShmem_s {
+    std::atomic<int> owners;
+    std::atomic<int> read_index;
+    std::atomic<int> write_index;
+    std::atomic<int> signal[IPC_SIGNALS_PER_EVENT];
+} ihipIpcEventShmem_t;
+
 
 struct ihipEventData_t {
     ihipEventData_t() {
@@ -693,18 +711,24 @@ struct ihipEventData_t {
         _stream = NULL;
         _timestamp = 0;
         _type = hipEventTypeIndependent;
+        _ipc_name = "";
+        _ipc_fd = 0;
+        _ipc_shmem = NULL;
     };
 
-    void marker(const hc::completion_future& marker) { _marker = marker; };
+    void marker(const hc::completion_future& marker) { _marker = marker; }
     hc::completion_future& marker() { return _marker; }
-    uint64_t timestamp() const { return _timestamp; };
-    ihipEventType_t type() const { return _type; };
+    uint64_t timestamp() const { return _timestamp; }
+    ihipEventType_t type() const { return _type; }
 
     ihipEventType_t _type;
     hipEventStatus_t _state;
     hipStream_t _stream;  // Stream where the event is recorded.  Null stream is resolved to actual
                           // stream when recorded
     uint64_t _timestamp;  // store timestamp, may be set on host or by marker.
+    std::string _ipc_name;
+    int _ipc_fd;
+    ihipIpcEventShmem_t *_ipc_shmem;
    private:
     hc::completion_future _marker;
 };
@@ -716,7 +740,7 @@ template <typename MUTEX_TYPE>
 class ihipEventCriticalBase_t : LockedBase<MUTEX_TYPE> {
    public:
     explicit ihipEventCriticalBase_t(const ihipEvent_t* parentEvent) : _parent(parentEvent) {}
-    ~ihipEventCriticalBase_t(){};
+    ~ihipEventCriticalBase_t() {}
 
     // Keep data in structure so it can be easily copied into snapshots
     // (used to reduce lock contention and preserve correct lock order)

src/hip_stream.cpp

@@ -63,11 +63,11 @@ hipError_t ihipStreamCreate(TlsData *tls, hipStream_t* stream, unsigned int flag
 
             // TODO - se try-catch loop to detect memory exception?
             //
-            // Note this is an execute_any_order queue, 
+            // Note this is an execute_any_order queue,
             // CUDA stream behavior is that all kernels submitted will automatically
-            // wait for prev to complete, this behaviour will be mainatined by 
-            // hipModuleLaunchKernel. execute_any_order will help 
-	    // hipExtModuleLaunchKernel , which uses a special flag
+            // wait for prev to complete, this behaviour will be mainatined by
+            // hipModuleLaunchKernel. execute_any_order will help
+            // hipExtModuleLaunchKernel , which uses a special flag
 
             {
                 // Obtain mutex access to the device critical data, release by destructor
@@ -130,18 +130,19 @@ hipError_t hipDeviceGetStreamPriorityRange(int* leastPriority, int* greatestPrio
 hipError_t hipStreamWaitEvent(hipStream_t stream, hipEvent_t event, unsigned int flags) {
     HIP_INIT_SPECIAL_API(hipStreamWaitEvent, TRACE_SYNC, stream, event, flags);
 
-    hipError_t e = hipSuccess;
+    if (!event) return ihipLogStatus(hipErrorInvalidHandle);
 
-    if (event == nullptr) {
-        e = hipErrorInvalidHandle;
-
-    } else {
-        auto ecd = event->locked_copyCrit(); 
+    auto ecd = event->locked_copyCrit();
+    if (event->_flags & hipEventInterprocess) {
+        // this is an IPC event
+        if (ecd._ipc_shmem->read_index >= 0) {
+            // we have at least one recorded event, so proceed
+            stream->locked_streamWaitEvent(ecd);
+        }
+    }
+    else {
         if ((ecd._state != hipEventStatusUnitialized) && (ecd._state != hipEventStatusCreated)) {
             if (HIP_SYNC_STREAM_WAIT || (HIP_SYNC_NULL_STREAM && (stream == 0))) {
-                // conservative wait on host for the specified event to complete:
-                // return _stream->locked_eventWaitComplete(this, waitMode);
-                //
                 ecd.marker().wait((event->_flags & hipEventBlockingSync) ? hc::hcWaitModeBlocked
                                                                          : hc::hcWaitModeActive);
             } else {
@@ -150,9 +151,9 @@ hipError_t hipStreamWaitEvent(hipStream_t stream, hipEvent_t event, unsigned int
                 stream->locked_streamWaitEvent(ecd);
             }
         }
-    }  // else event not recorded, return immediately and don't create marker.
+    }
 
-    return ihipLogStatus(e);
+    return ihipLogStatus(hipSuccess);
 };
 
 

src/trace_helper.h

@@ -71,6 +71,11 @@ inline std::string ToString(hipEvent_t v) {
     ss << v;
     return ss.str();
 };
+//  hipIpcEventHandle_t specialization. TODO
+template <>
+inline std::string ToString(hipIpcEventHandle_t v) {
+    return std::string{};
+};
 //  hipStream_t
 template <>
 inline std::string ToString(hipStream_t v) {

tests/src/runtimeApi/event/hipEventIpc.cpp

@@ -0,0 +1,112 @@
+/*
+Copyright (c) 2015-2016 Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+// Test hipEventRecord serialization behavior.
+// Through manual inspection of the reported timestamps, can determine if recording a NULL event
+// forces synchronization : set
+
+/* HIT_START
+ * BUILD: %t %s ../../test_common.cpp EXCLUDE_HIP_PLATFORM nvcc
+ * TEST: %t --iterations 10
+ * HIT_END
+ */
+
+#include "hip/hip_runtime.h"
+#include "test_common.h"
+
+int main(int argc, char* argv[]) {
+    HipTest::parseStandardArguments(argc, argv, true);
+
+    unsigned blocks = (N + threadsPerBlock - 1) / threadsPerBlock;
+    if (blocks > 1024) blocks = 1024;
+    if (blocks == 0) blocks = 1;
+
+    printf("N=%zu (A+B+C= %6.1f MB total) blocks=%u threadsPerBlock=%u iterations=%d\n", N,
+           ((double)3 * N * sizeof(float)) / 1024 / 1024, blocks, threadsPerBlock, iterations);
+    printf("iterations=%d\n", iterations);
+
+    size_t Nbytes = N * sizeof(float);
+
+    float *A_h, *B_h, *C_h;
+    float *A_d, *B_d, *C_d;
+    HipTest::initArrays(&A_d, &B_d, &C_d, &A_h, &B_h, &C_h, N);
+
+    hipEvent_t start, stop;
+
+    // NULL stream check:
+    HIPCHECK(hipEventCreateWithFlags(&start, hipEventDisableTiming|hipEventInterprocess));
+    HIPCHECK(hipEventCreateWithFlags(&stop, hipEventDisableTiming|hipEventInterprocess));
+
+
+    HIPCHECK(hipMemcpy(A_d, A_h, Nbytes, hipMemcpyHostToDevice));
+    HIPCHECK(hipMemcpy(B_d, B_h, Nbytes, hipMemcpyHostToDevice));
+
+
+    for (int i = 0; i < iterations; i++) {
+        //--- START TIMED REGION
+        long long hostStart = HipTest::get_time();
+        // Record the start event
+        HIPCHECK(hipEventRecord(start, NULL));
+
+        hipLaunchKernelGGL(HipTest::vectorADD, dim3(blocks), dim3(threadsPerBlock), 0, 0,
+                        static_cast<const float*>(A_d), static_cast<const float*>(B_d), C_d, N);
+
+
+        HIPCHECK(hipEventRecord(stop, NULL));
+        HIPCHECK(hipEventSynchronize(stop));
+        HIPCHECK(hipEventQuery(stop));
+        long long hostStop = HipTest::get_time();
+        //--- STOP TIMED REGION
+
+
+        float eventMs = 1.0f;
+        // should fail
+        HIPASSERT(hipSuccess != hipEventElapsedTime(&eventMs, start, stop));
+        float hostMs = HipTest::elapsed_time(hostStart, hostStop);
+
+        printf("host_time (gettimeofday)          =%6.3fms\n", hostMs);
+        printf("kernel_time (hipEventElapsedTime) =%6.3fms\n", eventMs);
+        printf("\n");
+
+    }
+
+    hipIpcEventHandle_t ipc_handle;
+    HIPCHECK(hipIpcGetEventHandle(&ipc_handle, start));
+
+    hipEvent_t ipc_event;
+    HIPCHECK(hipIpcOpenEventHandle(&ipc_event, ipc_handle));
+
+    HIPCHECK(hipEventSynchronize(ipc_event));
+
+    HIPCHECK(hipEventDestroy(ipc_event));
+    HIPCHECK(hipEventDestroy(start));
+    HIPCHECK(hipEventDestroy(stop));
+
+    HIPCHECK(hipMemcpy(C_h, C_d, Nbytes, hipMemcpyDeviceToHost));
+
+
+    printf("check:\n");
+
+    HipTest::checkVectorADD(A_h, B_h, C_h, N, true);
+
+
+    passed();
+}