Refactor for stream->_av.

- move _av into stream critical section. ( HCC accelerator_view is not thread-safe but HIP steram is. ) - Refactored many places in code that need to acquire critical section. some were previously thread races, ie enqueueing marker. -remove support for GRID_LAUNCH_VERSION < 20 -Enable USE_AV_COPY based on HCC work-week. - Review hipModule docs, some calrity/editing. Change-Id: I3ce7c25ece048c3504f55ecd4683e506bb1fc8b6
2016-08-30 17:29:50 -05:00
@@ -32,7 +32,7 @@ THE SOFTWARE.
 // #define USE_MEMCPYTOSYMBOL
 //
 //Use the new HCC accelerator_view::copy instead of am_copy
-#define USE_AV_COPY 0
+#define USE_AV_COPY (__hcc_workweek__ >= 16351) 

 // Compile peer-to-peer support.
 // >= 2 : use HCC hc:accelerator::get_is_peer
@@ -353,18 +353,36 @@ struct LockedBase {
 };


+class ihipModule_t{
+public:
+  hsa_executable_t executable;
+  hsa_code_object_t object;
+  std::string fileName;
+  void *ptr;
+  size_t size;
+};
+
+
+class ihipFunction_t{
+public:
+  hsa_executable_symbol_t kernel_symbol;
+  uint64_t kernel;
+};
+
+
 template <typename MUTEX_TYPE>
 class ihipStreamCriticalBase_t : public LockedBase<MUTEX_TYPE>
 {
 public:
-    ihipStreamCriticalBase_t() :
+    ihipStreamCriticalBase_t(hc::accelerator_view av) :
        _last_command_type(ihipCommandCopyH2H),
        _last_copy_signal(NULL),
        _signalCursor(0),
        _oldest_live_sig_id(1),
        _streamSigId(0),
        _kernelCnt(0),
-        _signalCnt(0)
+        _signalCnt(0),
+        _av(av)
    {
        _signalPool.resize(HIP_STREAM_SIGNALS > 0 ? HIP_STREAM_SIGNALS : 1);
    };
@@ -395,27 +413,14 @@ public:
                                               //   2 are required if a barrier packet is inserted.
    uint32_t                    _kernelCnt;    // Count of inflight kernels in this stream.  Reset at ::wait().
    SIGSEQNUM                   _streamSigId;      // Monotonically increasing unique signal id.
+
+    hc::accelerator_view        _av;
 };


 typedef ihipStreamCriticalBase_t<StreamMutex> ihipStreamCritical_t;
 typedef LockedAccessor<ihipStreamCritical_t> LockedAccessor_StreamCrit_t;

-class ihipModule_t{
-public:
-  hsa_executable_t executable;
-  hsa_code_object_t object;
-  std::string fileName;
-  void *ptr;
-  size_t size;
-};
-
-
-class ihipFunction_t{
-public:
-  hsa_executable_symbol_t kernel_symbol;
-  uint64_t kernel;
-};

 // Internal stream structure.
 class ihipStream_t {
@@ -431,16 +436,23 @@ typedef uint64_t SeqNum_t ;

    void copyAsync(void* dst, const void* src, size_t sizeBytes, unsigned kind);

+    int                  preCopyCommand(LockedAccessor_StreamCrit_t &crit, ihipSignal_t *lastCopy, hsa_signal_t *waitSignal, ihipCommand_t copyType);
+
    //---
-    // Thread-safe accessors - these acquire / release mutex:
-    bool                 lockopen_preKernelCommand();
+    // Member functions that begin with locked_ are thread-safe accessors - these acquire / release the critical mutex.
+    LockedAccessor_StreamCrit_t  lockopen_preKernelCommand();
    void                 lockclose_postKernelCommand(hc::completion_future &kernel_future);

-    int                  preCopyCommand(LockedAccessor_StreamCrit_t &crit, ihipSignal_t *lastCopy, hsa_signal_t *waitSignal, ihipCommand_t copyType);

    void                 locked_reclaimSignals(SIGSEQNUM sigNum);
    void                 locked_wait(bool assertQueueEmpty=false);
-    SIGSEQNUM            locked_lastCopySeqId() {LockedAccessor_StreamCrit_t crit(_criticalData); return lastCopySeqId(crit); };
+
+    hc::accelerator_view* locked_getAv() { LockedAccessor_StreamCrit_t crit(_criticalData); return &(crit->_av); };
+
+    void                 locked_waitEvent(hipEvent_t event);
+    void                 locked_recordEvent(hipEvent_t event);
+
+    //---

    // Use this if we already have the stream critical data mutex:
    void                 wait(LockedAccessor_StreamCrit_t &crit, bool assertQueueEmpty=false);
@@ -462,7 +474,6 @@ public:
    //---
    //Public member vars - these are set at initialization and never change:
    SeqNum_t                    _id;   // monotonic sequence ID
-    hc::accelerator_view        _av;
    unsigned                    _flags;


@@ -50,9 +50,8 @@ THE SOFTWARE.

 #if defined (GRID_LAUNCH_VERSION) and (GRID_LAUNCH_VERSION >= 20)
 // Use field names for grid_launch 2.0 structure, if HCC supports GL 2.0.
-#define USE_GRID_LAUNCH_20 1
 #else
-#define USE_GRID_LAUNCH_20 0
+#error (HCC must support GRID_LAUNCH_20)
 #endif

 #define HIP_LAUNCH_PARAM_BUFFER_POINTER ((void*) 0x01)
@@ -633,7 +632,6 @@ extern void ihipPostLaunchKernel(hipStream_t stream, grid_launch_parm &lp);
 #define KNRM  "\x1B[0m"
 #define KGRN  "\x1B[32m"

-#if USE_GRID_LAUNCH_20
 #define hipLaunchKernel(_kernelName, _numBlocks3D, _blockDim3D, _groupMemBytes, _stream, ...) \
 do {\
  grid_launch_parm lp;\
@@ -645,21 +643,6 @@ do {\
  _kernelName (lp, ##__VA_ARGS__);\
  ihipPostLaunchKernel(trueStream, lp);\
 } while(0)
-#else
-#define hipLaunchKernel(_kernelName, _numBlocks3D, _blockDim3D, _groupMemBytes, _stream, ...) \
-do {\
-  grid_launch_parm lp;\
-  lp.groupMemBytes = _groupMemBytes; \
-  hipStream_t trueStream = (ihipPreLaunchKernel(_stream, _numBlocks3D, _blockDim3D, &lp)); \
-    if (HIP_TRACE_API) {\
-        fprintf(stderr, KGRN "<<hip-api: hipLaunchKernel '%s' gridDim:(%d,%d,%d) groupDim:(%d,%d,%d) groupMem:+%d stream=%p\n" KNRM, \
-                #_kernelName, lp.gridDim.x, lp.gridDim.y, lp.gridDim.z, lp.groupDim.x, lp.groupDim.y, lp.groupDim.z, lp.groupMemBytes, (void*)(_stream));\
-    }\
-  _kernelName (lp, ##__VA_ARGS__);\
-  ihipPostLaunchKernel(trueStream, lp);\
-} while(0)
-
-#endif


 #elif defined (__HCC_C__)
@@ -1307,7 +1307,7 @@ hipError_t hipDriverGetVersion(int *driverVersion) ;
 hipError_t hipModuleLoad(hipModule_t *module, const char *fname);

 /**
- * @brief Freeing the module
+ * @brief Frees the module
 *
 * @param [in] module
 *
@@ -1319,7 +1319,7 @@ hipError_t hipModuleLoad(hipModule_t *module, const char *fname);
 hipError_t hipModuleUnload(hipModule_t module);

 /**
- * @brief Function with kname will be extracted present in module
+ * @brief Function with kname will be extracted if present in module
 *
 * @param [in] module
 * @param [in] kname
@@ -1330,19 +1330,20 @@ hipError_t hipModuleUnload(hipModule_t module);
 hipError_t hipModuleGetFunction(hipFunction_t *function, hipModule_t module, const char *kname);

 /**
- * @brief returns device memory pointer and size of the kernel present in the module with symbol - name
+ * @brief returns device memory pointer and size of the kernel present in the module with symbol @p name
 *
- * @param [in] moodule
- * @param [in] name
 * @param [out] dptr
 * @param [out[ bytes
+ * @param [in] hmod
+ * @param [in] name
 *
 * @returns hipSuccess, hipErrorInvalidValue, hipErrorNotInitialized
 */
 hipError_t hipModuleGetGlobal(hipDeviceptr_t *dptr, size_t *bytes, hipModule_t hmod, const char *name);

+
 /**
- * @brief builds module from code object which resides in host memory. And image is pointer to that location.
+ * @brief builds module from code object which resides in host memory. Image is pointer to that location.
 *
 * @param [in] image
 * @param [out] module
@@ -1351,8 +1352,9 @@ hipError_t hipModuleGetGlobal(hipDeviceptr_t *dptr, size_t *bytes, hipModule_t h
 */
 hipError_t hipModuleLoadData(hipModule_t *module, const void *image);

+
 /**
- * @brief launches kernel f with launch parameters and shared memory on stream with arguments passed to kerneelparams or extra
+ * @brief launches kernel f with launch parameters and shared memory on stream with arguments passed to kernelparams or extra
 *
 * @param [in[ f
 * @param [in] gridDimX