Add support for GT-Pin Callbacks [3/n]

Change-Id: Iea4b49efc9a666fde310ece15a9c69686d22f627

Author: hjnapiatek

Jenkinsfile

@@ -2,4 +2,4 @@
 neoDependenciesRev='735095-769'
 strategy='EQUAL'
 allowedF=43
-allowedCD=341
+allowedCD=340

runtime/api/api.cpp

@@ -359,7 +359,6 @@ cl_int CL_API_CALL clReleaseContext(cl_context context) {
     Context *pContext = castToObject<Context>(context);
     if (pContext) {
         pContext->release();
-        gtpinNotifyContextDestroy(context);
         return CL_SUCCESS;
     }
 
@@ -1294,7 +1293,6 @@ cl_kernel CL_API_CALL clCreateKernel(cl_program clProgram,
     Program *pProgram = nullptr;
     cl_kernel kernel = nullptr;
     cl_int retVal = CL_SUCCESS;
-
     DBG_LOG_INPUTS("clProgram", clProgram, "kernelName", kernelName);
 
     do {
@@ -1350,6 +1348,9 @@ cl_int CL_API_CALL clCreateKernelsInProgram(cl_program clProgram,
                     program,
                     *kernelInfo,
                     nullptr);
+                if (kernels[ordinal] != nullptr) {
+                    gtpinNotifyKernelCreate(kernels[ordinal]);
+                }
             }
         }
 
@@ -2559,6 +2560,11 @@ cl_int CL_API_CALL clEnqueueNDRangeKernel(cl_command_queue commandQueue,
         return retVal;
     }
 
+    auto pKernel = castToObjectOrAbort<Kernel>(kernel);
+    TakeOwnershipWrapper<Kernel> kernelOwnership(*pKernel, gtpinIsGTPinInitialized());
+
+    gtpinNotifyKernelSubmit(kernel, pCommandQueue);
+
     retVal = pCommandQueue->enqueueKernel(
         kernel,
         workDim,
@@ -3767,6 +3773,9 @@ cl_kernel CL_API_CALL clCloneKernel(cl_kernel sourceKernel,
     if (errcodeRet) {
         *errcodeRet = retVal;
     }
+    if (pClonedKernel != nullptr) {
+        gtpinNotifyKernelCreate(pClonedKernel);
+    }
 
     return pClonedKernel;
 }

runtime/command_queue/command_queue.cpp

@@ -28,6 +28,7 @@
 #include "runtime/device_queue/device_queue.h"
 #include "runtime/event/event.h"
 #include "runtime/event/event_builder.h"
+#include "runtime/gtpin/gtpin_notify.h"
 #include "runtime/helpers/aligned_memory.h"
 #include "runtime/helpers/array_count.h"
 #include "runtime/helpers/get_info.h"

runtime/command_queue/enqueue_common.h

@@ -27,6 +27,7 @@
 #include "runtime/command_queue/dispatch_walker.h"
 #include "runtime/command_stream/command_stream_receiver.h"
 #include "runtime/event/event_builder.h"
+#include "runtime/gtpin/gtpin_notify.h"
 #include "runtime/helpers/kernel_commands.h"
 #include "runtime/helpers/dispatch_info_builder.h"
 #include "runtime/mem_obj/buffer.h"
@@ -549,6 +550,8 @@ CompletionStamp CommandQueueHw<GfxFamily>::enqueueNonBlocked(
 
     DEBUG_BREAK_IF(taskLevel >= Event::eventNotReady);
 
+    gtpinNotifyPreFlushTask(this);
+
     CompletionStamp completionStamp = commandStreamReceiver.flushTask(
         commandStream,
         commandStreamStart,

runtime/command_stream/command_stream_receiver.cpp

@@ -24,6 +24,7 @@
 #include "runtime/command_stream/command_stream_receiver.h"
 #include "runtime/command_stream/preemption.h"
 #include "runtime/device/device.h"
+#include "runtime/gtpin/gtpin_notify.h"
 #include "runtime/memory_manager/memory_manager.h"
 #include "runtime/helpers/cache_policy.h"
 #include "runtime/os_interface/os_interface.h"
@@ -192,6 +193,7 @@ bool CommandStreamReceiver::waitForCompletionWithTimeout(bool enableTimeout, int
         }
     }
     if (*getTagAddress() >= taskCountToWait) {
+        gtpinNotifyTaskCompletion(taskCountToWait);
         return true;
     }
     return false;

runtime/command_stream/command_stream_receiver_hw.inl

@@ -23,6 +23,7 @@
 #include "runtime/command_stream/command_stream_receiver_hw.h"
 #include "runtime/command_stream/linear_stream.h"
 #include "runtime/device/device.h"
+#include "runtime/gtpin/gtpin_notify.h"
 #include "runtime/helpers/cache_policy.h"
 #include "runtime/helpers/preamble.h"
 #include "runtime/helpers/ptr_math.h"
@@ -377,6 +378,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
         engineType};
 
     this->taskLevel += levelClosed ? 1 : 0;
+
+    gtpinNotifyFlushTask(completionStamp.taskCount);
+
     return completionStamp;
 }
 

runtime/context/context.cpp

@@ -25,6 +25,7 @@
 #include "runtime/device/device.h"
 #include "runtime/device_queue/device_queue.h"
 #include "runtime/mem_obj/image.h"
+#include "runtime/gtpin/gtpin_notify.h"
 #include "runtime/helpers/get_info.h"
 #include "runtime/helpers/ptr_math.h"
 #include "runtime/platform/platform.h"
@@ -71,6 +72,7 @@ Context::~Context() {
     if (memoryManager && memoryManager->isAsyncDeleterEnabled()) {
         memoryManager->getDeferredDeleter()->removeClient();
     }
+    gtpinNotifyContextDestroy((cl_context)this);
 }
 
 DeviceQueue *Context::getDefaultDeviceQueue() {

runtime/gen8/gtpin_setup_gen8.cpp

@@ -22,6 +22,7 @@
 
 #include "gtpin_ocl_interface.h"
 #include "runtime/gtpin/gtpin_hw_helper.h"
+#include "runtime/gtpin/gtpin_hw_helper.inl"
 
 namespace OCLRT {
 

runtime/gen9/gtpin_setup_gen9.cpp

@@ -22,6 +22,7 @@
 
 #include "gtpin_ocl_interface.h"
 #include "runtime/gtpin/gtpin_hw_helper.h"
+#include "runtime/gtpin/gtpin_hw_helper.inl"
 
 namespace OCLRT {
 

runtime/gtpin/CMakeLists.txt

@@ -26,9 +26,11 @@ if(GTPIN_HEADERS_DIR)
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_helpers.h
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_hw_helper.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_hw_helper.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_hw_helper.inl
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_init.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_init.h
     ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_notify.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/gtpin_defs.h
     PARENT_SCOPE
   )
 else()

runtime/gtpin/gtpin_callback_stubs.cpp

@@ -32,4 +32,29 @@ void gtpinNotifyContextDestroy(cl_context context) {
 
 void gtpinNotifyKernelCreate(cl_kernel kernel) {
 }
+
+void gtpinNotifyKernelSubmit(cl_kernel kernel, void *pCmdQueue) {
+}
+
+void gtpinNotifyPreFlushTask(void *pCmdQueue) {
+}
+
+void gtpinNotifyFlushTask(uint32_t flushedTaskCount) {
+}
+
+void gtpinNotifyTaskCompletion(uint32_t completedTaskCount) {
+}
+
+void gtpinNotifyMakeResident(void *pKernel, void *pCommandStreamReceiver) {
+}
+
+void gtpinNotifyUpdateResidencyList(void *pKernel, void *pResidencyVector) {
+}
+
+void gtpinNotifyPlatformShutdown() {
+}
+
+bool gtpinIsGTPinInitialized() {
+    return false;
+}
 }