add multi-tile support for completion diagnostic

Related-To: NEO-6871

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

shared/source/command_stream/command_stream_receiver.cpp

@@ -323,6 +323,11 @@ void CommandStreamReceiver::cleanupResources() {
 }
 
 WaitStatus CommandStreamReceiver::waitForCompletionWithTimeout(const WaitParams &params, uint32_t taskCountToWait) {
+    bool printWaitForCompletion = DebugManager.flags.LogWaitingForCompletion.get();
+    if (printWaitForCompletion) {
+        printTagAddressContent(taskCountToWait, params.waitTimeout, true);
+    }
+
     uint32_t latestSentTaskCount = this->latestFlushedTaskCount;
     if (latestSentTaskCount < taskCountToWait) {
         if (!this->flushBatchedSubmissions()) {
@@ -331,7 +336,11 @@ WaitStatus CommandStreamReceiver::waitForCompletionWithTimeout(const WaitParams
         }
     }
 
-    return baseWaitFunction(getTagAddress(), params, taskCountToWait);
+    auto retCode = baseWaitFunction(getTagAddress(), params, taskCountToWait);
+    if (printWaitForCompletion) {
+        printTagAddressContent(taskCountToWait, params.waitTimeout, false);
+    }
+    return retCode;
 }
 
 WaitStatus CommandStreamReceiver::baseWaitFunction(volatile uint32_t *pollAddress, const WaitParams &params, uint32_t taskCountToWait) {
@@ -865,4 +874,21 @@ bool CommandStreamReceiver::createPerDssBackedBuffer(Device &device) {
     return perDssBackedBuffer != nullptr;
 }
 
+void CommandStreamReceiver::printTagAddressContent(uint32_t taskCountToWait, int64_t waitTimeout, bool start) {
+    auto postSyncAddress = getTagAddress();
+    if (start) {
+        PRINT_DEBUG_STRING(true, stdout,
+                           "\nWaiting for task count %u at location %p with timeout %llx. Current value:",
+                           taskCountToWait, postSyncAddress, waitTimeout);
+    } else {
+        PRINT_DEBUG_STRING(true, stdout,
+                           "%s", "\nWaiting completed. Current value:");
+    }
+    for (uint32_t i = 0; i < activePartitions; i++) {
+        PRINT_DEBUG_STRING(true, stdout, " %u", *postSyncAddress);
+        postSyncAddress = ptrOffset(postSyncAddress, this->postSyncWriteOffset);
+    }
+    PRINT_DEBUG_STRING(true, stdout, "%s", "\n");
+}
+
 } // namespace NEO

shared/source/command_stream/command_stream_receiver.h

@@ -342,6 +342,7 @@ class CommandStreamReceiver {
     void checkForNewResources(uint32_t submittedTaskCount, uint32_t allocationTaskCount, GraphicsAllocation &gfxAllocation);
     bool checkImplicitFlushForGpuIdle();
     void downloadTagAllocation(uint32_t taskCountToWait);
+    void printTagAddressContent(uint32_t taskCountToWait, int64_t waitTimeout, bool start);
     MOCKABLE_VIRTUAL std::unique_lock<MutexType> obtainHostPtrSurfaceCreationLock();
 
     std::unique_ptr<FlushStampTracker> flushStamp;

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -871,10 +871,6 @@ inline WaitStatus CommandStreamReceiverHw<GfxFamily>::waitForTaskCountWithKmdNot
     const auto params = kmdNotifyHelper->obtainTimeoutParams(useQuickKmdSleep, *getTagAddress(), taskCountToWait, flushStampToWait, throttle, this->isKmdWaitModeActive(),
                                                              this->isAnyDirectSubmissionEnabled());
 
-    PRINT_DEBUG_STRING(DebugManager.flags.LogWaitingForCompletion.get(), stdout,
-                       "\nWaiting for task count %u at location %p. Current value: %u\n",
-                       taskCountToWait, getTagAddress(), *getTagAddress());
-
     auto status = waitForCompletionWithTimeout(params, taskCountToWait);
     if (status == WaitStatus::NotReady) {
         waitForFlushStamp(flushStampToWait);
@@ -887,15 +883,13 @@ inline WaitStatus CommandStreamReceiverHw<GfxFamily>::waitForTaskCountWithKmdNot
         return status;
     }
 
-    UNRECOVERABLE_IF(*getTagAddress() < taskCountToWait);
+    for (uint32_t i = 0; i < this->activePartitions; i++) {
+        UNRECOVERABLE_IF(*(ptrOffset(getTagAddress(), (i * this->postSyncWriteOffset))) < taskCountToWait);
+    }
 
     if (kmdNotifyHelper->quickKmdSleepForSporadicWaitsEnabled()) {
         kmdNotifyHelper->updateLastWaitForCompletionTimestamp();
     }
-
-    PRINT_DEBUG_STRING(DebugManager.flags.LogWaitingForCompletion.get(), stdout,
-                       "\nWaiting completed. Current value: %u\n", *getTagAddress());
-
     return WaitStatus::Ready;
 }
 

shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp

@@ -2001,3 +2001,43 @@ TEST(CreateWorkPartitionAllocationTest, givenEnabledBlitterWhenInitializingWorkP
     EXPECT_TRUE(retVal);
     EXPECT_EQ(0u, memoryManager->copyMemoryToAllocationBanksCalled);
 }
+
+HWTEST_F(CommandStreamReceiverTest, givenMultipleActivePartitionsWhenWaitLogIsEnabledThenPrintTagValueForAllPartitions) {
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.LogWaitingForCompletion.set(true);
+
+    auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    csr.activePartitions = 2;
+
+    volatile uint32_t *tagAddress = csr.tagAddress;
+    constexpr uint32_t tagValue = 2;
+    *tagAddress = tagValue;
+    tagAddress = ptrOffset(tagAddress, csr.postSyncWriteOffset);
+    *tagAddress = tagValue;
+
+    WaitParams waitParams;
+    waitParams.waitTimeout = std::numeric_limits<int64_t>::max();
+    constexpr uint32_t taskCount = 1;
+
+    testing::internal::CaptureStdout();
+
+    WaitStatus status = csr.waitForCompletionWithTimeout(waitParams, taskCount);
+    EXPECT_EQ(WaitStatus::Ready, status);
+
+    std::string output = testing::internal::GetCapturedStdout();
+
+    std::stringstream expectedOutput;
+
+    expectedOutput << std::endl
+                   << "Waiting for task count " << taskCount
+                   << " at location " << const_cast<uint32_t *>(csr.tagAddress)
+                   << " with timeout " << std::hex << waitParams.waitTimeout
+                   << ". Current value: " << std::dec << tagValue
+                   << " " << tagValue
+                   << std::endl
+                   << std::endl
+                   << "Waiting completed. Current value: " << tagValue
+                   << " " << tagValue << std::endl;
+
+    EXPECT_STREQ(expectedOutput.str().c_str(), output.c_str());
+}