Revert "performance: set 1ms timeout for ulls controller since xe2"

This reverts commit 28c008e.

Signed-off-by: Compute-Runtime-Validation <[email protected]>

Author: Compute-Runtime-Validation

shared/source/debug_settings/debug_variables_base.inl

@@ -447,6 +447,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, EnableDirectSubmissionController, -1, "Enable di
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerTimeout, -1, "Set direct submission controller timeout, -1: default 5000 us, >=0: timeout in us")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerBcsTimeoutDivisor, -1, "If >=1, divide controller timeout to stop BCS only engines faster than others")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerMaxTimeout, -1, "Set direct submission controller max timeout - timeout will increase up to given value, -1: default 5000 us, >=0: max timeout in us")
+DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerDivisor, -1, "Set direct submission controller timeout divider, -1: default 1, >0: divider value")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionControllerAdjustOnThrottleAndAcLineStatus, -1, "Adjust controller timeout settings based on queue throttle and ac line status, -1: default, 0: disabled, 1: enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, DirectSubmissionForceLocalMemoryStorageMode, -1, "Force local memory storage for command/ring/semaphore buffer, -1: default - for all engines, 0: disabled, 1: for multiOsContextCapable engine, 2: for all engines")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableRingSwitchTagUpdateWa, -1, "-1: default, 0 - disable, 1 - enable. If enabled, completionFences wont be updated if ring is not running.")

shared/source/direct_submission/direct_submission_controller.cpp

@@ -10,7 +10,6 @@
 #include "shared/source/command_stream/command_stream_receiver.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/execution_environment/root_device_environment.h"
-#include "shared/source/helpers/gfx_core_helper.h"
 #include "shared/source/helpers/sleep.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/source/os_interface/os_thread.h"
@@ -26,6 +25,9 @@ DirectSubmissionController::DirectSubmissionController() {
     if (debugManager.flags.DirectSubmissionControllerTimeout.get() != -1) {
         timeout = std::chrono::microseconds{debugManager.flags.DirectSubmissionControllerTimeout.get()};
     }
+    if (debugManager.flags.DirectSubmissionControllerDivisor.get() != -1) {
+        timeoutDivisor = debugManager.flags.DirectSubmissionControllerDivisor.get();
+    }
     if (debugManager.flags.DirectSubmissionControllerBcsTimeoutDivisor.get() != -1) {
         bcsTimeoutDivisor = debugManager.flags.DirectSubmissionControllerBcsTimeoutDivisor.get();
     }
@@ -46,7 +48,7 @@ DirectSubmissionController::~DirectSubmissionController() {
 void DirectSubmissionController::registerDirectSubmission(CommandStreamReceiver *csr) {
     std::lock_guard<std::mutex> lock(directSubmissionsMutex);
     directSubmissions.insert(std::make_pair(csr, DirectSubmissionState()));
-    csr->getGfxCoreHelper().overrideDirectSubmissionTimeouts(this->timeout, this->maxTimeout);
+    this->adjustTimeout(csr);
 }
 
 void DirectSubmissionController::setTimeoutParamsForPlatform(const ProductHelper &helper) {
@@ -215,6 +217,21 @@ SteadyClock::time_point DirectSubmissionController::getCpuTimestamp() {
     return SteadyClock::now();
 }
 
+void DirectSubmissionController::adjustTimeout(CommandStreamReceiver *csr) {
+    if (EngineHelpers::isCcs(csr->getOsContext().getEngineType())) {
+        for (size_t subDeviceIndex = 0u; subDeviceIndex < csr->getOsContext().getDeviceBitfield().size(); ++subDeviceIndex) {
+            if (csr->getOsContext().getDeviceBitfield().test(subDeviceIndex)) {
+                ++this->ccsCount[subDeviceIndex];
+            }
+        }
+        auto curentMaxCcsCount = std::max_element(this->ccsCount.begin(), this->ccsCount.end());
+        if (*curentMaxCcsCount > this->maxCcsCount) {
+            this->maxCcsCount = *curentMaxCcsCount;
+            this->timeout /= this->timeoutDivisor;
+        }
+    }
+}
+
 void DirectSubmissionController::recalculateTimeout() {
     const auto now = this->getCpuTimestamp();
     const auto timeSinceLastTerminate = std::chrono::duration_cast<std::chrono::microseconds>(now - this->lastTerminateCpuTimestamp);

shared/source/direct_submission/direct_submission_controller.h

@@ -98,6 +98,7 @@ class DirectSubmissionController {
     MOCKABLE_VIRTUAL bool sleep(std::unique_lock<std::mutex> &lock);
     MOCKABLE_VIRTUAL SteadyClock::time_point getCpuTimestamp();
 
+    void adjustTimeout(CommandStreamReceiver *csr);
     void recalculateTimeout();
     void applyTimeoutForAcLineStatusAndThrottle(bool acLineConnected);
     void updateLastSubmittedThrottle(QueueThrottle throttle);

shared/source/helpers/gfx_core_helper.h

@@ -16,7 +16,6 @@
 #include "aubstream/aubstream.h"
 #include "igfxfmid.h"
 
-#include <chrono>
 #include <cstdint>
 #include <memory>
 #include <string>
@@ -200,8 +199,6 @@ class GfxCoreHelper {
     virtual uint32_t getDeviceTimestampWidth() const = 0;
     virtual void alignThreadGroupCountToDssSize(uint32_t &threadCount, uint32_t dssCount, uint32_t threadsPerDss, uint32_t threadGroupSize) const = 0;
 
-    virtual void overrideDirectSubmissionTimeouts(std::chrono::microseconds &timeout, std::chrono::microseconds &maxTimeout) const = 0;
-
     virtual ~GfxCoreHelper() = default;
 
   protected:
@@ -440,8 +437,6 @@ class GfxCoreHelperHw : public GfxCoreHelper {
 
     uint32_t getDeviceTimestampWidth() const override;
 
-    void overrideDirectSubmissionTimeouts(std::chrono::microseconds &timeout, std::chrono::microseconds &maxTimeout) const override;
-
     ~GfxCoreHelperHw() override = default;
 
   protected:

shared/source/helpers/gfx_core_helper_base.inl

@@ -833,8 +833,4 @@ uint32_t GfxCoreHelperHw<Family>::getInternalCopyEngineIndex(const HardwareInfo
     return std::min(defaultInternalCopyEngineIndex, highestAvailableIndex);
 }
 
-template <typename Family>
-void GfxCoreHelperHw<Family>::overrideDirectSubmissionTimeouts(std::chrono::microseconds &timeout, std::chrono::microseconds &maxTimeout) const {
-}
-
 } // namespace NEO

shared/source/helpers/gfx_core_helper_xe2_and_later.inl

@@ -82,16 +82,4 @@ void MemorySynchronizationCommands<Family>::encodeAdditionalTimestampOffsets(Lin
     EncodeStoreMMIO<Family>::encode(commandStream, RegisterOffsets::globalTimestampUn, globalAddress + sizeof(uint32_t), false, nullptr, isBcs);
 }
 
-template <>
-void GfxCoreHelperHw<Family>::overrideDirectSubmissionTimeouts(std::chrono::microseconds &timeout, std::chrono::microseconds &maxTimeout) const {
-    timeout = std::chrono::microseconds{1'000};
-    maxTimeout = std::chrono::microseconds{1'000};
-    if (debugManager.flags.DirectSubmissionControllerTimeout.get() != -1) {
-        timeout = std::chrono::microseconds{debugManager.flags.DirectSubmissionControllerTimeout.get()};
-    }
-    if (debugManager.flags.DirectSubmissionControllerMaxTimeout.get() != -1) {
-        maxTimeout = std::chrono::microseconds{debugManager.flags.DirectSubmissionControllerMaxTimeout.get()};
-    }
-}
-
 } // namespace NEO

shared/source/xe2_hpg_core/lnl/os_agnostic_product_helper_lnl.inl

@@ -45,6 +45,11 @@ bool ProductHelperHw<gfxProduct>::isCachingOnCpuAvailable() const {
     return false;
 }
 
+template <>
+bool ProductHelperHw<gfxProduct>::isAdjustDirectSubmissionTimeoutOnThrottleAndAcLineStatusEnabled() const {
+    return true;
+}
+
 template <>
 TimeoutParams ProductHelperHw<gfxProduct>::getDirectSubmissionControllerTimeoutParams(bool acLineConnected, QueueThrottle queueThrottle) const {
     TimeoutParams params{};

shared/test/common/test_files/igdrcl.config

@@ -125,6 +125,7 @@ DirectSubmissionMaxRingBuffers = -1
 USMEvictAfterMigration = 0
 EnableDirectSubmissionController = -1
 DirectSubmissionControllerTimeout = -1
+DirectSubmissionControllerDivisor = -1
 DirectSubmissionMonitorFenceInputPolicy = -1
 UseVmBind = -1
 EnableNullHardware = 0