Add Debug flags to force Pipe and Device Enqueue support

New flags are ForcePipeSupport and ForceDeviceEnqueueSupport.
Flag DisableDeviceEnqueue has been removed.

Change-Id: I8794222ee26dc8001af29c45cb9f63381e18e5c4
Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>

opencl/source/cl_device/cl_device.cpp

@@ -181,13 +181,16 @@ DeviceBitfield ClDevice::getDeviceBitfield() const {
 }
 
 bool ClDevice::isDeviceEnqueueSupported() const {
-    if (DebugManager.flags.DisableDeviceEnqueue.get()) {
+    if (DebugManager.flags.ForceDeviceEnqueueSupport.get() != -1) {
-        return false;
+        return DebugManager.flags.ForceDeviceEnqueueSupport.get();
     }
     return device.getHardwareInfo().capabilityTable.supportsDeviceEnqueue;
 }
 
 bool ClDevice::arePipesSupported() const {
+    if (DebugManager.flags.ForcePipeSupport.get() != -1) {
+        return DebugManager.flags.ForcePipeSupport.get();
+    }
     return device.getHardwareInfo().capabilityTable.supportsPipes;
 }
 

opencl/source/cl_device/cl_device_caps.cpp

@@ -260,7 +260,7 @@ void ClDevice::initializeCaps() {
     deviceInfo.minDataTypeAlignSize = 128;
 
     deviceInfo.deviceEnqueueSupport = isDeviceEnqueueSupported();
-    if (isDeviceEnqueueSupported() || (enabledClVersion == 21)) {
+    if (isDeviceEnqueueSupported()) {
         deviceInfo.maxOnDeviceQueues = 1;
         deviceInfo.maxOnDeviceEvents = 1024;
         deviceInfo.queueOnDeviceMaxSize = 64 * MB;

opencl/source/platform/extensions.cpp

@@ -7,6 +7,7 @@
 
 #include "opencl/source/platform/extensions.h"
 
+#include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/hw_info.h"
 #include "shared/source/helpers/string.h"
 
@@ -112,12 +113,16 @@ void getOpenclCFeaturesList(const HardwareInfo &hwInfo, StackVec<cl_name_version
         }
     }
 
-    if (hwInfo.capabilityTable.supportsDeviceEnqueue) {
+    auto forceDeviceEnqueueSupport = DebugManager.flags.ForceDeviceEnqueueSupport.get();
+    if ((hwInfo.capabilityTable.supportsDeviceEnqueue && (forceDeviceEnqueueSupport == -1)) ||
+        (forceDeviceEnqueueSupport == 1)) {
         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_device_enqueue");
         openclCFeatures.push_back(openClCFeature);
     }
 
-    if (hwInfo.capabilityTable.supportsPipes) {
+    auto forcePipeSupport = DebugManager.flags.ForcePipeSupport.get();
+    if ((hwInfo.capabilityTable.supportsPipes && (forcePipeSupport == -1)) ||
+        (forcePipeSupport == 1)) {
         strcpy_s(openClCFeature.name, CL_NAME_VERSION_MAX_NAME_SIZE, "__opencl_c_pipes");
         openclCFeatures.push_back(openClCFeature);
     }

opencl/test/unit_test/api/cl_create_command_queue_with_properties_tests.cpp

@@ -343,7 +343,7 @@ TEST_F(clCreateCommandQueueWithPropertiesApi, GivenMedPriorityWhenCreatingOoqCom
 
 TEST_F(clCreateCommandQueueWithPropertiesApi, GivenDeviceEnqueueDisabledWhenCreatingDeviceQueueThenNullQueueAndInvalidQueuePropertiesErrorIsReturned) {
     DebugManagerStateRestore restorer;
-    DebugManager.flags.DisableDeviceEnqueue.set(true);
+    DebugManager.flags.ForceDeviceEnqueueSupport.set(0);
 
     auto pClDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
     MockContext context{pClDevice.get()};
@@ -357,7 +357,7 @@ TEST_F(clCreateCommandQueueWithPropertiesApi, GivenDeviceEnqueueDisabledWhenCrea
 
 TEST_F(clCreateCommandQueueWithPropertiesApi, GivenDeviceEnqueueDisabledWhenCreatingDefaultDeviceQueueThenNullQueueAndInvalidQueuePropertiesErrorIsReturned) {
     DebugManagerStateRestore restorer;
-    DebugManager.flags.DisableDeviceEnqueue.set(true);
+    DebugManager.flags.ForceDeviceEnqueueSupport.set(0);
 
     auto pClDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
     MockContext context{pClDevice.get()};

opencl/test/unit_test/api/cl_set_default_device_command_queue_tests.inl

@@ -90,7 +90,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, clSetDefaultDeviceCommandQueueApiTest, GivenNullDevi
 
 TEST_F(clSetDefaultDeviceCommandQueueApiTest, GivenDeviceNotSupportingDeviceEnqueueWhenSettingDefaultDeviceQueueThenClInvalidOperationErrorIsReturned) {
     DebugManagerStateRestore dbgRestorer;
-    DebugManager.flags.DisableDeviceEnqueue.set(true);
+    DebugManager.flags.ForceDeviceEnqueueSupport.set(0);
 
     retVal = clSetDefaultDeviceCommandQueue(pContext, testedClDevice, nullptr);
     ASSERT_EQ(CL_INVALID_OPERATION, retVal);

opencl/test/unit_test/device/device_caps_tests.cpp

@@ -1152,7 +1152,7 @@ TEST(DeviceGetCaps, givenDebugFlagToDisableDeviceEnqueuesWhenCreatingDeviceThenD
     }
 
     DebugManagerStateRestore dbgRestorer;
-    DebugManager.flags.DisableDeviceEnqueue.set(true);
+    DebugManager.flags.ForceDeviceEnqueueSupport.set(0);
 
     auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
     const auto &caps = device->getDeviceInfo();
@@ -1299,6 +1299,88 @@ TEST_F(DeviceGetCapsTest, givenOcl21DeviceWhenCheckingPipesSupportThenPipesAreSu
     }
 }
 
+TEST_F(DeviceGetCapsTest, givenDeviceEnqueueSupportForcedWhenCheckingDeviceEnqueueSupportThenDeviceEnqueueIsCorrectlyReported) {
+    DebugManagerStateRestore dbgRestorer;
+    int32_t forceDeviceEnqueueSupportValues[] = {-1, 0, 1};
+    auto hwInfo = *defaultHwInfo;
+
+    for (auto isDeviceEnqueueSupportedByHw : ::testing::Bool()) {
+        hwInfo.capabilityTable.supportsDeviceEnqueue = isDeviceEnqueueSupportedByHw;
+
+        for (auto forceDeviceEnqueueSupport : forceDeviceEnqueueSupportValues) {
+            DebugManager.flags.ForceDeviceEnqueueSupport.set(forceDeviceEnqueueSupport);
+            auto pClDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&hwInfo));
+            auto &caps = pClDevice->getDeviceInfo();
+
+            size_t deviceEnqueueFeaturesCount = 0;
+            for (auto &openclCFeature : caps.openclCFeatures) {
+                if (0 == strcmp(openclCFeature.name, "__opencl_c_device_enqueue")) {
+                    deviceEnqueueFeaturesCount++;
+                }
+            }
+
+            bool expectedDeviceEnqueueSupport =
+                ((forceDeviceEnqueueSupport == -1) ? isDeviceEnqueueSupportedByHw : forceDeviceEnqueueSupport);
+            if (expectedDeviceEnqueueSupport) {
+                EXPECT_TRUE(pClDevice->isDeviceEnqueueSupported());
+                EXPECT_EQ(1024u, caps.maxOnDeviceEvents);
+                EXPECT_EQ(1u, caps.maxOnDeviceQueues);
+                EXPECT_EQ(64u * MB, caps.queueOnDeviceMaxSize);
+                EXPECT_EQ(128 * KB, caps.queueOnDevicePreferredSize);
+                EXPECT_EQ(static_cast<cl_command_queue_properties>(CL_QUEUE_PROFILING_ENABLE | CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE),
+                          caps.queueOnDeviceProperties);
+                EXPECT_EQ(1u, deviceEnqueueFeaturesCount);
+            } else {
+                EXPECT_FALSE(pClDevice->isDeviceEnqueueSupported());
+                EXPECT_EQ(0u, caps.maxOnDeviceEvents);
+                EXPECT_EQ(0u, caps.maxOnDeviceQueues);
+                EXPECT_EQ(0u, caps.queueOnDeviceMaxSize);
+                EXPECT_EQ(0u, caps.queueOnDevicePreferredSize);
+                EXPECT_EQ(static_cast<cl_command_queue_properties>(0), caps.queueOnDeviceProperties);
+                EXPECT_EQ(0u, deviceEnqueueFeaturesCount);
+            }
+        }
+    }
+}
+
+TEST_F(DeviceGetCapsTest, givenPipeSupportForcedWhenCheckingPipeSupportThenPipeIsCorrectlyReported) {
+    DebugManagerStateRestore dbgRestorer;
+    int32_t forcePipeSupportValues[] = {-1, 0, 1};
+    auto hwInfo = *defaultHwInfo;
+
+    for (auto isPipeSupportedByHw : ::testing::Bool()) {
+        hwInfo.capabilityTable.supportsPipes = isPipeSupportedByHw;
+
+        for (auto forcePipeSupport : forcePipeSupportValues) {
+            DebugManager.flags.ForcePipeSupport.set(forcePipeSupport);
+            auto pClDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&hwInfo));
+            auto &caps = pClDevice->getDeviceInfo();
+
+            size_t pipeFeaturesCount = 0;
+            for (auto &openclCFeature : caps.openclCFeatures) {
+                if (0 == strcmp(openclCFeature.name, "__opencl_c_pipes")) {
+                    pipeFeaturesCount++;
+                }
+            }
+
+            bool expectedPipeSupport = ((forcePipeSupport == -1) ? isPipeSupportedByHw : forcePipeSupport);
+            if (expectedPipeSupport) {
+                EXPECT_TRUE(pClDevice->arePipesSupported());
+                EXPECT_EQ(16u, caps.maxPipeArgs);
+                EXPECT_EQ(1024u, caps.pipeMaxPacketSize);
+                EXPECT_EQ(1u, caps.pipeMaxActiveReservations);
+                EXPECT_EQ(1u, pipeFeaturesCount);
+            } else {
+                EXPECT_FALSE(pClDevice->arePipesSupported());
+                EXPECT_EQ(0u, caps.maxPipeArgs);
+                EXPECT_EQ(0u, caps.pipeMaxPacketSize);
+                EXPECT_EQ(0u, caps.pipeMaxActiveReservations);
+                EXPECT_EQ(0u, pipeFeaturesCount);
+            }
+        }
+    }
+}
+
 TEST(Device_UseCaps, givenCapabilityTableWhenDeviceInitializeCapsThenVmeVersionsAreSetProperly) {
     HardwareInfo hwInfo = *defaultHwInfo;
 

opencl/test/unit_test/test_files/igdrcl.config

@@ -114,7 +114,6 @@ UseBindlessBuffers = 0
 UseBindlessImages = 0
 MakeAllBuffersResident = 0
 EnableCopyOnlyCommandListsAndCommandQueues = -1
-DisableDeviceEnqueue = 0
 EnableIntelVme = -1
 EnableIntelAdvancedVme = -1
 EnableBlitterOperationsSupport = -1
@@ -140,6 +139,8 @@ RenderCompressedBuffersEnabled = -1
 EnableSharedSystemUsmSupport = -1
 EnablePassInlineData = -1
 ForceFineGrainedSVMSupport = -1
+ForceDeviceEnqueueSupport = -1
+ForcePipeSupport = -1
 ForceSystemMemoryPlacement = 0
 ForceNonSystemMemoryPlacement = 0
 ForceOCLVersion = 0

shared/source/debug_settings/debug_variables_base.inl

@@ -152,6 +152,8 @@ DECLARE_DEBUG_VARIABLE(int32_t, RenderCompressedBuffersEnabled, -1, "-1: default
 DECLARE_DEBUG_VARIABLE(int32_t, EnableSharedSystemUsmSupport, -1, "-1: default, 0: shared system memory disabled, 1: shared system memory enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, EnablePassInlineData, -1, "-1: default, 0: Do not allow to pass inline data 1: Enable passing of inline data")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceFineGrainedSVMSupport, -1, "-1: default, 0: Do not report Fine Grained SVM capabilties 1: Report SVM Fine Grained capabilities if device supports SVM")
+DECLARE_DEBUG_VARIABLE(int32_t, ForceDeviceEnqueueSupport, -1, "-1: default, 0: disabled, 1: enabled")
+DECLARE_DEBUG_VARIABLE(int32_t, ForcePipeSupport, -1, "-1: default, 0: disabled, 1: enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, UseAsyncDrmExec, -1, "-1: default, 0: Disabled 1: Enabled. If enabled, pass EXEC_OBJECT_ASYNC to exec ioctl.")
 DECLARE_DEBUG_VARIABLE(bool, EnablePackedYuv, true, "Enables cl_packed_yuv extension")
 DECLARE_DEBUG_VARIABLE(bool, EnableDeferredDeleter, true, "Enables async deleter")
@@ -169,7 +171,6 @@ DECLARE_DEBUG_VARIABLE(bool, EnableFreeMemory, false, "Enable freeMemory in memo
 DECLARE_DEBUG_VARIABLE(bool, ForceSamplerLowFilteringPrecision, false, "Force Low Filtering Precision Sampler mode")
 DECLARE_DEBUG_VARIABLE(bool, UseBindlessBuffers, false, "Force compiler to use bindless buffer addressing instead of stateful one")
 DECLARE_DEBUG_VARIABLE(bool, UseBindlessImages, false, "Force compiler to use bindless image addressing instead of stateful one")
-DECLARE_DEBUG_VARIABLE(bool, DisableDeviceEnqueue, false, "Disable support for device enqueue")
 
 /*DRIVER TOGGLES*/
 DECLARE_DEBUG_VARIABLE(bool, ForcePerDssBackedBufferProgramming, false, "Always program per-DSS memory backed buffer in preamble")