Add ze_eu_count_t to get total number of EUs

Related-To: LOCI-2667

Signed-off-by: Jaime Arteaga <[email protected]>

Author: Jaime Arteaga

level_zero/core/source/device/device_imp.cpp

@@ -562,6 +562,19 @@ ze_result_t DeviceImp::getProperties(ze_device_properties_t *pDeviceProperties)
     std::string name = getNEODevice()->getDeviceInfo().name;
     memcpy_s(pDeviceProperties->name, name.length(), name.c_str(), name.length());
 
+    if (pDeviceProperties->pNext) {
+        ze_base_desc_t *extendedDesc = reinterpret_cast<ze_base_desc_t *>(pDeviceProperties->pNext);
+        if (extendedDesc->stype == ZE_STRUCTURE_TYPE_EU_COUNT_EXT) {
+            ze_eu_count_ext_t *ze_eu_count_desc = reinterpret_cast<ze_eu_count_ext_t *>(extendedDesc);
+            uint32_t numTotalEUs = hardwareInfo.gtSystemInfo.MaxEuPerSubSlice * hardwareInfo.gtSystemInfo.SubSliceCount * hardwareInfo.gtSystemInfo.SliceCount;
+
+            if (isImplicitScalingCapable()) {
+                numTotalEUs *= neoDevice->getNumGenericSubDevices();
+            }
+            ze_eu_count_desc->numTotalEUs = numTotalEUs;
+        }
+    }
+
     return ZE_RESULT_SUCCESS;
 }
 

level_zero/core/test/black_box_tests/zello_world_gpu.cpp

@@ -149,7 +149,7 @@ int main(int argc, char *argv[]) {
 
     SUCCESS_OR_TERMINATE(zeContextDestroy(context));
 
-    std::cout << "\nZello World Results validation " << (outputValidationSuccessful ? "PASSED" : "FAILED") << "\n";
+    std::cout << "\nZello World GPU Results validation " << (outputValidationSuccessful ? "PASSED" : "FAILED") << "\n";
 
     return 0;
 }

level_zero/core/test/unit_tests/sources/device/test_device.cpp

@@ -966,6 +966,34 @@ TEST_F(DeviceTest, givenDevicePropertiesStructureWhenDevicePropertiesCalledThenA
     EXPECT_NE(deviceProperties.maxMemAllocSize, devicePropertiesBefore.maxMemAllocSize);
 }
 
+TEST_F(DeviceTest, WhenRequestingZeEuCountThenExpectedEUsAreReturned) {
+    ze_device_properties_t deviceProperties = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_eu_count_ext_t ze_eu_count_desc = {ZE_STRUCTURE_TYPE_EU_COUNT_EXT};
+    deviceProperties.pNext = &ze_eu_count_desc;
+
+    uint32_t maxEuPerSubSlice = 48;
+    uint32_t subSliceCount = 8;
+    uint32_t sliceCount = 1;
+
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.MaxEuPerSubSlice = maxEuPerSubSlice;
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.SubSliceCount = subSliceCount;
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.SliceCount = sliceCount;
+
+    device->getProperties(&deviceProperties);
+
+    uint32_t expectedEUs = maxEuPerSubSlice * subSliceCount * sliceCount;
+
+    EXPECT_EQ(expectedEUs, ze_eu_count_desc.numTotalEUs);
+}
+
+TEST_F(DeviceTest, WhenRequestingZeEuCountWithIncorrectStypeThenPNextIsIgnored) {
+    ze_device_properties_t deviceProperties = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_eu_count_ext_t ze_eu_count_desc = {ZE_STRUCTURE_TYPE_SCHEDULING_HINT_EXP_PROPERTIES};
+    deviceProperties.pNext = &ze_eu_count_desc;
+    device->getProperties(&deviceProperties);
+    EXPECT_EQ(0u, ze_eu_count_desc.numTotalEUs);
+}
+
 TEST_F(DeviceTest, WhenGettingDevicePropertiesThenSubslicesPerSliceIsBasedOnSubslicesSupported) {
     ze_device_properties_t deviceProperties = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
     deviceProperties.type = ZE_DEVICE_TYPE_GPU;
@@ -1640,6 +1668,27 @@ TEST_F(MultipleDevicesDisabledImplicitScalingTest, whenCallingGetMemoryPropertie
     EXPECT_EQ(memProperties.totalSize, device0->getNEODevice()->getDeviceInfo().globalMemSize / numSubDevices);
 }
 
+TEST_F(MultipleDevicesEnabledImplicitScalingTest, WhenRequestingZeEuCountThenExpectedEUsAreReturned) {
+    ze_device_properties_t deviceProperties = {ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES};
+    ze_eu_count_ext_t ze_eu_count_desc = {ZE_STRUCTURE_TYPE_EU_COUNT_EXT};
+    deviceProperties.pNext = &ze_eu_count_desc;
+
+    uint32_t maxEuPerSubSlice = 48;
+    uint32_t subSliceCount = 8;
+    uint32_t sliceCount = 1;
+
+    L0::Device *device = driverHandle->devices[0];
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.MaxEuPerSubSlice = maxEuPerSubSlice;
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.SubSliceCount = subSliceCount;
+    device->getNEODevice()->getRootDeviceEnvironment().getMutableHardwareInfo()->gtSystemInfo.SliceCount = sliceCount;
+
+    device->getProperties(&deviceProperties);
+
+    uint32_t expectedEUs = maxEuPerSubSlice * subSliceCount * sliceCount;
+
+    EXPECT_EQ(expectedEUs * numSubDevices, ze_eu_count_desc.numTotalEUs);
+}
+
 TEST_F(MultipleDevicesEnabledImplicitScalingTest, whenCallingGetMemoryPropertiesWithSubDevicesThenCorrectSizeReturned) {
     L0::Device *device0 = driverHandle->devices[0];
     uint32_t count = 1;