Fix CTS api sub test failure for negative_get_command_queue_info#2634
Fix CTS api sub test failure for negative_get_command_queue_info#2634jujiang-del wants to merge 1 commit intoKhronosGroup:mainfrom
Conversation
| const cl_queue_properties properties[] = { | ||
| CL_QUEUE_PROPERTIES, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, 0 | ||
| CL_QUEUE_ON_DEVICE, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, 0 | ||
| }; | ||
| cl_int err = CL_INVALID_VALUE; | ||
| clCommandQueueWrapper cmd_queue = clCreateCommandQueueWithProperties( | ||
| context, device, properties, &err); | ||
| test_error(err, "clCreateCommandQueueWithProperties"); | ||
| cl_uint queue_size = -1; | ||
| err = clGetCommandQueueInfo(cmd_queue, CL_QUEUE_SIZE, | ||
| sizeof(queue_size), &queue_size, nullptr); | ||
| test_failure_error_ret(err, CL_INVALID_COMMAND_QUEUE, |
There was a problem hiding this comment.
I don't think this is a correct change. Note that this is a negative test, so it is checking that the proper error code is returned. Specifically, it is testing that CL_INVALID_COMMAND_QUEUE is returned when querying CL_QUEUE_SIZE for a command-queue that is not a device command-queue, therefore the command-queue should not be created with the CL_QUEUE_ON_DEVICE property.
IMHO there are some improvements we could make to this test - for example, we could run this scenario on all OpenCL 2.x or newer device, not just those that support device command-queues - but I do think the test is fundamentally correct. Please let me know if you still believe otherwise - thanks1
There was a problem hiding this comment.
If keep the original CL_QUEUE_PROPERTIES, the API clCreateCommandQueueWithProperties will fail and won't proceed to the following calls.
There was a problem hiding this comment.
No, the call with CL_QUEUE_PROPERTIES should succeed: It's creating a command queue with an array of properties, and the only property that is being passed is the CL_QUEUE_PROPERTIES bitfield with value CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE.
In other words, this is the equivalent of this call without using the array of properties:
clCommandQueueWrapper cmd_queue = clCreateCommandQueue(
context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err);edit: fixed copy-paste error
I ran this test on our CPU device (which supports device queues) and it behaved as expected:
>>>> clCreateCommandQueueWithProperties: context = 0x5cea69aec7c8, device = 12th Gen Intel(R) Core(TM) i9-12900K (CL_DEVICE_TYPE_CPU) (0x5cea69c1dbe8), properties = [ CL_QUEUE_PROPERTIES = CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE ]
<<<< clCreateCommandQueueWithProperties: returned 0x5cea6b7a8098 -> CL_SUCCESS
>>>> clGetCommandQueueInfo: command_queue = 0x5cea6b7a8098, param_name = CL_QUEUE_SIZE (00001094)
<<<< clGetCommandQueueInfo -> CL_INVALID_COMMAND_QUEUE
There was a problem hiding this comment.
The test code queries clGetDeviceInfo for the property CL_DEVICE_QUEUE_ON_DEVICE_PROPERTIES. When creating a command queue for the device, the correct flag to use is CL_QUEUE_ON_DEVICE. The runtime must validate that the device actually supports this capability. This requires carefully checking the value returned by clGetDeviceInfo.
When CL_QUEUE_PROPERTIES is incorrectly passed to clCreateCommandQueue, the call fails, and the subsequent negative test clGetCommandQueueInfo does not execute.
Additionally, several negative API test cases validate CL_QUEUE_PROPERTIES with the flag CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, including:
negative_create_command_queue
negative_create_command_queue_with_properties
After fully investigating and verifying the behavior, including applying the CTS patch and updating the OpenCL runtime validation, all 109 out of 109 tests now pass.
3 participants
To fix a test failure in api sub-test "negative_get_command_queue_info".
The test code gets clGetDeviceInfo for the property CL_DEVICE_QUEUE_ON_DEVICE_PROPERTIES, when it create command queue for the device, "CL_QUEUE_ON_DEVICE" should be the correct input, instead of CL_QUEUE_PROPERTIES.