#### SYCL exceptions
* In SYCL errors are handled by throwing exceptions.
* It is crucial that these errors are handled,
otherwise your application could fail in unpredictable ways.
* In SYCL there are two kinds of error:
* Synchronous errors (thrown in user thread) .
* Asynchronous errors (thrown by the SYCL scheduler).
#### Handling errors
int main() {
queue q();
/* Synchronous code */
q.submit([&](handler &cgh) {
/* Synchronous code */
cgh.parallel_for<add>(bufO.get_range(), [=](id<1> i) {
/* Asynchronous code */
});
});
}
* Kernels run asynchronously on the device, and will throw asynchronous errors.
* Everything else runs synchronously on the host, and will throw synchronous errors.
#### SYCL exceptions
#### Handling errors
class add;
int main() {
queue q();
/* Synchronous code */
q.submit([&](handler &cgh) {
/* Synchronous code */
cgh.single_task<add>([=](id<1> i) {
/* Asynchronous code */
});
}).wait();
}
* Code on the device runs asynchronously
* If errors are not handled, the application can fail:
* SYCL 1.2.1 application will fail silently.
* SYCL 2020 provides a default async handler that will call `std::terminate`
when an asynchronous error is thrown.
class add;
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try {
queue gpuQueue(gpu_selector{});
buffer bufA{dA};
buffer bufB{dB};
buffer bufO{dO};
gpuQueue.submit([&](handler &cgh) {
auto inA = accessor{bufA, cgh, read_only};
auto inB = accessor{bufB, cgh, read_only};
auto out = accessor{bufO, cgh, write_only};
cgh.parallel_for<add>(bufO.get_range(), [=](id<1> i) {
out[i] = inA[i] + inB[i];
});
}).wait();
} catch (...) { /* handle errors */ }
}
* Synchronous errors are typically thrown by SYCL API functions.
* In order to handle all SYCL errors you must wrap everything in a try-catch block.
class add;
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try{
queue gpuQueue(gpu_selector{}, async_handler{});
buffer bufA{dA};
buffer bufB{dB};
buffer bufO{dO};
gpuQueue.submit([&](handler &cgh) {
auto inA = accessor{bufA, cgh, read_only};
auto inB = accessor{bufB, cgh, read_only};
auto out = accessor{bufO, cgh, write_only};
cgh.parallel_for<add>(bufO.get_range(), [=](id<1> i) {
out[i] = inA[i] + inB[i];
});
}).wait();
gpuQueue.throw_asynchronous();
} catch (...) { /* handle errors */
}
* Asynchronous errors errors that may have occurred will be thrown after a command group has been submitted to a `queue`.
* To handle these errors you must provide an async handler when constructing the queue object.
* Then you must also call the `throw_asynchronous` or `wait_and_throw` member functions of the `queue` class.
* This will pass the exceptions to the async handler in the user thread so they can be thrown.
class add;
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try{
queue gpuQueue(gpu_selector{}, [=](exception_list eL) {
for (auto e : eL) { std::rethrow_exception(e); }
});
buffer bufA{dA};
buffer bufB{dB};
buffer bufO{dO};
gpuQueue.submit([&](handler &cgh) {
auto inA = accessor{bufA, cgh, read_only};
auto inB = accessor{bufB, cgh, read_only};
auto out = accessor{bufO, cgh, write_only};
cgh.parallel_for<add>(bufO.get_range(), [=](id<1> i) {
out[i] = inA[i] + inB[i];
});
}).wait();
gpuQueue.throw_asynchronous();
} catch (...) { /* handle errors */ }
}
* The async handler is a C++ lambda or function object that takes as a parameter an ``exception_list``
* The exception_list class is a wrapper around a list of ``exception_ptrs`` which can be iterated over
* The exception_ptrs can be rethrown by passing them to ``std::rethrow_exception``
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try {
queue gpuQueue(gpu_selector{}, [=](exception_list eL) {
for (auto e : eL) { std::rethrow_exception(e); }
});
...
gpuQueue.throw_asynchronous();
} catch (const std::exception& e) {
std::cout << “Exception caught: ” << e.what()
<< std::endl;
}
}
* Once rethrown and caught, a SYCL exception can provide information about the error
* The ``what`` member function will return a string with more details
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try {
queue gpuQueue(gpu_selector{}, [=](exception_list eL) {
for (auto e : eL) { std::rethrow_exception(e); }
});
...
gpuQueue.throw_asynchronous();
} catch (const sycl::exception& e) {
std::cout << “Exception caught: ” << e.what();
std:: cout << “ With OpenCL error code: ”
<< e.get_cl_code() << std::endl;
}
}
* In SYCL 1.2.1, if the exception has an OpenCL error code associated with it
this can be retrieved by calling the `get_cl_code` member function
* If there is no OpenCL error code this will return `CL_SUCCESS`
* SYCL 2020 provides the `error_category_for` templated free function
that allows checking for the category of the exception
depending on the backend used (e.g. `backend::opencl`),
and `e.code().value()` will correspond to the backend error code.
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
queue gpuQueue(gpu_selector{}, [=](exception_list eL) {
for (auto e : eL) { std::rethrow_exception(e); }
});
context gpuContext = gpuQueue.get_context();
try {
...
gpuQueue.wait_and_throw();
} catch (const sycl::exception& e) {
if (e.has_context()) {
if (e.get_context() == gpuContext) {
/* handle error */
}
}
}
}
* The `has_context` member function will tell you if there is a SYCL context associated with the error
* If that returns true then the `get_context` member function will return the associated SYCL context object
## Exception Types
* In SYCL 1.2.1 there are a number of different exception types that inherit from `std::exception`
* E.g. `runtime_error`, `kernel_error`
* SYCL 2020 only has a single `sycl::exception` type
which provides different error codes
* E.g. `errc::runtime`, `errc::kernel`
## Debugging SYCL Kernel Functions
* Every SYCL 1.2.1 implementation is required to provide a host device
* This device executes native C++ code but is guaranteed to emulate the SYCL execution and memory model
* This means you can debug a SYCL kernel function by switching to the host device and using a standard C++ debugger
* For example gdb
* SYCL 2020 only guarantees that a device will always be available,
and users can query the `host_debuggable` device aspect
to check whether they can use the same functionality
as the SYCL 1.2.1 host device
class add;
int main() {
std::vector<float> dA{ 7, 5, 16, 8 }, dB{ 8, 16, 5, 7 }, dO{ 0, 0, 0, 0 };
try{
queue hostQueue(aspect_selector<aspect::host_debuggable>(), async_handler{});
buffer bufA{dA};
buffer bufB{dB};
buffer bufO{dO};
hostQueue.submit([&](handler &cgh) {
auto inA = accessor{bufA, cgh, read_only};
auto inB = accessor{bufB, cgh, read_only};
auto out = accessor{bufO, cgh, write_only};
cgh.parallel_for<add>(bufO.get_range(), [=](id<1> i) {
out[i] = inA[i] + inB[i];
});
});
hostQueue.wait_and_throw();
} catch (...) { /* handle errors */ }
}
* Any SYCL application can be debugged on the host device by switching the queue for a host queue
* Replacing the device selector for the `aspect_selector`
will ensure that the queue submits all work to the device
with the requested aspects,
in this case a host debuggable device
* In SYCL 1.2.1, `host_selector` would be used instead, deprecated in SYCL 2020
## Questions
#### Exercise
Code_Exercises/Exercise_4_Handling_Errors/source
Add error handling to a SYCL application for both synchronous and asynchronous errors.