Memory Consistency for Memory Objects Created from SVM Allocations

[bashbaug]

Found while considering memory objects created from SVM allocations for the unified SVM extension:

What are the memory consistency expectations for memory objects created from SVM allocations?

Some specific questions to answer are:

1. If a memory object is created from a coarse-grain SVM allocation:
   • a. When are writes to the memory object visible through the coarse-grain SVM allocation on the device?
   • b. When are writes to the coarse-grain SVM allocation visible through the memory object on the device?
   • c. Is it safe to concurrently write to non-overlapping addresses through the memory object and the coarse-grain SVM allocation?
   • d. Presumably any host access is reconciled when the memory object or SVM allocation is mapped?
   • e. What is the expected behavior if the memory object is mapped for reading and the coarse-grain SVM allocation is used on the device? Does it matter if the device is only reading from the coarse-grain SVM allocation?
   • f. What is the expected behavior if the memory object is mapped for writing and the coarse-grain SVM allocation is used on the device? Does it matter if the device is only reading from the coarse-grain SVM allocation?
   • g. Confirm: It is undefined behavior to read from or write to the memory object on the device while the coarse-grain SVM allocation is mapped for writing (description of clEnqueueSVMUnmap).
   • h. What is the expected behavior if both the memory object and the coarse-grain SVM allocation are mapped for writing? Is this an error? Is this undefined behavior?
   • i. Confirm: It is undefined behavior to write to the memory object on the device while the coarse-grain SVM allocation is mapped for reading (description of clEnqueueSVMUnmap).
2. If a memory object is created from a fine-grain SVM allocation:
   • a. When are writes to the memory object visible through the fine-grain SVM allocation on the device?
   • b. When are writes to the fine-grain SVM allocation visible through the memory object on the device?
   • c. Is it safe to concurrently write to non-overlapping addresses through the memory object and the fine-grain SVM allocation?
   • d. If the fine-grain SVM allocation supports SVM atomics, is it safe to concurrently write to the same memory addresses atomically through the memory object and the fine-grain SVM allocation? What if the fine-grain SVM allocation does not support SVM atomics?
   • e. When are writes to the fine-grain SVM allocation on the host visible through the memory object on the device?
   • f. What is the expected behavior if the memory object is mapped for reading and the fine-grain SVM allocation is used on the device? Does it matter if the device is only reading from the fine-grain SVM allocation?
   • g. What is the expected behavior if the memory object is mapped for writing and the fine-grain SVM allocation is used on the device? Does it matter if the device is only reading from the fine-grain SVM allocation?
   • h. What is the expected behavior if both the memory object and the fine-grain SVM allocation are mapped for writing? Is this an error? Is this undefined behavior?
   • i. Confirm: It is OK to use the memory object on the device while the fine-grain SVM allocation is mapped for writing.

[bashbaug]

It's probably most intuitive to think of the "memory object" in the descriptions above as being a buffer memory object, but for added "fun", what if it is an image memory object?

[bashbaug]

We introduced this issue in the May 6th memory subgroup but we did not have any time to discuss in detail.

The key difference between the two scenarios is that a coarse-grain SVM allocation must be mapped to access it on the host, whereas no mapping is required for a fine-grain SVM allocation.

[bashbaug]

I started writing some tests to see whether implementations have address equivalence when a buffer memory object is created from an SVM allocation. Let's discuss next steps in our meeting today.

https://github.com/KhronosGroup/OpenCL-CTS/pull/2408