Add experimental extension to query mem properties of sorting kernels

[jandres742]

No description provided.

[jandres742]

@wdamon-intel please review

[andreyfe1]

I'm ok with changes. Thanks!

[andreyfe1]

Minor: seems, empty line at the end is missed.

[wdamon-intel]

Overall, looks good; just a few minor questions/comments.

[wdamon-intel]

Would it beneficial to have an explanation of this API with an example usage after the API overview section?

[jandres742]

@andreyfe1 : Please let me know if the following text is correct:

This extension allows applications to query for memory requirements of
the underlying compiler when selecting a particular group algorithm in
their kernels. Three types of algorithms are available:

$X_GROUP_ALGORITHM_TYPE_EXP_SORT: sorting algorithm

$X__GROUP_ALGORITHM_TYPE_EXP_SCAN: scan algorithm

$X__GROUP_ALGORITHM_TYPE_EXP_REDUCE: reduce algorithm

To know what amount of memory is needed to run use algorithms,
application would call this extension as follows:


$x_device_group_algorithm_memory_exp_properties_t groupAlgorithmMemoryProperties {];
groupAlgorithmMemoryProperties.stype = $X_STRUCTURE_TYPE_DEVICE_GROUP_ALGORITHM_MEMORY_EXP_PROPERTIES;
groupAlgorithmMemoryProperties.algorithm = $X_GROUP_ALGORITHM_TYPE_EXP_SORT;
groupAlgorithmMemoryProperties.memoryScope = $X_GROUP_ALGORITHM_MEMORY_SCOPE_EXP_SUBGROUP;

$x_result_t res = $xDeviceGetGroupAlgorithmMemoryPropertiesExp(hDevice, numOfElements, &groupAlgorithmMemoryProperties);

Upon return, groupAlgorithmMemoryProperties will contain in
globalMemorySize and sharedLocalMemorySize the global and
shared local memory needed, respectively, for executing the selected
algorithm in the target device.

[wdamon-intel]

@jandres742 Is the double underscore intentional? i.e. should $X__ be $X_ ?

[jandres742]

@wdamon-intel thanks. that's just an error. Once I write it in the scripts, I will correct it. Text in comment above is just a draft to confirm it is good.

[andreyfe1]

This extension allows applications to query for memory requirements of the underlying compiler when selecting a particular group algorithm in their kernels

@jandres742, the goal for this extension is to let users know how much temporary memory they need to allocate. When they know the value, they can allocate a memory and put it as a parameter to a group algorithm that requires additional memory. So, regarding wording. Maybe we can say something like this

This extension allows applications to query for memory requirements of
the underlying compiler when selecting a particular group algorithm in
their kernels. It gives applications the memory size that is required for
memory allocations needed for group algorithms.

[jandres742]

i just realized, this needs to be scalar values, as the struct is already pointer.

[MichalMrozek]

are we adding some algorithms to Level Zero ?
Is there somewhere more data on the request ?

[jandres742]

thanks @MichalMrozek . Sent email to you on this.

[jandres742]

@wdamon-intel , @andreyfe1 : Added content for programmers guide.

[andreyfe1]

@wdamon-intel , @andreyfe1 : Added content for programmers guide.

Looks good to me. Thanks

[rwmcguir]

Closing as duplicate of another issue pending discussion internally.