Enable GPU->CPU transfers #5593
Conversation
| // Creating the graph | ||
|
|
||
| for (auto& name_op_spec : op_specs_) { | ||
| string& inst_name = name_op_spec.instance_name; | ||
| OpSpec op_spec = name_op_spec.spec; | ||
| PrepareOpSpec(&op_spec, name_op_spec.logical_id); | ||
| try { | ||
| graph_builder_.Add(inst_name, op_spec); | ||
| } catch (...) { | ||
| PropagateError({std::current_exception(), | ||
| "Critical error when building pipeline:\n" + GetErrorContextMessage(op_spec), | ||
| "\nCurrent pipeline object is no longer valid."}); | ||
| } | ||
| } | ||
|
|
There was a problem hiding this comment.
This part has been moved after the outputs because when processing outputs we add MakeContiguous operators.
mzient
force-pushed
the
enable_gpu2cpu
branch
3 times, most recently
from
016ef1e
to
33fca28
Compare
| @@ -91,6 +94,20 @@ def gpu(self) -> DataNode: | |||
| return transferred_node | |||
| return DataNode(self.name, "gpu", self.source) | |||
|
|
|||
| def cpu(self) -> DataNode: | |||
There was a problem hiding this comment.
The logic here is copied from .gpu(), see above.
There was a problem hiding this comment.
If the code is identical it can be abstracted out to a function
There was a problem hiding this comment.
Something like:
def cpu(self):
return self._to_backend("cpu")
def gpu(self):
return self._to_backend("gpu")
def _to_backend(self, backend) -> DataNode:
...?
It could be done.
There was a problem hiding this comment.
Yes, this would be great! Thanks
There was a problem hiding this comment.
I support above^^
And maybe _to_backend should not be internal, maybe people would like to use it directly (transfer by parameter and not the function name).
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
mzient
force-pushed
the
enable_gpu2cpu
branch
from
33fca28
to
1826e37
Compare
…CPU `shapes` taking GPU input. Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [18299035]: BUILD STARTED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [18299035]: BUILD FAILED |
|
CI MESSAGE: [18301349]: BUILD STARTED |
|
CI MESSAGE: [18301349]: BUILD PASSED |
| * If `separated_execution == 0`, this value is ignored | ||
| * @param enable_memory_stats Enable memory stats. | ||
| */ | ||
| DLL_PUBLIC void |
There was a problem hiding this comment.
I added exec_flags so that we can (hopefully) get away without adding even more parameters when we add more flavors of the executor.
| @@ -91,6 +94,20 @@ def gpu(self) -> DataNode: | |||
| return transferred_node | |||
| return DataNode(self.name, "gpu", self.source) | |||
|
|
|||
| def cpu(self) -> DataNode: | |||
There was a problem hiding this comment.
If the code is identical it can be abstracted out to a function
| pipe = pdef() | ||
| pipe.build() | ||
| for i in range(10): | ||
| gpu, cpu = pipe.run() |
There was a problem hiding this comment.
Maybe you can add check for the TL backend here as well. check_batch can handle anything.
dali/test/python/test_pipeline.py
Outdated
| pipe = pdef() | ||
| pipe.build() | ||
| for i in range(10): | ||
| peek, gpu, cpu = pipe.run() |
| * @param enable_memory_stats Enable memory stats. | ||
| */ | ||
| DLL_PUBLIC void | ||
| daliCreatePipeline3(daliPipelineHandle *pipe_handle, const char *serialized_pipeline, int length, |
There was a problem hiding this comment.
Shouldn't you add a test to c_api_test.cc?
There was a problem hiding this comment.
Will do (perhaps in a follow-up).
Signed-off-by: Michal Zientkiewicz <michalz@nvidia.com>
|
CI MESSAGE: [18333257]: BUILD STARTED |
|
CI MESSAGE: [18333257]: BUILD PASSED |
Category:
New feature
Refactoring
Description:
GPU->CPU transfer is made possible via
.cpu()function inDataNode.Some refactoring of pipeline class.
The checks have been removed from Pipeline class. The old executor still raises an error when GPU->CPU transfer occurs.
The check for GPU arguments to CPU ops have been removed from Python front-end.
TODO: Extend InputDevice in Schema and use it for python-side checks.
Additional information:
Affected modules and functionalities:
Pipeline
Python front-end
Copy operator
Key points relevant for the review:
Some old tests were removed or reduced to allow for the new capability.
New tests were added which test:
.cpu()for transfers (including with conditionals)fn.shapes()with CPU backend and GPU inputChecklist
Documentation
DALI team only
Requirements
REQ IDs: EXE.06
JIRA TASK: DALI-4030