Overhaul backend function execution for improved performance and flexibility #270
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
jhalakpatel
commented
Oct 15, 2024
jhalakpatel
commented
Oct 15, 2024
jhalakpatel
force-pushed
the
jhalakp-use-non-dps-exec-func
branch
from
ba2dd98
to
b745b6a
Compare
jhalakpatel
force-pushed
the
jhalakp-use-non-dps-exec-func
branch
from
b745b6a
to
780e18b
Compare
jhalakpatel
changed the title
jhalakpatel
force-pushed
the
jhalakp-use-non-dps-exec-func
branch
2 times, most recently
from
ceeeaf6
to
19e841d
Compare
jhalakpatel
commented
Nov 4, 2024
| @@ -188,7 +188,7 @@ def eval(self) -> runtime.MemRefValue: | |||
| executor = Executor(executable) | |||
| # Upon computing the value of this tensor, we switch it to have a `Storage` | |||
| # parameter so that it does not need to be computed again. | |||
| data = executor.execute([out.device for out in flat_ir.outputs]) | |||
| data = executor.execute() | |||
| executor.stream.synchronize() | |||
| assert len(data) == 1, "Expects only one output from mlir_tensorrt.compiler executor" | |||
| data = data[0] | |||
There was a problem hiding this comment.
Could we also remove the hack a few lines below?
self.trace_tensor.device = flat_ir.outputs[0].deviceThere was a problem hiding this comment.
There are still a few issues:
- MLIR-TensorRT still allocates inputs only on the device.
- This PR fixes the copy operation only.
There are still issues since the input get always allocated on device. There will be a device mismatch between inputs (which are always on device) vs output which could be now on (host as well as device).
def infer_devices(self):
"""
Infers devices for the operation and updates output tensor devices accordingly.
"""
assert (
> self.inputs and len(self.outputs) == 1 and all(inp.device == self.inputs[0].device for inp in self.inputs)
), "Default implementation cannot handle cases where there are no inputs, multiple outputs, or multiple inputs with different devices. Please override."
E AssertionError: Default implementation cannot handle cases where there are no inputs, multiple outputs, or multiple inputs with different devices. Please override.
5247906
to
1914775
Compare
jhalakpatel
force-pushed
the
jhalakp-use-non-dps-exec-func
branch
2 times, most recently
from
a44e199
to
4986057
Compare
…ibility This PR replaces the DPS-style calling convention with a non-DPS approach, eliminating the requirement for call sites to preallocate output buffers. This change enables us to bypass the computation of output shapes and advance allocation of output buffers, laying the groundwork for supporting data-dependent shapes where network outputs can have dynamic dimensions. The underlying compiler stack has been enhanced to avoid allocating oversized buffers and eliminate an extra device-to-device copy operation from TensorRT-allocated memory to MLIR-TRT managed memory. Additionally, we've improved the copy operation to support copying to host memory. This enhancement removes the need to track output device allocations for device-to-host copies. Previously, copy outputs were restricted to device allocations; now they can be allocated on both device and host. Tests have been updated to align with the new calling convention, ensuring compatibility and correctness. Other changes: Fix type constraints tests Address review comments
jhalakpatel
force-pushed
the
jhalakp-use-non-dps-exec-func
branch
from
4986057
to
383c182
Compare
jhalakpatel
commented
Nov 11, 2024
| @@ -186,10 +186,11 @@ def eval(self) -> runtime.MemRefValue: | |||
|
|
|||
| compiler = Compiler(trt_builder_opt_level=0) | |||
| executable = compiler.compile(mlir, flat_ir=flat_ir) | |||
| # Ensure that session and client are available as long as tensor lives. | |||
There was a problem hiding this comment.
Remove this comment.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR replaces the DPS-style calling convention with a non-DPS approach, eliminating the requirement for call sites to preallocate output buffers. This change enables us to bypass the computation of output shapes and advance allocation of output buffers, laying the groundwork for supporting data-dependent shapes where network outputs can have dynamic dimensions.
The underlying compiler stack has been enhanced to avoid allocating oversized buffers and eliminate an extra device-to-device copy operation from TensorRT-allocated memory to MLIR-TRT managed memory.
Additionally, we've improved the copy operation to support copying to host memory. This enhancement removes the need to track output device allocations for device-to-host copies. Previously, copy outputs were restricted to device allocations; now they can be allocated on both device and host.
Tests have been updated to align with the new calling convention, ensuring compatibility and correctness.