[TensorRT] Fix perf issue for DDS nodes run by TRT 10 #23424
+101
−8
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
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Comment on lines
2483
to
2485
|
|
||
| std::set<std::string> exclude_ops_set; | ||
| std::set<std::string> exclude_ops_set; // currently not support to exclude ops | ||
|
|
There was a problem hiding this comment.
Suggested change
| std::set<std::string> exclude_ops_set; | |
| std::set<std::string> exclude_ops_set; // currently not support to exclude ops | |
| std::set<std::string> exclude_ops_set; // currently not support to exclude ops | |
Comment on lines
2670
to
+2675
|
|
||
| #if NV_TENSORRT_MAJOR >= 10 | ||
| // TRT EP will take appropriate actions later to prevent performance degradation if the graph has DDS op that run by TRT 10. | ||
| is_dds_op_in_graph_ = IsDDSOpInSubGraph(graph, result, dds_op_set_); | ||
| #endif | ||
|
|
There was a problem hiding this comment.
Suggested change
| #if NV_TENSORRT_MAJOR >= 10 | |
| // TRT EP will take appropriate actions later to prevent performance degradation if the graph has DDS op that run by TRT 10. | |
| is_dds_op_in_graph_ = IsDDSOpInSubGraph(graph, result, dds_op_set_); | |
| #endif | |
| #if NV_TENSORRT_MAJOR >= 10 | |
| // TRT EP will take appropriate actions later to prevent performance degradation if the graph has DDS op that run by TRT 10. | |
| is_dds_op_in_graph_ = IsDDSOpInSubGraph(graph, result, dds_op_set_); | |
| #endif | |
Comment on lines
2775
to
+2778
| std::vector<NodeComputeInfo>& node_compute_funcs) { | ||
|
|
||
| #if NV_TENSORRT_MAJOR >= 10 | ||
| // There is a known performance issue with the DDS ops (NonMaxSuppression, NonZero and RoiAlign) when running TRT EP with TRT 10. |
There was a problem hiding this comment.
Suggested change
| std::vector<NodeComputeInfo>& node_compute_funcs) { | |
| #if NV_TENSORRT_MAJOR >= 10 | |
| // There is a known performance issue with the DDS ops (NonMaxSuppression, NonZero and RoiAlign) when running TRT EP with TRT 10. | |
| std::vector<NodeComputeInfo>& node_compute_funcs) { | |
| #if NV_TENSORRT_MAJOR >= 10 | |
| // There is a known performance issue with the DDS ops (NonMaxSuppression, NonZero and RoiAlign) when running TRT EP with TRT 10. |
Comment on lines
+2786
to
2788
| #endif | ||
|
|
||
| for (auto& fused_node_graph : fused_nodes_and_graphs) { |
There was a problem hiding this comment.
Suggested change
| #endif | |
| for (auto& fused_node_graph : fused_nodes_and_graphs) { | |
| #endif | |
| for (auto& fused_node_graph : fused_nodes_and_graphs) { |
Comment on lines
+643
to
+645
| * Check if DDS op is in the ComputeCapability/subgraph. | ||
| */ | ||
| bool IsDDSOpInSubGraph(const GraphViewer& graph, |
There was a problem hiding this comment.
Suggested change
| * Check if DDS op is in the ComputeCapability/subgraph. | |
| */ | |
| bool IsDDSOpInSubGraph(const GraphViewer& graph, | |
| * Check if DDS op is in the ComputeCapability/subgraph. | |
| */ | |
| bool IsDDSOpInSubGraph(const GraphViewer& graph, |
Comment on lines
+263
to
+266
| bool TensorrtExecutionProvider::IsDDSOpInSubGraph(const GraphViewer& graph, | ||
| std::vector<std::unique_ptr<ComputeCapability>>& compute_capabilities, | ||
| std::unordered_set<std::string>& dds_op_set) const { | ||
| auto is_dds_op = [&](const auto& node) { |
There was a problem hiding this comment.
Suggested change
| bool TensorrtExecutionProvider::IsDDSOpInSubGraph(const GraphViewer& graph, | |
| std::vector<std::unique_ptr<ComputeCapability>>& compute_capabilities, | |
| std::unordered_set<std::string>& dds_op_set) const { | |
| auto is_dds_op = [&](const auto& node) { | |
| bool TensorrtExecutionProvider::IsDDSOpInSubGraph(const GraphViewer& graph, | |
| std::vector<std::unique_ptr<ComputeCapability>>& compute_capabilities, | |
| std::unordered_set<std::string>& dds_op_set) const { | |
| auto is_dds_op = [&](const auto& node) { |
Comment on lines
+269
to
+271
| }; | ||
|
|
||
| for (auto& compute_capability : compute_capabilities) { |
There was a problem hiding this comment.
Suggested change
| }; | |
| for (auto& compute_capability : compute_capabilities) { | |
| }; | |
| for (auto& compute_capability : compute_capabilities) { |
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.
There is a known performance issue with the DDS ops (NonMaxSuppression, NonZero and RoiAlign) when running TRT EP with TRT 10. The issue arises because when cudaStreamSynchronize is called after inference, GPU memory is released back to the OS. As a result, for the next inference run, TRT reallocates GPU memory again from the OS, introducing overhead and leading to performance degradation.
The solution is to increase the memory pool threshold, allowing TRT to retain the allocated memory and reduce this overhead.