[tmva][sofie] Add support for optimal memory allocation of dynamic tensors #20434
+902
−385
Conversation
This file contains hidden or 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
Add missing support for Dynamic tensors for some operators. With this commit a full support for dynamic tensor is available for ParticleNet model. Fix also a bug in Concat operator when the concat axis is not the first one
Since we use now for boolean tensors a std::vector<uint8_t> it is not needed to have a special treatment when the output ttype of the operator is a boolean (e.g. in Comparison)
…ensors Add a new function in SOFIE_common OrganizeMemory which computes the total memory and the offset for each tensor given tensor begin /end life and size. Fix also some small issue with dynamic tensor. One is for the bias of Gemm and Conv. The broadcasting of bias is done for dynamic tensor in the Session constructor only if needed. For the broadcasted tensor there is no need to create a new tensor, but the existing one is resized to the broadcasted needed size using vector::resize
… broadcasting The assert that was generated when broadcasting dynamic tensors was not correct
Test Results 22 files 22 suites 3d 18h 17m 35s ⏱️ For more details on these failures, see this check. Results for commit 21f3675. |
sanjibansg
reviewed
Nov 20, 2025
Contributor
sanjibansg
left a comment
sanjibansg
left a comment
There was a problem hiding this comment.
Looks good to me overall, just some questions:
| for (size_t i = 0; i < fNBroadcastedInputs.size(); i++) { | ||
| inputs[i] = fNBroadcastedInputs[i] + "[id]"; | ||
|
|
||
| // implement operator without broadcasting, but using loos on all indices |
Contributor
There was a problem hiding this comment.
Suggested change
| // implement operator without broadcasting, but using loos on all indices | |
| // implement operator without broadcasting, but using loops on all indices |
| std::copy(inputData, inputData + inputLength, outputData.begin() + offset ); | ||
| offset += inputLength; | ||
| // data do not need to be written as a weight | ||
| // data do not need to be written in teh generated code |
Contributor
There was a problem hiding this comment.
Suggested change
| // data do not need to be written in teh generated code | |
| // data do not need to be written in the generated code |
Comment on lines
+778
to
+788
| //fGC += "std::vector<float> fTensor_" + i.first + ";\n"; | ||
| fGC += "float * tensor_" + i.first + " = nullptr;\n"; | ||
| } else if (i.second.type == ETensorType::DOUBLE) { | ||
| fGC += "std::vector<double> fTensor_" + i.first + ";\n"; | ||
| //fGC += "std::vector<double> fTensor_" + i.first + ";\n"; | ||
| fGC += "double * tensor_" + i.first + " = nullptr;\n"; | ||
| } else if (i.second.type == ETensorType::INT64) { | ||
| fGC += "std::vector<int64_t> fTensor_" + i.first + ";\n"; | ||
| //fGC += "std::vector<int64_t> fTensor_" + i.first + ";\n"; | ||
| fGC += "int64_t * tensor_" + i.first + " = nullptr;\n"; | ||
| } else if (i.second.type == ETensorType::BOOL) { | ||
| //fGC += "std::vector<uint8_t> fTensor_" + i.first + ";\n"; | ||
| fGC += "uint8_t * tensor_" + i.first + " = nullptr;\n"; |
Contributor
There was a problem hiding this comment.
maybe we remove the commented out code?
| bool modelHasWeights = false; | ||
| for (auto &i : fInitializedTensors) { | ||
| if (i.second.type() == ETensorType::FLOAT) { | ||
| if (i.second.IsWeightTensor()) { |
Contributor
There was a problem hiding this comment.
Will it be an issue if we do not make type checks here?
Comment on lines
+878
to
+884
| // for (auto &i : fDynamicTensorInfos) { | ||
| // auto length = ConvertDynamicShapeToLength(i.second.shape); | ||
| // out << SP << "if (" << length << " > 0) {\n"; | ||
| // out << SP << SP << "fTensor_" << i.first << ".resize(" << length << ");\n"; | ||
| // out << SP << SP << "tensor_" << i.first << " = fTensor_" << i.first << ".data();\n"; | ||
| // out << SP << "}\n"; | ||
| // } |
Contributor
There was a problem hiding this comment.
maybe we can remove this commented code?
|
|
||
| struct MemoryEvent { | ||
| int t; // time (i.e. operator index) | ||
| int type; // 0 = END first, 1 = START |
Contributor
There was a problem hiding this comment.
what does the tensor index signify here?
Comment on lines
+247
to
+258
| // /d to add a new intermediate tensor for broadcasted bias tensor | ||
| // fNC2 = fNC + "bcast"; | ||
| // if (!fIsDynamic) { | ||
| // model.AddIntermed/ In case of session add broadcasting code in Session constructor and in GenerateInitCode | ||
| // // we neeiateTensor(fNC2, model.GetTensorType(fNC), shapeY); | ||
| // } | ||
| // else | ||
| // model.AddDynamicTensor(fNC2,model.GetTensorType(fNC), fShapeY); | ||
| // // do not add to lists of input/output tensors since broadcasted tensors are special | ||
| // // and we manage their memory separatly | ||
| // //fInputTensorNames.emplace_back(fNC2); | ||
| // //fOutputTensorNames.emplace_back(fNC2); |
Contributor
There was a problem hiding this comment.
if this else block is not needed anymore, maybe we can remove the if-else branching completely?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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 Pull request provides support for optimal memory allocation of dynamic tensor.
A function to compute the total size and the optimal offset for each tensor given the dynamic input parameters (e.g. batch_size, number of input features, etc..) is added in SOFIE_Common.