呢哇 by Xiaobanli-new · Pull Request #12 · InfiniTensor/TinyInfiniTensor

Xiaobanli-new · 2026-02-03T15:17:06Z

No description provided.

Added new includes and optimized graph processing rules to eliminate redundant transpose operations and merge them into matrix multiplication attributes.

Implement shape inference for Transpose operation.

Implement shape inference for matmul operation

Fix missing newline at end of file in graph.cc

Add newline at the end of graph.cc file

Add missing newline at end of file for graph.cc

Removed TODO comments for memory allocation and deallocation.

Updated inferShape method to correctly compute the concatenated shape.

Added a missing newline at the end of the file and updated comments.

Update transpose function to reverse dimensions

Removed TODO comments and added clarification about shape.

Copilot

Pull request overview

This PR implements homework assignments for a deep learning framework, completing several operator shape inference functions, broadcast logic, graph optimization passes, and memory allocation. The PR title "呢哇" (Chinese characters) doesn't describe the changes.

Changes:

Implements shape inference for Clip, Cast, Transpose, MatMul, and Concat operators
Implements bidirectional broadcasting shape inference (infer_broadcast)
Adds graph optimization passes to eliminate redundant transposes and fuse transposes into matmul operations
Implements memory allocation with first-fit strategy and block coalescing

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
src/utils/operator_utils.cc	Implements broadcast shape inference and applies formatting changes to namespace braces
src/operators/unary.cc	Implements shape inference for Clip and Cast operators
src/operators/transpose.cc	Fixes default permutation to reverse dimensions per ONNX spec and implements shape inference with validation
src/operators/matmul.cc	Implements shape inference with batch broadcasting and transpose support
src/operators/concat.cc	Implements shape inference with dimension validation
src/core/graph.cc	Implements graph optimization passes and memory allocation with lifetime analysis
src/core/allocator.cc	Implements memory allocator with first-fit allocation and block coalescing
include/core/allocator.h	Adds data structures and helper methods for memory management

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-03T15:22:21Z

src/operators/concat.cc

+            {
+                if (r == dim)
+                    continue;
+                IT_ASSERT(cur[r] == dims[r], "Concat dims mismatch on non-concat axis");


Similar to the other assertion, this message could be more informative by including which permutation element caused the issue. Consider: "Concat dims mismatch on non-concat axis " + std::to_string(r) + ": expected " + std::to_string(dims[r]) + ", got " + std::to_string(cur[r]).

Copilot · 2026-02-03T15:22:21Z

src/operators/transpose.cc

+        {
+            int p = transposePermute[i];
+            IT_ASSERT(p >= 0 && p < rank);
+            IT_ASSERT(seen[p] == 0);


The error message for duplicate permutation indices lacks helpful details. Consider: "Duplicate permutation index " + std::to_string(p) + " at position " + std::to_string(i) + " in transpose permutation".

Copilot · 2026-02-03T15:22:21Z

src/core/allocator.cc

+    std::map<size_t, size_t>::iterator Allocator::findFreeBlock(size_t size)
+    {
+        // first-fit: 找到第一个 size 足够的空闲块
+        for (auto it = freeBlocks.begin(); it != freeBlocks.end(); ++it)
+        {
+            if (it->second >= size)
+                return it;
+        }
+        return freeBlocks.end();


The findFreeBlock function uses a linear search (first-fit) through all free blocks. For workloads with many allocations and deallocations, this could become a performance bottleneck as it has O(n) complexity where n is the number of free blocks. Consider using a more efficient data structure, such as maintaining free blocks sorted by size (in addition to address) or using a best-fit strategy with size-indexed structures for better performance. However, if the number of tensors is small (typical in most ML graphs), this may not be a practical concern.

Copilot · 2026-02-03T15:22:22Z

src/operators/concat.cc

+        Shape dims = inputs[0]->getDims();
+        // =================================== 作业 ===================================
+        // TODO：修改 dims，返回正确的 concat 后的 shape
+        // REF: https://onnx.ai/onnx/operators/onnx__Concat.html#concat-13
+        // =================================== 作业 ===================================

-    // =================================== 作业 ===================================
-    // TODO：修改 dims，返回正确的 concat 后的 shape
-    // REF: https://onnx.ai/onnx/operators/onnx__Concat.html#concat-13
-    // =================================== 作业 ===================================
+        IT_ASSERT(!inputs.empty());


The assertion that inputs is not empty (line 22) comes after dims is already initialized from inputs[0] (line 16). If inputs were empty, line 16 would access an invalid element before the assertion could catch it. Move the IT_ASSERT(!inputs.empty()) check to before line 16 to ensure the precondition is validated first.

Copilot · 2026-02-03T15:22:22Z

src/core/graph.cc

+                    auto tr = as<TransposeObj>(src);
+                    if (tr->getOutput() != in)
+                        continue;
+                    if (!isSwapLast2Permute(tr->getPermute()))
+                        continue;
+
+                    auto trIn = tr->getInputs(0);
+                    mm->replaceInput(in, trIn);
+                    if (inputIdx == 0)
+                        mm->setTransA(!mm->getTransA());
+                    else
+                        mm->setTransB(!mm->getTransB());
+                    toRemove.insert(src.get());
+                    changed = true;


Potential issue: When a transpose operator is marked for removal and added to the toRemove set, the code doesn't verify that the transpose output is only consumed by this matmul operator. If the transpose output is used by multiple consumers, removing it would break other parts of the graph. Add a check to ensure tr->getOutput()->getTargets().size() == 1 before adding to toRemove, similar to the check in rule 1 at line 157.

Copilot · 2026-02-03T15:22:22Z

src/operators/matmul.cc

+        m = aM;
+        n = bN;
+        k = aK;


The matmul m, n, k member variables are being modified during inferShape, but this method is marked as non-const and is called during shape inference. These member variables should be considered as operator attributes that are derived from the inputs during inference. However, the issue is that m, n, k are being set as side effects of inferShape, which could lead to issues if inferShape is called multiple times or in different contexts. Consider whether these should be computed on-demand via getter methods instead of being cached as mutable state, or document that these are only valid after inferShape has been called.

Copilot · 2026-02-03T15:22:23Z

src/utils/operator_utils.cc

+            else if (dimB == 1)
+                out[i] = dimA;
+            else
+                IT_ASSERT(false, "Broadcast shape mismatch");


The assertion message "Broadcast shape mismatch" on line 31 could be more informative. Consider including the actual dimension values and indices to help with debugging. For example: "Broadcast shape mismatch at dimension " + std::to_string(i) + ": dimA=" + std::to_string(dimA) + ", dimB=" + std::to_string(dimB).

Suggested change

IT_ASSERT(false, "Broadcast shape mismatch");

{

std::string msg = "Broadcast shape mismatch at dimension " +

std::to_string(i) + ": dimA=" +

std::to_string(dimA) + ", dimB=" +

std::to_string(dimB);

IT_ASSERT(false, msg);

}

Xiaobanli-new added 17 commits February 3, 2026 21:34

习题一

c03c42c

习题一

3ecf247

习题一

5ddc1fe

Added new includes and optimized graph processing rules to eliminate redundant transpose operations and merge them into matrix multiplication attributes.

习题二

a7a23f0

Implement shape inference for Transpose operation.

习题三

351e8d0

作业五

4704fb6

习题六

5307d0e

习题七

d7423b5

Implement shape inference for matmul operation

习题八

f2114d7

Fix missing newline at end of file in graph.cc

习题八

6041ff6

Add newline at the end of graph.cc file

习题一

5ceb539

Fix missing newline at end of graph.cc

339cc12

Add missing newline at end of file for graph.cc

Remove TODO comments in allocator.cc

b6b6339

Removed TODO comments for memory allocation and deallocation.

Fix shape inference in ConcatObj::inferShape

26ec13c

Updated inferShape method to correctly compute the concatenated shape.

Fix formatting and comments in matmul.cc

bc26b8d

Added a missing newline at the end of the file and updated comments.

Modify transposePermute to reverse order of indices

a72ca89

Update transpose function to reverse dimensions

Refactor inferDataType and inferShape methods

40272e2

Removed TODO comments and added clarification about shape.

Copilot AI review requested due to automatic review settings February 3, 2026 15:17

Copilot started reviewing on behalf of Xiaobanli-new February 3, 2026 15:17 View session

Copilot AI reviewed Feb 3, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

呢哇#12

呢哇#12
Xiaobanli-new wants to merge 17 commits intoInfiniTensor:mainfrom
Xiaobanli-new:main

Xiaobanli-new commented Feb 3, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

-                IT_ASSERT(false, "Broadcast shape mismatch");
+            {
+                std::string msg = "Broadcast shape mismatch at dimension " +
+                                  std::to_string(i) + ": dimA=" +
+                                  std::to_string(dimA) + ", dimB=" +
+                                  std::to_string(dimB);
+                IT_ASSERT(false, msg);
+            }

Conversation

Xiaobanli-new commented Feb 3, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant