fix: free Content object in error paths of loadModelFile/createFromBuffer

[srpatcha]

Fix memory leaks in model loading error paths and typo

Problem

Two memory leaks exist in source/core/Interpreter.cpp:

1. In loadModelFile(): new Content is allocated but never freed if loader->merge() fails.
2. In createFromBuffer(): new Content is allocated but never freed if buffer.reset() returns nullptr.

Additionally, the error message "Memory not enought!" contains a typo.

Root Cause

Both functions allocate a Content object with new on the heap, then check a subsequent operation for failure. On failure, both return nullptr without calling delete on the allocated object, leaking memory.

Fix

• Added delete net; before return nullptr; in both error paths.
• Fixed typo: "Memory not enought!" → "Memory not enough!"

Testing

• Load an invalid/corrupt model file that triggers the loader->merge() failure path.
• Create an interpreter from a buffer with insufficient memory.
• Both paths should now properly clean up the allocated Content object.
• Run under Valgrind or AddressSanitizer to verify no leaks.

Impact

Affects MNN users loading models, especially in long-running services or on memory-constrained devices where repeated failed model loads would accumulate leaked memory.

[CLAassistant]

All committers have signed the CLA.

[wangzhaode]

Hi @srpatcha, friendly reminder — the core fixes (memory leak in Interpreter.cpp, bilinear boundary check, and QNN error handling) all look solid and correct. The only remaining item is consolidating the test file: please move the valuable new test cases (SinglePixelResizeTest, StrideMismatchTest) into the existing test/cv/ImageProcessTest.cpp and remove the separate ImageOpTest.cpp file to keep the test suite consolidated. Once that's done, this is ready to merge. Thanks for the contribution!

[wangzhaode]

Please delete this file.

[wangzhaode]

We already have comprehensive image processing tests in test/cv/ImageProcessTest.cpp covering bilinear/nearest resize, colorspace conversions (RGB↔BGR, RGBA→Gray, BGR→Gray, etc.), YUV blitting, and transform-based resize.

Most of the tests in this new file overlap with existing ones. However, SinglePixelResizeTest and StrideMismatchTest are genuinely new edge-case scenarios that would be valuable additions.

Could you move just the useful new test cases (e.g., single-pixel resize, stride mismatch) into the existing test/cv/ImageProcessTest.cpp instead of creating a separate file? This keeps the test suite consolidated and avoids confusion between ImageProcessTest.cpp and ImageOpTest.cpp.

[srpatcha]

Hi @wangzhaode, done! Consolidated the test cases per your request:

• Moved SinglePixelResizeTest and StrideMismatchTest into test/cv/ImageProcessTest.cpp
• Deleted test/cv/ImageOpTest.cpp
• Updated test suite names from cv/image_op/* to cv/image_process/* to match existing convention

Commit: daf87a3. Ready for final review!

[wangzhaode]

Hi @srpatcha, thanks for the contribution! The core Interpreter.cpp memory leak fixes look solid and correct.

However, the PR currently fails to compile with -DMNN_BUILD_TEST=ON. The two new test cases introduced in test/cv/ImageProcessTest.cpp use unqualified enum names that are ambiguous in this translation unit:

test/cv/ImageProcessTest.cpp:1299:29: error: reference to 'NEAREST' is ambiguous
test/cv/ImageProcessTest.cpp:1337:29: error: reference to 'BILINEAR' is ambiguous

Both MNN::CV::Filter::BILINEAR and MNN::Express::BILINEAR are in scope, causing the ambiguity. The existing test cases in the same file all use fully qualified names.

Fix: please update the two lines to use the fully qualified enum:

// line 1299
config.filterType = MNN::CV::Filter::NEAREST;

// line 1337
config.filterType = MNN::CV::Filter::BILINEAR;

Build command used:

cmake .. -DMNN_BUILD_LLM=ON -DMNN_LOW_MEMORY=ON -DMNN_BUILD_TEST=ON -DMNN_SUPPORT_TRANSFORMER_FUSE=ON -GNinja && ninja

Please push the fix and we'll re-verify. Thanks!

[srpatcha]

Hi @wangzhaode, fixed! Updated both lines to use fully qualified enum names:

• \config.filterType = MNN::CV::Filter::NEAREST;\ (line 1299)
• \config.filterType = MNN::CV::Filter::BILINEAR;\ (line 1337)

Commit: a1b0b20. Should now compile cleanly with -DMNN_BUILD_TEST=ON. Thanks for catching this!

[wangzhaode]

Hi @srpatcha, thanks for this PR! The memory leak fixes in Interpreter.cpp, the bilinear clamp in ImageProcessFunction.cpp, and the QNN null-check hardening all look correct and well-motivated.

However, the two newly added tests (stride_mismatch and single_pixel_resize) are currently failing on our local build (macOS arm64, CPU backend). Here is the test output:

running cv/image_process/stride_mismatch.
Error for run, 1318

running cv/image_process/single_pixel_resize.
Error for run, 1358

TEST_CASE_AMOUNT_UNIT: {"blocked":0,"failed":2,"passed":18,"skipped":0}

The issue appears to be an incorrect call to the second overload of ImageProcess::convert:

// Current (incorrect):
process->convert(src.data(), W, H, srcStride, dst.data(), W, H, channels * W, RGB);

// Signature:
// convert(source, iw, ih, stride, dest, ow, oh, outputBpp=0, outputStride=0, type=halide_type_of<float>())

• outputBpp is receiving channels * W (should be channels or 0 to auto-detect from destFormat)
• outputStride is receiving the RGB/RGBA enum value (which is unrelated)
• type defaults to halide_type_of<float>(), but the test reads the output buffer as uint8_t

Could you please:

1. Fix the convert() call parameters (set outputBpp = 0 or the correct channel count, outputStride = 0 or ow * channels, and pass halide_type_of<uint8_t>() as the type)
2. Run the tests locally to verify they pass before updating the PR

The rest of the fixes (Interpreter leaks, bilinear clamp, QNN null checks) are solid — just need the test code corrected. Thanks!