[LIB/CLI] Butteraugli uses too much memory

First of all, thanks for making this awesome tool!

Describe the bug When using “auto” to minify a sufficiently complex image, the butteraugli encode step will use more memory than is available, and will fail with a memory acces out of bounds error. It also won’t properly exit the process, as described in #983.

This can be reproduced with any image, but it seems like it has to not only be a large pixel size, but also be visually complex. An image with a flat, simple motive doesn’t seem to be a problem.

To Reproduce

• Use a computer with at most 16GB of RAM (mine has 16GB, not sure if having more will invalidate this example).
• Download the original 3872x2592 “photo.jpg” test image from squoosh.app:
• Run npx @squoosh/cli --mozjpeg "auto" "/path/to/photo.jpg" (it doesn’t matter what format you use, as long as you encode it with "auto")
• Observe that the process fails with the following message:

Worker error:  Error [RuntimeError]: memory access out of bounds
    at <anonymous>:wasm-function[69]:0x82f8
    at <anonymous>:wasm-function[27]:0x29a3
    at <anonymous>:wasm-function[71]:0xa1f5
    at <anonymous>:wasm-function[116]:0xc600
    at <anonymous>:wasm-function[115]:0xc4f5
    at VisDiff$distance [as distance] (eval at Kb (/[...]/@squoosh/cli/build/index.js:1:1), <anonymous>:9:10)
    at /[...]/@squoosh/cli/build/index.js:7:519788
    at async binarySearch (/[...]/@squoosh/cli/build/index.js:7:519087)
    at async autoOptimize (/[...]/@squoosh/cli/build/index.js:7:519585)

Expected behavior Expected the process to complete without errors.

Version:

• OS w/ version: Like Ubuntu 20.10
• Node version: 14.16.1, 15.14.0 (12.22.1 also fails, but error message is botched)
• npm version: 6.4.12, 7.10.0

Additional context, screenshots, screencasts I understand that at some point, the image becomes so large that you can’t expect Squoosh to work properly, but I think it is reasonable to expect that a relatively normal 4K photo can be encoded on a modern machine with above-average specs.

[jcao219]

Maybe this might help with debugging: I recompiled the visdif module with -g2 and -s ASSERTIONS=1 and got this stack trace:

RuntimeError: abort(Assertion failed: reinterpret_cast<uintptr_t>(aligned) % kCacheLineSize == 0, at: node_modules/butteraugli/butteraugli/butteraugli.cc,82,Free) at Error
    at jsStackTrace (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:31265)
    at stackTrace (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:31441)
    at abort (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:25898)
    at ___assert_fail (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:31599)
    at butteraugli::CacheAligned::Free(void*) (<anonymous>:wasm-function[42]:0x46bf)
    at butteraugli::Blur(butteraugli::Image<float> const&, float, float) (<anonymous>:wasm-function[49]:0x57fa)
    at butteraugli::OpsinDynamicsImage(std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&) (<anonymous>:wasm-function[51]:0x5adc)
    at butteraugli::ButteraugliDiffmap(std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&, std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&, double, butteraugli::Image<float>&) (<anonymous>:wasm-function[41]:0x1b7f)
    at VisDiff::distance(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >) (<anonymous>:wasm-function[37]:0xddd)
    at emscripten::internal::MethodInvoker<double (VisDiff::*)(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >), double, VisDiff*, std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> > >::invoke(double (VisDiff::* const&)(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >), VisDiff*, emscripten::internal::BindingType<std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >, void>::'unnamed'*) (<anonymous>:wasm-function[38]:0x1012)
    at abort (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:25929)
    at ___assert_fail (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:31599)
    at butteraugli::CacheAligned::Free(void*) (<anonymous>:wasm-function[42]:0x46bf)
    at butteraugli::Blur(butteraugli::Image<float> const&, float, float) (<anonymous>:wasm-function[49]:0x57fa)
    at butteraugli::OpsinDynamicsImage(std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&) (<anonymous>:wasm-function[51]:0x5adc)
    at butteraugli::ButteraugliDiffmap(std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&, std::__2::vector<butteraugli::Image<float>, std::__2::allocator<butteraugli::Image<float> > > const&, double, butteraugli::Image<float>&) (<anonymous>:wasm-function[41]:0x1b7f)
    at VisDiff::distance(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >) (<anonymous>:wasm-function[37]:0xddd)
    at emscripten::internal::MethodInvoker<double (VisDiff::*)(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >), double, VisDiff*, std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> > >::invoke(double (VisDiff::* const&)(std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >), VisDiff*, emscripten::internal::BindingType<std::__2::basic_string<char, std::__2::char_traits<char>, std::__2::allocator<char> >, void>::'unnamed'*) (<anonymous>:wasm-function[38]:0x1012)
    at VisDiff$distance [as distance] (eval at new_ (C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:26:56942), <anonymous>:9:10)
    at C:\Users\jcao2\dev\squoosh\libsquoosh\build\index.js:34:175

Edit: I investigated that error above (misaligned block of memory) and it appears it might be due to some compiler bug. When I recompiled with -O0, the assertion error disappeared, and I was left with:

RuntimeError: memory access out of bounds
    at free (<anonymous>:wasm-function[673]:0x130e7)
    at operator delete(void*) (<anonymous>:wasm-function[60]:0x1547)
    at void std::__2::__libcpp_operator_delete<void*>(void*) (<anonymous>:wasm-function[126]:0x268c)
    at void std::__2::__do_deallocate_handle_size<>(void*, unsigned long) (<anonymous>:wasm-function[124]:0x263f)
    at std::__2::__libcpp_deallocate(void*, unsigned long, unsigned long) (<anonymous>:wasm-function[121]:0x25c2)
    at std::__2::allocator<float>::deallocate(float*, unsigned long) (<anonymous>:wasm-function[272]:0x5073)
    at std::__2::allocator_traits<std::__2::allocator<float> >::deallocate(std::__2::allocator<float>&, float*, unsigned long) (<anonymous>:wasm-function[268]:0x4f6c)
    at std::__2::__vector_base<float, std::__2::allocator<float> >::~__vector_base() (<anonymous>:wasm-function[260]:0x4df0)
    at std::__2::vector<float, std::__2::allocator<float> >::~vector() (<anonymous>:wasm-function[253]:0x4a27)
    at butteraugli::Convolution(butteraugli::Image<float> const&, std::__2::vector<float, std::__2::allocator<float> > const&, float) (<anonymous>:wasm-function[307]:0x5c01)

[surma]

Just tested with CLI v0.7.1. The error persists, although I just see a bus error. I’ll have to investigate deeper into Butteraugli.