Spot the Buggy MatMul

Spot the Buggy MatMul

Have a look at this code for a matrix-vector multiplication using vector dot product. It took me a long time to realize what was wrong with this. Can you spot the bug?

    void aussie_matmul_vector_basic1_buggy(ymatrix m, float v[], int n)
    {
        // Basic matrix-by-vector using vector dot products..
        for (int i = 0; i < n; i++) {
                float* rowvector = &m[i][0];
                float sum = aussie_vecdot_basic(rowvector, v, n);  // Dot product
                v[i] = sum;
        }
    }

The bug is a kind of aliasing problem here:

    v[i] = sum;  // Bug!

It looks correct, but it's wrong. The computation of v[i] is setting its value in the middle of the loop, and then going around for the next matrix row, which will then use that newly calculated v[i] value as if it was part of the input vector. Because I'm misusing “v” as both the input and output vector, parts of the output vector will get used as the input vector. It's a very insidious type of aliasing bug, and many of my simple unit tests with zero matrices and identity matrices were still succeeding. It's my fault for trying to do matrix-vector multiplication as an element-wise vector method. The solution is simple: matrix-vector multiplication needs a third operand for the output vector.

• Next: • Up: Table of Contents • Buy: Generative AI in C++: Coding Transformers and LLMs

The new AI programming book by Aussie AI co-founders: AI coding in C++ Transformer engine speedups LLM models Phone and desktop AI Code examples Research citations Get your copy from Amazon: Generative AI in C++