MATLAB Performance Optimizer

Optimize MATLAB code performance with vectorization, memory management, and profiling tools.

When to Use This Skill

• Optimizing slow or inefficient MATLAB code
• Converting loops to vectorized operations
• Reducing memory usage
• Improving algorithm performance
• When user mentions: slow, performance, optimize, speed up, efficient, memory
• Profiling code to find bottlenecks
• Parallelizing computations

Core Optimization Principles

1. Vectorization (Most Important)

Replace loops with vectorized operations whenever possible.

SLOW - Using loops:

% Slow approach
n = 1000000;
result = zeros(n, 1);
for i = 1:n
    result(i) = sin(i) * cos(i);
end

FAST - Vectorized:

% Fast approach
n = 1000000;
i = (1:n).';
result = sin(i) .* cos(i);

2. Preallocate Arrays

Always preallocate arrays before loops.

SLOW - Growing arrays:

% Very slow - array grows each iteration
result = [];
for i = 1:10000
    result(end+1) = i^2;
end

FAST - Preallocated:

% Fast - preallocated array
n = 10000;
result = zeros(n, 1);
for i = 1:n
    result(i) = i^2;
end

3. Use Built-in Functions

MATLAB built-in functions are highly optimized.

SLOW - Manual implementation:

% Slow
sum_val = 0;
for i = 1:length(x)
    sum_val = sum_val + x(i);
end

FAST - Built-in function:

% Fast
sum_val = sum(x);

Vectorization Techniques

Element-wise Operations

Use .*, ./, .^ for element-wise operations:

% Instead of this:
for i = 1:length(x)
    y(i) = x(i)^2 + 2*x(i) + 1;
end

% Do this:
y = x.^2 + 2*x + 1;

Logical Indexing

Replace conditional loops with logical indexing:

% Instead of this:
count = 0;
for i = 1:length(data)
    if data(i) > threshold
        count = count + 1;
        filtered(count) = data(i);
    end
end
filtered = filtered(1:count);

% Do this:
filtered = data(data > threshold);

Matrix Operations

Use matrix multiplication instead of nested loops:

% Instead of this:
C = zeros(size(A, 1), size(B, 2));
for i = 1:size(A, 1)
    for j = 1:size(B, 2)
        for k = 1:size(A, 2)
            C(i,j) = C(i,j) + A(i,k) * B(k,j);
        end
    end
end

% Do this:
C = A * B;

Cumulative Operations

Use cumsum, cumprod, cummax, cummin:

% Instead of this:
running_sum = zeros(size(data));
running_sum(1) = data(1);
for i = 2:length(data)
    running_sum(i) = running_sum(i-1) + data(i);
end

% Do this:
running_sum = cumsum(data);

Memory Optimization

Use Appropriate Data Types

% Instead of default double (8 bytes)
data = rand(1000, 1000);  % 8 MB

% Use single precision when appropriate (4 bytes)
data = single(rand(1000, 1000));  % 4 MB

% Use integers when applicable
indices = uint32(1:1000000);  % 4 MB instead of 8 MB

Sparse Matrices

For matrices with mostly zeros:

% Dense matrix (wastes memory)
A = zeros(10000, 10000);
A(1:100, 1:100) = rand(100);  % 800 MB

% Sparse matrix (efficient)
A = sparse(10000, 10000);
A(1:100, 1:100) = rand(100);  % Only stores non-zeros

Clear Unused Variables

% Process large data
largeData = loadData();
processedData = processData(largeData);

% Clear when no longer needed
clear largeData;

% Continue with processed data
results = analyze(processedData);

In-Place Operations

% Instead of creating copies
A = A + 5;  % In-place when possible

% Avoid unnecessary copies
B = A;      % Creates copy if A is modified later
B = A + 0;  % Forces copy

Profiling and Benchmarking

Using the Profiler

% Profile code execution
profile on
myFunction(inputs);
profile viewer
profile off

The profiler shows:

• Time spent in each function
• Number of calls to each function
• Lines that take the most time

Timing Comparisons

% Time single execution
tic;
result = myFunction(data);
elapsedTime = toc;

% Benchmark with timeit (more accurate)
timeit(@() myFunction(data))

% Compare multiple approaches
time1 = timeit(@() approach1(data));
time2 = timeit(@() approach2(data));
fprintf('Approach 1: %.6f s\nApproach 2: %.6f s\n', time1, time2);

Common Optimization Patterns

Pattern 1: Replace find with Logical Indexing

% SLOW
indices = find(x > 5);
y = x(indices);

% FAST
y = x(x > 5);

Pattern 2: Use Implicit Expansion Instead of repmat

% SLOW - repmat to match dimensions
A = rand(1000, 5);
B = rand(1, 5);
C = A - repmat(B, size(A, 1), 1);

% FAST - implicit expansion (R2016b+)
C = A - B;

Pattern 3: Avoid Repeated Calculations

% SLOW - recalculates each iteration
for i = 1:n
    result(i) = data(i) / sqrt(sum(data.^2));
end

% FAST - calculate once
norm_factor = sqrt(sum(data.^2));
for i = 1:n
    result(i) = data(i) / norm_factor;
end

% EVEN FASTER - vectorize
result = data / sqrt(sum(data.^2));

Pattern 4: Efficient String Operations

% SLOW - concatenating in loop
str = '';
for i = 1:1000
    str = [str, sprintf('Line %d\n', i)];
end

% FAST - cell array + join
lines = cell(1000, 1);
for i = 1:1000
    lines{i} = sprintf('Line %d', i);
end
str = strjoin(lines, '\n');

% FASTEST - vectorized sprintf
str = sprintf('Line %d\n', 1:1000);

Pattern 5: Use Table for Mixed Data Types

% Instead of separate arrays
names = cell(1000, 1);
ages = zeros(1000, 1);
scores = zeros(1000, 1);

% Use table
data = table(names, ages, scores);
% Faster access and better organization

Algorithm-Specific Optimizations

Convolution and Filtering

% Use built-in functions
filtered = conv(signal, kernel, 'same');
filtered = filter(b, a, signal);

% For 2D
filtered = conv2(image, kernel, 'same');
filtered = imfilter(image, kernel);

% FFT-based for large kernels (zero-pad for linear convolution)
nfft = length(signal) + length(kernel) - 1;
filtered = ifft(fft(signal, nfft) .* fft(kernel, nfft));

Distance Calculations

% Instead of nested loops for pairwise distances
% SLOW
n = size(points, 1);
distances = zeros(n, n);
for i = 1:n
    for j = 1:n
        distances(i,j) = norm(points(i,:) - points(j,:));
    end
end

% FAST - vectorized
distances = pdist2(points, points);

Sorting and Searching

% Presort for multiple searches
sortedData = sort(data);

% Binary search on sorted data
idx = find(sortedData >= value, 1, 'first');

% Use ismember for set operations
[isPresent, locations] = ismember(searchValues, data);

% Use unique for removing duplicates
uniqueData = unique(data);

Parallel Computing

Simple Parallel Loops (parfor)

% Convert for to parfor for independent iterations
parfor i = 1:n
    results(i) = expensiveFunction(data(i));
end

Requirements for parfor:

• Iterations must be independent
• Loop variable must be consecutive integers
• Variables must be classified as loop, sliced, broadcast, or reduction

Parallel Array Operations

% Create parallel pool
parpool('local', 4);  % 4 workers

% Use parfeval for asynchronous parallel execution
futures = parfeval(@expensiveFunction, 1, data);
result = fetchOutputs(futures);

% GPU arrays for massive parallelization
gpuData = gpuArray(data);
result = arrayfun(@myFunction, gpuData);
result = gather(result);  % Bring back to CPU

Advanced Optimizations

MEX Functions for Critical Sections

Convert performance-critical code to C/C++:

% Create MEX file for bottleneck function
% Write myFunction.c, then compile:
% mex myFunction.c

% Call like regular MATLAB function
result = myFunction(inputs);

Persistent Variables for Cached Results

function result = expensiveComputation(input)
    persistent cachedData cachedInput

    if isequal(input, cachedInput)
        % Return cached result
        result = cachedData;
        return;
    end

    % Compute and cache
    result = computeExpensiveOperation(input);
    cachedData = result;
    cachedInput = input;
end

JIT Acceleration Best Practices

MATLAB's JIT (Just-In-Time) compiler optimizes:

• Simple for-loops with scalar operations
• Functions without dynamic features

JIT-friendly code:

function result = jitFriendly(n)
    result = 0;
    for i = 1:n
        result = result + i;
    end
end

JIT-unfriendly code (avoid):

function result = jitUnfriendly(n)
    result = 0;
    for i = 1:n
        eval(['x' num2str(i) ' = i;']);  % Dynamic code
    end
end

Performance Checklist

Before finalizing optimized code, verify:

• Loops are vectorized where possible
• Arrays are preallocated before loops
• Built-in functions used instead of manual implementations
• Logical indexing used instead of find + indexing
• Appropriate data types used (single vs double, integers)
• Sparse matrices used for sparse data
• Repeated calculations moved outside loops
• String concatenation uses efficient methods
• Code profiled to identify actual bottlenecks
• Matrix operations used instead of element-wise loops
• Parallel computing considered for independent operations
• Memory-intensive operations optimized
• Caching implemented for repeated expensive calls

Profiling Workflow

1. Measure First: Profile before optimizing

   profile on
   myScript;
   profile viewer

2. Identify Bottlenecks: Focus on functions taking most time
3. Optimize: Apply appropriate techniques
4. Measure Again: Verify improvement

   % Before
   time_before = timeit(@() myFunction(data));
   
   % After optimization
   time_after = timeit(@() myFunctionOptimized(data));
   
   fprintf('Speedup: %.2fx\n', time_before/time_after);

5. Iterate: Repeat for remaining bottlenecks

Common Performance Pitfalls

Pitfall 1: Premature Optimization

• Profile first, optimize second
• Focus on actual bottlenecks, not assumptions

Pitfall 2: Over-vectorization

• Sometimes loops are clearer and fast enough
• Balance readability with performance

Pitfall 3: Ignoring Memory Access Patterns

% SLOW - inner loop over columns (row-major traversal in column-major MATLAB)
for i = 1:rows
    for j = 1:cols
        A(i,j) = process(i, j);
    end
end

% FAST - inner loop over rows (column-major traversal, contiguous memory)
for j = 1:cols
    for i = 1:rows
        A(i,j) = process(i, j);
    end
end

% FASTEST - vectorized
[I, J] = ndgrid(1:rows, 1:cols);
A = process(I, J);

Pitfall 4: Unnecessary Data Type Conversions

% SLOW - repeated conversions
for i = 1:n
    x = double(data(i));
    result(i) = sin(x);
end

% FAST - convert once
x = double(data);
result = sin(x);

Optimization Examples

Example 1: Image Processing

% SLOW
[rows, cols] = size(image);
output = zeros(rows, cols);
for i = 2:rows-1
    for j = 2:cols-1
        output(i,j) = mean(image(i-1:i+1, j-1:j+1), 'all');
    end
end

% FAST
kernel = ones(3,3) / 9;
output = conv2(image, kernel, 'same');

Example 2: Statistical Analysis

% SLOW
n = size(data, 1);
means = zeros(n, 1);
for i = 1:n
    means(i) = mean(data(i, :));
end

% FAST
means = mean(data, 2);

Example 3: Time Series Processing

% SLOW
n = length(signal);
movingAvg = zeros(size(signal));
window = 10;
for i = window:n
    movingAvg(i) = mean(signal(i-window+1:i));
end

% FAST - trailing window: [window-1 past samples, 0 future samples]
movingAvg = movmean(signal, [window-1 0]);

Troubleshooting Performance

Issue: Code still slow after vectorization

• Solution: Profile to find new bottlenecks; consider algorithm complexity

Issue: Out of memory errors

• Solution: Use smaller data types, process in chunks, use sparse matrices

Issue: parfor slower than for loop

• Solution: Check if overhead outweighs benefits; ensure iterations are expensive enough

Issue: GPU computation slower than CPU

• Solution: Data transfer overhead may exceed computation time; use for large arrays

Additional Resources

• Use profile viewer to analyze performance
• Use memory to check memory usage
• Use doc with: timeit, tic/toc, parfor, gpuArray, sparse
• Check MATLAB Performance and Memory documentation