bibek198/computeedge

How bibek198/computeedge works

ComputeEdge is an instrumented C++ tensor engine for measuring CPU–GPU execution costs on Apple Silicon. A client creates tensors, invokes matrix multiplication, and can run it through the CPU implementations or the Metal backend; instrumentation records allocations and execution events. The benchmark harness and documented comparisons measure performance and environment details. The README also identifies storage-mode and dispatch/synchronization characterization as planned work, not a completed dispatch subsystem. Several benchmark and Metal implementation files were not sampled, so those relationships are included only where documented.

Main entry points: matmul.cpp CPU matmul entry · tensor.h Tensor API · benchmark.cpp Benchmark runner · tensor.cpp Tensor storage

Components 16

Tensor core

Tensor API include/computeedge/tensor.h
Tensor storage src/core/tensor.cpp
Host allocator src/core/allocator.cpp
Matmul validation src/core/validate.cpp

CPU backend

CPU matmul entry src/cpu/matmul.cpp
Accelerate kernel src/cpu/matmul_accelerate.cpp
Naive kernel src/cpu/matmul_naive.cpp

Metal backend

Metal API include/computeedge/metal.h
Metal backend src/metal/metal_backend.mm
Matmul shader src/metal/kernels/matmul.metal

Measurement

Event log src/core/instrumentation.cpp
Benchmark runner src/benchmark/benchmark.cpp
Benchmark statistics src/benchmark/stats.cpp
Environment recorder src/benchmark/env_recorder.mm
MPS and MLX benchmarks

Other components

Library client

Connections 16

Library client → Tensor API: uses README.md
Tensor API → CPU matmul entry: calls matmul
Tensor API → Metal API: calls matmul README.md
Tensor API → Tensor storage: creates storage src/core/tensor.cpp
Tensor storage → Host allocator: allocates through src/core/tensor.cpp
Host allocator → Event log: emits events src/core/allocator.cpp
CPU matmul entry → Matmul validation: validates with src/cpu/matmul.cpp
CPU matmul entry → Accelerate kernel: dispatches to src/cpu/matmul.cpp
CPU matmul entry → Naive kernel: dispatches to src/cpu/matmul.cpp
CPU matmul entry → Event log: emits events src/cpu/matmul.cpp
Metal API → Metal backend: implemented by README.md
Metal backend → Matmul shader: runs shader
Library client → Benchmark runner: runs benchmark README.md
Benchmark runner → Benchmark statistics: reports statistics
Benchmark runner → Environment recorder: records environment
Benchmark runner → MPS and MLX: compares against README.md

Drawn by GitDiagram on November 13, 2026 from the repository's file tree, README and source files. Read as Markdown ·