mini-npu-stack

mini-npu-stack is a small C project for learning AI accelerator system software. It begins as a user-space runtime simulator with a minimal API and command-line tool. It does not perform neural network inference.

The project is intended to grow in small, testable stages while keeping the boundary between applications, the runtime, and a future device interface clear.

Build

Requirements:

• GCC
• Make

make
./tools/mininpu_run
./examples/add_relu_demo
make test

Use make clean to remove generated files.

Public API

Include mininpu.h to use the runtime:

struct mininpu_ctx *ctx = NULL;
int32_t a[] = {1, -2};
int32_t b[] = {3, 4};
int32_t output[2];

if (mininpu_open(&ctx) == 0) {
    mininpu_execute(ctx, MININPU_OP_ADD, a, b, output, 2);
    mininpu_close(ctx);
}

• mininpu_open creates an opaque runtime context. It returns 0 on success or a negative error code for invalid arguments or allocation failure.
• mininpu_close releases a context. Passing NULL is allowed.
• mininpu_version returns the runtime version as a static string.
• mininpu_execute runs ADD or RELU over plain int32_t buffers.

The context internals and operator table are private to the runtime implementation.

Layout

• include/: public runtime API
• src/: runtime implementation
• tools/: command-line programs
• examples/: small runtime examples
• tests/: runtime tests
• docs/: design notes and roadmap

Milestones

1. Establish a C11 user-space runtime skeleton and test harness.
2. Model a small simulated device, command queue, and memory buffers.
3. Add deterministic command execution and error handling.
4. Define a user/kernel interface suitable for a character driver.
5. Optionally implement a Linux pseudo accelerator driver.

See docs/roadmap.md for scope and progression.