Raster

Fast numerical computing for Clojure.

Write math with deftm, get world-class numerical performance on the JVM — with full REPL interactivity, automatic differentiation, and GPU compilation.

(require '[raster.core :refer [deftm ftm]])
(require '[raster.numeric :refer [+ *]])
(require '[raster.math :refer [sqrt]])

;; Typed function — compiles to primitive JVM bytecode, no boxing
(deftm norm [x :- Double, y :- Double] :- Double
  (sqrt (+ (* x x) (* y y))))

(norm 3.0 4.0) ;; => 5.0

Try It

The fastest way to see Raster in action:

# Clone and start a REPL
git clone https://github.com/replikativ/raster.git
cd raster
clojure -M:nREPL

Then open one of the interactive notebooks in your editor (Kindly/Clay compatible):

Notebook	What you'll learn
Getting Started	deftm, typed dispatch, value types
Automatic Differentiation	Forward-mode, reverse-mode, value+grad
ODE Solvers	Lorenz attractor, adaptive solvers, events
Linear Algebra	Vec/Mat types, LU, Cholesky, SVD
Optimization	L-BFGS, Nelder-Mead, Newton's method
Deep Learning	MLP training with compiled AD

Game Examples

Three games built on the Raster Vulkan engine — playable demos of typed dispatch, parallel primitives, and procedural generation:

Geometric Asteroids — polygon shapes via defvalue + deftm dispatch. Add new shape types from the REPL during gameplay.	Valley — voxel survival with procedural terrain, crafting, and mobs. Uses raster.noise for terrain generation.

There's also a Doom-style renderer with WAD loading and BSP traversal.

See examples/README.md for running instructions.

Why Raster?

Clojure is great for data-driven applications, but numerical computing has traditionally required dropping to Java or calling out to Python. Raster fills that gap with three ideas:

1. Typed multiple dispatch (inspired by Julia) — define the same function for different types and the most specific method is selected automatically. The compiler resolves dispatch at call sites when the consumer's argument types are known — either by selecting the closest concrete method or by specializing a parametric template — and emits JVM bytecode with zero overhead.

2. Parallel combinators (inspired by Futhark) — express data parallelism with par/map, par/reduce, and par/scan. The compiler treats these as first-class IR nodes, fuses producer-consumer chains (SOAC fusion), and lowers them to SIMD loops, OpenCL kernels, or Vulkan compute — from the same source code.

3. End-to-end compilation — a transparent nanopass compiler that you own and can inspect. Because we compile from Clojure source to JVM bytecode directly (no LLVM dependency), we control every optimization pass and can target multiple backends: JVM bytecode with SIMD vectorization, OpenCL C, Vulkan SPIR-V, or Intel Level Zero. The pipeline is fully inspectable via explain-pipeline. We can also emit C or potentially LLVM IR when needed, but the default path avoids external toolchain dependencies entirely.

This combination — typed dispatch for abstraction, parallel combinators for structure, and a self-contained compiler for performance — is how Raster aims to match JAX-level deep learning performance while remaining a regular Clojure library with full REPL interactivity.

What this enables

• Clojure-native — deftm is a macro, not a DSL. Your functions are regular Clojure values that work at the REPL, in tests, with your editor.
• Fast — the compiler fuses array operations across function boundaries, hoists buffer allocations, and vectorizes loops. Competitive with Julia and JAX on numerical workloads.
• Composable — functions compose naturally. Write a loss function, take its gradient, compile the training step — all with the same deftm functions. compile-aot inlines entire call chains into a single JVM method with zero heap allocations in the hot path.
• Multi-backend — the same par/map expression runs as a sequential loop at the REPL, a SIMD-vectorized loop in compiled JVM code, or a GPU kernel on OpenCL/Vulkan/Level Zero.

Performance

All benchmarks on Valhalla JDK 27, single-threaded CPU unless noted. GPU benchmarks on Intel Arc A770 (Level Zero).

Workload	Raster	Reference	vs Reference
ODE solve (DP5 Lorenz, t=100)	432 µs	Julia DiffEq 583 µs	1.4x faster
MLP 784-128-10 train step (f64)	136 µs	JAX CPU 86 µs	JAX 1.6x faster
MLP 784-128-10 train step (f32)	77 µs	JAX CPU 50 µs	JAX 1.5x faster
LeNet-5 train step (f64)	222 µs	JAX CPU 370 µs	1.7x faster
LeNet-5 train step (f32)	148 µs	JAX CPU 356 µs	2.4x faster
AD sensitivity (Lotka-Volterra, Dual4)	15 µs	Julia ForwardDiff 16 µs	1.1x faster
ABM 10M agents/period (GPU)	172 ms	CPU-parallel 459 ms	2.7x faster

The DL numbers include the full compiled pipeline: forward pass, reverse-mode AD backward pass, and SGD parameter update — compiled to a single JVM method via compile-aot. Zero heap allocations in the hot path. Dense layers use fused par/map + par/reduce SOACs (no BLAS calls) — the same strategy as Futhark and XLA's CPU backend. BLAS via Panama FFI (cblas_dgemv/cblas_dgemm) is available for explicit use in large batched GEMM workloads.

Core Concepts

Typed Functions: deftm and ftm

deftm (typed multi) defines functions with type annotations and multiple dispatch:

;; Multiple dispatch — same name, different types
(deftm norm [x :- Double, y :- Double] :- Double
  (sqrt (+ (* x x) (* y y))))

(deftm norm [x :- Float, y :- Float] :- Float
  (sqrt (+ (* x x) (* y y))))

(norm 3.0 4.0)              ;; => 5.0 (Double path)
(norm (float 3) (float 4))  ;; => 5.0 (Float path, no boxing)

;; Typed lambdas
(deftm apply-twice [f :- (Fn [Double] Double), x :- Double] :- Double
  (f (f x)))

(apply-twice (ftm [x :- Double] :- Double (* x x)) 3.0) ;; => 81.0

Use deftm/ftm for all numerical code. Use plain defn for glue code, configuration, and I/O.

Parallel Combinators

raster.par provides declarative parallel primitives that the compiler can fuse, vectorize, and lower to different backends:

(require '[raster.par :as par])

;; Element-wise map — returns a new array
(deftm relu [xs :- (Array double)] :- (Array double)
  (par/map [i (alength xs)]
    (Math/max 0.0 (aget xs i))))

;; Dot product — fused map + reduce
(deftm dot [xs :- (Array double), ys :- (Array double)] :- Double
  (par/reduce + 0.0 [i (alength xs)]
    (* (aget xs i) (aget ys i))))

When composed, the compiler fuses adjacent par/map and par/reduce operations into single loops — no intermediate arrays are allocated. This is the same SOAC (Second-Order Array Combinator) fusion strategy pioneered by Futhark, applied here within a general-purpose host language rather than a standalone array language.

Polymorphic Arithmetic

raster.numeric provides +, -, *, / and friends that dispatch on concrete types — Long stays Long, Float stays Float:

(+ 1 2)                    ;; Long + Long -> Long
(+ 1.0 2.0)                ;; Double + Double -> Double
(+ (float 1) (float 2))    ;; Float + Float -> Float

Automatic Differentiation

Forward-mode (Dual numbers) and reverse-mode (IR source transformation):

(require '[raster.ad :refer [value+grad grad]])

(deftm rosenbrock [x :- Double, y :- Double] :- Double
  (let [a (- y (* x x))
        b (- 1.0 x)]
    (+ (* 100.0 a a) (* b b))))

;; value+grad returns [f(x), ∇f(x)]
(value+grad rosenbrock 1.0 1.0) ;; => [0.0 [0.0 0.0]]

Compilation

For maximum performance, compile-aot inlines an entire call chain — forward pass, AD, optimizer update — into a single JVM method:

(require '[raster.compiler.pipeline :as pipeline])

(def fast-step (pipeline/compile-aot #'train-step!))

The compiler performs buffer fusion (reusing dead arrays), SOAC fusion (merging parallel combinator chains), and emits SIMD-vectorized loops — all automatically from your deftm source.

Use (pipeline/explain-pipeline #'my-fn) to see what each compiler pass does.

What's Included

Scientific Computing

Module	Description
raster.ode	ODE solvers (Euler, RK4, DP5, Tsit5), PDE (method-of-lines), SDE
raster.sci.optim	L-BFGS, Nelder-Mead, Newton, gradient descent
raster.linalg	Dense linear algebra, LU, Cholesky, SVD, QR, eigendecomposition
raster.linalg.iterative	Krylov methods (CG, GMRES, BiCGSTAB)
raster.linalg.sparse	Sparse vectors and matrices
raster.sci.interpolation	Linear, cubic spline, Akima, PCHIP, 2D bilinear/bicubic
raster.sci.fft	Fast Fourier Transform
raster.sci.quadrature	Numerical integration (Gauss-Kronrod, Simpson)
raster.sci.roots	Root finding (bisection, Brent, Newton)
raster.sci.distributions	Normal, Uniform, Exponential, Gamma, Poisson
raster.sci.stats	t-tests, chi-squared, KS test, correlation
raster.sci.special	Gamma, beta, error functions

Deep Learning

Module	Description
raster.dl.nn	Linear, conv1d/2d, maxpool, normalization, activations
raster.dl.attention	Scaled dot-product and multi-head attention
raster.dl.loss	MSE, cross-entropy, Huber, L1 (with AD rules)
raster.dl.optim	SGD, Adam, AdamW, learning rate schedulers
raster.dl.einsum	Einstein summation and einops-style rearrangement
raster.dl.diffusion	DDPM noise schedules and sampling

All layers are deftm functions on flat arrays — the compiler sees through layer boundaries and fuses operations end-to-end. See the LeNet-5 and GPT-2 examples.

Symbolic and Algebraic

Module	Description
raster.sym	Symbolic expressions, differentiation, Taylor series
raster.ga	Geometric algebra Cl(p,q,r) with compiled multivector types
raster.types.complex	Complex number arithmetic

GPU Computing

The same parallel combinators (par/map, par/reduce, par/scan) that run as SIMD-vectorized loops on CPU compile to GPU kernels through backend passes. Backends: OpenCL (NVIDIA, AMD, Intel), Intel Level Zero, Vulkan compute, JDK Vector API SIMD.

See GPU Computing for the session API and backend details.

Simulation

raster.abm — a GPU-accelerated agent-based model of firm formation and labor markets. Same Clojure source runs on CPU and GPU. Includes a differentiable variant for gradient-based parameter calibration.

Under the Hood

Raster's compiler is a nanopass pipeline: each pass transforms a small, well-defined IR dialect. Dispatch resolution, type inference, AD expansion, SOAC fusion, buffer allocation, and backend code generation are all separate passes that you can inspect individually.

Document	Description
Compiler Pipeline	Nanopass architecture, passes, compile-aot, diagnostics
Automatic Differentiation	Forward/reverse AD, rrules, sensitivity analysis
GPU Computing	Parallel primitives, session API, backends, SoA layout
Deep Learning	Layers, loss, optimizers, compiled training

Built on TypedClojure for type inference and beichte for purity analysis.

Project Structure

src/raster/
  core.clj                -- deftm, ftm, defvalue, specialize macros
  numeric.clj             -- polymorphic +, -, *, /, comparisons
  math.clj                -- sin, cos, exp, log, sqrt, fma, ...
  arrays.clj              -- polymorphic aget/aset/alength
  par.clj                 -- parallel combinators (map, reduce, scan)

  ad/                     -- automatic differentiation
  compiler/               -- nanopass compiler pipeline
  ode/                    -- ODE/PDE/SDE solvers
  linalg/                 -- linear algebra, LAPACK via Panama FFI
  sci/                    -- special functions, distributions, optimization
  gpu/                    -- unified GPU session (OpenCL ICD + Level Zero)
  dl/                     -- deep learning layers, optimizers, training
  sym/                    -- symbolic computation
  ga/                     -- geometric algebra
  vk/                     -- Vulkan rendering engine

Requirements

• JDK 24+ — for the ClassFile API (bytecode emission) and Panama FFM (native bindings)
• Clojure 1.12+
• OpenCL (optional) — for GPU computing
• Dependencies resolve automatically via deps.edn

For Valhalla value types (Dual, Float16), use JDK 27 early-access.

Running

# Tests
clojure -M:test

# REPL
clojure -M:nREPL

# With Valhalla JDK 27
source valhalla-env.sh
clojure -J--enable-preview \
        -J--add-exports=java.base/jdk.internal.vm.annotation=ALL-UNNAMED \
        -J--enable-native-access=ALL-UNNAMED \
        -M:test:valhalla

Documentation

• doc/ — technical guides (compiler, AD, GPU, deep learning)
• notebooks/raster/ — interactive notebooks
• examples/ — code examples and game demos
• design/ — architecture and design documents
• CHANGELOG.md — release notes

License

MIT License. See LICENSE for the full text.

Third-party notices are in THIRD_PARTY_NOTICES.md.